text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
1
Robust Real-Time Multi-View Eye Tracking
arXiv:1711.05444v2 [cs.CV] 3 Jan 2018
Nuri Murat Arar, Student Member, IEEE, and Jean-Philippe Thiran, Senior Member, IEEE
Abstract—Despite significant advances in improving the gaze
tracking accuracy under controlled conditions, the tracking
robustness under real-world conditions, such as large head pose
and movements, use of eyeglasses, illumination and eye type variations, remains a major challenge in eye tracking. In this paper,
we revisit this challenge and introduce a real-time multi-camera
eye tracking framework to improve the tracking robustness.
First, differently from previous work, we design a multi-view
tracking setup that allows for acquiring multiple eye appearances
simultaneously. Leveraging multi-view appearances enables to
more reliably detect gaze features under challenging conditions,
particularly when they are obstructed in conventional single-view
appearance due to large head movements or eyewear effects.
The features extracted on various appearances are then used for
estimating multiple gaze outputs. Second, we propose to combine
estimated gaze outputs through an adaptive fusion mechanism to
compute user’s overall point of regard. The proposed mechanism
firstly determines the estimation reliability of each gaze output
according to user’s momentary head pose and predicted gazing
behavior, and then performs a reliability-based weighted fusion.
We demonstrate the efficacy of our framework with extensive
simulations and user experiments on a collected dataset featuring
20 subjects. Our results show that in comparison with stateof-the-art eye trackers, the proposed framework provides not
only a significant enhancement in accuracy but also a notable
robustness. Our prototype system runs at 30 frames-per-second
(fps) and achieves ∼1◦ accuracy under challenging experimental
scenarios, which makes it suitable for applications demanding
high accuracy and robustness.
I. I NTRODUCTION
Gaze movements provide cues indicating observer’s visual
attention, emotional state and cognitive processes [1], [2].
Therefore, tracking gaze movements, also known as eye tracking, is essential for human behaviour research and diagnostics
applied in a wide variety of disciplines, including among many
others, sociology, psychology, cognitive science, neuroscience,
and marketing research. Besides, it is an important modality
to enhance human-computer interaction for controlling and
navigation, such as in virtual reality, augmented reality, and
gaming research. The popularity of eye tracking technology
has recently been increasing, owing to its large spectrum of
applications as well as the promising technical advancements.
Despite valuable efforts, high-accuracy (≤1◦ ) eye tracking
systems still suffer from various factors, such as high cost,
complex and inflexible setup configurations, and more importantly, low tolerance to varying real-world conditions, which
hinder them from being widely used. Hence, there is room
for further research efforts, particularly towards improving the
tracking robustness under unconstrained conditions.
Remote video-oculography, in which users’ eyes are nonintrusively captured by remote sensors, is the focus of this
paper since it provides the most natural and convenient interaction for the users. As described in a recent survey [3],
N. M. Arar and J.-P. Thiran are with the Signal Processing Laboratory
(LTS5), École Polytechnique Fédérale de Lausanne (EPFL), Switzerland.
remote sensor-based eye tracking methods can be classified
into two categories, namely, appearance-based and featurebased. Appearance-based methods use the image content as
the input. They learn a mapping from the image features
directly to the gaze points. On the other hand, feature-based
methods utilize local features extracted on eye images, such
as pupil center and reflections on the cornea (aka glints), to
determine the gaze. They mostly require particular hardware
configuration and leverage the features that are formally related to the gaze points through the geometry of the system
and eye physiology. Feature-based methods can further be
categorized into 3D model-based, regression-based, and cross
ratio-based methods. Each category has its own advantages
and disadvantages regarding the estimation accuracy, robustness, and system implementation complexity, as summarized
in Table I. 3D model-based methods [4]–[8] compute the gaze
from the features obtained from a 3D geometric eye model,
whereas regression-based methods [9]–[12] assume a direct
mapping from the features to the gaze points. On the other
hand, cross ratio-based methods [13]–[18] compute the gaze
by leveraging the cross ratio property of the projective space.
Over more than two decades, numerous works from each
of the aforementioned categories have been presented. Among
these, the main emphasis has been mostly given to the estimation accuracy improvements through introducing different
gaze models [6], [10], [13], [19] or developing effective user
calibration techniques [17], [18], [20]–[22]. Consequently,
high accuracies (<1◦ ) are reported under controlled conditions. Nevertheless, research on the tracking robustness under
real-world conditions, such as large head movements, use of
eyewear, variations in illumination and eye type, have been
largely neglected. Thus, these remain major concerns in eye
tracking.
In this paper, extending our previous efforts that focus on
improving the estimation accuracy [18], [19], we address the
major robustness concerns of existing eye tracking systems.
We present a real-time multi-view gaze estimation methodology to improve the tracking robustness to real-world conditions. Instead of tracking the gaze from a single view as performed by previous work, we design a multi-view framework
to leverage multiple eye appearances simultaneously acquired
from various views. In conventional single-view setups, there
exists a single appearance, on which the features may be
obstructed due to challenging conditions such as large head
movements or occlusions caused by glasses. Whereas in our
framework, the main benefit is to simultaneously perform
feature detection on multi-view appearances. For each frame,
our approach enables to compute multiple gaze outputs using
the extracted features. Furthermore, these gaze outputs are
effectively combined by a novel adaptive fusion mechanism to
compute user’s overall point of regard (PoR). In this context,
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
2
TABLE I: A generic comparison of gaze estimation techniques.
Evaluation Criteria
Setup complexity:
System calibration:
Hardware requirements (# of cameras):
Hardware requirements (# of lights):
Gaze estimation accuracy error:
Implicit robustness to head movements:
Implicit robustness to varying illumination:
Implicit robustness to use of eyewear:
3D Model-based
Regression-based
Cross Ratio-based
Appearance-based
high
fully-calibrated
2+ infrared (stereo)
2+ infrared
< 1◦
medium-high
medium-high
low
medium
×
1+ infrared
2+ infrared
∼ 1 − 2◦
low-medium
medium-high
low
medium
×
1+ infrared
4+ infrared
∼ 1 − 2◦
low-medium
medium-high
low
low
×
1+ ordinary
×
> 2◦
low
low
medium
the proposed mechanism firstly determines the estimation
reliability of each gaze output according to certain gaze
reliability indicators, e.g., user’s predicted gazing behavior,
momentary head pose with respect to the camera. Then, it
performs a reliability-based weighted fusion, which leads to
improved accuracy and robustness. Extensive evaluations on
both simulated and real-world data were performed to validate
the proposed methodology. In addition to thorough simulations, a database featuring 20 users performing 8 experiments
under varying illumination conditions and head movements
was collected. In these experiments, natural human-computer
interaction was targeted. Users followed some conventional
experimental scenarios as well as a newly introduced one. The
results demonstrate that in comparison to conventional singleview eye tracking, the multi-view approach provides a significantly better performance both in accuracy and robustness to
aforementioned challenging conditions.
The proposed framework is highly flexible and can easily
adapt to hardware and software modifications. Depending on
the application type and desired tracking performance, the
employed gaze estimation method, number of cameras and
their configuration can simply be alternated, even without
requiring any system adjustments (e.g., camera or geometric
scene calibration). Our current uncalibrated prototype relies on
a cross ratio-based gaze estimation method, and operates with
low-resolution eye data. The system’s overall hardware setup
and computational complexities are lower than those of fullycalibrated systems. Hence, it enables a fast and accurate eye
tracking without requiring any cumbersome camera and geometric scene calibrations. Our three-camera prototype outputs
PoRs with an accuracy of ∼1◦ at 30 fps and also obtains nearly
100% estimation availability under challenging scenarios.
The outline of the paper is described as follows: Section II
gives an overview of the related work. Section III describes
the proposed framework. Evaluations on the simulated and real
data are presented in Section IV and V. Section VI discusses
the acquired insights, and conclusions are given in Section VII.
II. R ELATED W ORK
A generic comparison of existing eye tracking solutions
regarding various evaluation criteria is given in Table I. The
following presents a detailed overview of the related work.
A. Gaze Estimation Accuracy & Setup Complexity
The majority of the existing work focus on improving the
estimation accuracy. There is no doubt that the accuracy is directly proportional with the setup complexity. 3D model-based
methods [4]–[7], [23] are widely preferred as they provide high
accuracy under generic head movements, owing to their explicit and fine 3D modeling. Most commercial high-accuracy
solutions rely on 3D model-based methods. However, they
have a significant drawback, that is to require fully-calibrated
systems. To acquire an accurate 3D eye model, a complex
setup that requires camera and geometric scene calibrations
(e.g., stereo, depth sensor) is needed. Alternatively, cross ratiobased [13]–[19] and regression-based methods [12], [24] have
mostly lower setup complexity and avoid setup calibrations.
However, they rely on approximations, and so, their performances are lower in accuracy and movement robustness.
On the contrary to feature-based methods, appearance-based
methods [25]–[28] simply require an ordinary camera. Yet,
they are restricted to particular applications due to their
limitations in accuracy and robustness.
The great majority of the existing eye trackers, regardless
of the gaze estimation method employed, rely on a singleview framework, which employs either a single-camera setup
[12], [14], [15], [26], or a multi-camera setup that is designed
to acquire 3D info (stereo, depth) [5], [10], [23], [29] or to
acquire high-resolution eye data using a pan-tilt unit [4], [7],
[13]. On the other hand, the efficacy of the multi-camera systems that perform multi-view tracking has not adequately been
investigated. In this regard, to the best of our knowledge, there
exists only one previous effort. [30] proposed a two-camera
setup mainly to obtain a wide observation area for a gazereactive signboard. Yet, the system was designed for a highly
coarse gaze tracking, which achieved ∼11◦ accuracy. Hence,
we are the first to exploit a multi-camera setup for multiview tracking so as to improve the estimation performance for
high-accuracy eye tracking [19]. In this paper, we extend our
previous work to also improve the tracking robustness under
real-world conditions by investigating various multi-camera
configurations as well as novel adaptive fusion mechanisms.
B. User Calibration
In addition to hardware setup calibration, user calibration
plays an important role in user experience and convenience.
User calibration is required for modelling the person-specific
eye parameters, which are crucial for the estimation bias
correction. The calibration quality improves, to a certain
extent, when the amount of calibration data increases. However, augmenting the data amount by increasing the number
of calibration points could be tedious and harms the user
experience. In this regard, the trade-off between the quality
and convenience of user calibration has been widely studied
in the literature. Significant advancements have been made,
for instance, better geometric eye models [21], [29], [31],
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
more effective bias correction models [14], [16], [18], [22],
and implicit calibration methods [32], [33] were developed.
C. Head Movement Robustness
3D model-based methods are theoretically more tolerant
to the changes in head pose and location due to explicit
parametrization of person-specific eye parameters. Yet, in
practice, they suffer from inaccuracy under large head movements. One of the main reasons is that most systems are faced
with the dilemma of trading off between the head movement
range and data resolution. In early efforts [4], [7], a wide
field-of-view (FoV) stereo system was employed to allow
free head movement together with one or more narrow FoV
stereo system to capture high-resolution eye images. These
systems were mostly interconnected through a pan-tilt unit
which mechanically reoriented the narrow FoV camera to
users’ eye. Despite enabling high accuracy and certain head
movement robustness, the use of a pan-tilt unit increased the
setup complexity and cost. Later, researchers avoided such
mechanical units and focused on introducing more robust
models, which eliminated the need for narrow FoV cameras.
For instance, [6] introduced a method that used the pupil
center and at least two glints, which were estimated from the
eye images captured by at least two cameras. Their system
achieved <1◦ accuracy by tolerating head movements in a
volume of 10×8×10 cm3 . In addition, [5] presented a single
camera non-stereo system that employed ray tracing rather
than depth from focus. Their system allowed an accurate
(<1◦ ) estimation in a volume of 14×12×20 cm3 . Recently,
[8] proposed a Kinect sensor-based technique, which used
a parametrized iris model. They reported 1.4−2.7◦ accuracy
error under head movements in a volume of 20×20×8 cm3 .
Regression-based methods indirectly model the eye physiology, geometry, and optical properties. When the user moves
away from the calibration position, the features non-linearly
change, therefore, the calibration mapping becomes less accurate causing inaccurate estimations. To address this, multiple
glints-based approaches have been suggested. [34] proposed to
use a second light source, which permitted differentiation of
head movement from eye rotation in the camera image. Using
two glints as points of reference and exploiting spatial symmetries, they proposed a spatially dynamic calibration method to
compensate for lateral head translation automatically. Later,
a thorough review of polynomial-based regression methods
using two glints was presented in [11]. They evaluated various
models using different pupil-glint vectors and polynomial
functions. In addition, [12] studied how binocular information can improve the accuracy and robustness against head
movements for the polynomial based systems using one or
two glints. Moreover, [24] suggested two calibration strategies
to reduce the errors caused by head movements. The results
of the experiments showed that both strategies achieved a
reduction in error by a factor of two for ±6 cm depth movements. From a different perspective, [10] proposed a stereobased system that achieved an ∼2◦ accuracy while allowing
for a significantly greater working volume (20×20×30 cm3 )
without using a chinrest. They estimated 3D optical eye axis
by directly applying triangulation techniques on the glints and
3
pupil center. They also suggested that 3D head pose can be
used to compensate for the bias caused by head movements.
However, the main drawback was that a multi-camera fullycalibrated stereo setup was required to obtain 3D information.
Cross ratio-based methods are sensitive mainly to the depth
movements. Various attempts have been made to enhance the
depth movement tolerance. Most of these focused on adapting
the user calibration to the changes in head movements. For
instance, [15] proposed dynamic calibration correction and
planarization of features, which achieved ∼0.5◦ accuracy
while tolerating up to ±12.5 cm depth changes. However,
their system required high-resolution (640×480 pixels) eye
images. Also, a chinrest was required to keep users’ eye
within camera’s FoV and to fix users’ head pose and location
during the experiments. [16] proposed a homography-based
calibration modeling with a binocular fixation constraint to
jointly estimate the homography matrix from both eyes. They
reported ∼0.6◦ using a much lower resolution while allowing
for ±5 cm head movements. A potential drawback of their
system was that the features from both eyes must be accurately
detected to compute the gaze, which constrains the estimation
availability due to the limited head pose allowance. Moreover, [17] proposed an adaptive homography calibration. They
learned an offline model on the simulated data by exploring
the relationship between the estimation bias and varying head
movements. They achieved promising results on the simulated
(±25 cm) and real data (±10 cm). An important limitation
in [16] and [17] is that they use a chinrest to fix the head
pose during the evaluations, similar to [15]. Although reporting
performances with chinrest leads to more stable results, it
causes the evaluations to discard the impact of continuous
head pose variations. Besides, it is impractical for real-world
applications and significantly harms user experience.
D. Eyewear (Glasses) Robustness
Eyewear robustness, particularly to glasses, has been a
challenging research problem since the reflection and refraction from glasses drastically obstruct the eye appearance and
features. As appearance-based methods [26]–[28] neither use
light sources nor rely directly on the detection of individual
gaze features, their performances are less affected by the
glasses in comparison with feature-based methods, which require explicit solutions. Unfortunately, glasses robustness has
largely been neglected by the majority of the previous work.
There exists only a limited number of attempts to address
the reflections and refraction on the glasses. [35] introduced a
robust pupil detection method by leveraging the bright-pupil
effect generated with a differential lighting scheme. He also
suggested a method for eliminating the reflections appearing
on the glasses. His method was successfully realized in [36]
for monitoring driver vigilance. In addition, [7] proposed a
dual illumination technique to avoid the reflections on the
glasses. When a specular reflection was detected, the system
deactivated the current illuminator and activated the alternative
illuminator on the opposite side, such that the reflection
can be avoided. Furthermore, [29], [37] demonstrated that
compensating for the refraction can increase the accuracy up
to ∼1◦ . Recently, [38] also simulated reflection and refraction
of glasses to study their impacts in connection with pupil and
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
4
Fig. 1: Overview of a single-camera system, which comprises of two gaze sensors, one for each eye.
User
Calibration
Single Camera
System
Single Camera
System
Single Camera
System
Bias
Correction
PoRleft
PoRright
Head pose
PoRleft
PoRright
Head pose
PoRleft
PoRright
Head pose
Gazing
Behaviour
Analysis
Adaptive Fusion
PoROverall
Fig. 2: Overview of the proposed multi-camera framework.
glint detection. In our work, from a different perspective, we
address the robustness to glasses by focusing on generating
and detecting more reliable gaze features. As the occlusion
and distortion of features on an eye appearance depends on
the relative positioning of a camera, light sources, and eye, we
propose to perform multi-view tracking to obtain alternative
eye appearances. Hence, the main benefit is that in case the
glasses effects obstruct the features from certain views, they
can still be recovered from alternative views.
III. P ROPOSED M ETHODOLOGY
Our methodology comprises of simultaneously operating
independent single-camera systems as illustrated in Fig. 2.
Each camera system firstly performs blink & gaze features
detection (Section III-A) followed by gaze estimation and user
calibration (Section III-B), as shown in Fig. 1. As the main
contribution, gaze outputs obtained from each eye view from
each camera are then fed into the proposed adaptive fusion
mechanism to output the overall PoR (Section III-C).
A. Blink & Gaze Features Detection
Our methodology starts with eye localization, in which the
existence of eyes is determined. We employ a robust non-rigid
face tracker based on supervised decent method (SDM) [39]
to localize and track the eyes. SDM method assumes that an
accurate final face shape with 66 landmarks can be estimated
with a cascade of regression models given an initial shape.
Once the shape is fitted accurately, we extract eye regions by
using the landmarks around eyes. Note that neither registration
nor alignment is required, i.e., no particular eye resolution
is ensured. We detect eye blinks by computing the ratio of
vertical eye opening to the eye width using the landmarks. If
there is no blink, we continue with detection and removal of
the specular reflections (glares) caused by glasses using wellknown image processing techniques. Next, gaze features, i.e.,
corneal reflections (glints) and pupil center, are localized.
For glint detection, we initially perform histogram equalization to improve the contrast on input eye region. We then apply
a thresholding to segment out the prospective glints. Here,
we utilize spatial adaptive thresholding to take into account
spatial variations in illumination. Instead of tuning a global
threshold value, spatial adaptive thresholding applies different
thresholds for small regions within the image, which leads to
more robust results under varying illumination. Morphological
opening and closing operations are then performed to get rid of
the small blobs caused by noise. In the resulting binary image,
we expect to find four blobs forming a trapezium since they
emerge from the reflections of four LEDs located around the
screen. Therefore, we perform connected component analysis
to determine the candidate glints. If there are four or more
candidate glints, we consider the shapes formed by any fourglints combination. The set of candidates whose convex hull
has the highest match with a template shape representing the
screen is considered as the final set of glints.
For pupil center detection, we follow a dark-pupil based
approach rather than a bright-pupil based one due to its improved robustness to glasses and the variations in illumination
and eye type [40]. More specifically, we first perform bilateral
filtering on input eye region with dark-pupil to smooth the
pupil region while still keeping the edges (pupil to iris) sharp.
We then equalize the histogram to enhance the contrast. We
approximate the average intensity within the pupil by the
surrounding regions of each glint and center of glints polygon.
We then remove the glints by filling them with the average
intensity. On the resulting image, we apply global thresholding
by considering the average intensity within the pupil. We then
invert the image to highlight the pupil blob. Nonetheless, a
few other blobs, which are as dark as the pupil region, such
as eye lashes, eye lids, shades, also remain in the binary image.
To distinguish the actual pupil region from the noisy blobs,
we perform morphological operations for the noise removal.
Among the remaining candidate blobs, we determine the final
pupil by considering the shape, size, and location of the blobs.
Its center of gravity is then used as the pupil center feature.
Further explanations and figures on aforementioned processes
can be found in [40].
B. Gaze Estimation & User Calibration
It is important to note that the proposed multi-view framework is independent of the gaze estimation algorithm used.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
Therefore, any gaze estimation technique described in Section II can be utilized within this framework. The optimal
method can be determined based on the application requirements, e.g., accuracy, robustness, setup complexity and flexibility. For instance, in our prototype system where the focus
is high-accuracy PoR estimation on desktop scenarios, we
employ a cross ratio-based method due to its particular advantages, such as enabling high-accuracy using an uncalibrated
and flexible setup.
The original method relies on the cross-ratio invariant of
the projective space [13]. More specifically, four light sources
are positioned around a screen to create glints on subject’s
cornea. The polygon formed by the glints on the cornea is
the projection of the screen. Another projection takes place
from the corneal plane to the image plane. As the virtual
tangent plane on the cornea has the same planar projective
transformation of the screen and image planes, the pupil center
on the image plane corresponds to the PoR on the screen plane.
Hence, the PoR can be computed by the equality of the crossratios on screen and camera image planes, as detailed in [40].
Nevertheless, a user calibration is essential to compensate for
the estimation bias caused by person-specific eye parameters,
such as the angular offset between the visual and optical
axis of the eyeball and the cornea radius and curvature. The
calibration is performed once, prior to the use of the system.
The users are asked to look at N calibration points on the
screen for K frames long. Person-specific bias correction,
F , can be learned by minimizing the distances between the
estimated gaze positions and corresponding calibration points
on the screen as follows:
N K
min ∑ ∑ kPi, j − F (Zi, j )k,
i
(1)
j
where Zi, j and Pi, j are the estimated PoRs on the screen and
the corresponding target calibration points, respectively.
In this work, we apply a regularized least-squares regression
based calibration technique [18]. The method has recently been
shown to have better modeling and generalization capabilities
than the state-of-the-art methods, owing to reduced model
parameters. It enables to model the bias more effectively,
particularly when the calibration data is limited in size and
quality. Therefore, it facilitates tracking with low-resolution
data and requires minimal user effort.
C. Adaptive Fusion Scheme
The proposed multi-view framework, which consists of
individual single-camera trackers, is designed to empower
a robust tracking under challenging conditions. Within this
framework, each tracker simultaneously estimates the gaze for
each eye (aka gaze sensor). In each frame, two distinct gaze
outputs can be computed per camera. Consequently, in a multicamera setup with C cameras, up to 2C gaze outputs can be
generated per frame. The overall PoR can then be computed by
the fusion of available gaze outputs obtained from all sensors.
Here, we propose an adaptive fusion mechanism to effectively
combine the gaze outputs towards achieving a higher overall
estimation accuracy and robustness.
5
We have investigated several algorithms to perform the
fusion. Among these, the most straightforward one is to
average all gaze outputs. Despite its simplicity, fusion by
simple averaging results in a significant improvement in estimation accuracy in comparison to the best of single-view
trackers, particularly when the majority of the sensors produce
reliable gaze outputs. In such cases, simple averaging provides
a more accurate and consistent overall estimation through
smoothing out the arbitrary noise. On the other hand, under
more challenging scenarios, in which a higher variance exists
among available gaze outputs, fusion by simple averaging is
far from the optimal fusion. To this effect, we propose to
combine the available gaze outputs in a weighted manner,
where the weights are determined according to the estimation
reliability of each sensor. Here, there is several factors that
affect the reliability, such as camera’s viewing angle, gaze
point location on the screen, eyewear effects, person-specific
gaze behaviours. Fig. 3 shows sample eye appearances captured from different camera views when users gaze at different
regions on the screen. While some views permit reliable
gaze features, some others do not even contain any available
features. The views capturing the best eye appearances continuously vary when users gaze at different regions on the
screen. In addition, when users wear glasses, reflections and
refraction occur on the glasses. When these effects distort or
overlap with the features, the estimation becomes unavailable
from that particular view. On the other hand, as there exists
simultaneously captured several other views, the features can
still be recovered from some of these views. This, in fact,
constitutes an important benefit of the multi-view approach in
comparison to conventional single-view approach, employed
by the majority of the previous work. Hence, an effective
fusion that accounts for the estimation reliability of individual
gaze outputs can significantly improve the overall estimation
accuracy and robustness.
The adaptive fusion is formulated as follows:
z∗
∑ ∑ wec
c
e
=
∑ ∑ zec wec ,
c
(2)
e
= 1, e ∈ {L, R},
c ∈ {1, 2, ..,C},
where z∗ is the overall PoR and, wRc and wLc are the weights
for the right and left eye’s gaze outputs from the cth camera, respectively. In case any of the gaze outputs cannot be
computed, then the weight of the missing one is set to zero.
We do not report an overall PoR in case there is no available
gaze outputs for a given frame. To determine the weights, we
propose two algorithms, namely, head pose-based fusion and
person-specific gazing behaviour-based fusion:
1) Head pose-based fusion: As can be depicted in Fig. 3
and psycho-visually evidenced in [41], prior to the fixation,
most users initially perform head rotation to find the most
comfortable viewing angle. Therefore, when users gaze at
different target points, the eye appearances continuously vary.
Here, the estimation accuracy is strongly correlated with the
quality of the eye appearance and gaze features, which relies
on user’s head pose relative to each camera view. For instance,
when users gaze at the upper left corner of the screen, the left
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
6
Fig. 3: Example eye appearances from three different camera views: (left) left side camera, (middle) bottom camera and (right)
right side camera while gazing at a stimulus point displayed on (top) upper-left and (bottom) upper-right region of the screen.
camera system often generates more accurate estimations than
the others. The main reason is obviously that the relative head
pose is more frontal from left camera’s view, and consequently
the corresponding eye appearances facilitate a more reliable
feature detection. Hence, our first adaptive fusion relies on
relative head pose angles estimated with respect to each
camera. We assign the weights inversely proportional to the
angles, as follows:
λce
=
wec
=
αmax − |αce |
,
αmax
λce
,
∑c ∑e λce
(3)
(4)
where αc is the head pose yaw angle, αmax is the maximum
angle allowed, e.g., 45◦ . We calculate head pose angles using
the landmarks obtained by the face tracker and point distribution model [42]. Both eyes are assigned with the same weights.
Normalization is then performed using Eq. (4) prior to fusion.
Note also that we investigated an alternative weighting
approach in our previous effort [19]. Instead of calculating
head pose angles, we first calculate an initial PoR using simple
averaging. We then iteratively refine the initial estimation
by weighting the cameras with respect to their distances to
the estimations. Although this approach performs an effective
fusion, it has two drawbacks: first, it requires camera locations
to be known to compute the distances, and second, its performance is affected by the quality of the initial PoR. A poor
initial PoR estimation leads to a less effective weighting.
2) Gazing behaviour-based fusion: Although head posebased weighting works well for most users, it does not take
person-specific gazing behaviors into account, and consequently, it may experience a performance drop when a user
has a particular gazing behaviour. For instance, although the
majority of the users perform head rotation prior to fixation to
have a comfortable viewing angle (frontal eyeball pose), some
users do not perform any head movements but rather rotate
their eye balls (non-frontal eyeball pose). In addition, head
pose-based approach weighs cameras rather than the eyes.
However, some users’ vision may rely more on one particular
eye than the other due to eye dominance or a physiological
reason (e.g., lazy eye, strabism). For such users, assigning
equal weights to both eyes may result in a low estimation
performance. Hence, we propose to determine person-specific
weights for each eye independently through leveraging user’s
calibration data. During user calibration, we generate fusion
weight maps in addition to learning the user calibration model.
Once the weight maps are obtained, our algorithm performs a
weighted averaging of individual gaze outputs as follows:
z∗
∑ ∑ Mec (x, y)
c
e
=
∑ ∑ zec Mec (zec .x, zec .y),
c
(5)
e
= 1, e ∈ {L, R},
c ∈ {1, 2, ..,C},
where z∗ is the overall PoR, zec are initial gaze outputs estiL
mated using simple averaging, and MR
c and Mc are the weight
maps of right and left eye of the cth camera, respectively.
For generating the weight maps (Mec ), we extract various
statistics that are relevant to the estimation reliability (e.g.,
accuracy, precision, availability, gain) from the calibration
data. For instance, as perhaps the most relevant and effective
indicator, we calculate the estimation accuracy of each sensor
(eye) on each calibration point. If a sensor’s calibration accuracy around a point is consistently higher than the others,
that sensor’s estimation performance during testing is expected
to be more reliable around the same point. Thus, higher
weights are assigned to the sensors with better calibration
performances for certain regions. More specifically, to calculate the weighting indicators (e.g., accec,k ) for each point, after
learning the calibration model on the whole calibration data,
we apply the learned model on the very same data. Then,
we compute the aforementioned statistics on the calibrated
samples, such as the accuracy by measuring how close the
calibrated samples are to their corresponding target points.
We perform this process for each calibration point of each
sensor independently, and obtain 2C values for each calibration
point. We then normalize these accuracy values to compute the
sensor weights (wec,k ) for each calibration point:
wec,k
Wec
accec,k
,
(6)
∑c ∑e ∑k accec,k
= {wec,k | e ∈ {L, R}, c ∈ {1, 2, ..,C}, 1 ≤ k ≤ K},
=
where K is the number of calibration points. Lastly, we
interpolate and extrapolate the weight set (Wec ) over the
whole screen to generate the weight maps (Mec ). Sample
generated weight maps from a three-camera setup are shown
in Fig. 4. In this paper, we use the estimation accuracy and
availability statistics as the weighting indicators. Nevertheless,
other alternatives can also be employed towards more robustly
determining the sensor weights. For example, the estimation
precision, which is the ability to reliably reproduce the same
estimation for a target calibration point, or the histogram of
the best performing sensor, which stores the information about
how often each sensor achieves the best estimation for a target
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
(a) ML
left cam
(b) MR
left cam
(c) ML
bottom cam
(d) MR
bottom cam
7
(e) ML
right cam
(f) MR
right cam
Fig. 4: Sample generated weight maps based on the calibration accuracy and gaze availability in a three-camera setup.
calibration point, could provide complementary evidence. We
plan to investigate such alternative indicators in our future
work. On the other hand, the main drawbacks of this method
are: first, it may be sensitive to large head movements since the
weights are estimated according to the calibration position, and
second, an initial estimation using simple averaging is required
to localize the weights on the generated maps.
IV. E VALUATION ON S IMULATED DATA
We conduct extensive simulations to primarily investigate
and understand how the overall tracking performance is affected when the number of cameras is increased in various
configurations. In real-world settings, the number of cameras
to be employed for real-time eye tracking is limited due to cost
data bandwidth constraints. Therefore, we start our evaluations
on the simulated data to analyze the efficacy and limits of the
proposed framework. The tracking performance is measured
as the gaze accuracy and estimation availability. The accuracy
is defined as the average displacement in degrees of visual
angle (◦ ) between the stimuli points and estimated PoRs, using
all raw samples, i.e., neither temporal smoothing nor postprocessing is applied. The availability is % of samples, which
the system is able to compute a PoR during the evaluation. In
other words, it indicates the system’s working volume.
A. Simulation Setup
Simulation data is generated using an open-source software
framework [43]. The simulator enables detailed modeling of
different components of the hardware setup and an eye in 3D,
and provides a realistic simulation framework. Nonetheless,
simulation of non-spherical cornea, eyelid occlusions, eyewear
effects, lens or sensor distortions, are currently not possible.
To simulate the impact of increasing the number of cameras,
we setup two configurations: i) single-view tracking by placing
multiple cameras densely at the bottom of a screen (case 0)
and ii) multi-view tracking by placing them uniformly around
a screen (case 1), as visualized in Fig. 5. We simulate an eye
using the typical eye parameters listed in [29]. In addition,
we simulate a similar environment that we have in the user
experiments, i.e., 24-inch screen, 4 light sources, cameras
with 1280×1024 pixels resolution, and lenses with 8mm focal
Fig. 5: Single-view (case 0) and multi-view (case 1) setups.
length (diagonal FoV=58◦ ) to allow for large head movements.
We acquire calibration and test data using the simulated environment. For the calibration, we generated data when gazing
at 9 uniformly distributed target stimuli points on the screen,
whereas for the test data acquisition, we randomly generate 18
test points in order to avoid reporting over-optimistic results
due to overfitting on the calibration point locations. The test
points are displayed in a 3×3 grid with 2 points per region
to cover the whole screen. In addition, to simulate realistic
test conditions, we alter the noise level to examine the impact
of noise-free and noisy data. For each point, we collect 100
samples, and introduce uniformly distributed feature position
errors with a maximum magnitude of 0.4 pixels per feature
(noise level ∈ {0, 0.1, 0.2, 0.4}).
The simulations are performed under two different scenarios, namely, Stationary Head (SH) and Moving Head (MH). In
SH scenario, the eye is located 60 cm to the screen and kept
at the same position during the experiments. Whereas, in MH
scenario, the eye location is changed along three directions, X,
Y, and Z as shown in Fig. 8. In both scenarios, the calibration
is performed at the default head position (0, 20, 60) cm.
B. Results on Stationary Head (SH) Scenario
In SH scenario, the main emphasis is given to the impact
of increased number of cameras on estimation accuracy as
there is no head movement. Fig. 6 shows the obtained results
under various setup configurations while also altering feature
detection noise amount to understand the theoretical and
practical impact of setup configurations. In case 0 (singleview tracking), when no noise is introduced, increasing the
number of cameras, even up to 25 cameras, does not provide
any estimation accuracy improvement (see Fig. 6a). Contrarily,
when a significant amount of noise is introduced, the more
cameras the system employs, the higher accuracies it achieves
since the noisy outputs are smoothed out.
The simulation results shown in Fig. 6b indicate that not
only the number of cameras, but also their configuration is
crucial for improving the tracking performance. For instance,
when there is no noise, a 3-camera multi-view setup (case
1) outperforms a 25-camera single-view setup (case 0). In
addition, the results suggest that when higher levels of noise
(e.g., 0.4) are introduced, the number of cameras has more
impact than the configuration of cameras since more cameras
can better filter the noise out. Under the real-world noise level
(∼0.2), the camera configuration is undoubtedly more effective
than the number of cameras. In other words, the multi-view
configuration always outperforms the single-view one when
employing the same number of cameras. Furthermore, Fig. 6c
shows the comparison of non-adaptive and adaptive fusion
mechanisms. The results indicate that the proposed adaptive
mechanisms perform better than the simple averaging.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
8
Simulations − Scenario SH − Case 0
Simulations − Scenario SH − Case 1
2.1
1.8
1.5
1.2
0.9
0.6
0.3
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Simulations − Scenario SH
2.4
2.1
1.8
1.5
1.2
0.9
0.6
0.3
0
0.55
mean accuracy error (°)
noise level: 0 (no noise)
noise level: 0.1
noise level: 0.2
noise level: 0.4
mean accuracy error (°)
mean accuracy error (°)
2.4
1 camera
3 cameras (case 0)
3 cameras (case 1)
9 cameras (case 0)
9 cameras (case 1)
25 cameras (case 0)
25 cameras (case 1)
36 cameras (case 1)
0
number of cameras
(a) Case 0 (single-view)
0.1
0.2
noise levels
25 cameras − simple averaging
0.5 25 cameras − head pose−based fusion
25 cameras − gazing behaviour−based fusion
0.45
0.4
0.35
0.4
0
(b) Case 0 vs Case 1
0.1
0.2
noise levels
0.4
(c) Case 1 (multi-view)
−30
−15
0
15
30
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
head movement
across
(cm)
Simulations
− Scenario
MHX −axis
Availability
mean accuracy error (°)
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
mean accuracy error (°)
mean accuracy error (°)
Fig. 6: SH scenario with varying feature detection noise levels. The impact of increasing number of cameras with (a) single-view
and (b) multi-view
setups, and (c) a comparison of the Simulations
investigated
fusion methods for multi-viewSimulations
tracking.
Simulations − Scenario MH
− Scenario MH
− Scenario MH
−10
0
10
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
−20
−10
0
10
20
Simulations
− Scenario
MH
− Availability
head movement
across
Z axis
(cm)
head movement across Y axis (cm)
Simulations
− Scenario MH − Availability
100
80
60
40
20
0
−50 −40 −30 −20 −10 0 10 20 30 40
head movement across X axis (cm)
50
mean availability (%)
100
mean availability (%)
mean availability (%)
100
80
60
40
20
0
−50 −40 −30 −20 −10 0 10 20 30 40
head movement across Y axis (cm)
50
80
60
1 camera
3 cameras (case 0)
3 cameras (case 1)
25 cameras (case 1)
36 cameras (case 1)
40
20
0
−50
−40
−30
−20
−10
0
10
20
30
head movement across Z axis (cm)
40
50
Fig. 7: MH scenario with real-world (0.2) noise level. The impact of number of cameras and their configurations (case 0 and 1)
on the head movement robustness (top row) and gaze availability (bottom row) when user moves from the default calibration
position (0, 20, 60) along X, Y, and Z directions. This figure best viewed in color. All subfigures have the same legend.
C. Simulation Results on Moving Head (MH) Scenario
In MH scenario, we examine the impact of single- and
multi-view setups, particularly on the estimation availability
and head movement robustness along X, Y, and Z directions.
For this scenario, the virtual eye is calibrated at the default
position (0, 20, 60) and real-world noise level (0.2) is introduced. The eye then is moved to various locations along three
directions as shown in Fig. 8 and the tracking is performed
in these locations using the learned calibration at the default
position. As depicted in Fig. 7 (top row), the tracking is
highly robust to head movements along X (horizontal) and Y
(vertical) directions. In fact, even the single-camera system
is highly tolerant owing to the employed gaze estimation
method and user calibration technique [18]. In these cases,
Fig. 8: Simulation setup. Calibration is at the black circle.
increasing the number of cameras simply enhances the overall
estimation accuracy. On the other hand, the robustness to
head movements along Z axis (depth translations) is very
challenging for cross ratio-based systems due to insufficient
bias correction. As the calibration is learned as an offset at
a fixed head location, the learned offset does not sufficiently
compensate for the bias when the user moves away from the
calibrated position, especially along Z axis. Therefore, such
movements cause a significant decay in estimation accuracy
for a single-camera system. Nonetheless, as clearly depicted
from the line slopes in Fig. 7 (top-right), multi-view setups
(case 1) yields a significant tolerance compared to singlecamera or single-view (case 0) setups. For instance, 3-camera
multi-view configuration improves the accuracy by 43% and
35% in comparison to single-camera and 3-camera single-view
configuration, respectively.
Furthermore, an important benefit of the multi-view tracking
is the increased estimation availability. Fig. 7 (bottom row)
demonstrates the impact of setup configurations on the gaze
availability, in %, when the user moves along X, Y, and Z
directions. The results clearly show that multi-view setups allow for significantly larger head movements (working volume)
in all three directions in comparison to single-view setups.
For instance, 3-camera multi-view setup (case 1) provides an
additional ±15 cm and ±10 cm head movement tolerance
along X and Y directions, respectively. The reason is that each
camera has a different FoV, and consequently, the overall FoV
increases with the fusion of all FoVs. Note also that increasing
the number of cameras from 3 to 36 does not drastically
improve the availability as their FoVs starts to overlap.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
9
TABLE II: Experimental configurations.
Fig. 9: Proposed 3-camera prototype setup.
V. E VALUATION ON U SER E XPERIMENTS
This section describes the evaluation of our approach on
real-world data obtained through user experiments.
A. Hardware Setup
Our prototype setup consists of three PointGrey Flea3
monochrome cameras, four groups of near-infrared (NIR)
LEDs for the illumination, and a controller unit for the
synchronization. Each camera has an image resolution of
1280×1024, and is equipped with an 8 mm manual focus lens
(diagonal FoV=58◦ ). The cameras are installed on a frame
around a 24-inch monitor as shown in Fig. 9. One camera
is located slightly below the screen, whereas the other two
are placed on the left and right sides of the screen. Note
that the setup corresponds to the 3-camera multi-view (case
1) configuration described in the simulations. In order to
create the glints, 850nm-wavelength LEDs are placed on the
corners of the monitor. A micro-controller is programmed to
synchronize all cameras, so that the images are simultaneously
captured from all cameras at 30 fps. In addition, we optimize
the light emissions regarding the eye safety by synchronizing
the cameras’ shutters with the emission duration of LEDs.
B. Dataset & Experimental Protocol
A series of user experiments were conducted to comprehensively evaluate the proposed framework regarding the estimation accuracy, availability, robustness against head movements,
varying illumination, use of eyewear, and between-subject
variations. In total, 20 subjects, most of whom had no previous
experience with any gaze tracking system, participated in our
user experiments. 11 participants did not have any eyewear,
while 5 and 4 participants wore glasses and contact lenses, respectively. The participants are from diverse origin. Therefore,
eye shapes and appearances exhibit a large variability.
Each participant was asked to follow 8 different experiments
as described in Table II. Experiment #2 being the default protocol, in the first three experiments, we analyzed the system’s
tolerance to varying ambient illumination conditions, i.e., sunlight, darkness, indoor light. The remaining five experiments
were designed to evaluate the system’s robustness to head
movements. Four of them were conventional experiments, in
which the subjects were asked to move along X (horizontal)
and Z (depth) axes. The remaining one stood for a novel scenario, in which the subjects were asked to continuously move
their head while still fixating on the displayed gaze points. The
goal of this experiment was to analyze the system’s sensitivity
Exp.
Lighting
Head (X, Z)
Experimental Variable
0
1
2
3
4
5
6
7
sunlight
darkness
indoor
indoor
indoor
indoor
indoor
indoor
0
0
0
0
0
0
+15
-15
illumination
illumination
illumination
continuous head movements (HM)
-10 cm HM along Z axis
+10 cm HM along Z axis
+15 cm HM along X axis
-15 cm HM along X axis
60
60
60
60
50
70
60
60
to continuous head movements, head pose changes, and slight
head translations during the fixations. In fact, this represents
better the real-world scenarios, e.g., free-head gazing while
listening music or talking on the phone. As our evaluation targeted natural human-computer interaction, we aimed to collect
the ground truth data as natural for the subjects as possible.
For instance, chinrest was avoided to keep the subject’s head
still and to keep the eye within the cameras’ FoV to capture
high-resolution eye data, as frequently performed by previous
work. In addition, the subjects were asked to gaze at the target
stimuli points in a natural and comfortable way. As a result, the
subjects had different head-pose and eye-pose characteristics,
facial expressions, and heights (along Y axis) while gazing.
Data acquisition and performance evaluations are done similar
to the simulations described in Section IV-A. The default
user-to-screen distance is set to 60 cm and user calibration
is performed only at this distance. The learned calibration is
then applied during testing for all configurations. Head pose
statistics, sample video frames from each experiment, and
an example video (Exp #2 vs Exp #3) are provided in the
supplementary material.
C. Results
Our framework starts with face tracking on the captured
frames, in which we extract eye regions of size ∼90×50
pixels. Feature detection is then performed to detect the pupil
center and four glints. The size of the polygon formed by the
glints is ∼9×5 pixels. Next, we apply cross ratio-based gaze
estimation with the detected gaze features to calculate raw
PoRs. We then apply the learned calibration models on raw
gaze outputs to compensate for the person-specific bias. Lastly,
calibrated PoRs obtained from each sensor are combined using
the adaptive fusion mechanism to output an overall PoR. In
the following subsections, we present and discuss the results
of various experiments.
1) Single-view vs Multi-view Tracking: Our first analysis
emphasizes on the benefits of multi-view tracking (case 1)
over single-view tracking (case 0) as implemented by the
majority of the existing trackers. In this regard, we conducted
TABLE III: Comparison of single- and multi-view setups.
Setup configuration
Single-cam left eye only
Single-cam right eye only
Single-cam both eyes
Multi-cam (case 0)
Multi-cam (case 1) pose-based best cam selection
Multi-cam (case 1) simple averaging
Multi-cam (case 1) head pose-based fusion
Multi-cam (case 1) gazing behavior-based fusion
Eye
Data
1
1
max
max
max
max
max
max
2
6
2
6
6
6
Estimation
(◦ )
(%)
1.4
1.35
1.25
1.04
1.17
0.89
0.76
0.74
77.6
69.6
93.4
97.2
100
100
100
100
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
Exp 7 (−15 cm)
Exp 2
Exp 6 (+15 cm)
(a) Horizontal movements (along X axis)
2.4
2.2
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
mean accuracy error (°)
2.4
2.2
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
10
User Experiments − HM Robustness (Continuous Movements)
User Experiments − HM Robustness (along Z)
mean accuracy error (°)
mean accuracy error (°)
User Experiments − HM Robustness (along X)
Exp 4 (−10 cm)
Exp 2
Exp 5 (+10 cm)
(b) Depth movements (along Z axis)
2.4
2.2
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
Single−camera (single eye)
Single−cam (both eyes)
Multi−view (head pose−based fusion)
Multi−view (gazing behaviour−based fusion)
Exp 3
Exp 2
(c) Continuous movements
Fig. 10: Performance comparison of single- and multi-view setups under varying head movement scenarios.
2) Head Movement Robustness: To evaluate our framework’s robustness to head movements, we analyze the results
on experiments #2-7 on all subjects. Experiments #2, #6,
and #7 account for the horizontal movements (along X axis),
whereas experiments #2, #4 and #5 account for the depth
movements (along Z axis). Note that vertical movements
(along Y axis) are not explicitly experimented as the subjects
were asked to freely adjust their heights for convenience.
Moreover, we introduced a new experimental scenario (experiment #3), in which the users were asked to perform continuous
head pose/location changes while still fixating on the target
points. The purpose of this experiment is to measure the
system’s sensitivity to sudden arbitrary changes during the
user interaction, which may frequently occur in real-world
conditions. Fig. 10 illustrates the results achieved on these
experiments and their cross comparisons.
For horizontal head movement robustness, the results
(Fig. 10a) are highly in line with the simulation results (Fig. 7),
such that the system is highly tolerant (1◦ vs 1.1◦ ) to head
movements along X axis up to ±15 cm movements. On
the other hand, along Z axis (depth translations), the results
partially differ from the simulation results. In simulations
(Fig. 7), the estimation accuracy is shown to be negatively
affected by the depth movements. The same result holds
for the user experiments. However, we also observe that the
system’s performance gets worse when users move away from
the screen, which contradicts the simulation results. In fact,
the main reason relates to the current hardware setup, which
employs manual focus lenses. Despite the aperture adjustments
to obtain a larger depth-of-field, out-of-focus still occurs when
user’s depth varies from the default position. This causes
blurry appearances, and consequently less precisely detected
features. In addition, the eye image resolution gets significantly lower when the user moves away from the camera,
which causes the features to be detected less accurately. Yet,
in overall, the multi-view framework provides more robustness, by about 25% in accuracy and 10% in availability,
to depth translations compared to the single-camera system.
Furthermore, continuous head movements results are shown in
Fig. 10c. The results indicate that the framework, as expected,
experiences an accuracy drop, yet it continues to output PoRs
with an acceptable accuracy (∼1.4◦ ) under such a challenging
scenario. As this new experimental scenario constitutes an
essential use case in real-world eye tracking, we recommend
future efforts to consider it in their evaluations.
3) Illumination Robustness: Fig. 11 illustrates the proposed
framework’s robustness under sunlight, darkness, and indoor
lighting. The results indicate that ambient illumination variations do not significantly influence the estimation performance.
As the suggested framework operates under active (NIR) illumination, the system is implicitly more robust to illumination
variations than natural light-based systems. For our system,
the robustness is in practice more related to how robust the
employed feature detection algorithms are to the changes in
features when the illumination varies (Fig. 12). In this respect,
User Experiments − Illumination Robustness
mean accuracy error (°)
experiments on a subset of our dataset (3 subjects, one subject
per eyewear category) using a single-camera, 3-camera singleview (case 0), and 3-camera multi-view (case 1) setups. Mean
estimation accuracy errors and availabilities obtained using
these setups on experiment #2 are shown in Table III. The
results are clearly in line with the findings of the simulations
(see Fig. 6), and demonstrate the efficacy of the proposed
multi-view framework in terms of both estimation accuracy
and availability. Firstly, significant accuracy improvements, by
about 41% and 29%, are achieved using the proposed 3-camera
multi-view setup in comparison to the conventional singlecamera setup (using both eyes) and 3-camera single-view
setup, respectively. In addition, the estimation availability is
also increased. Yet, the availability analysis is more interesting
when considering head movements and eyewear robustness in
Sections V-C2 and V-C4. Furthermore, the results show the
impacts of the adaptive fusion mechanisms. Although a simple
averaging standalone achieves a significant performance improvement, employing the proposed adaptive fusion algorithms
further enhances the accuracy.
2.5
2.3
2.1
1.9
1.7
1.5
1.3
1.1
0.9
0.7
0.5
Single−camera (single eye)
Single−cam (both eyes)
Multi−view (head pose−based fusion)
Multi−view (gazing behaviour−based fusion)
Exp 0
Exp 1
Exp 2
Fig. 11: Performance comparison under varying illumination.
Fig. 12: Sample appearances of eye and gaze features (glints
and pupil) under varying illumination conditions: (left) sunlight, (center) darkness, and (right) indoor lighting.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
11
TABLE IV: Performance comparison on the same subject.
Eyewear
Fig. 13: Sample impacts of glasses on eye appearance.
we proposed illumination-robust feature detection algorithms,
as described in detail in [40]. We also note that our results
under indoor lighting slightly outperforms the others because
the feature detection is mainly optimized for this scenario.
4) Eyewear Robustness: Considering that about 30% of
young adults and more than half of elders in industrial nations
need eyewear [44], any intolerance to glasses or contact lenses
undoubtedly harms the user experience. Still, it is undoubtedly
one of the most challenging issues in eye tracking. Unfortunately, it has been neglected by the great majority of the
previous efforts. The main challenges stem from the reflection
and refraction effects on the glasses, which can significantly
affect the tracking performance. Example glasses impacts on
eye appearance, such as distorted features due to the refraction
and coating, lost features due to the reflection, challenging
feature detection due to multiple reflections, which were encountered during the user experiments can be seen in Fig. 13.
As some of the impacts are unrecoverable, conventional singleview approaches are likely to fail under such circumstances.
On the other hand, the proposed multi-view approach leverages
various eye appearances per frame, in such a way to more
reliably detect the features from various views.
We evaluate the efficacy of our proposed method with two
separate analysis. First, we categorize the subjects into 4
groups according to their eyewear and vision quality such as
the ones who wear glasses, who wear contact lenses, who do
not wear glasses, and who have perfect vision. Fig. 14 shows
the performance comparison across these groups. The results
clearly depict the improvements achieved using the multi-view
setup for both the generic scenario (experiment #2) and over
all scenarios (experiments #0-#7). Among all groups, the best
performance (∼0.8◦ ) is obtained on the subjects with perfect
vision (6 subjects) and contact lenses (4 subjects). As the
group who do not wear glasses (15 subjects) includes some
subjects with lower vision quality (who actually needs slight
vision correction), a small accuracy drop is observed. Lastly,
the subjects with glasses (5 subjects) achieves a lower accuracy
(1.38◦ with 91.9% availability) in comparison with the other
groups. However, the performance improvement, by about 0.6◦
and 10%, compared to the single-camera setup highlights the
benefits of the multi-view approach.
Furthermore, to eliminate between-subject variations, we
Contacts
Glasses
Single-cam view
Exp 2
Exp [0-7]
(◦ )
(%)
(◦ )
(%)
0.99
2.1
96.1
84
1.18
2.31
95.1
82
Multi-view
Exp 2
Exp [0-7]
(◦ )
(%)
(◦ )
(%)
0.76
1.08
100
100
0.97
1.53
99.9
99
compare the tracking performance on the same subject, who
completed the user experiments firstly by wearing glasses and
then once again by wearing contact lenses. The results shown
in Table IV shows the efficacy of the multi-view system over
the single-view one. For glasses, it provides a substantial
improvement in accuracy by about 50% and 40% for the
generic scenario and all scenarios, respectively. In addition,
it brings ∼17% enhancement in estimation availability. For
contact lenses, single-camera system standalone yields a high
accuracy and availability. Yet, a multi-view system further
enhances the accuracy and availability.
5) Eye Type Robustness: Variations in eye type, e.g., iris
color, eye shape, pupil response, may significantly affect the
performance of eye trackers [45]. As our dataset contains eye
type variations (Fig. 15), we evaluate the proposed framework’s robustness to iris color and eye shape. Firstly, since
the iris color has a great influence on both the pupil size and
opening of eyelids when exposed to various illumination conditions, we categorized the subjects into two groups according
to the iris color such as dark-eyes (10 subjects) and lighteyes (10 subjects). On average over all experiments, darkand light-eyes groups achieve 1.15◦ with 96.2% availability
and 1.44◦ with 93.3% performances, respectively. However,
the results may be biased towards the dark-eyes group since
most of its subjects do not wear glasses. An interesting result
is that the performance difference between the light-eyes
(1.33◦ ) and dark-eyes groups (0.92◦ ) is especially large under
sunlight (experiment #0). The reason is that the pupil size and
eye opening are more significantly affected for light-eyes in
comparison with the dark-eyes due to their higher sensitivity
to the sunlight. Table S-II in the supplementary information
shows average estimation accuracy and estimation availability
in detail for each experiment in various categories.
It is also important to note that the pupil detection method
has an influence on the robustness to eye color variations.
As mentioned earlier and discussed in detail in [40], brightpupil based method is frequently employed by the previous
work as the feature detection is simpler compared to darkpupil based one. However, in our preliminary experiments,
mean accuracy error (°)
User Experiments − Eye Wear Robustness
2.4
2.2
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
Single−cam (Exp [0−7])
Single−cam (Exp 2)
Multi−view (Exp [0−7])
Multi−view (Exp 2)
w/ glasses
w/o glasses
w/ contacts w/ perfect vision
Fig. 14: Performance comparison of varying eyewear groups.
(a)
(b)
(c)
(d)
Fig. 15: Sample eye appearances: (a) Asian dark eyes without
glasses, (b) Asian dark eyes with glasses, (c) Caucasian dark
eyes without glasses, (d) Caucasian dark eyes with glasses.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
we observed that dark-pupil based feature detection is less
sensitive to the variations in eye color. In bright-pupil based
method, the accuracy of the pupil detection heavily relies on
the pupil response (brightness), which is highly affected by
users’ momentary pupil size that varies according to the eye
color, ethnicity, and ambient illumination. Therefore, in the
final framework, we employ dark-pupil based feature detection
for becoming less sensitive to eye type and illumination
factors. In our feature work, we aim to give a more structured
quantitative comparison on pupil detection methods.
Furthermore, we categorized the subjects by their eye shape
into two groups: Asian eyes (2 subjects) and non-Asian eyes
(18 subjects) to analyze the impact of the eye shape. Our
results show that Asian eyes (1.58◦ with 93.35% availability) perform worse than non-Asian eyes (1.25◦ with 97.7%
availability). Yet, the system can still accurately estimate the
gaze for our Asian subjects. The decrease in the availability
may indicate that the feature detection for them might be
more challenging due to the eye shape. Nevertheless, it is
difficult to make a strong conclusion as the two sets are
imbalanced. In addition, we note that there is a significant
variation across Asian eyes including round, narrow, almond,
hooded, triangular, prominent, or deep-set eye shapes [46]. In
addition, the eyes can be a single eyelid, low/incomplete eyelid
crease, and double eyelid. For some of these eye shapes (e.g.,
narrow, hooded), eye tracking could be highly challenging
as the creation and detection of the gaze features could be
exigent. Our dataset currently do not contain sufficient eye
type variation. Hence, we plan to recruit a larger number of
Asian subjects and increase the variation in eye type to obtain
a more reliable analysis.
D. Real-time Implementation
The proposed gaze estimation system and methods are
developed in C++1 . OpenCV library is mainly used for image
processing and computer vision algorithms. Localization of
facial landmarks were performed using a supervised descent
method (SDM)-based face tracker. Furthermore, to achieve
real-time tracking performance, OpenMP application programming interface was utilized for the parallellization of our
library implementation. The computational complexity of the
system is lower than 3D model-based methods as the gaze
estimation relies on perspective geometry transformations and
computation of cross ratios. This enables to achieve a real-time
implementation without requiring any particular performance
optimization. In our implementation, the most computationally
expensive process is face tracking. Cross ratio-based gaze
estimation on both eyes, user calibration, and adaptive fusion processes require much lower computational effort. For
instance, these three processes take only ∼8 ms on a PC
with Intel i7 3.2GHz processor, whereas face tracking solely
takes ∼24 ms. Our current three-camera prototype system can
simultaneously output PoRs for both eyes as well as an overall
PoR at ∼30 fps with a mean estimation accuracy error of
∼1◦ of visual angle. Yet, we note that there is much room for
computationally improving our implementation to reach higher
frame rates. As future work, we plan to replace the current
1 https://lts5www.epfl.ch/eyetracking
12
computationally expensive face tracker with a faster one, such
as local binary features (LBF)-based face tracking [48], one
millisecond face alignment with an ensemble of regression
trees [49]. In addition, we plan to further optimize feature
detection and adaptive fusion implementations.
VI. D ISCUSSION
Future directions in eye tracking research, towards becoming a pervasive technology, should not only focus on achieving
high estimation accuracies, but also on having robustness
against real-world settings such as natural head pose changes,
large head movements, varying illumination conditions, use
of eyewear, and between-subject eye type variations. Besides,
having a convenient user calibration, flexible hardware setup,
minimal setup calibration, low complexity and cost should
be taken into consideration as important evaluation criteria.
In this regard, in Section II, we describe various eye tracking techniques, analyze their pros and cons with respect to
each other, and discuss whether they satisfy some of the
aforementioned criteria. Therefore, the best, in other words,
the most appropriate, approach depends on the application
requirements. In this work, we mainly target eye tracking
scenarios that require high-accuracy (∼1◦ ) estimation and
robustness, e.g., gaze-based controlling, typing and navigation.
To achieve our accuracy and robustness goals, we design
a novel multi-camera framework, which tracks users’ gaze
simultaneously from various views, and then combines the
acquired gaze information from all sensors using an adaptive
fusion mechanism to output an overall PoR. In comparison
with conventional single-view approaches, multi-view tracking
enables a more reliable gaze features detection even under
challenging scenarios. In addition, owing to the proposed
adaptive fusion mechanisms, the framework achieves high
accuracies and robustnmess under real-world conditions.
A comparison of previous work in various aspects such
as hardware setup and calibration requirements, accuracy,
robustness, and working volume, is given in Table V. Since the
majority of existing efforts requires particular hardware and
system setups, e.g., additional light sources, setup calibration,
use of 3D or depth information, we could not reproduce and
validate the reported performances for all. Instead, for these,
we reported the performances directly from the corresponding
references. Although a direct numerical comparison would not
be completely fair, the provided information can still help us to
make the following inferences. First of all, we observe that the
popularity of appearance-based methods, which have lower
hardware and calibration requirements, have been increasing
recently in parallel with the recent advancements in machine
learning, e.g., convolutional neural networks (CNNs), and in
the synthesizing and rendering technology. Despite the fact
that their accuracies and head movement tolerances are still not
comparable to those of feature-based methods, their potential
is likely to be exploited in the foreseeable future.
On the other hand, even though feature-based methods
outperform appearance-based, the bothersome hardware requirements, e.g., NIR cameras and light sources, remain an
important concern. The setup complexity is especially high for
3D model-based systems, such that fully-calibrated setups consisting of multiple cameras or a Kinect-like sensor are required
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
Cross ratio
Regression
3D Model
Appearance
Method
Zhang et al (2015) [26]
Wood et al (2016) [28]
Lu et al (2015) [25]
Mora & Odobez (2016) [47]
Krafka et al (2016) [27]
Beymer & Flickner (2003) [4]
Hennessey et al (2006) [5]
Guestrin & Eizenman (2007) [6]
Park (2007) [7]
Lai et al (2015) [23]
Sun et al (2015) [8]
Zhu & Ji (2007) [10]
Cerrolaza et al (2008) [11]
Sesma-sanchez et al (2012) [12]
Cerrolaza et al (2012) [24]
Yoo & Chung (2005) [13]
Hansen et al (2010) [14]
Coutinho & Morimoto (2013) [15]
Zhang & Cai (2014) [16]
Huang et al (2014) [17]
Arar et al (2015) [19] (multi-view)
Proposed multi-view framework
Hardware Setup
Cam(s)
Light(s)
Calib.
1
1
1
1+Kinect
5
pre
1
4*
2
fully
1+1
3
fully
2
4
fully
1+2*
4
fully
2
2
fully
Kinect
?
pre
2
2
fully
1
2
1
2
1
2
1+1*
4+1
1
4
1
4+1
1
8
1
8
3
4+1
3
4
-
SH(◦ )
6.31
9.952
2.5
1.93
∼3.54
∼0.6
∼1
∼1.5
∼1.1
∼1
∼1
∼0.9
∼1.6
∼1
∼0.4
∼0.4
∼0.8
0.86
0.99
Accuracy
MH(◦ ) Eye Data
?
Both
?
?
9.65
Single
3.5
Both
?
Both
Single
<1
Single
∼1
Single
∼1
Both
∼1
Single
∼2
Single
∼1.8
Single
∼1
Single
∼1
Both
∼1.3
Single
Single
Single
∼0.5
Single
∼0.6
Both
∼1.6
Single
Both
1.27
Both
13
HP
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Fixed
Fixed
Fixed
Free
Free
Fixed
Fixed
Fixed
Free
Free
Robustness
Head Movement Eyewear
?
3
?
3
(40 x ? x ? )
?
?
3
Limited
(14 x 12 x 20)
(10 x 8 x 10)
(? x ? x 20)
3
(10 x 5 x 10)
(20 x 20 x 8)
(20 x 20 x 30)
(0 x 0 x 10)
(0 x 0 x 6)
(0 x 0 x 12)
Limited
Limited
(25 x ? x 25)
(10 x ? x 20)
(? x ? x 20)
Limited
(30 x ? x 20)
3
Illum.
3
3
3
3
FoV
FL
?
?
?
6.1
?
4.8
32
35
?
37
6.1
?
35
35
16
?
?
> 35
13
13
125
85
TABLE V: Comparison of existing eye tracking systems. In ”Cam(s)” column, ∗ indicates that a pan-tilt unit is employed.
”Calib.” column indicates whether explicit camera and scene geometry calibrations are required: ”fully” means both are
required, ”pre” means the sensor is pre-calibrated. In ”Accuracy”, ”SH” and ”MH” correspond to stable and moving head
scenarios, respectively. The results refer to, unless stated otherwise, person-specific scenarios on within-dataset evaluations.
”HP” column indicates whether users’ head pose were fixed, e.g., using a chinrest. In ”FoV” column, the systems’ working
volume is presented by ”FL”, focal length in mm. The smaller the focal length, the larger the FoV.
for accurate 3D modelling. Cross ratio-based systems and
most of regression-based systems have an important advantage
over 3D model-based ones. They require uncalibrated setups
and less complex (2D) eye models while providing competitive
accuracies to 3D model-based systems. Among these, it is also
clear that there is an accuracy gap between fixed-head (using
chinrest) and free-head eye tracking since the approximated
eye models are sensitive to head movements.
Moreover, the results indicate that tracking performances
significantly benefit from high-resolution eye data. For example, Coutinho & Morimoto [15] reported an impressive
accuracy, about 0.5◦ , under large head movements through planarization of gaze features. Nonetheless, their system required
eye resolution of 640×480 pixels, that is 7-fold of ours. They
captured eye data using a narrow FoV lens and used a chinrest
to keep users’ eye within the FoV of the camera. In addition,
Huang et al. [17] and Zhang & Cai [16] proposed two alternative methods that are highly effective to compensate for the
head movements, while requiring relatively lower resolution
eye data, i.e., 13-mm focal lenses were used. However, similar
to [15], they both utilized a chinrest during their evaluation.
Since use of chinrest is unnatural for users and represents an
unrealistic tracking scenario, it remains an important limitation
of their evaluations. On the contrary, our methodology allows
for not only head translations but also head rotations while
requiring lower resolution eye data (∼90×50 pixels) captured
using 8-mm lenses. Lower resolution data naturally results in a
lower accuracy, yet, the proposed adaptive fusion mechanism
successfully closes the accuracy gap by effectively combining
the gaze outputs obtained by multiple sensors. Besides, our
system accounts for eyewear and illumination robustness,
which have largely been neglected by the majority of the
previous efforts.
As depicted from Table V, the proposed multi-view ap-
proach and implemented prototype system achieves a competitive accuracy while offering more robustness to aforementioned real-world conditions. Still, further improvements
on accuracy, robustness or setup complexity can be achieved
through certain hardware and software modifications. For
instance, explicit head movement compensation techniques,
such as learning an adaptive homography from simulated data
[17] or planarization of features [15], can be employed to
further improve the head movement robustness. In addition,
as utilized by most commercial eye trackers, certain hardware
solutions, e.g., auto-focus lenses or smart dynamic illumination techniques, can greatly enhances the estimation accuracy
and availability.
As the proposed multi-view framework is independent of
the gaze estimation method used, alternative gaze estimation
methods can also be integrated with regard to the application
requirements. Although cross ratio-based method is suggested
for high-accuracy estimation on desktop scenarios due to its
particular advantages, such as enabling high-accuracy using an
uncalibrated and flexible setup, an appearance (CNN)-based
method (e.g., Zhang et al. [26], Krafka et al. [27]) can be
used to lower the setup complexity, or a 3D model-based
method can be implemented to achieve even higher accuracies.
Furthermore, the number of cameras and their configuration
can be varied according to the application scenario without
requiring further system adjustments, particularly for uncalibrated settings. For instance, the current prototype can easily
be configured to work under challenging tracking scenarios,
such as in-car driving scenarios, children’s eye tracking, or
customized eye trackers for disabled people.
1 Person-independent
within-dataset evaluation on MPIIGaze dataset [26].
cross-dataset evaluation on MPIIGaze dataset [26].
3 Person-specific within-dataset evaluation on Eyediap dataset [50].
4 Person-specific within-dataset evaluation on GazeCapture dataset [27].
5 Single camera property. Multi-view setup has a significantly larger FoV.
2 Person-independent
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
VII. C ONCLUSIONS
This paper presents a novel multi-view eye tracking framework to revisit the robustness concerns in eye tracking, particularly to head movements and glasses. Instead of computing
the user gaze from a single view as performed by the previous
work, leveraging multiple eye appearances simultaneously acquired from various views provides with enhanced estimation
accuracy and robustness under challenging real-world conditions. The main advantage of the multi-view approach is that
for each frame, we calculate multiple gaze outputs using the
features extracted from various eye appearances. This enables
to extract the features more reliably even under challenging
conditions, where they are obstructed in the conventional
single-camera view. Under large head movements and use
of glasses, our evaluations show that the proposed approach
improves the tracking performance of a single-camera setup
by about 20% (0.2-0.6◦ ) in estimation accuracy and 10-20%
in estimation availability. The results also demonstrate that our
approach is highly tolerant to illumination and between-subject
eye type variations. In addition to the improved robustness to
challenging conditions, the system’s overall accuracy greatly
benefits from the multi-view setup under normal conditions.
The proposed methodology provides by about 30% improvement in accuracy, owing to the adaptive fusion mechanisms,
which account for the reliability of estimations determined
from user’s predicted gazing behavior and momentary head
poses with respect to each camera. The current implementation runs at 30 fps, obtains ∼1◦ estimation accuracy error
and nearly 100% estimation availability under challenging
experimental scenarios, which makes it appropriate for highaccuracy demanding applications.
R EFERENCES
[1] G. Underwood, Cognitive processes in eye guidance. Oxford Pr., 2005.
[2] A. T. Duchowski, Eye tracking methodology. Springer, 2007.
[3] D. W. Hansen and Q. Ji, “In the eye of the beholder: a survey of models
for eyes and gaze,” PAMI, vol. 32, no. 3, pp. 478–500, 2010.
[4] D. Beymer and M. Flickner, “Eye gaze tracking using an active stereo
head,” in CVPR, 2003, pp. 451–458.
[5] C. Hennessey, B. Noureddin, and P. Lawrence, “A single camera
eye-gaze tracking system with free head motion,” Measurement, 2006.
[6] E. D. Guestrin and M. Eizenman, “Remote Point-of-Gaze Estimation
with Free Head Movements Requiring a Single-Point Calibration,” in
Proc. Int. Conf. Eng. Med. Biol. Soc., 2007, pp. 4556–4560.
[7] K. R. Park, “A real-time gaze position estimation method based on a
3-D eye model.” TSMC Part B, vol. 37, no. 1, pp. 199–212, 2007.
[8] L. Sun, Z. Liu, and M. T. Sun, “Real time gaze estimation with a
consumer depth camera,” Inf. Sci. (Ny)., vol. 320, pp. 346–360, 2015.
[9] Z. Zhu, Q. Ji, and K. P. Bennett, “Nonlinear eye gaze mapping function
estimation via support vector regression,” in ICPR, 2006, pp. 1132–1135.
[10] Z. Zhu and Q. Ji, “Novel Eye Gaze Tracking Techniques Under Natural
Head Movement,” TBE, vol. 54, no. 12, 2007.
[11] J. J. Cerrolaza, A. Villanueva, and R. Cabeza, “Taxonomic study of
polynomial regressions applied to the calibration of video-oculographic
systems,” in ETRA, 2008.
[12] L. Sesma-sanchez, A. Villanueva, and R. Cabeza, “Gaze Estimation
Interpolation Methods Based on Binocular Data,” TBE, vol. 59, no. 8,
pp. 2235–2243, 2012.
[13] D. H. Yoo and M. J. Chung, “A novel non-intrusive eye gaze estimation
using cross-ratio under large head motion,” CVIU, vol. 98, no. 1, 2005.
[14] D. Hansen, J. Agustin, and A. Villanueva, “Homography normalization
for robust gaze estimation in uncalibrated setups,” in ETRA, 2010.
[15] F. L. Coutinho and C. Morimoto, “Improving head movement tolerance
of cross-ratio based eye trackers,” IJCV, vol. 101, no. 3, 2013.
[16] Z. Zhang and Q. Cai, “Improving cross-ratio-based eye tracking
techniques by leveraging binocular fixation constraint,” in ETRA, 2014.
14
[17] J.-B. Huang, Q. Cai, Z. Liu, N. Ahuja, and Z. Zhang, “Towards
accurate and robust cross-ratio based gaze trackers through learning
from simulation,” in ETRA, 2014, pp. 75–82.
[18] N. M. Arar, H. Gao, and J. P. Thiran, “A regression-based user
calibration framework for real-time gaze estimation,” TCSVT, vol. 27,
no. 12, pp. 2623–2638, 2017.
[19] ——, “Robust gaze estimation based on adaptive fusion of multiple
cameras,” in IEEE Int. Conf. Aut. Face & Gesture Recogn. (FG), 2015.
[20] J. J. Kang, E. D. Guestrin, W. J. Maclean, and M. Eizenman, “Simplifying the cross-ratios method of point-of-gaze estimation,” in Can. Med.
Biol. Eng. Conf., 2007.
[21] A. Villanueva and R. Cabeza, “A novel gaze estimation system with one
calibration point,” TSMC Part B, vol. 38, no. 4, pp. 1123–1138, 2008.
[22] N. M. Arar, H. Gao, and J. P. Thiran, “Towards convenient calibration
for cross-ratio based gaze estimation,” in WACV, 2015, pp. 642–648.
[23] C. C. Lai, S. W. Shih, and Y. P. Hung, “Hybrid method for 3-D gaze
tracking using glint and contour features,” TCSVT, vol. 25, no. 1, 2015.
[24] J. Cerrolaza, A. Villanueva, M. Villanueva, and R. Cabeza, “Error
characterization compensation in eye tracking systems,” in ETRA, 2012.
[25] F. Lu, Y. Sugano, T. Okabe, and Y. Sato, “Gaze Estimation From Eye
Appearance: A Head Pose-Free Method via Eye Image Synthesis,”
TIP, vol. 24, no. 11, pp. 3680–3693, 2015.
[26] X. Zhang, Y. Sugano, M. Fritz, and A. Bulling, “Appearance-based
gaze estimation in the wild,” in CVPR, 2015, pp. 4511–4520.
[27] K. Krafka, A. Khosla, P. Kellnhofer, H. Kannan, S. Bhandarkar,
Matusik, and A. Torralba, “Eye tracking for everyone,” in CVPR, 2016.
[28] E. Wood, T. Baltrušaitis, L.-P. Morency, P. Robinson, and A. Bulling,
“Learning an appearance-based gaze estimator from one million
synthesised images,” in ETRA, 2016, pp. 131–138.
[29] E. Guestrin and M. Eizenman, “General Theory of Remote Gaze
Estimation Using the Pupil Center and Corneal Reflections,” TBE,
vol. 53, no. 6, pp. 1124–1133, jun 2006.
[30] A. Utsumi, K. Okamoto, N. Hagita, and K. Takahashi, “Gaze tracking
in wide area using multiple camera observations,” in ETRA, 2012.
[31] T. Nagamatsu, R. Sugano, Y. Iwamoto, J. Kamahara, and N. Tanaka,
“User-calibration-free gaze estimation method using a binocular 3D eye
model,” IEICE Trans. Inf. Syst., vol. E94-D, no. 9, pp. 1817–1829, 2011.
[32] L. Sun, M. Song, Z. Liu, and M. Sun, “Real-time gaze estimation with
online calibration,” IEEE Multimedia, vol. 21, no. 4, pp. 28–37, 2014.
[33] J. Chen and Q. Ji, “A probabilistic approach to online eye gaze tracking
without explicit personal calibration,” TIP, vol. 24, no. 3, 2015.
[34] K. White, T. Hutchhinson, and J. Carley, “Spatially Dynamic Calibration
of an Eye-Tracking System,” TSMC, vol. 23, no. 4, pp. 1162–1168, 1993.
[35] Y. Ebisawa, “Improved video-based eye-gaze detection method,” IEEE
Trans. Instrum. Meas., vol. 47, no. 4, pp. 948–955, 1998.
[36] Q. Ji and X. Yang, “Real-time eye, gaze, and face pose tracking for
monitoring driver vigilance,” Real-Time Imaging, vol. 8, no. 5, 2002.
[37] A. Villanueva and R. Cabeza, “Models for gaze tracking systems,”
Eurasip J. Image Video Process., vol. 2007, pp. 1–16, 2007.
[38] T. C. Kübler, T. Rittig, E. Kasneci, J. Ungewiss, and C. Krauss,
“Rendering refraction and reflection of eyeglasses for synthetic eye
tracker images,” in ETRA, 2016, pp. 143–146.
[39] X. Xiong and F. De la Torre, “Supervised Descent Method and Its
Applications to Face Alignment,” in CVPR, 2013, pp. 532–539.
[40] N. M. Arar, “Robust eye tracking based on adaptive fusion of multiple
cameras,” Ph.D. dissertation, EPFL, 2017, thèse no 7933.
[41] W. H. Zangemeister and L. Stark, “Types of gaze movement: variable
interactions of eye and head movements.” Experimental neurology,
vol. 77, no. 3, pp. 563–77, 1982.
[42] J. M. Saragih, S. Lucey, and J. F. Cohn, “Face alignment through
subspace constrained mean-shifts,” in ICCV, 2009, pp. 1034–1041.
[43] M. Böhme, M. Dorr, M. Graw, T. Martinetz, and E. Barth, “A software
framework for simulating eye trackers,” in ETRA, 2008, p. 251.
[44] F. Schaeffel, “Myopia: The Importance of Seeing Fine Detail,” Curr.
Biol., vol. 16, no. 7, pp. 257–259, apr 2006.
[45] K. Nguyen, C. Wagner, D. Koons, and M. Flickner, “Differences in the
infrared bright pupil response of human eyes,” in ETRA, 2002, p. 133.
[46] K. Kiranantawat, J. Suhk, and A. Nguyen, “The asian eyelid: Relevant
anatomy,” Seminars in Plastic Surgery, vol. 29, no. 3, 2015.
[47] K. A. Mora and J.-M. Odobez, “Gaze Estimation in the 3D Space
Using RGB-D Sensors,” IJCV, vol. 118, no. 2, pp. 194–216, jun 2016.
[48] S. Ren, X. Cao, Y. Wei, and J. Sun, “Face Alignment at 3000 FPS via
Regressing Local Binary Features,” in CVPR, 2014, pp. 1685–1692.
[49] V. Kazemi and J. Sullivan, “One millisecond face alignment with an
ensemble of regression trees,” in CVPR, 2014, pp. 1867–1874.
[50] K. A. Mora and J.-M. Odobez, “Geometric generative gaze estimation
(G3E) for remote RGB-D cameras,” in CVPR, 2014, pp. 1773–1780.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. XX, NO. XX, 2018
1
Robust Real-Time Multi-View Eye Tracking
S UPPLEMENTARY I NFORMATION
Nuri Murat Arar, Student Member, IEEE, and Jean-Philippe Thiran, Senior Member, IEEE
Subject #18
Subject #8
TABLE SI: Head pose statistics (in ◦ ) of two subjects from the dataset. The head pose angles are estimated with respect to
the bottom camera view separately on calibration and six individual test sessions relevant to head movements.
Session
Exp No
Min
Yaw
Max
Std
Calibration
#2
#2
#3
#4
#5
#6
#7
#2
#2
#3
#4
#5
#6
#7
-13.3
-11.3
-20.9
-21.9
-15.8
-12.9
-15.2
-14.1
-15.8
-24.5
-18.1
-15.7
-18.1
-17.2
2.6
6.2
31.2
10.5
10.2
13.2
9.9
22.9
22.9
19.4
22.9
19.1
18.9
28.2
Test
Calibration
Test
Mean
Min
Pitch
Max Std
Mean
Min
Roll
Max Std
-4.8
-2.6
-1.5
-3.4
-3.6
-2.4
-3.5
2.6
3.7
-0.9
0.2
0.6
1.9
3.6
-13.6
-16.5
-21.3
-15.9
-5.7
-16.9
-8.2
-21.1
-23.1
-20.8
-24.3
-21.3
-22.9
-22.6
4.9
3.8
6.7
7.2
5.6
8.4
6.4
0.9
-0.7
7.9
-1.9
-2.3
3.6
4.9
-3.6
-5.3
-3.4
-1.8
-0.2
-2.3
-0.4
-9.6
-10.2
-7.4
-12.6
-10.2
-8.4
-7.8
-6.2
-6.7
-20.1
-12.8
-9.4
-11.9
-3.7
-4.4
-4.2
-25.5
-6.3
-5.4
-14.2
1.1
0.1
0.1
13.5
0.7
0.2
-2.6
4.5
4.4
2.4
16.7
1.9
1.9
-6.5
7.4
4.7
4.6
11.3
9.1
6.2
6.3
5.3
13.5
12.3
10.4
12.7
10.5
10.2
11.9
5.3
5.1
5.6
5.6
2.9
4.9
2.9
6.9
7.1
6.2
6.2
4.7
7
6.8
1.7
1.5
5.7
3.4
2.1
2.2
1.8
2.4
1.6
8.7
2.2
1.4
1.6
1.6
Mean
-3.2
-2.7
-3.9
-5.3
-4.6
-8.5
0.5
-0.6
-1.1
-2.7
-1.6
-2
-9.4
4.4
TABLE SII: Mean estimation accuracy errors (◦ ) and gaze availabilities (%) achieved by different setup configurations.
multi-view
single-camera
Configurations
camera eye
right
left
right
right
right
both
left
left
left
right
left
both
bottom left
bottom right
bottom both
overall
w/ perfect vis.
w/ contacts
w/o glasses
w/ glasses
dark-eyed
light-eyed
(◦ )
Exp 0
(%)
1.71
1.5
1.55
1.64
1.55
1.5
1.68
1.65
1.46
1.07
0.86
0.83
0.98
1.39
0.92
1.33
50.6
60.4
74.4
62.1
51.4
73.1
77
74.6
91.2
96.2
98
96.8
97
93.4
96.7
95.4
(◦ )
Exp 1
(%)
1.68
1.86
1.6
2.02
1.86
1.8
1.86
1.86
1.51
1.22
1.02
1.05
1.11
1.63
1.16
1.33
51.6
68.1
74.9
68.9
56.1
78.3
78.1
79.5
94.6
98.4
98.5
99.7
99.1
95.6
97.8
99.3
(◦ )
Exp 2
(%)
1.59
1.52
1.41
1.54
1.49
1.36
1.50
1.54
1.30
0.99
0.77
0.76
0.89
1.38
0.86
1.22
57.3
69.6
77.6
65.8
56.9
79.4
81.8
79.2
93.6
97.2
98.8
98.9
98.6
91.9
98.2
95.7
(◦ )
Exp 3
(%)
2.04
2.16
2
2.07
1.86
1.88
1.93
1.90
1.67
1.45
1.20
1.23
1.38
1.68
1.29
1.70
46.8
58.6
69.7
57
44.2
68.3
72.1
68.3
87
92.2
94.8
92.5
94.2
84.4
93.8
89.5
(◦ )
Exp 4
(%)
1.87
2.27
2.11
2.1
1.83
1.79
1.96
2.05
1.73
1.52
1.21
1.41
1.40
1.95
1.42
1.68
36.3
41.7
52
47.4
44.7
61.8
64.7
64.8
82.4
92.4
94.9
97.5
96.5
76.8
93.5
90.5
(◦ )
Exp 5
(%)
2.06
2.26
2.07
2.18
2.28
2.07
1.99
2.32
1.81
1.60
1.18
1.17
1.44
2.19
1.46
1.85
49.6
63.2
75.2
53.1
51.2
71.9
67.7
67.4
83.9
93.7
98.8
99.1
96.8
82
96.4
89.1
(◦ )
Exp 6
(%)
1.84
1.83
1.77
1.88
2.08
1.77
1.79
1.76
1.33
1.10
0.79
0.88
0.92
1.77
1.05
1.18
51.1
66.7
73.7
47.6
33.8
58.7
77.6
75.5
92.5
95.2
98.8
99
98.9
81.4
95.9
94
(◦ )
Exp 7
(%)
1.53
1.55
1.35
1.68
1.75
1.51
1.63
1.58
1.32
1.12
0.81
1.02
1.01
1.51
1.03
1.26
37.5
65.1
69.4
70.4
62.1
83.2
73.4
75.1
90.8
95.5
98.9
98.5
98.3
85
97
93
2
(a) Illumination variations, i.e., experiments #2, #0, #1.
(b) Depth movements, i.e., experiments #2, #5, #4.
(c) Horizontal movements, i.e., experiments #2, #6, #7.
Fig. S1: Sample images from the collected dataset: (left column) right camera view, (middle column) bottom camera view,
and (right column) left camera view.
| 1cs.CV
|
Can CNNs Construct Highly Accurate Model Efficiently with
Limited Training Samples?
Yu Lia, Hu Wanga, Juanjuan Liua
a. State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body, Hunan
University, Changsha, 410082, PR China
Highlights
·CNN is applied to model high dimensional and strong nonlinear problems.
·Different intervals, dimensions, and training sample sizes are considered.
·Several highly nonlinear and dimensional mathematical functions are modeled.
·CNN is applied to the optimization of complex geometric structure.
Abstract
It is well known that metamodel or surrogate modeling techniques have been widely
applied in engineering problems due to their higher efficiency. However, with the
increase of the linearity and dimensions, it is difficult for the present popular
metamodeling techniques to construct reliable metamodel and apply to more and more
complicated high dimensional problems. Recently, neural networks (NNs), especially
deep neural networks (DNNs) have been widely recognized as feasible and effective
tools for multidiscipline. Actually, some popular NNs, such as back propagation neural
networks (BPNNs) can be regarded as a kind of metamodeling techniques. However,
for high dimensional problems, it seems difficult for a BPNN to construct a metamodel.
In this study, to construct the high accurate metamodel efficiently, another powerful NN,
convolutional neural networks (CNNs) are introduced to construct metamodels.
Considering the distinctive characteristic of the CNNs, the CNNs are considered to be
a potential modeling tool to handle highly nonlinear and dimensional problems with the
limited training samples.
Keywords: CNN; Metamodel; High dimension; Strong nonlinear; Small sample set
1. Introduction
At present, many engineering analyses ask the requirement of complicated analyses and
calculation codes, e.g. finite element analysis (FEA). Despite the increasing of
computers, the cost of manpower and material resources for calculation limit
development. How to provide a computationally cheap code and how to solve the highdimensional and strong nonlinear analyses efficiently and accurately have become
unresolved hot spots of concern.
A popular method to solve the above issues is to apply a cheap approximate model to
describe the complex analysis. And this approximate model, i.e. metamodels, provide a
"model of the model" [1]. Mathematically, assume the input information is vector x,
and the output information is vector y, the code actually calculates
y f (x)
(1)
where f(x) is the function for the real model. The function for the approximate model is
yˆ g (x)
(2)
y yˆ
(3)
where
where ԑ includes both approximation and random errors.
Today, there are many metamodels help g(x) approximate f(x), i.e. polynomial response
surface models (PRSMs), multivariate adaptive regression splines (MARSs), radial
basis functions (RBFs), kriging interpolationv (KI) and support vector regression (SVR)
[2, 3]. However, it appears that few of them are powerful when modeling highdimensional and strong nonlinear analyses.
In this study, deep neural networks (DNNs) is employed for complex analyses. In 2006,
Professor Hinton [4], one of the founders of the deep learning (DL), through the
analysis of human brain's decision making power, proposed to build deep neural
networks (DNNs) by increasing the artificial neural networks(ANNs)' layers and the
neurons of each layer. He started the wave of applications of DL and large data. Until
2017, the team DeepMind of Google published the latest research Mastering the game
of Go without human knowledge on nature. DNNs and large data have always gathered
people's focus.
DNNs have a significant advantage for analysis of unstructured, unspecified patterns
and cross-domain large data. Generally, the classical NNs can be classified into
feedforward NNs and feedback NNs. The former has been applied widely for it strong
ability of approximating nonlinear mappings [5-10]. The most popular feedforward NN
is back propagation neural network (BPNN). The classical BPNNs [11] make full use
of the advantage of multilayer forward network structure [12], and their algorithm is
relatively mature. Infinite deep neural networks (IDNNs), i.e. recurrent neural networks
(RNNs) [13], are dynamical systems essentially. Their essential character is that the
states of the networks change with the time and couple the time. Hence they are suitable
for extracting the features of the time. Radial basis function neural network (RBFNN)
[14] is a special kind of single hidden layer NNs, which has a simple structure, and it is
easily trained. Meanwhile, it is at present widely applied in engineering fields such as
pattern recognition, target prediction and function interpolation [15]. Moreover, the
RBFNN greatly reduces the training samples and improves the accuracy through
continuous optimization from scholars, such as the changing width factor RBFNNs
(CWF-RBFNNs) [16]. Adaptive wavelet neural networks (AWNNs) [17] which are
combined with wavelet analysis (VA) make full use of the advantages of BPNNs and
RBFNNs. Moreover, they have a faster convergence rate while overcome the RBFNNs'
shortcoming of local minimal. Since, in many case, the BPNN of one hidden layer or
three layers is sufficient to analyze most problems [18-21], the BPNNs have become
the preferred NN models.
BPNNs' characteristics are presented in Table 1. However, due to the bottleneck of their
algorithms, the computationally expensive of computers and time consuming, the
number of design variables is limited.
Table 1 Characteristics of the BPNN.
BPNNs
Advantages
Characteristics
i. Strong ability of non-linear mapping.
ii. Higher capability of self-learning and self-adapt.
Shortcomings
iii. Good generalization ability to adapt to a new sample.
iv. Strong fault tolerance.
i. A large number of necessary training samples.
ii. The convergence rate is slow.
iii. The training process has randomness, and it means that under the same
configuration parameters, the training results may appear quit differences.
For the characteristics mentioned above, we seek to provide another NN to construct a
high-dimensional model. In this study, convolutional neural networks (CNNs) are
considered. Commonly, CNNs are widely applied in image recognition, and the input
variables are pixels of an image. Generally, if the image's pixel is 28×28=784, it means
the input dimension to the CNN is 784. However, the input dimension may increase to
million level for larger images such as 1024×1024 pixels. Surprisingly, in the field of
image classification, CNN's accuracy is more than 95%. Therefore, CNN has its own
advantages in analyzing high-dimensional problems. This is why we seek to employ
the CNNs to analyze the high-dimensional (the input dimension is greater than 100)
and strong nonlinear problems which are bottleneck for the aforementioned metamodels.
In this study, we explore, evaluate and analyze CNN's capability (when modeling highly
nonlinear and dimensional problems) and transfer learning from image recognition to
engineering application. The remainder of this paper is organized as follows. The basic
theories of BPNNs are summarized, and an evaluation scheme for BPNNs' ability is
provided next. In Section 3, CNNs are applied to analyze the high-dimensional and
strong nonlinear problems. In Section 4, the results of BPNNs and CNNs are compared
and analyzed. After a series of observations and analyses, CNNs are applied to the
optimization of the complex geometric structure in Section 5. The summaries are given
in the final section.
2. Back propagation neural network
2.1. Overall framework
BPNN, which is the most widely employed model among the ANN models, was
introduced by Rumelhart and McClelland [22] in 1986, and it is a kind of multi-layer
feedforward neural networks (MFNNs). Its overall framework is shown in Fig. 1,
which consists of N layers including the input layer, several hidden layers and the
output layer. The core characteristic of the BPNN is to feed back the error from output
to the each layer's weight in order to correct each weight [23].
Fig. 1. Flow chart of BPNNs’ signal.
(where X is the input matrix, V and W are weight matrix, f(x) is the activation function, f '(x) is the derivative of
the activation function, O is the prediction of BPNN and d is the actual value, δo is the gradient of error d-O to W
and δy is the gradient of error d-O to V, and η is the learning rate.)
2.2. Application
There are following three aims should be taken into account about BPNN.
I.
What is the maximum input interval used in BPNNs?
II.
What is the maximum input dimension BPNN can model?
III.
What is the cost of training samples' preparation for BPNN?
2.3. A multidimensional case
In this section, we seek to apply the BPNN to a typical nonlinear multidimensional case.
Because the Griewank function has many local maxima, the Griewank function is
employed to test the BPNN's capability. The Griewank function shown in Fig. 2 is given
by
d
xi2
x
f x
cos i 1
i
i 1 4000
i 1
d
where d is the number of design variables.
(4)
[0, 600]
[0, 50]
[0, 10]
[0, 5]
Fig. 2. The Griewank function.
In the classical BNPP, the Sigmoid function are usually employed as the activation
functions in the input and hidden layers [24], while a linear function is employed as the
output layer's. The Sigmoid function can be categorized into two groups based on
whether the output value contains a negative: Log-Sigmoid function and Tan-Sigmoid
function. The Log-Sigmoid function is employed to this study and is calculated by
f x
1
1 e x
(5)
A feedforward neural network can construct the mapping between the input variables
and the output value through iterative training. The loss function is employed to
determine if the training should be terminated, and calculated by
total _ losses MSE y, y ' R w
(6)
where
n
y y '
MSE y, y '
i 1
i
n
i
2
(7)
R w wi 1 wi2
(8)
i
where yi is the actual value, yi' is the predictive value, α is the proportion of L1
regularization to the total regularization, and n is the number of training samples.
Learning rate, usually, is designated initially as 0.7 [25], and subsequently adjusted
according to the training process. In this study, a more flexible method is introduced
to determine the learning rate -- exponential decay method (EDM):
0 decay _ rate global _ step / decay _ step
(9)
where η0 is the initial learning rate, decay_rate is the attenuation coefficient,
global_step is the current step in the training process, and decay_step is the decay rate.
In order to clearly master the accuracy of the trained model, the relative errors (REs)
of the testing samples are shown to us and calculated by
RE
yi yi '
yi
(10)
If each RE value is close to 0, it indicates that the goodness-of-fit of the NN model is
satisfied and the model has a strong capability.
The interval of each xi is clustered into [0, 5], [0, 10], [0, 50], [0, 600] and dimension is
to 2 to observe BPNNs' capabilities increased with the interval of each xi, structural
parameters of BPNN are shown in Table 2. Usually some rules are employed to
determine the number of hidden layer's nodes.
I.
The number of hidden layer's nodes is 70% to 90% of the input layer's. [26]
II.
The number of hidden layer's nodes is less than twice of the input layer's. [27]
III.
The number of the hidden layer's nodes should be between the input and output
layers'. [28]
IV.
The nodes of the first and second hidden layers should be approximately equal.
[21]
Actually, the number of hidden layers and each layer's node are not absolutely certain.
Therefore, we comply the rules mentioned above incompletely based on actual training
process.
Table 2 The structural parameters of BPNN under different input intervals.
Interval
Number of
hidden layers
Nodes’ number of
hidden layers
Training
dataset size
Testing
dataset size
[0,5]
1
6
3000
100
[0,10]
2
8000
100
[0,50]
4
10000
100
[0,600]
4
5000
100
No.1
No.2
No.1
No.2
No.1
No.2
No.3
No.4
6
6
6
8
4
8
8
6
RE
≤30%
If the BPNN analyzed above is equipped with accuracy requirement, the BPNN can be
used for further analysis. When the interval is set as [0, 100], the input dimension should
be increased to test BPNN's capability increased with the input dimension. In this study,
the dimensions of Griewank function are set as 5, 10 and 20, respectively. The structural
parameters of the BPNN under different input dimensions are presented in Table 3.
Table 3 The structural parameters of BPNN under different input dimensions.
Dimension
Number of
hidden layers
5
3
10
3
20
4
Nodes’ number of
hidden layers
No.1
No.2
No.3
No.1
No.2
No.3
No.1
No.2
No.3
No.4
10
8
8
20
8
10
8
18
16
12
Training
dataset size
Testing
dataset size
5000
100
8000
100
14000
100
RE
≤30%
As is known to all, with the increase of dimension and expand of interval, the number
of necessary training samples increases rapidly. It is necessary to know the most
reasonable number of training samples once the interval, dimension and the maximum
RE are determined, in order to avoid wasting time of redundant training samples.
3. Convolutional neural network
It is well known to all that the more the NN's layers are, the stronger the NN is. However,
it is difficult to get a "deep" BPNN by the gradient descent optimizer (GDO) because
BPNN's gradient is difficult to exceed three layers. This is partly because BPNN's
gradient is difficult to exceed three layers. Therefore, a "deep" BPNN's model is
difficult to be achieved. What's more, when a BPNN is employed to model a highdimensional problem, the critical problem is that a large number of training samples are
commonly necessary, and a great deal of manpower and material resources for samples
are consumed. At the same time, the parameters in a BPNN's full connection layer are
too many. For example, if the input dimension is 900 and there are 500 nodes in the
first hidden layer, then the full connection layer obtains 900×500+500=450500
parameters, where 500 additional parameters are the bias corresponding to each node.
Moreover, too many parameters will easily cause the over-fitting problem. Therefore, a
more reasonable and effective network structure for high-dimensional and strong
nonlinear problems is demanded. In our opinion, the Convolutional Neural Network
(CNN) [29, 30] might be a useful tool for analyzing the aforementioned problems, even
the CNN is seldom applied to modeling.
Learning method of CNNs is the supervise learning (SL). Similar to the full connection
NN, the CNN is composed through layers of nodes. And it has been applied widely,
such as natural language processing [31-34], medical development [35], disaster
discovery [36] and even the human intelligence program [37].
3.1. Overall framework
As shown in Fig. 3, the training and testing procedures of CNN are divided into three
parts. Firstly, the training and testing samples are randomly generated in design interval.
And then, each training sample is submitted to the CNN to generate a function
approximation. At last, additional testing samples are applied to test the trained model's
accuracy.
Fig. 3. The procedure of model construction using CNN.
There exists many CNN architectures, such as CifarNet [38, 39], AlexNet [30],
GoogLeNet [40] and so on, and a more complex CNN models among them have only
about 150K or 15K parameters [39, 41]. Where does CNN's capability come from when
modeling high-dimensional and strong nonlinear problems? CNNs illustrated in Fig. 4
combine three properties [42, 43]:
I.
Local connection. This is the most straightforward characteristic of CNNs.
Each neuron is not connected to all neurons from the upper layer, but connected
to a small part. This enable the model to reduce parameters effectively.
II.
Weight sharing. Each network layer shares a same weight vector wi, rather than
each local connection mentioned above has a different weight vector. This also
reduces a lot of parameters.
III.
Local pooling. The pooling layer is used to reduce each layer's nodes. It not
only can reduce the CNN's parameters, but also can enhance the CNN's
robustness.
Fig. 4. Architecture diagram of a CNN.
The key features of the CNN architecture are described below:
I.
Input layer. The Input layer is the entrance to the whole CNN. It takes an image
of pixels W×H×3 (RGB image) or W×H×1 (monochrome image) as input.
Starting from this layer, the transfer vector of each layer is converted into a
three-dimensional matrix.
II.
Convolutional layer. The convolutional layer is the most important part of a
CNN. Compared with the classical full connection layer, the input to each node
of a convolution layer is a small part from the upper layer's nodes. The
convolutional layer analyzes each node's input to obtain a higher degree of
abstract understanding.
III.
Pooling layer. The pooling layer reduces the computation for deeper layers
while preserving the important activations.
IV.
Full connection layer. After several convolutional layers and pooling layers, it
can be considered that the information of the input variables has been abstracted
into a very high level. After feature extraction and analysis of the front layers, a
full connection layer is still necessary to complete the model's construction.
V.
Softmax layer. The softmax layer is mainly used for the problems of
classification, if the problem is regression, the softmax layer is not the essential
element.
The forward propagation process of a convolutional layer mainly depends on a filter or
a kernel. As shown in Fig. 5 and Eq. (11), where ax,y,z is the feature vector at spatial
location (x, y, z) in the input vector, w cx , y , z is the weight vector from c-th filter in a
convolutional layer and bc is the bias vector, then the output from a filter is
g x f a x i , y j , z k w ic, j ,k b c
i , j ,k
(11)
where * represents dot product.
Fig. 5. The forward propagation process of the convolutional layers.
A pooling layer can reduce the size of the input matrix very effectively. The pooling
layer can not only speed up the calculation, but also prevent the over-fitting problem.
Similar to the convolutional layer, the forward propagation process of the pooling layer
is done by moving a structure which is similar to a filter. However, the pooling layer's
filter does not calculate the weighted sum of the input, but the maximum value or
average value. The pooling layer, which calculates the maximum value, is the maximum
pooling, and the pooling layer, which is averaged, is the average pooling.
3.2. LeNet-5
The model of CNN applied in this study is LeNet-5, which was developed by Yuan [44]
in 1998. And better results has been achieved by the LeNet-5 in the image classification
problem. As shown in Fig. 6, in order to apply the LeNet-5 to this study better, the
architecture of LeNet-5 is optimized. Commonly, LeNet-5 is applied to classify images
and the input variables are pixels of the image. When analyzing a high-dimensional
mathematical problem, design variable vector is converted into a matrix of M×N×1
which is similar to the distribution of a monochrome image's pixels. Then the matrix is
regarded as a "monochrome image", and each design variable is regarded as an "image's
pixel". At last LeNet-5 analyzes the "image". The key features of this LeNet-5
architecture are described in Table 4.
Fig.6. LeNet-5 of this study.
Table 4 LeNet-5 architecture used in this study.
Number of
hidden layer
Type of layer
1st
Convolutional layer
2nd
Pooling layer
3rd
Convolutional layer
4th
Pooling layer
5th
Full connection layer
6th
Full connection layer
Parameter & characteristic
Size of filter 5×5, depth of filter is 32, moving step
of filter is 1
Maximum pooling, length of filter is 2, moving
step of filter is 2
Size of filter 5×5, depth of filter is 64, moving step
of filter is 1
Maximum pooling, length of filter is 2, moving
step of filter is 2
Dropout [45] is added to avoid over-fitting
phenomenon
512 nodes
The Maximum Pooling is applied in 2nd and 4th hidden layers for two reasons.
I.
By eliminating non-maximal values, it reduces computation for upper layers.
II.
It provides a form of translation invariance.
Because the essential of this study is a regression problem, the output from 6th hidden
layer will be the output from the whole CNN.
The activation function, Rectified Linear Units Function (ReLU), used in LeNet-5 can
speed up the training process, and ReLU is calculated by
f x max x, 0
(12)
The loss function for this study is the mean square error (MSE).
n
y y '
MSE y, y '
i 1
i
n
i
2
(13)
In order to avoid the over-fitting problem, the regularization is added to the loss
function, it means that an indicator of CNN's complexity is added to the loss function.
Assuming that the loss function is J(θ), CNN doesn't optimize J(θ), but optimize
J(θ)+λR(w), where R(w) is the CNN's complexity, λ is the proportion of the complex
loss to the total losses.
There are two ways to evaluate the complexity, one is L1 regularization, known as
Lasso Regularization.
R w w 1 wi
(14)
i
Another is L2 regularization, known as Weight Decay.
2
R w w 2 wi2
(15)
i
Firstly, when using L1 regularization, if an element of the weight vector is a positive
number, the element becomes smaller, conversely, becomes larger. It means that the
purpose of L1 regularization is making the weight vector change toward 0 and making
the weights "sparser". Therefore, L1 regularization can achieve the function similar to
the feature selection. Otherwise, L2 regularization cannot make the weights "sparser"
because that when an element is small, such as 0.001, the square of the element is
smaller and can be ignored, so the model will not change the element into 0. Secondly,
the formula of L1 regularization cannot be derived while L2 regularization's can. Since
the derivative of the loss function needs to be calculated during optimization, the
optimization of L2 regularization's loss function is easier. After comprehensive analysis,
both L1 regularization and L2 regularization are used in this study.
R w wi 1 wi2
(16)
i
where α is the proportion of L1 regularization to the total regularization. And the loss
function is
total _ losses MSE y, y ' R w
(17)
In the reverse propagation process, the loss function is used to update weights and
biases.
wi wi wi
(18)
b b b
(19)
where
wi y ' y xi
(20)
b y ' y
(21)
where wi is the initial weight, b is the initial bias, y is the actual value, y' is the predicted
value, and η is the learning rate.
Learning rate η determines the update rate of each parameter, it's usually designated at
a constant. However, once η is too large, η might cause that the prediction swings on
both sides of the optimal value and CNN can't converge. On the contrary, if η is too
small, although the CNN can converge eventually, the training speed will be greatly
reduced. Therefore, the η can't be too large, nor too small. Meanwhile, a more flexible
method for learning rate's determination is employed -- exponential decay method, and
its mathematical function is Eq. (22). As shown in Fig. 7, through this method, the CNN
can obtain a larger learning rate so that it can be quickly trained initially, and then the
CNN gradually reduces the learning rate to ensure that the model can converge to the
optimal value at the end of the training.
=0 decay _ rate global _ step / decay _ step
(22)
Fig. 7. Learning rate by exponential decay method.
(where η0 is 0.1, decay_rate is 0.9, and decay_step is 50.)
As Hush and Horne [46] explained, network learning can be terminated under three
different scenarios:
I.
When the MSE is less than the pre-designated value.
II.
When the pre-designated training steps is achieved.
III.
When a valid MSE sample has increased.
For the first two scenarios, it would easily over-fit for basing upon pre-designated value.
Smith [47] suggested using cross-validation to prevent over-fitting. Therefore, the
cross-validation rule is employed to this study.
4. Experiments and analysis
4.1. Test functions
This section illustrates the application of a NN to some typical high-dimensional and
strong nonlinear mathematical models shown in Table 5. Subsequently, NN's capability
is evaluated. For selecting mathematical models, we try to cover the operations
including index, trigonometric function, absolute value, tired ride and accumulate to
ensure we can achieve a comprehensive assessment of the NNs.
Table 5 Mathematical functions employed by this study.
Models' Name
Model1:
Griewank
Model 2:
Ackley
Model 3:
Schwefel
Model 4:
Levy (Modified)
Model 5:
Sum of Different
Powers
Function
d
f x
i 1
Description
d
xi2
x
cos i 1
4000 i 1
i
1 d 2
f x 20 exp 0.2
xi
d i 1
1 d
exp i 1 cos 2xi 20 exp(1)
d
d
f x 418.9829d xi sin
xi
i 1
d
2
f x xi 1 1 sin 2 3xi
i 1
d
f x xi
i 1
i 1
Griewank function has many
widespread local minima which
are regularly distributed. The
function is usually evaluated on
the hypercube [-600, 600]
Ackley function is widely used for
testing optimization algorithms.
The function is usually evaluated
on the hypercube [-32.768, 32.768]
Schwefel function with many local
minima is complex. The function
is usually evaluated on the
hypercube [-500, 500]
This function is usually evaluated
on the hypercube [-10, 10]
This is unimodal and is usually
evaluated on the hypercube [-1, 1]
d
Model 6:
Rastrigin
Model 7:
Schaffer N2
(Modified)
Model 8:
Schaffer N4
(Modified)
f x 10d xi2 10 cos2xi
i 1
f x 0.5
f x 0.5
sin 2
d
i 1
1i xi2 0.5
1 0.001 x
d
2
i 1 i
d
i
cos sin 1 xi2
i 1
2
This has several local minima. It is
highly multimodal, but locations of
the minima are regularly
distributed. The function is usually
evaluated on the hypercube [-5.12,
5.12]
Schaffer function. This function is
usually evaluated on the hypercube
[-100, 100]
0.5 Schaffer function. This function is
d
1
0
.
001
xi2
i 1
2
usually evaluated on the hypercube
[-100, 100]
4.2. Performance criteria
In this study, the performance of each NN is measured from accuracy and robustness.
The accuracy is indicated by the model's predict ability during the design space. While
the robustness is indicated by the high-precision acquisition capabilities during
different problems.
For accuracy, the goodness-of-fit of training data is not enough to illustrate the model's
accuracy. Therefore, additional testing samples are necessary for further verify the
accuracy of the model trained by NN. In order to evaluate the training results
comprehensively, several expressions of error in Table 6 [2, 3] are employed, where
STD is standard deviation. The RMSE indicates the overall accurate approximation,
while the MAE can reveal the presence of regional areas of poor approximation. The
RAAE is usually related with MSE and a large RMAE indicates a though the RAAE
may be great, so a small RMAE is preferred. However, because that the input interval
and dimension from different NNs and situations are different, these statistics enable
comparisons only within each NN.
Table 6 Criteria for error.
Name
Expression
Maximal Absolute Error (MAE)
max yi yi '
Average Absolute Error (AAE)
n
i 1
yi yi ' n
n
y y '
Root Mean Square Error (RMSE)
i 1
Relative Average Absolute Error (RAAE)
n
i 1
i
yi ' yi
i
n
n STD
max yi ' yi
Relative Maximum Absolute Error (RMAE)
2
STD
However, errors shown above cannot clearly present the difference between the actual
value and the predict value of each testing sample, the χi is calculated by Eq. (23). In
order to better display and compare the testing sample's χis including maximum value
and minimum value under different situations, testing points are re-ordered and
displayed according to the levels of their χis.
i 1
yi ' yi
1 RE
yi
(23)
The robustness of NN is indicated by the variance between its χi [2] and is calculated
by Eq. 24. A small σ2 means a good performance in robustness, on the contrary, a big
σ2 means bad.
2
=
i
n
2
(24)
where μ is the average of χi.
4.3. Distribution of samples
As is known to all, with the increase of dimension and expand of interval, the number
of necessary training samples increases rapidly. It is necessary to know the most
reasonable number of training samples once the interval, dimension and the maximum
RE are determined, in order to avoid the wasting time of redundant training samples.
By comparing the results for training different sample's sizes under a same NN, we
select the minimum training samples which meets the requirement of the NN's training.
4.4. Result and discussion
As mentioned above, the closer the χi to 1 is, the higher precision the model is. For both
RAAE and RMAE, a smaller value indicates a higher accuracy. And for variance, a
smaller value indicates a stronger robustness.
4.4.1. Performance for different input intervals. As shown in Fig. 8, the errors and
variances of Griewank function show the robustness for different input intervals by the
BPNN when the input dimension is 2D. As shown in Fig. 9, only very few sample's χis
are less than 60% when each xi belongs to [0, 50]. And more than 80% sample's χis in
all intervals are more than 80%.
It is noted that BPNN performs satisfactory in terms of both accuracy and robustness
for a high-order nonlinear and low-dimensional problem. However, the robustness,
although small for BPNN, becomes larger when the interval expands.
3.0E+00
2.5E+00
2.0E+00
1.5E+00
1.0E+00
5.0E-01
0.0E+00
MAE
AAE
RMSE
Variance
RAAE
RMAE
[0,5]
9.1E-02
5.1E-03
1.3E-02
1.4E-02
1.8E-01
6.9E-01
[0,10]
4.2E-01
3.4E-02
9.3E-02
1.9E-01
1.7E-01
9.8E-01
[0, 50]
1.5E+00
9.0E-01
6.0E-01
6.4E-01
1.3E-01
9.2E-01
[0, 600]
1.8E+01
1.1E+00
2.5E+00
9.6E-01
1.3E-01
1.5E+00
Fig. 8. Comparison of errors and variances of testing samples between different intervals when the
input dimension is 2 by BPNN.
Fig. 9. Comparison of testing sample's χis between different intervals when the input dimension
is 2 by BPNN.
For CNNs, as shown in Figs. 10 and 11, the changes of error, variance and χi with the
increase of input dimension are presented when modeling the Griewank function. It is
found that with expand of the interval, the variance is still very small and χis from all
intervals are more than 80%. It indicates that CNN's robustness and calculation ability
are excellent when analyzing high-dimensional and strong nonlinear problems.
4.0E+01
3.5E+01
3.0E+01
2.5E+01
2.0E+01
1.5E+01
1.0E+01
5.0E+00
0.0E+00
MAE
AAE
RMSE
Variance
RAAE
RMAE
[0, 5]
1.4E-01
3.5E-02
4.3E-02
3.7E-04
1.8E+00
5.9E+00
[0, 40]
3.8E+00
8.9E-01
1.1E+00
1.1E-03
2.8E+01
9.5E+01
[0, 100]
1.8E+01
5.1E+00
6.4E+00
9.9E-04
1.6E+02
6.2E+02
[0, 1000]
1.6E+03
3.7E+02
4.8E+02
6.7E-04
1.6E+04
7.3E+04
Fig. 10. Comparison of errors and variances of testing samples between different intervals when
the input dimension is 144 by CNN.
Fig. 11. Comparison of testing sample's χis between different intervals when the input dimension
is 144 by CNN.
Overall, in terms of the accuracy and robustness, with the increase of input intervals,
CNN is proved to be the best choice compared to the commonly used BPNN when
handling high-dimensional and strong nonlinear problems.
4.4.2. Performance for different input dimensions. As shown in Figs. 12 and 13, the
errors, variance and χis of the Griewank function are presented for different dimensions
by BPNN when the input variables are in [0, 100]. The variance rapidly increases and
more than 40% testing sample's χis are less than 80%. Even when the dimension is 10
or 20, some χis are negative. It indicated that the BPNN performs powerless in
robustness and accuracy. And it is noted that the dimension which the BPNN can model
is 20 at most. And with the increase of input dimension, BPNN's robustness trends
toward sharp decline. Therefore, with the increase of input dimension, BPNN performs
powerless.
4.0E+01
3.5E+01
3.0E+01
2.5E+01
2.0E+01
1.5E+01
1.0E+01
5.0E+00
0.0E+00
MAE
AAE
RMSE
Variance
RAAE
RMAE
5d
2.1E+01
5.8E+00
4.7E+00
1.3E+00
1.6E-01
4.0E+00
10d
3.7E+01
7.9E+00
9.7E+00
1.3E+01
9.9E-02
3.0E+00
20d
3.1E+01
1.1E+01
1.3E+01
1.9E+01
9.4E-01
1.9E+00
Fig. 12. Comparison of errors and variance of testing samples between different dimensions when
the interval is [0, 100] by BPNN.
Fig. 13. Comparison of testing sample's χis between different dimensions when the interval is [0,
100] by BPNN.
As shown in Figs. 14 and 15, the training results of the Griewank function by CNN are
presented when the interval is [0, 1000]. From the Fig. 15, almost all testing sample's
χis are more than 90%, and it is amazing that the variances are less than 10-3. Thus, the
calculation ability and robustness of CNN are very surprised to us when analyzing a
high-dimensional and strong nonlinear problem.
1.0E+04
9.0E+03
8.0E+03
7.0E+03
6.0E+03
5.0E+03
4.0E+03
3.0E+03
2.0E+03
1.0E+03
0.0E+00
MAE
AAE
RMSE
Variance
RAAE
RMAE
144d
1.6E+03
3.7E+02
4.8E+02
6.7E-04
1.6E+04
7.3E+04
784d
3.2E+03
8.5E+02
1.1E+03
1.0E-04
8.9E+04
3.1E+05
1600d
4.7E+03
1.4E+03
1.8E+03
7.3E-05
1.9E+05
6.3E+05
3600d
1.1E+04
3.9E+03
3.9E+03
6.0E-05
3.8E+05
1.6E+06
Fig. 14. Comparison of errors and variance of testing samples between different dimensions when
the interval is [0, 1000] by CNN.
Fig. 15. Comparison of testing sample's χis between different dimensions when the interval is [0,
1000] by CNN.
4.4.3. Performance for modeling other mathematical models. To further judge the
CNNs' ability, some other mathematical models shown in Table 5 are calculated.
Moreover, Table 7 shows the errors and variances and Fig. 16 presents the χis. The
interval is [-100, 100] for Ackley function, [-1000, 1000] for Schwefel function, [-10,
10] for Levy function, [-1, 1] for Sum of Different Powers function, [-10, 10] for
Rastrigin function, [-100, 100] for Schaffer N2 function and [-100, 100] for Schaffer
N4 function. The column of variance whose values are generally less than 10-2 in Table
7 can further illustrate the performance of CNN's robustness. As shown in Fig. 16,
although the χis' interval of the Sum of Different Powers function, Schaffer N2 function
and Schaffer N4 function are a little worse than other functions', the results calculated
by CNN are very satisfactory compared with existing methods.
Table 7 Variances and errors between different functions by CNN.
Function
Ackley
Schwefel
Levy
Sum of
Different
Powers
Rastrigin
Schaffer
N2
Schaffer
N4
Dimension
144D
784D
1600D
3600D
144D
784D
1600D
3600D
144D
784D
1600D
144D
784D
1600D
3600D
144D
784D
1600D
3600D
144D
784D
1600D
3600D
144D
784D
1600D
Variance
-4
7.2×10
5.5×10-4
1.3×10-3
1.6×10-3
3.7×10-3
4.7×10-4
2.6×10-4
1.1×10-4
8.0×10-4
1.4×10-4
9.2×10-5
3.1×10-2
5.1×10-2
2.9×10-2
5.3×10-2
4.5×10-4
5.2×10-5
3.4×10-5
3.4×10-6
1.8×10-2
1.6×10-2
1.7×10-2
1.0×10-2
1.5×10-2
6.5×10-3
6.2×10-3
MAE
AAE
0
2.4×10
2.5×100
3.8×100
4.9×100
1.8×100
4.1×100
5.5×100
7.2×100
9.9×100
2.4×105
4.1×105
3.8×100
5.3×100
6.3×100
6.9×100
2.9×102
1.2×103
2.0×103
4.4×103
6.7×10-1
7.5×10-1
4.6×10-1
2.3×10-1
6.5×10-1
5.2×10-1
2.0×10-1
RMSE
-1
7.7×10
7.3×10-1
1.1×10-1
1.3×100
4.9×103
9.4×103
1.5×104
2.2×104
3.3×104
9.6×104
1.5×105
9.9×10-1
1.4×100
1.5×100
1.8×100
7.6×101
3.1×102
5.6×102
2.2×103
5.1×10-2
6.4×10-2
6.2×10-2
7.2×10-2
1.1×10-1
1.3×10-1
1.4×10-1
-1
9.7×10
8.9×10-1
1.3×100
1.4×100
6.0×103
1.2×104
1.9×104
2.7×104
3.8×104
9.9×104
1.5×105
1.2×100
1.8×100
1.5×100
2.2×100
9.4×101
3.9×102
7.0×102
2.3×103
8.4×10-2
9.0×10-2
9.0×10-2
8.8×10-2
1.2×10-1
1.0×10-1
1.1×10-1
RAAE
RAME
1
8.3×101
1.2×102
1.3×102
3.6×102
3.0×105
1.8×106
8.3×106
6.5×106
3.5×106
2.0×107
4.3×107
2.2×101
2.4×101
3.7×101
3.0×101
1.4×104
1.6×105
3.4×105
7.4×105
5.1×100
5.9×100
3.5×100
2.3×100
5.3×100
6.4×100
2.4×100
3.1×10
2.9×101
2.8×101
5.0×101
8.1×104
4.3×105
9.0×105
2.0×106
8.7×105
5.3×106
1.0×107
5.5×100
6.3×100
9.2×100
8.0×100
3.5×103
4.2×104
9.6×104
3.7×105
3.9×10-1
5.0×10-1
4.7×10-1
7.1×10-1
3.6×10-1
6.3×10-1
7.0×10-1
1.0
1.0
0.8
0.8
Dimension:144d
Dimension:784d
Dimension:1600d
Dimension:3600d
i
0.4
0.4
0.2
0.2
0.0
0
50
100
150
Dimension:144d
Dimension:784d
Dimension:1600d
Dimension:3600d
0.6
i
0.6
0.0
200
0
Number of testing points
50
100
150
200
Number of testing points
(a) Ackley function
(b) Schwefel function
1.0
1.0
Dimension:144d
Dimension:784d
Dimension:1600d
0.8
0.8
0.6
Dimension:144d
Dimension:784d
Dimension:1600d
Dimension:3600d
i
i
0.6
0.4
0.4
0.2
0.2
0.0
0.0
0
50
100
150
Number of testing points
0
200
50
100
150
200
Number of testing points
(c) Levy function
(d) Sum of Different Powers function
1.0
1.0
Dimension:144d
Dimension:784d
Dimension:1600d
Dimension:3600d
0.8
0.8
i
i
Dimension:144d
Dimension:784d
Dimension:1600d
Dimension:3600d
0.6
0.6
0.4
0.4
0.2
0.2
0.0
0.0
50
100
Number of testing points
150
0
200
50
100
(e) Rastrigin function
(f) Schaffer N2 function
1.0
0.8
Dimension:144d
Dimension:784d
Dimension:1600d
0.6
0.4
0.2
0.0
0
150
Number of testing points
i
0
50
100
Number of testing points
150
200
(g) Schaffer N4 function
Fig. 16. Model's χis under different dimensions by CNN.
200
We also observe another interesting law about CNN, when the dimension is low, the
accuracy and robustness are generally worse than those of a higher dimension. This rule
further confirms the absolute advantage of CNN's capability in high-dimensional
problems.
4.4.4. Performance under different sample sizes. We end our analysis with a comparison
of training different sizes of training samples. The training results by BPNN are
presented in Fig. 17, the interval is [0, 100] and dimension is 2, it can be seen that there
is a clear decline when the training sample's size is less than 3000. Therefore, under the
current parameters, BPNN's minimum training sample's size should beyond 3000. As
shown in Figs. 18- 25, the results calculated by the CNN indicate that only 2000
necessary training samples can CNN get the results with high accuracy. Moreover, we
also find that BPNNs gradually converge after 20000 training steps, and CNNs only
need 3000 training steps to get a convergence result. So the efficiency of CNN is better.
1.0
180
Dataset size:4000
Dataset size:3000
Dataset size:2000
150
0.8
Dataset size:4000
Dataset size:3000
Dataset size:2000
0.6
i
Total losses
120
90
0.4
60
0.2
30
0
0
20000
40000
60000
80000
Training steps
100000
0.0
0
20
40
60
80
100
Number of testing points
a. total losses
b. χis
Fig. 17. Training results of Griewank function under different dataset sizes by BPNN when the
interval is [0, 100] and dimension is 2.
1.00E+008
1.0
Dataset size:600
Dataset size:1000
Dataset size:2000
Dataset size:600
Dataset size:1000
Dataset size:2000
0.8
6.00E+007
0.6
i
Total losses
8.00E+007
4.00E+007
0.4
2.00E+007
0.2
0.0
0.00E+000
0
500
1000
1500
Training steps
2000
2500
3000
0
50
100
150
Number of testing points
a. total losses
b. χis
Fig. 18. Training results of Griewank function under different dataset sizes by CNN when the
interval is [0, 1000] and dimension is 3600.
200
1.0
200
180
Dataset size:600
Dataset size:1000
Dataset size:2000
160
Dataset size:600
Dataset size:1000
Dataset size:2000
0.8
Total losses
140
0.6
i
120
100
0.4
80
60
0.2
40
20
0.0
0
500
1000
1500
Training steps
2000
2500
0
3000
50
100
150
200
Number of testing points
a. total losses
b. χis
Fig. 19. Training results of Ackley function under different dataset sizes by CNN when the
interval is [-1000, 1000] and dimension is 3600.
1.0
5.00E+009
Dataset size: 600
Dataset size:1000
Dataset size:2000
Total losses
4.00E+009
Dataset size:600
Dataset size:1000
Dataset size:2000
0.8
0.6
i
3.00E+009
0.4
2.00E+009
0.2
1.00E+009
0
400
800
1200
1600
2000
0.0
0
Training steps
50
100
Number of testing points
150
200
a. total losses
b. χis
Fig. 20. Training results of Schwefel function under different dataset sizes by CNN when the
interval is [-500, 500] and dimension is 3600.
1.0
10000000
Dataset size:600
Dataset size:1000
Dataset size:2000
9000000
8000000
Dataset size:600
Dataset size:1000
Dataset size:2000
0.8
Total losses
7000000
0.6
i
6000000
5000000
4000000
0.4
3000000
2000000
0.2
1000000
0
0
400
800
1200
Training steps
1600
2000
0.0
0
50
100
Number of testing points
150
200
a. total losses
b. χis
Fig. 21. Training results of Levy function under different dataset sizes by CNN when the interval
is [-10, 10] and dimension is 1600.
50
1.0
Dataset size:600
Dataset size:1000
Dataset size:2000
0.8
0.6
i
Total losses
40
Dataset size:600
Dataset size:1000
Dataset size:2000
30
0.4
20
0.2
0.0
10
0
500
1000
Training steps
1500
0
2000
50
100
Number of testing points
150
200
a. total losses
b. χis
Fig. 22. Training results of Sum of Different Powers function under different dataset sizes by
CNN when the interval is [-1, 1] and dimension is 3600.
1.0
20000000
Dataset size:600
Dataset size:1000
Dataset size:2000
18000000
16000000
Dataset size:600
Dataset size:1000
Dataset size:2000
0.8
Total losses
14000000
0.6
i
12000000
10000000
8000000
0.4
6000000
4000000
0.2
2000000
0
0
500
1000
Training steps
1500
2000
0.0
0
50
100
150
200
Number of testing points
a. total losses
b. χis
Fig. 23. Training results of Rastrigin function under different dataset sizes by CNN when the
interval is [-10, 10] and dimension is 3600.
1.0
100
Dataset size:1000
Dataset size:2000
Dataset size:1000
Dataset size:2000
0.8
0.6
60
i
Total losses
80
0.4
40
0.2
20
0.0
0
500
1000
Training steps
1500
2000
0
50
100
Number of testing points
150
200
a. total losses
b. χis
Fig. 24. Training results of Schaffer N2 function under different dataset sizes by CNN when the
interval is [-100, 100] and dimension is 3600.
100
1.0
Dataset size:1000
Dataset size:2000
Dataset size:3000
0.8
60
Dataset size:1000
Dataset size:2000
Dataset size:3000
0.6
i
Total losses
80
40
0.4
20
0.2
0.0
0
0
200
400
600
800
1000
1200
1400
Training steps
1600
1800
2000
0
50
100
Number of testing points
150
200
a. total losses
b. χis
Fig. 25. Training results of Schaffer N4 function under different dataset sizes by CNN when the
interval is [-100, 100] and dimension is 1600.
5. Complex geometric modeling by CNN
It is well known that the CNN can handle many problems. While in our views, on
geometric issues, the CNN has its unique advantages.
Isogeometric analysis [48-51] (IGA) is a popular method of numerical solution of the
physical field recent years. The main advantage of IGA based on Non-uniform Rational
B-splines (NURBS) is the integration of Computer Aided Design (CAD) and Computer
Aided Engineering (CAE). Moreover, IGA can provide a new method that can combine
design, analysis and optimization. Non-uniform Rational B-splines (NURBSs) are a
standard basis function used in CAD and CG, and the most preferred computational
technology in IGA. Recent applications [52-55] of IGA include solids, structures, fluids,
and fluid–structure interaction. It is known that NURBSs are built from B-splines
through a projective transformation [56] and inherit their superior approximation power,
and are constructed by lots of control points and knot vectors whose modification
provides a feasible way to optimization. In the IGA and the NURBS-enhanced FEA,
where NURBS is adopted as the basis function, the numerical quadrature needs more
and more computational load with the increase of the model's scale, and then researches
for faster quadrature rules have been very active [57-61]. And this issue is more obvious
in the structural optimization problems. Due to the limitations of computers'
computational ability, we cannot solve issues of large design variables using the IGA.
The method in this study utilizes CNN to extract rules inherent in an IGA application,
which usually are implicit and sometimes too complicated to grasp from the large
amount of available data to solve the problem that IGA's design variables cannot be too
many.
In this study, Rhinoceros software is applied to provide the input information of a
tubular structure shown in Fig. 26. Subsequently, NURBS is constructed and applied to
geometry model and analysis model. As shown in Figs. 27(b) and (c), the Degrees of
Freedom (DOFs) on bottom surface marked in red are fixed, and as shown in Fig. 27(a),
three concentrated forces which are equal to 3×105N in y direction are enforced to the
control points marked in orange. Finally, the stress of the structure is analyzed by IGA.
The initial control points and the weight factors are used as the input variables of the
model, and the set of maximum stress after 2500 iterations is used as the output
information. By modifying the values of input variables, 2000 results as training
samples and 500 results as testing samples are calculated for CNN.
Fig. 26. The control lattice for the tubular structure.
Fig. 27. Constraints and load of the tubular structure.
For each sample, design variables include 324 control points, and each control point
includes cylindrical coordinate (r, φ, z) which calculates by Eq. 25 and a weight factor
w, so the input dimension is 324×(3+1)=1296, the output is the maximum stress of the
tubular structure. As shown in Figs. 28 and 29, the results by CNN are presented, it can
be seen that the errors, variance and χis are surprised to us.
x r cos
y r sin
z z
(25)
1.4E+06
1.2E+06
1.0E+06
8.0E+05
6.0E+05
4.0E+05
2.0E+05
0.0E+00
Errors & variance
Variance
2.4E-04
MAE
1.4E+05
AAE
3.9E+04
RMSE
4.3E+04
Fig. 28. Errors and variance of testing samples.
RAAE
1.0E+06
RMAE
9.3E+06
Fig. 29. Testing sample's χis.
Then the CNN model is optimized using genetic algorithm (GA) reanalysis methods.
The tubular structure is optimized by changing the coordinate of the control points.
And the objective function to be minimized is the stress of the tubular structure. The
shape of tubular structure is chosen as the design variable. In particular, the tubular
structure changes only in the direction of the r axis. Therefore, the optimization
problem is stated as
min y CNN B r1 ,r2 ,,r324
(26)
ri 0.5, 2.5
(27)
s.t.
where CNN is the maximal stress of the tubular structure, B is the control points
which is set as design variables, and △ri is the amplification factors of control points.
The optimized shape and corresponding stress analysis results are shown in Fig. 30. It
can be found that only 2000 training samples from the IGA can CNN get a very
satisfactory model, and CNN model can combine the reanalysis methods to further
improve the efficiency of optimization. This section further illustrates the application
of CNN to high dimensional problems.
(a)
(b)
Fig. 30. The optimized shape and stress analysis results of the tubular structure.
Conclusions
In conclusion, this study focuses on the rule extraction capability of CNN, showing the
possibility of its new application to the high dimensional and strong nonlinear problems
in which existing methods have never been successful
In this study, we investigate convolutional neural networks (CNNs) as alternative
technique for high-dimension and strong nonlinear problems using limited training
samples. The computationally efficient theory behind CNN is reviewed, and CNN
approximations are compared against the aforementioned BPNN using several complex
mathematical functions. CNN achieves more accurate and more robust function
approximations than the BPNN, and shows great potential for modeling applications.
In terms of the accuracy and robustness, can provide the observations and conclusions
from Section 4 as follows. CNN excels in majority mathematical models. In terms of
the input interval and dimension, accuracy and robustness of CNN are far superior to
other methods. For the most difficult problems, i.e. high-dimensional and strong
nonlinear problems, we compared the ability between CNN and BPNN, the training
results represent that CNN can construct extremely high-precision models with limited
samples. While BPNN requires a large number of samples, and its result is not
satisfactory. From the above observations, we can conclude that CNN is the most
dependable model for highly nonlinear and dimensional problems.
Finally, an application of CNN to optimization of complex geometric structure is also
proposed. CNN, which is based on the characteristics of location connection, weight
sharing and local pooling has overstepped the limits of the present popular
metamodeling techniques, widening its application scope to a large extent.
Acknowledgments
This work has been supported by Project of the Key Program of National Natural
Science Foundation of China under the Grant Numbers 11572120, 11302266 and
61232014.
References
[1] Oron G. Statistical tools for simulation practitioners : Jack P.C. Kleijnen Volume 76
in: Statistics: Textbooks and Monographs, Marcel Dekker, New York, 1987, 448
pages, $69.75. European Journal of Operational Research. 1988;34(2):240-1.
[2] Jin R, Chen W, Simpson TW. Comparative studies of metamodelling techniques
under multiple modelling criteria. Structural & Multidisciplinary Optimization.
2000;23(1):1-13.
[3] Clarke SM, Griebsch JH, Simpson TW, editors. Analysis of Support Vector
Regression for Approximation of Complex Engineering Analyses. ASME 2003
International Design Engineering Technical Conferences and Computers and
Information in Engineering Conference; 2005.
[4] Hinton A, Kwiatkowska M, Norman G, Parker D, editors. PRISM: A Tool for
Automatic Verification of Probabilistic Systems. International Conference on
Tools and Algorithms for the Construction and Analysis of Systems; 2006.
[5] Lin CT, Lee CSG. Neural-Network-Based Fuzzy Logic Control and Decision
System. IEEE Transactions on Computers. 1991;40(12):1320-36.
[6] Sanger TD. Optimal unsupervised learning in a single-layer linear feedforward
neural network. Neural Networks. 1989;2(6):459-73.
[7] Venkat V, Chan K. A neural network methodology for process fault diagnosis. Aiche
Journal. 2010;35(12):1993-2002.
[8] Qian L, Winfree E, Bruck J. Neural network computation with DNA strand
displacement cascades. Nature. 2011;475(7356):368-72.
[9] Tsai CC, Huang HC, Lin SC. Adaptive Neural Network Control of a Self-Balancing
Two-Wheeled
Scooter.
IEEE
Transactions
on
Industrial
Electronics.
2010;57(4):1420-8.
[10] Sudheer KP, Gosain AK, Ramasastri KS. A data-driven algorithm for constructing
artificial neural network rainfall-runoff models. Hydrological Processes.
2010;16(6):1325-30.
[11] Liu C, Shu T, Chen S, Wang S, Lai KK, Gan L. An improved grey neural network
model for predicting transportation disruptions. Expert Systems with Applications.
2016;45(C):331-40.
[12] Yang Z, Liu HC, Amp TE. A Sweeping Robot Path Pathfinding Algorithm Based
on Bp Neural Network. Computer Knowledge & Technology. 2017.
[13] Williams RJ, Zipser D. A Learning Algorithm for Continually Running Fully
Recurrent Neural Networks. Neural Computation. 2014;1(2):270-80.
[14] Zheng XJ, Zhang GZ, Xie ZY. Research on a High-Precision Algorithm of RBF
Neural Network and Its Applications. Journal of Northeastern University.
2009;30(9):1314-7+45.
[15] Krishnan V, Ulrich KT. Product Development Decisions: A Review of the
Literature. Management Science. 2001;47(1):1-21.
[16] Bing W. A Classification Method for Large Data Sets Based on the Changing Width
Factor RBF Networks. Microelectronics & Computer. 2015:32(6).
[17] Liu J, Zhao H, Zhou R, Wang P. Exploration of High-Precision Adaptive Wavelet
Neural Network Artificial Intelligence Method. Journal of Frontiers of Computer
Science and Technology. 2016;10(8):1122-32.
[18] Khaw JFC, Lim BS, Lim LEN. Optimal design of neural networks using the
Taguchi method. Neurocomputing. 1995;7(3):225-45.
[19] Kim YH. Study on Impact Mechanism for Beef Cattle Farming and Importance of
Evaluating Agricultural Information in Korea Using DEMATEL, PCA and AHP.
Agricultural Information Research. 2006;15(3):267-79.
[20] Kaastra I, Boyd M. Designing a neural network for forecasting financial and
economic time series. Neurocomputing. 1996;10(3):215-36.
[21] Karsoliya S. Approximating Number of Hidden layer neurons in Multiple Hidden
Layer BPNN Architecture. International Journal of Engineering Trends &
Technology. 2012;3(6).
[22] Rumelhart DE, editor McClelland and the PDP Research Group. Symposium on
Parallel and Distributed Processing; 1986.
[23] Rumelhart DE, Hinton GE, Williams RJ. Learning representations by backpropagating errors. Nature. 1986;323(6088):533-6.
[24] Lecun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521(7553):436-44.
[25] Hu HY, Lee YC, Yen TM, Tsai CH. Using BPNN and DEMATEL to modify
importance–performance analysis model – A study of the computer industry.
Expert Systems with Applications. 2009;36(6):9969-79.
[26] Boger Z, Guterman H, editors. Knowledge extraction from artificial neural
network models. IEEE International Conference on Systems, Man, and
Cybernetics, 1997 Computational Cybernetics and Simulation; 1997.
[27] Han J, editor Data mining techniques. ACM SIGMOD International Conference
on Management of Data, Montreal, Quebec, Canada, June; 1996.
[28] Blum. Neural Networks in C++. NY:wiley. 1992.
[29] Cun YL, Boser B, Denker JS, Howard RE, Habbard W, Jackel LD, et al., editors.
Handwritten digit recognition with a back-propagation network. Advances in
Neural Information Processing Systems; 1990.
[30] Krizhevsky A, Sutskever I, Hinton GE, editors. ImageNet classification with deep
convolutional neural networks. International Conference on Neural Information
Processing Systems; 2012.
[31] Shen Y, He X, Gao J, Deng L, editors. Learning semantic representations using
convolutional neural networks for web search. Companion Publication of the
International Conference on World Wide Web Companion; 2014.
[32] Grefenstette E, Blunsom P, Freitas ND, Hermann KM. A Deep Architecture for
Semantic Parsing. Computer Science. 2014;30(5):1-15.
[33] Kalchbrenner N, Grefenstette E, Blunsom P. A Convolutional Neural Network for
Modelling Sentences. Eprint Arxiv. 2014;1.
[34] Kim Y. Convolutional Neural Networks for Sentence Classification. Eprint Arxiv.
2014.
[35] Wallach I, Dzamba M, Heifets A. AtomNet: A Deep Convolutional Neural
Network for Bioactivity Prediction in Structure-based Drug Discovery.
Mathematische Zeitschrift. 2015;47(1):34-46.
[36] Liu Y, Racah E, Prabhat, Correa J, Khosrowshahi A, Lavers D, et al. Application
of Deep Convolutional Neural Networks for Detecting Extreme Weather in
Climate Datasets. 2016.
[37] Clark C, Storkey A. Teaching Deep Convolutional Neural Networks to Play Go.
Eprint Arxiv. 2015:1766-74.
[38] Krizhevsky A. Learning Multiple Layers of Features from Tiny Images. 2009.
[39] Roth HR, Lu L, Liu J, Yao J, Seff A, Cherry K, et al. Improving Computer-Aided
Detection Using Convolutional Neural Networks and Random View Aggregation.
IEEE Transactions on Medical Imaging. 2016;35(5):1170.
[40] Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, et al., editors. Going
deeper with convolutions. Computer Vision and Pattern Recognition; 2015.
[41] Cireşan DC, Giusti A, Gambardella LM, Schmidhuber J, editors. Mitosis detection
in breast cancer histology images with deep neural networks2013.
[42] Lecun Y, Bottou L, Bengio Y, Haffner P, editors. Gradient-Based Learning Applied
to Document Recognition. IEEE; 1998.
[43] Saxe AM, Pang WK, Chen Z, Bhand M, Suresh B, Ng AY, editors. On random
weights and unsupervised feature learning. International Conference on
International Conference on Machine Learning; 2012.
[44] Lécun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to
document recognition. Proceedings of the IEEE. 1998;86(11):2278-324.
[45] Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a
simple way to prevent neural networks from overfitting. Journal of Machine
Learning Research. 2014;15(1):1929-58.
[46] Hush DR, Horne BG. Progress in supervised neural networks. IEEE Signal
Processing Magazine. 2002;10(1):8-39.
[47] Smith M. Neural Networks for Statistical Modeling: Van Nostrand Reinhold; 1993.
[48] Hughes TJR, Cottrell JA, Bazilevs Y. Isogeometric analysis: CAD, finite elements,
NURBS, exact geometry and mesh refinement. Computer Methods in Applied
Mechanics & Engineering. 2005;194(39–41):4135-95.
[49] Cottrell JA, Hughes TJR, Bazilevs Y. Isogeometric Analysis: Toward Integration
of CAD and FEA: Wiley Publishing; 2009. 299-325 p.
[50] Bazilevs Y, Hughes TJR. NURBS-based isogeometric analysis for the computation
of flows about rotating components. Computational Mechanics. 2008;43(1):14350.
[51] Abedi R, Petracovici B, Haber RB. A space–time discontinuous Galerkin method
for linearized elastodynamics with element-wise momentum balance. Computer
Methods in Applied Mechanics & Engineering. 2006;195(25–28):3247-73.
[52] Bazilevs Y, Calo VM, Cottrell JA, Hughes TJR, Reali A, Scovazzi G. Variational
multiscale residual-based turbulence modeling for large eddy simulation of
incompressible flows. Computer Methods in Applied Mechanics & Engineering.
2007;197(1–4):173-201.
[53] Bazilevs Y, Calo VM, Hughes TJR, Zhang Y. Isogeometric fluid-structure
interaction: theory, algorithms, and computations. Computational Mechanics.
2008;43(1):3-37.
[54] Bazilevs Y, Calo VM, Zhang Y, Hughes TJR. Isogeometric Fluid–structure
Interaction Analysis with Applications to Arterial Blood Flow. Computational
Mechanics. 2006;38(4-5):310-22.
[55]
BAZILEVS
Y,
COTTRELL
JA,
HUGHES
TJR,
SANGALLI
G.
ISOGEOMETRIC ANALYSIS: APPROXIMATION, STABILITY AND ERROR
ESTIMATES FOR h-REFINED MESHES. Mathematical Models & Methods in
Applied Sciences. 2006;16(07):1031-90.
[56] Farin GE. NURB curves and surfaces: from projective geometry to practical use:
A. K. Peters, Ltd.; 1995.
[57] Sevilla R, Fernández-Méndez S. Numerical integration over 2D NURBS-shaped
domains with applications to NURBS-enhanced FEM. Finite Elements in Analysis
& Design. 2011;47(10):1209-20.
[58] Schillinger D, Hossain SJ, Hughes TJR. Reduced Bézier element quadrature rules
for quadratic and cubic splines in isogeometric analysis. Computer Methods in
Applied Mechanics & Engineering. 2014;277:1-45.
[59] Antolin P, Buffa A, Calabrò F, Martinelli M, Sangalli G. Efficient matrix
computation for tensor-product isogeometric analysis: The use of sum factorization.
Computer Methods in Applied Mechanics & Engineering. 2015;285:817-28.
[60] Hughes TJR, Reali A, Sangalli G. Efficient quadrature for NURBS-based
isogeometric analysis ☆ . Computer Methods in Applied Mechanics &
Engineering. 2010;199(5):301-13.
[61] Auricchio F, Calabrò F, Hughes TJR, Reali A, Sangalli G. A simple algorithm for
obtaining nearly optimal quadrature rules for NURBS-based isogeometric analysis.
Computer Methods in Applied Mechanics & Engineering. 2012;s 249–252(2):1527.
| 1cs.CV
|
arXiv:1704.06024v1 [math.GR] 20 Apr 2017
Ovoidal fibrations in P G(3, q), q even
N.S. Narasimha Sastry and R.P. Shukla
Stat.- Math. Unit; Indian Statistical Institute;
8th Mile Mysore Road; R.V. College Post;
Bangalore - 560 059 (India)
and
Department of Mathematics; University of Allahabad;
Allahabad - 211 002 (India)
Abstract
We prove that, given a partition of the point-set of P G(3, q), q =
> 2, by ovoids {θi }qi=0 of P G(3, q) and a line ℓ of P G(3, q), not
tangent to θ0 if ℓ⊥ denotes the polar of ℓ relative to the symplectic
form on P G(3, q) whose isotropic lines are the tangent lines to θ0 , then
ℓ and ℓ⊥ are tangent to distinct ovoids θj , θk , both distinct from θ0
(Theorem 1.2). This uses the fact that the radical of the linear code
generated by the dual duals ℓ ∪ ℓ⊥ of the hyperbolic quadrics , with ℓ
and ℓ⊥ as above, is of codimension 1 (Theorem 1.4)
2n
Keywords: Dual grid; Elliptic quadric; Frobenius reciprocity; Tits
Ovoid
MSC (2010): 05B25, 51E22, 94B05
1 Introduction and Statement of the
Result
An ovoid in P G(3, q) is a set of q 2 + 1 points, no three collinear.
Elliptic quadrics and the Suzuki-Tits ovoids ([24], See [13], 16.4 ),
1
which exist, if and only if, n is odd, are the only known ovoids in
P G(3, q) and these are the only ovoids if n ≤ 6 [17]. Classification of
ovoids in P G(3, q) is a fundamental problem in Incidence Geometry.
We mention in passing that the elliptic quadrics are the only ovoid
in P G(3, q) if q is odd (independently due to Barlotti [4] and Panella
[18], see [17], Theorem 2.1, p.178). We recall that a general linear
complex in P G(3, q) is the set of all absolute lines with respect to a
symplectic polarity of P G(3, q) (that is, a polarity for which all points
are absolute). If P is the set of all points of P G(3, q) and L is a general
linear complex in P G(3, q), then the incidence system (P, L), which we
denote by W (q), is a generalized quadrangle of order (q, q) (see [19],
p.37). Since the collineation group P GL(4, q) acts by conjugation
transitively on the set of all symplectic polarities of P G(3, q), W (q) is
uniquely defined, up to a collineation of P G(3, q). An ovoid of W (q) is
a set O of q 2 + 1 points, pairwise noncollinear in W (q). Then, any line
l of P G(3, q) meets O in a point if it is a line of W (q) and in zero or
two points if it is not a line of W (q) as shown by a simple count. Thus
any ovoid of W (q) is an ovoid of P G(3, q). The following observation,
due to Segre [21], proves the converse: let O be an ovoid of P G(3, q), q
even, and W (O) be the incidence system whose points are the points
of P G(3, q) and lines are the tangent lines of O. For each x ∈ O, the
union of the set of the q + 1 tangent lines to O through x is a plane
πx . The correspondence x 7→ πx defines a symplectic polarity (= null
polarity) of P G(3, q) [9] whose absolute lines are precisely the lines
of W (O). Thus, W (O) = g(W (q)) for some g ∈ P GL(4, q) and the
ovoid O of P G(3, q) is an ovoid of (a unique copy of) W (q). Hence,
the classification of ovoids of P G(3, q), q even, is equivalent to the
classification of ovoids in W (q).
It is easy to see that any three mutually skew lines in P G(3, q) have
exactly q + 1 transversals (lines meeting each of the three given lines).
Such a set of q + 1 transversals is called a regulus in P G(3, q). The
transversals to the lines in any regulus form another regulus, called
the opposite of the given regulus. Thus, any three mutually skew lines
of P G(3, q) are in a unique regulus. A spread of P G(3, q) is a set of
q 2 + 1 mutually skew lines. A spread S is said to be regular (see [9],
5.1) if the unique regulus containing three lines in S is contained in
S.
An ovoidal fibration F = {θi }qi=0 of P G(3, q) is a set of q + 1 ovoids
partitioning the point set of P G(3, q). We recall that a Singer group in
P GL(4, q) is a cyclic group of order (q 2 + 1)(q + 1) acting transitively
2
on the point-set of P (3, q). Let S be a Singer group in P GL(4, q), and
T, K be its unique subgroups of orders q 2 + 1 and q + 1, respectively.
Then, the point T -orbits {Ei }qi=0 in P G(3, q) is an ovoidal fibration
by elliptic quadrics (see [1], Lemma 2, p.141); also [10], Theorem 3
and [7]); the set S of common tangent lines to the Ei ’s is a regular
spread in P G(3, q); T acts regularly on the set S and K acts regularly
on each member of S ([12]; see [1], Lemma 2, p.141). More generally,
if θ is an ovoid of P G(3, q) and S is contained in the general linear
complex L = L(θ) consisting of all tangent lines to θ (see, Segre [21])
and if K is the group of collineations fixing each line in S, then K is a
subgroup of order q + 1 contained in a Singer subgroup of P GL(4, q)
(see [12]) and the set {x(θ) : x ∈ K} of ovoids in P G(3, q) is an ovoidal
fibration of P G(3, q) ( see [5], Theorem 3.1, p.161). Further, θ is the
only ovoid of the symplectic generalized quadrangle (P, L) ≃ W (q),
where P is the point-set of P G(3, q) (see [1], Lemma 2, p.141).
Let θ be an ovoid of W (q) and ‘⊥’ denote the orthogonality relative
to the nondengerate symplectic form on P G(3, q) whose isotropic lines
are the tangent lines to θ. Since the symplectic polarity defined by
θ maps an external line to θ to a secant line to θ and vice-versa, if
m is a line of P G(3, q) which is not tangent to θ, then one of the
lines m, m⊥ shares two points with θ and the other is disjoint from
it ([13], Corollary 1 of Theorem 16.1.8). Consequently, θ is an ovoid
of P G(3, q) also. Thus, the classification of ovoids in P G(3, q) and
in W (q) are equivalent problems. We do not know any examples of
ovoidal fibration containing a pair of projectively nonequivalent ovoids
(see ([6]) for a discussion of the case q = 8); and of any examples of
fibrations consisting of projectively equivalent ovoids, but with no
transitive action of a subgroup of P GL(4, q) on its constituent ovoids
.
The subset m∪m⊥ of P , with m, m⊥ as above, is called a dual grid
of W (q). With the lines of W (q) incident with it as ‘lines’, m ∪ m⊥ is
a (1, q)-subgeneralized quadrangle of W (q). We denote by D the set
of dual grids in W (q).
Proposition 1.1 Let {θi }qi=0 be an ovoidal fibration of P G(3, q) and
Li denote the general linear complex consisting of all tangent lines to
θi , 0 ≤ i ≤ q. Then, the following hold:
(i) The set S of common tangent lines to the ovoids θi is a regular
spread in P G(3, q). Further, {Li }qi=0 are the only general linear
3
complexes in P G(3, q) containing S; and Li ∩ Lj = S for all
0 ≤ i 6= j ≤ q.
(ii) Each line ℓ of P G(3, q) not in S is tangent to a unique ovoid θi ,
secant to q/2 ovoids θj , j 6= i, and is disjoint from each of the
remaining q/2 ovoids θk of the fibration.
Our object in this note is to prove
Theorem 1.2 Let {θi }qi=0 be an ovoidal fibration of P G(3, q), W (q)
be the generalized quadrangle whose line set is the set L(θ0 ) of tangent
lines to θ0 and m be a line of P G(3, q) not in L(θ0 ). Then, m and its
perp m⊥ in W (q) are tangent to distinct ovoids θi , each distinct from
θ0 .
An intrinsic description of the ovoids θi the lines m and m⊥ are
tangent to may be interesting. We would like to view this note as
a contribution towards understanding the packings of P G(3, q) by
ovoids.
Our proof of Theorem 1.2 uses a property of the binary code
we now define. Let G denote the projective symplectic subgroup
P Sp(4, q) of P GL(4, q) defined by W (q). Let FP2 and FD
2 denote the
F2 G-permutation modules on P and D, respectively,Pand η denote the
F2 G-module homomorphism taking m∪m⊥ ∈ D to x∈m∪m⊥ x ∈ FP2 .
We identify the characteristic function of a subset of P , considered as
an element of FP2 , with the subset itself. Let C and D denote the
F2 G-submodules of FP2 whose generators are, respectively, the lines
of W (q) and the dual grids of W (q). Then, η(FD
2 ) = D. Since any
line of W (q) meets a dual grid in zero or two points, D is contained
in the dual code C ⊥ of C. Further, it is generated by words of C ⊥ of
minimum weight ([2], Theorem 1.4 ).
Corollary 1.3 D ⊂ C ⊥ .
+
We need the following
Theorem 1.4 The F2 G-radical U of D is of codimension one in D.
Consequently, the sum of two dual grids of W (q) is in the radical U .
Lemma 1.5 The group G contains a unique conjugacy class of subgroups T of order q 2 + 1 and T is cyclic. Let {Ei }qi=0 be the point
T -orbits in P G(3, q) and S denote the set of all common tangent lines
4
to Ei ’s. Let ℓ be a line of P G(3, q). Then,
P
t∈T
t(ℓ) ∈ FP2 is P (i.e.,
the ‘all-one’ vector) or the unique ovoid Ei the line ℓ is tangent to,
according as ℓ ∈ S or ℓ 6∈ S.
2
Preliminaries
Let V (q) be the vector space of dimension four over Fq ; ~ be a nondegenerate symplectic bilinear form on it; and W (q) be the incidence
system with the set P = P G(3, q) of all one dimensional subspaces of
V (q) as its point-set, the set L of all two dimensional subspaces of V (q)
which are isotropic with respect to ~ as its line-set and symmetrized
inclusion as the incidence. Then, W (q) is a regular generalized quadrangle of order q ([19], p.37) and the symplectic group G defined by ~
acts as incidence preserving permutations on the sets P and L.
Let k be an algebraically closed extension field of Fq . Let N =
{0, 1, · · · , 2n − 1}. Addition in N is always taken modulo 2n. Let N
denote the set of all subsets I of N containing no consecutive elements.
Let V = V (q) ⊗ k. The natural extension of the symplectic form ~
to V defined above is also denoted by ~. Then, G is the subgroup
of the algebraic group Sp(V ) ≃ Sp(4, k) fixed by the n-th power of
the Frobenius map σ (which is the algebraic group endomorphism
of GL(4, k) raising each entry of a matrix to its 2nd -power). It is
well known that Sp(V ) has an algebraic group endomorphism τ with
τ 2 = σ ([23], Theorem 28, p.146). For any non-negative integer i, we
denote by Vi the Sp(V )- module whose k- vector space structure is the
same as that of V and an element g of Sp(V ) acts on Vi as τ i (g) would
act on V . For I ⊆ N , let VI denote the kG- module ⊗i∈I V i (with
V∅ = k). Then, by Steinberg’s tensor product Theorem ([22], §11),
{VI : I ⊆ N } is a complete set of inequivalent simple kG -modules.
For a kG-module M , we denote by [M : VI ] the multiplicity of VI in
a composition series of M . We denote by rad(M ) the radical of M
(that is, the smallest submodule of M with semisimple quotient). We
refer to M/rad(M ) as the head of M .
We now describe a graph automorphism τ of Sp(V ), following ([11],
pp.58-60). (The argument presented in loc. cit. constructs a graph
automorphism for G = Sp(V (q)), however the arguments are valid
for Sp(V ) also.) Let {e1 , e2 , e3 , e4 } be an ordered basis of V and Q
denote the nondegenerate quadratic form on the exterior square Λ2 V
5
of V defined by
Q (Σ1≤i<j≤4 λij ei ∧ ej ) = λ12 λ34 + λ13 λ24 + λ14 λ23
(whose zero set in P (Λ2 V ) is the well-known Klein quadric). Let β
denote the polarization of Q and γ = e1 ∧ e4 + e2 ∧ e3 ∈ Λ2 V . Then,
the restriction of Q to the hyperplane U = {x ∈ Λ2 V : β (x, γ) = 0} of
Λ2 V is a nondegenerate quadratic form; and the restriction of β to U is
an alternating form with radical kγ. The alternating form β induced
U
is nondegenerate. So the symplectic space U , β
by β on U = kγ
is isometric to (V, ~) . Let p : U → V be the isometric isomorphism
induced by the linear map p : U → V defined by
p (e1 ∧ e2 ) = e1 , p (e1 ∧ e3 ) = e2 , p (e2 ∧ e4 ) = e3 ,
p (e3 ∧ e4 ) = e4 , p (γ) = 0.
Then the map taking g ∈ Sp(V ) to p(∧2 (g)) p−1 ∈ Sp(V ) is a graph
automorphism τ of Sp(V ) which, on restriction to G, gives a graph
automorphism of G.
We use the following results:
Lemma 2.1 ([8], Corollary 7.11, p.148) Let K be a field of characteristic p > 0, let E be a field extension of K and let X be a finite
group. Then, for each simple KX -module M, M E = M ⊗K E is a
direct sum of simple EX -modules, no two of which are isomorphic.
Lemma 2.2 ([20], Corollary 6) Let K, K ′ ∈ N be distinct and f~ be a
nondegenerate quadratic form on V (q) of index 2 which polarizes to ~.
Then, VK and VK ′ are semi-simple kΩ(f~ )-module with no irreducible
factors in common.
3
Proofs
Proof of Theorem 1.4: Let g ∈ D and L = StabG (g). We view
FD
2 as the induced module of the trivial F2 G- module F2 , and write
D
IndG
L (F2 ) = F2 . Then,
D
[FD
2 /Rad(F2 ) : F2 ] = 1
6
because, by Frobenius reciprocity ([16], p.689),
dimF2 (HomF2 G (IndG
L (F2 )), F2 ) = dimF2 (HomF2 L (F2 , F2 )) = 1.
This shows that the trivial module F2 is a summand of the head (that
is, the largest semi-simple quotient) of the F2 G-module FD
2 . Let H
denote the set of all hyperbolic quadrics of W (q) and f~ be a nondegenerate quadratic form on V (q) of index 2 which polarizes to ~.
Then, the variety defined by f~ is a member of H. Let O(f~ ) ⊆ G
denote the orthogonal group of f~ and Ω(f~ ) denote its commutator
subgroup. Let I be a nonempty subset of N . Then by Lemma 2.2,
HomkΩ(f~ ) (k, VI ) ∼
= HomkΩ(f~ ) (VIe , VIo ) = {0}.
This proves that VI has no nonzero fixed points for the action of O(f~ ).
Since G acts transitively on H and since the automorphism τ maps
the stabilizer of a hyperbolic quadric in W (q) to the stabilizer of a dual
grid in W (q) and vice-versa,
G
dimk (HomkG (IndG
L (k), VI )) = dimk (HomkG (IndO(f~ ) (k), VI )).
Again using Frobenius reciprocity, we have
G
dimk (HomkG (IndG
L (k), VI )) = dimk (HomkG (IndO(f~ ) (k, VI ))
= dimk (HomkO(f~ ) (k, VI ))
≤ dimk (HomkΩ(f~ ) (k, VI )) = 0
This proves that the head of the kG-module kD ∼
= IndG
L (k) is k.
Now the Lemma 2.1 completes the proof of Theorem 1.4.
Proof of Proposition 1.1: The statement (i) follows from ([3],
Theorem 2.2) and ([5], Lemma 2.4 ). Since |Li | = (q 2 + 1)(q + 1)
and Li ∩ Lj = S which has q 2 + 1 elements for all i, j, i 6= j (see [3],
Corollary 3.3), each line ℓ of P G(3, q) not in S is in Li for a unique i
and |ℓ ∩ θj | ∈ {0, 2} for all j 6= i. Since {θi } partitions P G(3, q), (ii)
follows.
Proof of Lemma 1.5: The first statement follows from ([11],
Lemma 3); the second from ([1], Lemma 2, p.141); and the third
follows by Proposition 1.1(ii) and the regular action of T on each Ei
(see [1], Lemma 4, p.142).
Proof of Theorem 1.2: Let S be the set of common tangent lines
to the θi ’s. Then, S is a regular spread in P G(3, q) by Proposition 1.1
7
(i). Let H be the subgroup of P GL(4, q) fixing each line in S. Then H
is contained in a Singer subgroup K of P GL(4, q) (see [12]). Consider
the unique subgroup T of K of order q 2 + 1. The point-orbits of T
are elliptic ovoids Ei , forming an ovoidal fibration of P G(3, q) and S
is also the set of common tangents to Ei ’s (See [1], Lemma 2, p.141).
Let Li be the set of all tangent lines to θi and W (q) = W (L0 ). Since S
is contained in exactly q + 1 general linear complexes (see Proposition
1.1 (i)), we may assume that Li is also the set of all tangent lines to
Ei . Thus θ0 and E0 are ovoids of W (q) = W (L0 ) and T is a subgroup
of G (= Sp(4, q) defined by W (q)). Now, assume that m ∪ m⊥ is a
dual grid of W (q) such that m amd m⊥ are tangents to θi for some
i 6= 0. Then, both are tangents to Ei also. Since m, m⊥ ∈
/ S, by
Lemma 1.5,
X
t(m + m⊥ ) = Ei + Ei = 0.
t∈T
On the other hand, since for each
+ m⊥ ) ∈ D and the
P t ∈ T , t(m
codimension of U in D is one, t∈T t(m + m⊥ ) = m + m⊥ (mod U ).
Since G is transitive on D, U can not contain any dual grid of W(q).
This completes the proof of the theorem.
Proof of Corollary 1.3: Let T be a cyclic subgroup of order q 2 +1
contained in the stabilizer in G of an elliptic ovoid θ0 of W (q). Let
{θi }qi=0 be the point T -orbits in P G(3, q) and S denote the set of all
common tangent lines to θi ’s (See [1], Lemma 2, P
p.141). Consider the
t(w). As we noted
F2 -linear map σ : FP2 −→ FP2 defined by σ(w) =
t∈T
earlier, D ⊆ C ⊥ . Assume that C ⊥ = D. Then C = (C ⊥ )⊥ = D ⊥ ⊇ D,
as any two dual grids of W (q) intersect in zero or two points. Now,
σ(C) = {∅, P, θ0 , P \ θ0 } (see Lemma 1.5); however if m ∪ m⊥ is a dual
grid of W (q), then by Lemma 1.5 and Theorem 1.2 σ(m+m⊥ ) = θi +θj
for distinct i and j (i > 0, j > 0), a contradiction.
References
[1] B. Bagchi and N.S.N. Sastry, Even order inversive planes, generalized quadrangles and Codes, Geom. Dedicata, 22 (1987) 137-147
[2] B. Bagchi and N.S.N. Sastry, One-step completely orthogonalizable codes from generalized quadrangles, Inform. and Comput.
77 (1988) 123-130
8
[3] B. Bagchi and N.S.N. Sastry, Ovoidal packing of P (3, q) for even
q, Discrete Math. 313 (2013) 2213-2217
[4] A. Barlotti, Un’estensione del teorema di Segre-Kustaanheimo,
Boll. Un. Mat. ltal. 10 (1955) 96-98.
[5] D.K.Butler, On the intersection of ovoids sharing a polarity,
Geom. Dedicata 135 (2008) 157-165
[6] A. Cossidente and Sam K.J. Vereecke, Some geometry of the
isomorphism Sp(4, q) ∼
= O(5, q), q even, J. Geom. 70 (2001) 28-37
[7] A. Cossidente and L. Storme, Caps on elliptic quadrics, Finite
Fields Appl. (1995) 412-420
[8] C.W. Curtis and I. Reiner, Methods of Representation Theory,
Vol I, Wiley-Interscience Pub. New York (1981)
[9] P. Demboswki, Finite Geometries, Springer-Verlag, Berlin, Heidelberg, New York, (1967)
[10] G.L. Ebert, Partitioning projective geometries into caps, Canad.
J.Math.37 (1985) 163-175
[11] D.E. Flesner, Maximal subgroups of PSp4 (2n ) containing central
elations or noncentered skew elatins, Illinois J. Math. 19 (1975),
247-268
[12] D.G. Glynn, On the set of lines of P G(3, q) corresponding to a
maximal cap contained in the Klein quadric of P G(5, q), Geom.
Dedicata 26 (1988) 273-280
[13] J.W.P. Hirschfeld, Projective spaces of three dimensions, Clarendon Press, Oxford (1985)
[14] J.W.P. Hirschfeld, Projective spaces of square size, Simon Stevin
65 (1991) 319-329
[15] B. Huppert, Enddliche Gruppen , Springer-Verlag, Berlin, Heidelberg, New York, (1967)
[16] S. Lang, Algebra, Third Edition, Addison-Wessley, 1999.
[17] C.M. O’Keefe, Ovoids in P G(3, q), a survey, Discrete Math.
151(1996) 171-188
[18] G.Panella, Caratterizzazione delle quadriche di uno spazio (tridimensionale) lineare sopra un corpo finito. Boll. Un. Mat. Ital. 10
(1955) 507-513
9
[19] S.E. Payne and J.A. Thas, Finite generalized quadrangles, Advance Publishing Program, Pitman, Boston 1984
[20] N.S.N. Sastry and R. P. Shukla, Structure of a code related
to Sp(4,q), q even, Proc. Indian Acad. Sci.(Math. Sci.) 117(4)
(2007), 457-470
[21] B. Segre, Introduction to Galois Geometries, Mem. Acad. Naz.
Lincei 8 (1967), 137-236
[22] R. Steinberg, Representations of Algebraic groups, Nagoya J.
Math. 22 (1963) 33-56
[23] R. Steinberg, Lectures on Chevalley Groups, Mimeographed
Notes, Yale Univ. Math. dept., New Haven, Conn., 1968
[24] J. Tits, Ovoides et groupes de Suzuki, Arch. math. 13(1962),
187-198
10
| 4math.GR
|
Neural Graph Machines: Learning Neural Networks Using Graphs
Thang D. Bui∗1
Sujith Ravi†
Vivek Ramavajjala†
∗
University of Cambridge, United Kingdom
Google Research, Mountain View, CA, USA
arXiv:1703.04818v1 [cs.LG] 14 Mar 2017
†
TDB 40@ CAM . AC . UK
SRAVI @ GOOGLE . COM
VRAMAVAJ @ GOOGLE . COM
Abstract
Label propagation is a powerful and flexible
semi-supervised learning technique on graphs.
Neural networks, on the other hand, have proven
track records in many supervised learning tasks.
In this work, we propose a training framework
with a graph-regularised objective, namely Neural Graph Machines, that can combine the power
of neural networks and label propagation. This
work generalises previous literature on graphaugmented training of neural networks, enabling
it to be applied to multiple neural architectures
(Feed-forward NNs, CNNs and LSTM RNNs)
and a wide range of graphs. The new objective allows the neural networks to harness both
labeled and unlabeled data by: (a) allowing the
network to train using labeled data as in the supervised setting, (b) biasing the network to learn
similar hidden representations for neighboring
nodes on a graph, in the same vein as label propagation. Such architectures with the proposed objective can be trained efficiently using stochastic gradient descent and scaled to large graphs,
with a runtime that is linear in the number of
edges. The proposed joint training approach convincingly outperforms many existing methods on
a wide range of tasks (multi-label classification
on social graphs, news categorization, document
classification and semantic intent classification),
with multiple forms of graph inputs (including
graphs with and without node-level features) and
using different types of neural networks.
1
Work done during an internship at Google.
1. Introduction
Semi-supervised learning is a powerful machine learning paradigm that can improve the prediction performance
compared to techniques that use only labeled data, by leveraging a large amount of unlabeled data. The need of semisupervised learning arises in many problems in computer
vision, natural language processing or social networks, in
which getting labeled datapoints is expensive or unlabeled
data is abundant and readily available.
There exist a plethora of semi-supervised learning methods. The simplest one uses bootstrapping techniques to
generate pseudo-labels for unlabeled data generated from
a system trained on labeled data. However, this suffers
from label error feedbacks (Lee, 2013). In a similar vein,
autoencoder based methods often need to rely on a twostage approach: train an autoencoder using unlabeled data
to generate an embedding mapping, and use the learnt embeddings for prediction. In practice, this procedure is often costly and inaccurate. Another example is transductive SVMs (Joachims, 1999), which is too computationally
expensive to be used for large datasets. Methods that are
based on generative models and amortized variational inference (Kingma et al., 2014) can work well for images
and videos, but it is not immediately clear on how to extend
such techniques to handle sparse and multi-modal inputs or
graphs over the inputs.
In contrast to the methods above, graph-based techniques
such as label propagation (Zhu & Ghahramani; Bengio
et al., 2006) often provide a versatile, scalable, and yet effective solution to a wide range of problems. These methods construct a smooth graph over the unlabeled and labeled data. Graphs are also often a natural way to describe
the relationships between nodes, such as similarities between embeddings, phrases or images, or connections between entities on the web or relations in a social network.
Edges in the graph connect semantically similar nodes or
datapoints, and if present, edge weights reflect how strong
such similarities are. By providing a set of labeled nodes,
Neural Graph Machines: Learning Neural Networks Using Graphs
such techniques iteratively refine the node labels by aggregating information from neighbours and propagate these labels to the nodes’ neighbours. In practice, these methods
often converge quickly and can be scaled to large datasets
with a large label space (Ravi & Diao, 2016). We build
upon the principle behind label propagation for our method.
Another key motivation of our work is the recent advances
in neural networks and their performance on a wide variety of supervised learning tasks such as image and speech
recognition or sequence-to-sequence learning (Krizhevsky
et al., 2012; Hinton et al., 2012; Sutskever et al., 2014).
Such results are however conditioned on training very large
networks on large datasets, which may need millions of labeled training input-output pairs. This begs the question:
can we harness previous state-of-the-art semi-supervised
learning techniques, to jointly train neural networks using
limited labeled data and unlabeled data to improve its performance?
Contributions: We propose a discriminative training objective for neural networks with graph augmentation, that
can be trained with stochastic gradient descent and efficiently scaled to large graphs. The new objective has a regularization term for generic neural network architectures
that enforces similarity between nodes in the graphs, which
is inspired by the objective function of label propagation.
In particular, we show that:
• Graph-augmented neural network training can work
for a wide range of neural networks, such as feedforward, convolutional and recurrent networks. Additionally, this technique can be used in both inductive and transductive settings. It also helps learning in
low-sample regime (small number of labeled nodes),
which cannot be handled by vanilla neural network
training.
• The framework can handle multiple forms of graphs,
either naturally given or constructed based on embeddings and knowledge bases.
• Using graphs and neighbourhood information alone as
direct inputs to neural networks in this joint training
framework permits fast and simple inference, yet provides competitive performance with current state-ofthe-art approaches which employ a two-step method
of first training a node embedding representation from
the graph and then using it as feature input to train a
classifer separately (see section 4.1).
• As a by-product, our proposed framework provides
a simple technique to finding smaller and faster neural networks that offer competitive performance with
larger and slower non graph-augmented alternatives
(see section 4.2).
We experimentally show that the proposed training framework outperforms state-of-the-art or perform favourably on
a variety of prediction tasks and datasets, involving text features and/or graph inputs and on many different neural network architectures (see section 4).
The paper is organized as follows: we first review some
background and literature, and relate them to our approach
in section 2; we then detail the training objective and its
properties in section 3; and finally we validate our approach
on a range of experiments in section 4.
2. Background and related works
In this section, we will lay out the groundwork for our proposed training objective in section 3.
2.1. Neural network learning
Neural networks are a class of non-linear mapping from inputs to outputs and comprised of multiple layers that can
potentially learn useful representations for predicting the
outputs. We will view various models such as feed-forward
neural networks, recurrent neural networks and convolutional networks under the same umbrella. Given a set of N
training input-output pairs {xn , yn }N
n=1 , such neural networks are often trained by performing maximum likelihood
learning, that is, tuning their parameters so that the networks’ outputs are close to the ground truth under some
criterion,
X
CNN (θ) =
c(gθ (xn ), yn ),
(1)
n
where gθ (·) denotes the overall mapping, parameterized by
θ, and c(·) denotes a loss function such as l-2 for regression
or cross entropy for classification. The cost function c and
the mapping g are typically differentiable w.r.t θ, which facilitates optimisation via gradient descent. Importantly, this
can be scaled to a large number of training instances by
employing stochastic training using minibatches of data.
However, it is not clear how unlabeled data, if available,
can be treated using this objective, or if extra information
about the training set, such as relational structures can be
used.
2.2. Graph-based semi-supervised learning
In this section, we provide a concise introduction to graphbased semi-supervised learning using label propagation
and its training objective. Suppose we are given a graph
G = (V, E, W ) where V is the set of nodes, E the set of
edges and W the edge weight matrix. Let Vl , Vu be the labeled and unlabeled nodes in the graph. The goal is to predict a soft assignment of labels for each node in the graph,
Ŷ , given the training label distribution for the seed nodes,
Neural Graph Machines: Learning Neural Networks Using Graphs
Y . Mathematically, label propagation performs minimization of the following convex objective function, for L labels,
CLP (Ŷ ) = µ1
X
2
Ŷv − Yv
v∈Vl
X
+ µ2
2
2
wu,v Ŷv − Ŷu
v∈V,u∈N (v)
+ µ3
X
v∈V
2
2
Ŷv − U
,
2
(2)
PL
subject to l=1 Ŷvl = 1, where N (v) is the neighbour
node set of the node v, and U is the prior distribution
over all labels, wu,v is the edge weight between nodes u
and v, and µ1 , µ2 , and µ3 are hyperparameters that balance the contribution of individual terms in the objective.
The terms in the objective function above encourage that:
(a) the label distribution of seed nodes should be close to
the ground truth, (b) the label distribution of neighbouring
nodes should be similar, and, (c) if relevant, the label distribution should stay close to our prior belief. This objective
function can be solved efficiently using iterative methods
such as the Jacobi procedure. That is, in each step, each
node aggregates the label distributions from its neighbours
and adjusts its own distribution, which is then repeated until convergence. In practice, the iterative updates can be
done in parallel or in a distributed fashion which then allows large graphs with a large number of nodes and labels
to be trained efficiently. Bengio et al. (2006) and Ravi &
Diao (2016) provide good surveys on the topic for interested readers.
There are many variants of label propagation that could be
viewed as optimising modified versions of eq. (2). For
example, manifold regularization (Belkin et al., 2006) replaces the label distribution Ŷ by a Reproducing Kernel
Hilbert Space mapping from input features. Similarly, Weston et al. (2012) also employs such mapping but uses a
feed-forward neural network instead. Both methods can
be classified as inductive learning algorithms; whereas the
original label propagation algorithm is transductive (Yang
et al., 2016).
These aforementioned methods are closest to our proposed
approach; however, there are key differences. Our work
generalizes previously proposed frameworks for graphaugmented training of neural networks (e.g., Weston et al.
(2012)) and extends it to new settings, for example, when
there is only graph input and no features are available.
Unlike the previous works, we show that the graph augmented training method can work with multiple neural network architectures (Feed-forward NNs, CNNs, RNNs) and
on multiple prediction tasks and datasets using natural as
well as constructed graphs. The experiment results (see
section 4) clearly validate the effectiveness of this method
in all these different settings, in both inductive and transductive learning paradigms. Besides the methodology, our
study also presents an important contribution towards assessing the effectiveness of graph combined neural networks as a generic training mechanism for different architectures and problems, which was not well studied in previous work.
More recently, graph embedding techniques have been used
to create node embedding that encode local structures of the
graph and the provided node labels (Perozzi et al., 2014;
Yang et al., 2016). These techniques target learning better
node representations to be used for other tasks such as node
classification. In this work, we aim to directly learn better
predictive models from the graph. We compare our method
to these two-stage (embedding + classifier) techniques in
several experiments in section 4.
Our work is also different and orthogonal to recent works
on using neural networks on graphs, for example: Defferrard et al. (2016) employs spectral graph convolution to
create a neural-network like classifier. However, these approaches requires many approximations to arrive at a practical implementation. Here, we advocate a training objective that uses graphs to augment neural network learning,
and works with many forms of graphs and with any type of
neural network.
3. Neural graph machines
In this section, we devise a discriminative training objective for neural networks, that is inspired by the label propagation objective and uses both labeled and unlabeled data,
and can be trained by stochastic gradient descent.
First, we take a close look at the two objective functions discussed in section 2. The label propagation objective equation 2 ensures the predicted label distributions of
neighbouring nodes to be similar, while those of labeled
nodes to be close to the ground truth. For example: if a cat
image and a dog image are strongly connected in a graph,
and if the cat node is labeled as animal, the predicted probability of the dog node being animal is also high. In contrast, the neural network training objective equation 1 only
takes into account the labeled instances, and ensure correct
predictions on the training set. As a consequence, a neural
network trained on the cat image alone will not make an
accurate prediction on the dog image.
Such shortcoming of neural network training can be rectified by biasing the network using prior knowledge about
the relationship between instances in the dataset. In particular, for the domains we are interested in, training instances (either labeled or unlabeled) that are connected in
a graph, for example, dog and cat in the above example,
Neural Graph Machines: Learning Neural Networks Using Graphs
yi
xi
xi
xi
yi
xj
yj
xj
yj
xj
xk
Graph input
xk
Node features
A
B
xi
yi
xj
yj
yi
xi
This
sounds
good
yj
xj
That
looks
awesome
D
C
Figure 1. A: An example of a graph and feature inputs. In this case, there are two labeled nodes (xi , xj ) and one unlabeled node (xk ),
and two edges. The feature vectors, one for each node, are used as neural network inputs. B, C and D: Illustration of Neural Graph
Machine for feed-forward, convolution and recurrent networks respectively: the training flow ensures the neural net to make accurate
node-level predictions and biases the hidden representations/embeddings of neighbouring nodes to be similar. In this example, we force
hi and hj to be similar as there is an edge connecting xi and xj nodes.
should have similar predictions. This can be done by encouraging neighboring data points to have a similar hidden representation learnt by a neural network, resulting in
a modified objective function for training neural network
architectures using both labeled and unlabeled datapoints.
We call architectures trained using this objective Neural
Graph Machines, and schematically illustrate the concept
in figure 1. The proposed objective function is a weighted
sum of the neural network cost and the label propagation
cost as follows,
CNGM (θ) =
Vl
X
c(gθ (xn ), yn )
n=1
+ α1
X
wuv d(hθ (xu ), hθ (xv ))
(u,v)∈ELL
+ α2
X
wuv d(hθ (xu ), hθ (xv ))
(u,v)∈ELU
+ α3
X
wuv d(hθ (xu ), hθ (xv ),
(3)
(u,v)∈EU U
where ELL , ELU , and EU U are sets of labeled-labeled,
labeled-unlabeled and unlabeled-unlabeled edges correspondingly, h(·) represents the hidden representations of
the inputs produced by the neural network, and d(·) is a distance metric, and {α1 , α2 , α3 } are hyperparameters. Note
that we have separated the terms based on the edge types,
as these can affect the training differently.
In practice, we choose an l-1 or l-2 distance metric for d(·),
and h(x) to be the last layer of the neural network. However, these choices can be changed, to a customized metric,
or to using an intermediate hidden layer instead.
3.1. Connections to previous methods
The graph-dependent α hyperparameters control the balance of the contributions of different edge types. When
{αi = 0}3i=1 , the proposed objective ignores the similarity
constraint and becomes a supervised-only neural network
objective as in equation 1. When only α1 6= 0, the training cost has an additional term for labeled nodes, that acts
as a regularizer. When gθ (x) = hθ (x) = ŷ, where ŷ is
the label distribution, the individual cost functions (c and
d) are squared l-2 norm, and the objective is trained using
ŷ directly instead of θ, we arrive at the label propagation
objective in equation 2. Therefore, the proposed objective could be thought of as a non-linear version of the label
propagation objective, and a graph-regularized version of
Neural Graph Machines: Learning Neural Networks Using Graphs
3.3. Optimization
the neural network training objective.
3.2. Network inputs and graph construction
Similar to graph-based label propagation, the choice of the
input graphs is critical, to correctly bias the neural network’s prediction. Depending on the type of the graphs and
nodes in the graph, they can be readily available to use such
as social networks or protein linking networks, or they can
be constructed (a) using generic graphs such as Knowledge
Bases, that consists of relationship links between entities,
(b) using embeddings learnt by an unsupervised learning
technique, or, (c) using sparse feature representations for
each vertex. Additionally, the proposed training objective
can be easily modified for directed graphs.
We have discussed using node-level features as inputs to
the neural network. In the absence of such inputs, our training scheme can still be deployed using input features derived from the graph itself. We show in figure 2 and in
experiments that the neighbourhood information such as
rows in the adjacency matrix are simple to construct, yet
powerful inputs to the network. These features can also be
combined with existing features.
yi
xi
yj
xj
xk
The proposed objective function in equation 3 has several
summations over the labeled points and edges, and can be
equivalently written as follows,
X
CNGM (θ) =
α1 wuv d(hθ (xu ), hθ (xv )) + cuv
(u,v)∈ELL
X
+
α2 wuv d(hθ (xu ), hθ (xv )) + cu
(u,v)∈ELU
X
+
α3 wuv d(hθ (xu ), hθ (xv ),
(4)
(u,v)∈EU U
where
1
1
c(gθ (xu ), yu ) +
c(gθ (xv ), yv )
|u|
|v|
1
c(gθ (xu ), yu ),
cu =
|u|
cuv =
xi
|u| and |v| are the number of edges incident to vertices u
and v, respectively. The objective in its new form enables
stochastic training to be deployed by sampling edges. In
particular, in each training iteration, we use a minibatch of
edges and obtain the stochastic gradients of the objective.
To further reduce noise and speedup learning, we sample
edges within a neighbourhood region, that is to make sure
some sampled edges have shared end nodes.
xj
3.4. Complexity
xk
Figure 2. Illustration of how we can construct inputs to the neural
network using the adjacency matrix. In this example, we have
three nodes and two edges. The feature vector created for each
node (shown on the right) has 1’s at its index and indices of nodes
that it’s adjacent to.
When the number of graph nodes is high, this construction
can have a high complexity and result in a large number
of input features. This can be avoided by several ways:
(i) clustering the nodes and using the cluster assignments
and similarities, (ii) learning an embedding function of
nodes (Perozzi et al., 2014), or (iii) sampling the neighbourhood/context (Yang et al., 2016). In practice, we observe that the input space can be bounded by a constant,
even for massive graphs, with efficient scalable methods
like unsupervised propagation (i.e., propagating node identity labels across the graph and selecting ones with highest
support as input features to neural graph machines).
The complexity of each epoch in training using equation
4 is O(M ) where M = |E| is the number of edges in
the graph. In the case where there is a large number of
unlabeled-unlabeled edges, they potentially do not help
learning and could be ignored, leading to a lower complexity. One strategy to include them is self-training, that is to
grow seeds or labeled nodes as we train the networks. We
experimentally demonstrate this technique in section 4.4.
Predictions at inference time can be made at the same cost
as that of vanilla neural networks.
4. Experiments
In this section, we provide several experiments showing the
efficacy of the proposed training objective on a wide range
of tasks, datasets and network architectures. All the experiments are done using a TensorFlow implementation (Abadi
et al., 2015).
4.1. Multi-label Classification of Nodes on Graphs
We first demonstrate our approach using a multi-label classification problem on nodes in a relationship graph. In particular, the BlogCatalog dataset (Agarwal et al., 2009), a
Neural Graph Machines: Learning Neural Networks Using Graphs
network of social relationships between bloggers is considered. This graph has 10,312 nodes, 333,983 edges and
39 labels per node, which represent the bloggers, their social connections and the bloggers’ interests, respectively.
Following previous approaches in the literature (Grover &
Leskovec, 2016; Agarwal et al., 2009), we train and make
predictions using multiple one-vs-rest classifiers.
Since there are no provided features for each node, we
use the rows of the adjacency matrix as input features, as
discussed in section 3.2. Feed-forward neural networks
(FFNNs) with one hidden layer of 50 units are employed
to map the constructed inputs to the node labels. As we use
the test set to construct the graph and augment the training
objective, the training in this experiment is transductive.
Critically, to combat the unbalanced training set, we employ weighted sampling during training, i.e. making sure
each minibatch has both positive and negative examples. In
this experiment, we fix αi to be equal, and experiment with
α = 0.1 and use the l-2 metric to compute the distance d
between the hidden representations of the networks. In addition, we create a range of train/test splits by varying the
number of training points being presented to the networks.
We compare our method (NGM-FFNN) against a two-stage
approach that first uses node2vec (Grover & Leskovec,
2016) to generate node embeddings and then uses a linear one-vs-rest classifier for classification. The methods
are evaluated using two metrics Macro F1 and Micro F1.
The average results for different train/test splits using our
method and the baseline are included in table 1. In addition,
we compare NGM-FFNN with a non-augmented FFNN in
which α = 0, i.e. no edge information is used during training. We observe that the graph-augmented training scheme
performs better (6% relative improvement on Macro F1
when the training set size is 20% and 50% of the dataset) or
comparatively (when the training size is 80%) compared to
the vanilla neural networks trained with no edge information. Both methods significantly outperform the approach
that uses node embeddings and linear classifiers. We observe the same improvement over node2vec on the Micro
F1 metric and NGM-FFNN is comparable to vanilla FFNN
(α = 0) but outperforms other methods on the recall metric.
Table 1. Macro F1 results for BlogCatalog dataset averaged over
10 random splits. The higher is better. Graph regularized neural
networks outperform node2vec embedding and a linear classifer
in all training size settings.
2
|Train| / |Dataset|
NGM-FFNN
node2vec
20%
50%
80%
0.191
0.242
0.262
0.168
0.174
0.177
These results demonstrate that using the graph itself as direct inputs to the neural network and letting the network
figure out a non-linear mapping directly from the raw graph
is more effective than the two-stage approach considered.
More importantly, the results also show that using the graph
information improves the performance in the limited data
regime (for example: when training set is only 20% or 50%
of the dataset).
4.2. Text Classification using Character-level CNNs
We evaluate the proposed objective function on a multiclass text classification task using a character-level convolutional neural network (CNN). We use the AG news
dataset from (Zhang et al., 2015), where the task is to classify a news article into one of 4 categories. Each category has 30,000 examples for training and 1,900 examples
for testing. In addition to the train and test sets, there are
111,469 examples that are treated as unlabeled examples.
As there is no provided graph structure linking the articles,
we create such a graph based on the embeddings of the articles. We restrict the graph construction to only the train set
and the unlabeled examples and keep the test set only for
evaluation. We use the Google News word2vec corpus to
calculate the average embedding for each news article and
use the cosine similarity of document embeddings as a similarity metric. Each node is restricted to have a maximum
of 5 neighbors.
We construct the CNN in the same way as (Zhang et al.,
2015) and pick their competitive “small CNN” as our baseline for a more reasonable comparison to our set-up. Our
approach employs the same network, but with significantly
smaller number of convolutional layers and layer sizes, as
shown in table 2.
Table 2. Settings of CNNs for the text classification experiment,
including the number of convolutional layers and their sizes. The
baseline model is the small CNN from (Zhang et al., 2015) and is
significantly larger than our model.
Setting
Baseline
Our “tiny CNN”
# of conv. layers
Frame size in conv. layers
# of FC layers
Hidden units in FC layers
6
256
3
1024
3
32
3
256
The networks are trained with the same hyper-parameters
as reported in (Zhang et al., 2015). We observed that the
2
These results are different compared to (Grover & Leskovec,
2016), since we treat the classifiers (one per label) independently.
Both methods shown here use the exact same setting and training/test data splits.
Neural Graph Machines: Learning Neural Networks Using Graphs
model converged within 20 epochs (the model loss did not
change much) and hence used this as a stopping criterion
for this task. Experiments also showed that running the network for longer also did not change the qualitative performance. We use the cross entropy loss on the final outputs
of the network, that is d = cross entropy(g(xu ), g(xv )),
to compute the distance between nodes on an edge. In addition, we also experiment with a data augmentation technique using an English thesaurus, as done in (Zhang et al.,
2015).
We compare the “tiny CNN” trained using the proposed objective function with the baseline using the accuracy on the
test set in table 3. Our approach outperforms the baseline
by provides a 1.8% absolute and 2.1% relative improvement in accuracy, despite using a much smaller network. In
addition, our model with graph augmentation trains much
faster and produces results on par or better than the performance of a significantly larger network, “large CNN”
(Zhang et al., 2015), which has an accuracy of 87.18 without using a thesaurus, and 86.61 with the thesaurus.
Table 3. Results for news article categorization using characterlevel CNNs. Our method gives better predictive accuracy, despite
using a much smaller CNN compared to the “small CNN” baseline from (Zhang et al., 2015)‡ .
Network
Accuracy %
Baseline‡
Baseline with thesaurus augmentation‡
Our “tiny” CNN
Our “tiny” CNN with NGM
84.35
85.20
85.07
86.90
4.3. Semantic Intent Classification using LSTM RNNs
We compare the performance of our approach for training
RNN sequence models (LSTM) for a semantic intent classification task as described in the recent work on SmartReply (Kannan et al., 2016) for automatically generating short
email responses. One of the underlying tasks in SmartReply is to discover and map short response messages to semantic intent clusters.3 We choose 20 intent classes and
created a dataset comprised of 5,483 samples (3,832 for
training, 560 for validation and 1,091 for testing). Each
sample instance corresponds to a short response message
text paired with a semantic intent category that was manually verified by human annotators. For example, “That
sounds awesome!” and “Sounds fabulous” belong to the
sounds good intent cluster.
We construct a sparse graph in a similar manner as the
news categorization task using word2vec embeddings over
3
For details regarding SmartReply and how the semantic intent
clusters are generated, refer (Kannan et al., 2016).
the message text and computing similarity to generate a response message graph with fixed node degree (k=10). We
use l-2 for the distance metric d(·) and choose α based on
the development set.
We run the experiments for a fixed number of time steps
and pick the best results on the development set. A multilayer LSTM architecture (2 layers, 100 dimensions) is
used for the RNN sequence model. The LSTM model and
its NGM variant are also compared against other baseline
systems—Random baseline ranks the intent categories
randomly and Frequency baseline ranks them in order of
their frequency in the training corpus. To evaluate the
intent prediction quality of different approaches, for each
test instance, we compute the rank of the actual intent
category ranki with respect to the ranking produced by
the method and use this to calculate the Mean Reciprocal
Rank:
MRR =
N
1 X 1
N i=1 ranki
We show in table 4 that LSTM RNNs with our proposed
graph-augmented training objective function outperform
standard baselines by achieving a better MRR.
Table 4. Results for Semantic Intent Classification using graphaugmented LSTM RNNs and baselines. Higher MRR is better.
Model
Mean Reciprocal Rank (MRR)
Random
Frequency
LSTM
NGM-LSTM
0.175
0.258
0.276
0.284
4.4. Low-supervision Document Classification
Finally, we compare our method on a task with very limited
supervision—the PubMed document classification problem
(Sen et al., 2008). The task is to classify each document
into one of 3 classes, with each document being described
by a TF-IDF weighted word vector. The graph is available
as a citation network: two documents are connected to each
other if one cites the other. The graph has 19,717 nodes and
44,338 edges, with each class having 20 seed nodes and
1000 test nodes. In our experiments we exclude the test
nodes from the graph entirely, training only on the labeled
and unlabeled nodes.
We train a feed-forward neural network (FFNN) with two
hidden layers with 250 and 100 neurons, using the l-2 distance metric on the last hidden layer. The NGM-FFNN
model is trained with αi = 0.2, while the baseline FFNN
is trained with αi = 0 (i.e., a supervised-only model). We
Neural Graph Machines: Learning Neural Networks Using Graphs
use self-training to train the model, starting with just the 60
seed nodes (20 per class) as training data. The amount of
training data is iteratively increased by assigning labels to
the immediate neighbors of the labeled nodes and retraining the model. For the self-trained NGM-FFNN model, this
strategy results in incrementally growing the neighborhood
and thereby, LL and LU edges in equation 4 objective.
We compare the final NGM-FFNN model against the
FFNN baseline and other techniques reported in (Yang
et al., 2016) including the Planetoid models (Yang et al.,
2016), semi-supervised embedding (Weston et al., 2012),
manifold regression (Belkin et al., 2006), transductive
SVM (Joachims, 1999), label propagation (Zhu et al.,
2003), graph embeddings (Perozzi et al., 2014) and a linear
softmax model. Full results are included in table 5.
Table 5. Results for document classification on the PubMed
dataset using neural networks. The top results are taken from
(Yang et al., 2016). The bottom two rows are ours, with the NGM
training outperforming all other baselines, except Planetoid-I.
Please see text for relevant references.
Method
Accuracy
Linear + Softmax
Semi-supervised embedding
Manifold regularization
Transductive SVM
Label propagation
Graph embedding
Planetoid-I
Planetoid-G
Planetoid-T
0.698
0.711
0.707
0.622
0.630
0.653
0.772
0.664
0.757
Feed-forward NN
NGM-FFNN
0.709
0.759
The results show that the NGM model (without any
tuning) outperforms many baselines including FFNN,
semi-supervised embedding, manifold regularization
and Planetoid-G/Planetoid-T, and compares favorably to
Planetoid-I. Most importantly, this result demonstrates the
graph augmentation scheme can lead to better regularised
neural networks, especially in low sample regime (20
samples per class in this case). We believe that with
tuning, NGM accuracy can be improved even further.
5. Conclusions
We have revisited graph-augmentation training of neural
networks and proposed Neural Graph Machines as a general framework for doing so. Its objective function encourages the neural networks to make accurate node-level predictions, as in vanilla neural network training, as well as
constrains the networks to learn similar hidden representations for nodes connected by an edge in the graph. Impor-
tantly, the objective can be trained by stochastic gradient
descent and scaled to large graphs.
We validated the efficacy of the graph-augmented objective on various tasks including bloggers’ interest, text category and semantic intent classification problems, using a
wide range of neural network architectures (FFNNs, CNNs
and LSTM RNNs). The experimental results demonstrated
that graph-augmented training almost always helps to find
better neural networks that outperforms other techniques
in predictive performance or even much smaller networks
that are faster and easier to train. Additionally, the nodelevel input features can be combined with graph features
as inputs to the neural networks. We showed that a neural network that simply takes the adjacency matrix of a
graph and produces node labels, can perform better than a
recently proposed two-stage approach using sophisticated
graph embeddings and a linear classifier. Our framework
also excels when the neural network is small, or when there
is limited supervision available.
While our objective can be applied to multiple graphs
which come from different domains, we have not fully explored this aspect and leave this as future work. We expect
the domain-specific networks can interact with the graphs
to determine the importance of each domain/graph source
in prediction. We also did not explore using graph regularisation for different hidden layers of the neural networks;
we expect this is key for the multi-graph transfer setting
(Yosinski et al., 2014). Another possible future extension
is to use our objective on directed graphs, that is to control
the direction of influence between nodes during training.
Acknowledgements
We would like to thank the Google Expander team for insightful feedback.
References
Abadi, Martı́n, Agarwal, Ashish, Barham, Paul, Brevdo,
Eugene, Chen, Zhifeng, Citro, Craig, Corrado, Greg S.,
Davis, Andy, Dean, Jeffrey, Devin, Matthieu, Ghemawat, Sanjay, Goodfellow, Ian, Harp, Andrew, Irving, Geoffrey, Isard, Michael, Jia, Yangqing, Jozefowicz,
Rafal, Kaiser, Lukasz, Kudlur, Manjunath, Levenberg,
Josh, Mané, Dan, Monga, Rajat, Moore, Sherry, Murray,
Derek, Olah, Chris, Schuster, Mike, Shlens, Jonathon,
Steiner, Benoit, Sutskever, Ilya, Talwar, Kunal, Tucker,
Paul, Vanhoucke, Vincent, Vasudevan, Vijay, Viégas,
Fernanda, Vinyals, Oriol, Warden, Pete, Wattenberg,
Martin, Wicke, Martin, Yu, Yuan, and Zheng, Xiaoqiang.
TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL http://tensorflow.
org/. Software available from tensorflow.org.
Neural Graph Machines: Learning Neural Networks Using Graphs
Agarwal, Nitin, Liu, Huan, Murthy, Sudheendra, Sen,
Arunabha, and Wang, Xufei. A social identity approach
to identify familiar strangers in a social network. In 3rd
International AAAI Conference on Weblogs and Social
Media (ICWSM09), 2009.
Belkin, Mikhail, Niyogi, Partha, and Sindhwani, Vikas.
Manifold regularization: A geometric framework for
learning from labeled and unlabeled examples. Journal
of machine learning research, 7(Nov):2399–2434, 2006.
Bengio, Yoshua, Delalleau, Olivier, and Le Roux, Nicolas.
Label propagation and quadratic criterion. In Chapelle,
O, Scholkopf, B, and Zien, A (eds.), Semi-supervised
learning, pp. 193–216. MIT Press, 2006.
Defferrard, Michaël, Bresson, Xavier, and Vandergheynst,
Pierre. Convolutional neural networks on graphs with
fast localized spectral filtering. In Advances in Neural
Information Processing Systems, pp. 3837–3845, 2016.
Grover, Aditya and Leskovec, Jure. node2vec: Scalable
feature learning for networks. In Proceedings of the 22nd
ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, San Francisco, CA, USA,
August 13-17, 2016, pp. 855–864, 2016.
Hinton, Geoffrey, Deng, Li, Yu, Dong, Dahl, George E,
Mohamed, Abdel-rahman, Jaitly, Navdeep, Senior, Andrew, Vanhoucke, Vincent, Nguyen, Patrick, Sainath,
Tara N, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four
research groups. IEEE Signal Processing Magazine, 29
(6):82–97, 2012.
Joachims, Thorsten. Transductive inference for text classification using support vector machines. In International
Conference on Machine Learning, 1999.
Kannan, Anjuli, Kurach, Karol, Ravi, Sujith, Kaufmann,
Tobias, Tomkins, Andrew, Miklos, Balint, Corrado,
Greg, Lukacs, Laszlo, Ganea, Marina, Young, Peter,
and Ramavajjala, Vivek. Smart reply: Automated response suggestion for email. In Proceedings of the ACM
SIGKDD Conference on Knowledge Discovery and Data
Mining (KDD), 2016.
Kingma, Diederik P, Mohamed, Shakir, Rezende,
Danilo Jimenez, and Welling, Max. Semi-supervised
learning with deep generative models. In Advances in
Neural Information Processing Systems, pp. 3581–3589,
2014.
Krizhevsky, Alex, Sutskever, Ilya, and Hinton, Geoffrey E.
Imagenet classification with deep convolutional neural
networks. In Advances in Neural Information Processing
Systems, pp. 1097–1105, 2012.
Lee, Dong-Hyun. Pseudo-label: The simple and efficient
semi-supervised learning method for deep neural networks. In ICML 2013 Workshop : Challenges in Representation Learning (WREPL), 2013.
Perozzi, Bryan, Al-Rfou, Rami, and Skiena, Steven. Deepwalk: Online learning of social representations. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pp.
701–710. ACM, 2014.
Ravi, Sujith and Diao, Qiming. Large scale distributed
semi-supervised learning using streaming approximation. In Proceedings of the 19th International Conference on Artificial Intelligence and Statistics, pp. 519–
528, 2016.
Sen, Prithviraj, Namata, Galileo, Bilgic, Mustafa, Getoor,
Lise, Galligher, Brian, and Eliassi-Rad, Tina. Collective
classification in network data. AI magazine, 29(3):93,
2008.
Sutskever, Ilya, Vinyals, Oriol, and Le, Quoc V. Sequence
to sequence learning with neural networks. In Advances
in Neural Information Processing Systems, pp. 3104–
3112, 2014.
Weston, Jason, Ratle, Frédéric, Mobahi, Hossein, and Collobert, Ronan. Deep learning via semi-supervised embedding. In Neural Networks: Tricks of the Trade, pp.
639–655. Springer, 2012.
Yang, Zhilin, Cohen, William, and Salakhudinov, Ruslan.
Revisiting semi-supervised learning with graph embeddings. In Proceedings of The 33rd International Conference on Machine Learning, pp. 40–48, 2016.
Yosinski, Jason, Clune, Jeff, Bengio, Yoshua, and Lipson,
Hod. How transferable are features in deep neural networks? In Advances in Neural Information Processing
Systems, pp. 3320–3328, 2014.
Zhang, Xiang, Zhao, Junbo, and LeCun, Yann. Characterlevel convolutional networks for text classification. In
Advances in Neural Information Processing Systems, pp.
649–657, 2015.
Zhu, X, Ghahramani, Z, and Lafferty, J. Semi-supervised
learning using gaussian fields and harmonic functions.
In Proceedings of the 20th International Conference on
Machine Learning (ICML-2003) Volume 2, volume 2,
pp. 912–919. AIAA Press, 2003.
Zhu, Xiaojin and Ghahramani, Zoubin. Learning from labeled and unlabeled data with label propagation. Technical report, School of Computer Science, Canegie Mellon
University.
| 9cs.NE
|
Evaluating Haskell expressions in a tutoring environment
Tim Olmer
Bastiaan Heeren
Open University of the Netherlands
Faculty of Management, Science and Technology
Heerlen, The Netherlands
[email protected]
[email protected]
Johan Jeuring
Universiteit Utrecht
Department of Information and Computing Sciences
Utrecht, The Netherlands
[email protected]
A number of introductory textbooks for Haskell use calculations right from the start to give the
reader insight into the evaluation of expressions and the behavior of functional programs. Many
programming concepts that are important in the functional programming paradigm, such as recursion,
higher-order functions, pattern-matching, and lazy evaluation, can be partially explained by showing
a stepwise computation. A student gets a better understanding of these concepts if she performs these
evaluation steps herself. Tool support for experimenting with the evaluation of Haskell expressions
is currently lacking. In this paper we present a prototype implementation of a stepwise evaluator for
Haskell expressions that supports multiple evaluation strategies, specifically targeted at education.
Besides performing evaluation steps the tool also diagnoses steps that are submitted by a student,
and provides feedback. Instructors can add or change function definitions without knowledge of the
tool’s internal implementation. We discuss some preliminary results of a small survey about the tool.
1
Introduction
Many textbooks that introduce the functional programming language Haskell begin with showing calculations that illustrate how expressions are evaluated, emphasizing the strong correspondence with mathematical expressions and the property of referential transparency of the language. For instance, Hutton
presents calculations for the expressions double 3 and double (double 2) on page 1 of his textbook [12],
and Bird and Wadler show the steps of reducing square (3 + 4) on page 5 of their book [3]. Similar
calculations can be found in the first chapter of Thompson’s The Craft of Functional Programming [28]
and Hudak’s The Haskell School of Expression. Textbooks that do not show these calculations [23, 18]
compensate for this by giving lots of examples with an interpreter.
Stepwise evaluating an expression on a piece of paper can give a student a feeling for what a program
does [4]. However, there is no simple way to view intermediate evaluation steps for a Haskell expression.
In this paper we present a prototype implementation of the Haskell Expression Evaluator (HEE) that can
show evaluation steps, and lets students practice with evaluating expressions on their own by providing
feedback and suggestions (see Figure 1).1 The tool supports multiple evaluation strategies and can handle
multiple (alternative) definitions for functions from the prelude. It is relatively easy for instructors to
change the granularity of the steps, or to customize the feedback messages that are reported by the tool.
Showing calculations can be a useful approach to let a student better understand some of the central
programming concepts behind the programming language Haskell, such as pattern-matching, recursion,
higher-order functions, and lazy evaluation. This approach is also used in textbooks on Haskell [12,
3]. For instance, Haskell’s lazy evaluation strategy is fundamentally different from other mainstream
programming languages (such as C, C++, C# and Java), and the tool can highlight these differences.
A novice functional programmer often faces difficulties in understanding the evaluation steps in a lazy
1 The
prototype is available via http://ideas.cs.uu.nl/HEE/.
J. Caldwell, Ph.K.F. Hölzenspies, P. Achten:
Trends in Functional Programming in Education (TFPIE 2014)
EPTCS 170, 2014, pp. 50–66, doi:10.4204/EPTCS.170.4
T. Olmer, B. Heeren & J. Jeuring
51
Figure 1: Prototype screen for student interaction
language, and even more experienced programmers find it hard to predict the space behavior of their
programs [2]. Another stumbling block is the very compact syntax that is used in Haskell, which makes
it sometimes hard to get an operational view of a functional program [27]. More generally, students often
do not clearly understand operator precedence and associativity and misinterpret expressions [14, 15].
Showing evaluation steps can partly alleviate these problems.
Contributions and scope. This paper presents a tool that enables students to practice with evaluation
steps for Haskell expressions. A student can not only inspect the evaluation steps of a program, but can
also provide evaluation steps as input, which can then be checked against various evaluation strategies.
Furthermore, the steps are presented at a level of abstraction typically expected in an educational setting.
Practicing with evaluation steps gives a student insight into how certain programming concepts such as
recursion, higher-order functions, and pattern matching work. It also gives a student insight into various
evaluation strategies.
Our evaluation tool only supports integers, list notation, recursion, higher order functions, and pattern
matching. The target audience for the evaluator is students taking an introductory course on functional
programming. Another limitation is that only small code fragments are considered, which is sufficient
for the intended audience. We assume that the expressions evaluated by the tool are well-typed and do
not contain compile-time errors, although we could let the tool check for errors before evaluating.
Related work mainly focuses on showing evaluation steps [17, 25], and does not offer the possibility
to let a student enter evaluation steps herself, or presents evaluation steps at a lower level of abstraction,
such as the lambda-calculus [26, 1].
52
Evaluating Haskell expressions in a tutoring environment
Roadmap. The rest of the paper is structured as follows. We start with an example that illustrates
different evaluation strategies in Section 2. Next, we define rewrite rules and rewrite strategies for a
simple expression language in Section 3, which are used for stepwise evaluating an expression and for
calculating feedback. We also discuss how rewrite rules and rewrite strategies can be generated for
arbitrary function definitions. Section 4 discusses the prototype in more detail, shows how we present
feedback, and describes the results of the survey. We conclude the paper with a discussion on related
work (Section 5) and present conclusions and future work (Section 6).
2
An example
We start by demonstrating some evaluation strategies that are supported by the tool. We use the expression sum ([3, 7] ++ [5]) as a running example in the rest of the paper. Because the tool is developed
for education, we use list notation when showing expressions and we make associativity explicit. The
evaluation steps in our examples are based on the following standard definitions for prelude functions:
sum :: [Int ] → Int
sum = foldl (+) 0
foldl :: (a → b → a) → a → [b] → a
foldl f v [ ]
=v
foldl f v (x : xs) = foldl f (f v x) xs
(++) :: [a] → [a] → [a]
[]
++ ys = ys
(x : xs) ++ ys = x : (xs ++ ys)
Figure 2 shows two different ways to evaluate the expression sum ([3, 7] ++ [5]). The evaluation
steps on the left-hand side of Figure 2 correspond to an outermost evaluation order (call-by-name). After
rewriting sum into a foldl, it is the list pattern in foldl’s definition (its third argument) that drives evaluation. The evaluation steps nicely show the accumulating parameter of foldl for building up the result, the
interleaving of steps for ++ (which produces a list) and foldl (which consumes list), and the additions
that are calculated at the very end. The evaluation steps on the right-hand side of Figure 2 illustrate the
left-most innermost evaluation order (call-by-value), which fully evaluates sub-expression [3, 7] ++ [5]
before using foldl’s definition. Observe that in contrast to call-by-name evaluation the additions are immediately computed. Also observe that sum is immediately rewritten into foldl. This might be surprising
behavior of an innermost evaluation strategy where arguments are completely evaluated before the function is evaluated. The reason for this behavior lies in the definition of sum. The definition of sum does
not have an explicitly specified parameter, but it applies foldl partially. Therefore, the evaluator does not
handle the sub-expression [3, 7] ++ [5] as a child of sum but as a neighbor of sum.
From an educational perspective it is interesting to allow for alternative definitions of prelude functions, e.g. sum defined with explicit recursion, or sum defined with the strict foldl0 function. With our
tool it is possible to switch between these alternative definitions and to observe the consequences for
evaluation.
It is important to keep in mind that the tool is capable of doing more than only showing evaluation
steps. The tool also lets students practice with evaluating expressions, and can diagnose intermediate
steps, suggest reducible expressions, and provide progress information by showing the number of eval-
T. Olmer, B. Heeren & J. Jeuring
=
=
=
=
=
=
=
=
=
=
=
sum ([3, 7] ++ [5])
{ definition sum }
foldl (+) 0 ([3, 7] ++ [5])
{ definition ++ }
foldl (+) 0 (3 : ([7] ++ [5]))
{ definition foldl }
foldl (+) (0 + 3) ([7] ++ [5])
{ definition ++ }
foldl (+) (0 + 3) (7 : ([ ] ++ [5]))
{ definition foldl }
foldl (+) ((0 + 3) + 7) ([ ] ++ [5])
{ definition ++ }
foldl (+) ((0 + 3) + 7) [5]
{ definition foldl }
foldl (+) (((0 + 3) + 7) + 5) [ ]
{ definition foldl }
((0 + 3) + 7) + 5
{ applying + }
(3 + 7) + 5
{ applying + }
10 + 5
{ applying + }
15
53
=
=
=
=
=
=
=
=
=
=
=
sum ([3, 7] ++ [5])
{ definition sum }
foldl (+) 0 ([3, 7] ++ [5])
{ definition ++ }
foldl (+) 0 (3 : ([7] ++ [5]))
{ definition ++ }
foldl (+) 0 (3 : (7 : ([ ] ++ [5])))
{ definition ++ }
foldl (+) 0 [3, 7, 5]
{ definition foldl }
foldl (+) (0 + 3) [7, 5]
{ applying + }
foldl (+) 3 [7, 5]
{ definition foldl }
foldl (+) (3 + 7) [5]
{ applying + }
foldl (+) 10 [5]
{ definition foldl }
foldl (+) (10 + 5) [ ]
{ applying + }
foldl (+) 15 [ ]
{ definition foldl }
15
Figure 2: Evaluating sum ([3, 7] ++ [5]) using the outermost (left-hand side) or innermost (right-hand
side) evaluation strategy
uation steps remaining. It is possible to train one particular evaluation strategy, or to allow any possible
reduction step.
3
Rewrite rules and strategies
We use I DEAS, a framework for developing domain reasoners that give intelligent feedback [9], for
rewriting expressions. Many exercises, such as solving a mathematical equation or a programming exercise, can be solved by following some kind of procedure. A procedure or strategy describes how basic
steps may be combined to solve a particular problem. Such a strategy is expressed in an embedded
domain-specific language. I DEAS interprets a strategy as a context-free grammar. The sentences of this
grammar are sequences of rewrite steps that are used to check if a student follows the strategy. The main
advantage of using I DEAS in our tool is that it is a generic framework that makes it possible to define
exercises that must be solved using some kind of strategy, and that it provides feedback to a student who
is solving an exercise. Feedback is added by means of labels at particular locations in the strategy.
To use the I DEAS framework, we construct three components: the domain of the exercise (an expression datatype), rules for rewriting terms in this domain (the evaluation steps), and a rewrite strategy that
54
Evaluating Haskell expressions in a tutoring environment
data Expr = App Expr Expr
| Abs String Expr
| Var String
| Lit Int
-- application
-- lambda abstraction
-- variable
-- integer
-- smart constructors
appN
= foldl App
-- n-ary application
nil
= Var "[]"
cons x xs = appN (Var ":") [x, xs]
Figure 3: Datatype for expressions
combines these rules. Other components, such as parsing, pretty-printing, and testing expressions for
equality, are omitted in this paper.
Figure 3 defines an expression datatype with application, lambda abstraction, variables, and integers,
together with some helper functions for constructing expressions. The Var constructor is also used to
represent datatype constructors (e.g., constructor : for building lists).
3.1
Rewrite rules
We introduce a rewrite rule for each function (and operator) from the prelude. The rewrite rules are
based on datatype-generic rewriting technology [21], where rules are constructed using operator .
This operator takes expressions on the left-hand side and the right-hand side. Based on sum’s definition,
we define the rewrite rule for sum as follows:
sumRule :: Rule Expr
sumRule = describe "Calculate the sum of a list of numbers" $
rewriteRule "eval.sum.rule" $
Var "sum" appN (Var "foldl") [Var "+", Lit 0]
Each rule has an identifier (here "eval.sum.rule") that is used for identifying the rewrite step, and
optionally also a description for explaining the step. The descriptions of the prelude functions are taken
from the appendix of Hutton’s textbook [12]. Note that functions such as describe, rewriteRule, operator
and type constructor Rule are provided by the I DEAS framework.
The rewrite rule for foldl’s definition is more involved since it uses pattern matching. The pattern
variables in foldl’s definition are turned into meta-variables of the rewrite rule by introducing these
variables in a lambda abstraction:
foldlRule :: Rule Expr
foldlRule =
describe "Process a list using an operator that associates to the left" $
rewriteRules "eval.foldl.rule"
[λ f v x xs → appN (Var "foldl") [f , v, nil] v
, λ f v x xs → appN (Var "foldl") [f , v, cons x xs]
appN (Var "foldl") [f , appN f [v, x], xs]
]
T. Olmer, B. Heeren & J. Jeuring
55
Combinator
Description
Combinator
Description
s <?> t
s <|> t
s.t
fix f
label ` s
succeed
first s, then t
either s or t
apply s, or else t
fixed point combinator
attach label ` to s
always succeeds
sequence xs
alternatives xs
repeat s
child n s
outermost s
spinebu s
checkCurrent p
generalizes sequence (<?>) to lists
generalizes choice (<|>) to lists
apply s as long as possible
apply s to the n-th child
apply s at left-most outermost position
apply s to the left-spine (bottom-up)
succeeds if predicate p holds
Figure 4: Strategy combinators
The rewrite rules sumRule, foldlRule, and appendRule (for operator ++) have an intensional representation with a left- and right-hand side, which does not only make the rules easier to define, but also
lets us generate documentation for the rule, take the inverse of the rule, or alter the matching algorithm
for the rule’s left-hand side (e.g., to take associativity of an operator into account).
Besides the rewrite rules, we also introduce a rule for the primitive addition function (addRule), and
a rule for beta-reduction. Adding two integers cannot be defined by a rewrite rule, and therefore we use
the function makeRule from the I DEAS framework to turn a function of type Expr → Maybe Expr into a
value of type Rule Expr.
addRule :: Rule Expr
addRule = describe "Add two integers" $ makeRule "eval.add.rule" f
where
f :: Expr → Maybe Expr
f (App (App (Var "+") (Lit x)) (Lit y)) = Just $ Lit (x + y)
f
= Nothing
In a similar way we define betaReduction :: Rule Expr that reduces expressions of the form (λ x → e1 ) e2
by using makeRule and implementing a capture-avoiding substitution.
3.2
Rewrite strategies
The embedded domain-specific language for specifying rewrite strategies in I DEAS defines several generic combinators to combine rewrite rules into a strategy [9]. We briefly introduce the combinators that
are used in this paper. The combinators are summarized in Figure 4.
Rewrite rules are the basic building block for composing rewrite strategies. All strategy combinators
in the I DEAS framework are overloaded and take rules or strategies as arguments. The sequence combinator (<?>) specifies the sequential application of two strategies. The choice combinator (<|>) defines
that either the first operand or the second operand is applied: combinator alternatives generalizes the
choice combinator to lists. The left-biased choice combinator (.) only tries the second strategy if the first
strategy fails. Combinator repeat is used for repetition: this combinator applies its argument strategy as
often as possible. The fixed point combinator fix is used to explicitly model recursion in the strategy. It
takes as argument a function that maps a strategy to a new strategy. We can use labels at any position in
the strategy to specialize the feedback that is generated.
The strategy language supports all the usual traversal combinators such as innermost and oncebu [30].
The strategy child n s applies strategy s to the n-th child and can be used to define other generic traversal
56
Evaluating Haskell expressions in a tutoring environment
isApp :: Expr → Bool
isApp (App
) = True
isApp
= False
isFun :: String → Int → Expr → Bool
isFun fn 0 (Var s) = fn == s
isFun fn n (App f ) = isFun fn (n − 1) f
isFun
= False
Figure 5: Predicates on expressions (helper functions)
combinators. Combinator checkCurrent takes a predicate and only succeeds if the predicate holds for the
current expression.
An evaluation strategy defines in which order sub-expressions are reduced. We can use the standard
left-most outermost (or innermost) rewrite strategy to turn the rewrite rules into an evaluation strategy:
rules :: [Rule Expr ]
rules = [sumRule, foldlRule, appendRule, addRule, betaReduction]
evalOutermost :: LabeledStrategy (Context Expr)
evalOutermost = label "eval.outer" $
outermost (alternatives (map liftToContext rules))
Note that we use the Context type from the I DEAS framework as a zipper [11] datatype for traversing
expressions. A zipper maintains a sub-expression that has the focus, and is used by traversal combinators
such as outermost. We lift the rules to the Context type with the function liftToContext :: Rule a →
Rule (Context a) that applies the rule to the sub-expression that currently has the focus.
The attentive reader will have noticed that the evalOutermost strategy does not result in the evaluation
that is shown on the left-hand side of Figure 2. Evaluation of a foldl application is driven by pattern
matching on the function’s third argument (the list). The expression at this position should first be
evaluated to weak-head normal form (whnf), after which we can decide which case to take. We define
a rewrite strategy that first checks that foldl is applied to exactly three arguments, then brings the third
argument to weak-head normal form, and finally applies the rewrite rule for foldl. We need a strategy
that can evaluate an expression to weak-head normal form, and pass this as an argument to the strategy
definition.
foldlS :: Strategy (Context Expr) → LabeledStrategy (Context Expr)
foldlS whnf = label "eval.foldl" $
checkCurrent (isFun "foldl" 3) -- check that foldl has exactly 3 arguments
<?> arg 3 3 whnf
-- bring the third argument (out of 3) to whnf
<?> liftToContext foldlRule
-- apply the rewrite rule for foldl’s definition
The predicate isFun (defined in Figure 5) tests whether an expression is a function application of a
specific function with an exact number of arguments. We define the strategy combinator arg i n s, used in
the definition of foldlS, to apply strategy s to the i-th argument of a function application with n arguments.
For the last argument (i == n), we apply s to the second sub-expression of an application. Otherwise, we
visit the first sub-expression and call arg recursively. The combinator is not defined for i > n.
T. Olmer, B. Heeren & J. Jeuring
57
arg :: Int → Int → Strategy (Context a) → Strategy (Context a)
arg i n s | i == n = child 1 s
| i < n = child 0 (arg i (n − 1) s)
We also give the evaluation strategy for the definition of the append function, to emphasize its similarity with the definition of foldlS. Other function definitions have a similar evaluation strategy.
appendS :: Strategy (Context Expr) → LabeledStrategy (Context Expr)
appendS whnf = label "eval.append" $
checkCurrent (isFun "++" 2) -- check that append (++) has exactly 2 arguments
<?> arg 1 2 whnf
-- bring the first argument (out of 2) to whnf
<?> liftToContext appendRule
-- apply the rewrite rule for append’s definition
Evaluating an expression to weak-head normal form is a fixed-point computation over the evaluation
strategies for the definitions, since each definition takes the whnf strategy as an argument. We combine
the evaluation strategies with the rewrite rule for beta-reduction, apply it to the left-spine of an application (in a bottom-up way), and repeat this until the strategy can no longer be applied. This brings the
expression in weak-head normal form.
prelude :: [Strategy (Context Expr) → LabeledStrategy (Context Expr)]
prelude = [sumS, foldlS, appendS, addS]
spinebu :: Strategy (Context Expr) → Strategy (Context Expr)
spinebu s = fix $ λ x → (checkCurrent isApp <?> child 0 x) . s
evalWhnf :: LabeledStrategy (Context Expr)
evalWhnf = label "eval.whnf" $ fix $ λ whnf →
repeat (spinebu (liftToContext betaReduction <|> alternatives [f whnf | f ← prelude]))
The repeat combinator in the definition of evalWhnf applies its argument strategy zero or more times.
For example, the strategy does not have to be applied for an expression that already is in weak-head
normal form, and is applied twice to rewrite the expression (id id) 3 into id 3 and then 3 (where id is
Haskell’s identity function).
The result of applying evalWhnf to the expression sum ([3, 7] ++ [5]) gives the calculation shown in
Figure 2 on the left-hand side. To fully evaluate a value (such as a list), we have to repeat the evalWhnf
strategy for the sub-parts of a constructor.
3.3
User-defined function definitions
From an educational viewpoint it is interesting to experiment with different function definitions for the
same function and to observe the implications. For example, the function sum can be defined using foldl,
foldr, the strict foldl0 , or as an explicit recursive definition. This feature is also interesting for instructors
who would like to incorporate the tool into their course on functional programming because they can
then use their own examples to explain programming concepts to their students.
Multiple functions definitions can be added by giving each definition a different name (for instance,
sum, sum0 , and sum00 ) and by manually adding rewrite rules and evaluation strategies for these definitions. However, this approach has some drawbacks. To define rewrite rules and evaluation strategies,
the maintainer of the prototype (probably the instructor) needs to have knowledge of the datatype for
58
Evaluating Haskell expressions in a tutoring environment
data Pat = PCon String [Pat ]
| PVar String
| PLit Int
-- pattern constructor
-- pattern variable
-- integer pattern
data Def = Def String [Pat ] Expr
-- function definition
Figure 6: Datatype for patterns and function definitions
Haskell expressions and of the internal implementation to construct rules and evaluation strategies. After
adding definitions, the prototype needs to be recompiled, and hence a maintainer also needs a complete
build environment (Haskell compiler and libraries used). Another disadvantage is that the translation of
function definitions to rewrite rules is a manual, error-prone process.
Function definitions have rewrite rules and evaluation strategies of a similar structure. An alternative
approach to manually adding rewriting strategies is therefore to generate rewrite rules and evaluation
strategies from function definitions. These function definitions can be defined in a Haskell source file,
and using annotations [7] we can add a description to every function definition. Annotations are multiline comments, so the file remains a valid Haskell source file. Annotations have an identifier, for instance
DESC for description, which is used by the prototype to interpret the string after DESC as the rule
description. With this approach there is a single file, located on the web server, that contains the function
definitions, and their corresponding descriptions. No knowledge of the tool’s internal implementation is
required to add or change function definitions. An example of such a file is given below:
{-# DESC sum defined with a foldr to sum up all elements of a list. #-}
sum0 = foldr (+) 0
{-# DESC sum defined recursively to sum up all elements of a list. #-}
sum00 [ ]
=0
sum00 (x : xs) = x + sum00 xs
{-# DESC double function to double a number. #-}
double x = x + x
We name this file the evaluator’s prelude because it defines the functions that can be used in the
evaluator. Notice that primitive functions (such as the operator +) cannot be defined in the prelude
file. Primitive functions are functions that cannot be implemented directly in Haskell and are provided
natively by the compiler.
The function definitions given in the example prelude file use pattern matching. The left-hand side
of the equal sign consists of the function name and zero or more patterns. If the function name and the
patterns match with a particular expression the expression is rewritten to the expression at the right-hand
side of the equal sign (after substituting the pattern variables). Figure 6 defines a datatype for patterns
and function definitions.
A function definition can have multiple function bindings. For example, the function foldl has a
binding for the empty list and the non-empty list. In the remainder of the paper we will only consider
function definitions with one binding. Definitions with more bindings can be supported by combining
generated rules and strategies. An evaluation strategy is generated for each function binding, and these
strategies are tried in order of appearance in the corresponding definition. For this, we use the strategy
combinator . from the I DEAS framework.
T. Olmer, B. Heeren & J. Jeuring
59
Generation of rewrite rules. We need a function that takes a rule identifier (that may contain a description) and a function definition and turns these into a rewrite rule for expressions.
genRule :: Id → Def → Rule Expr
Rewrite rules are constructed with operator
that expects two expressions as its operands. Hence, we
need a function to convert the patterns on the left-hand side of a definition into an expression.
patToExpr :: Pat → Expr
patToExpr (PCon s ps) = appN (Var s) (map patToExpr ps)
patToExpr (PVar s)
= Var s
patToExpr (PLit n)
= Lit n
Note that the pattern variables of a function definition act as meta-variables in the rewrite rule, and the
corresponding variables on the right-hand side have to be substituted when the rewrite rule is applied.
We omit the definition of genRule that uses patToExpr.
Generation of evaluation strategies. We can generate an evaluation strategy for a function definition
used in the outermost evaluation strategy based on pattern matching. Such an evaluation strategy consists
of three steps: check the name of the function, bring some arguments to weak-head normal form, and
apply the definition’s rewrite rule. We first focus on the second step of bringing arguments to weak-head
normal form and define strategy combinator args for applying a list of strategies to the arguments of a
function application. We reuse combinator arg:
args :: [Strategy (Context a)] → Strategy (Context a)
args xs = sequence [arg i (length xs) s | (i, s) ← zip [1 . .] xs]
The function patS constructs an evaluation strategy for a given pattern. For pattern variables, nothing
has to be evaluated. For a pattern constructor, we bring the expression to weak-head normal form, check
the name of the constructor, and then recursively deal with the constructor’s sub-patterns. For a literal
pattern, we bring the expression to weak-head normal form and then check whether it is the same number.
patS :: Strategy (Context Expr) → Pat → Strategy (Context Expr)
patS whnf (PCon s ps) = whnf <?> patListS whnf s ps
patS whnf (PVar s)
= succeed
patS whnf (PLit n)
= whnf <?> checkCurrent (==Lit n)
patListS :: Strategy (Context Expr) → String → [Pat ] → Strategy (Context Expr)
patListS whnf s ps = checkCurrent (isFun s (length ps))
<?> args (map (patS whnf ) ps)
Function genEvalStrat takes a rule identifier, a function definition and the weak-head normal form strategy and returns an evaluation strategy for that function definition. First, arguments are evaluated by
patListS. Second, the generated rewrite rule for the definition is applied.
genEvalStrat :: Id → Def → Strategy (Context Expr) → Strategy (Context Expr)
genEvalStrat rId (Def s ps e) whnf = patListS whnf s ps
<?> liftToContext (genRule rId (Def s ps e))
60
Evaluating Haskell expressions in a tutoring environment
Figure 7: Component diagram of the prototype
The generated evaluation strategy defined above does not behave correctly for multiple or nested patterns.
The strategy s <?> t only succeeds if both s and t succeed. In the context of an evaluation strategy this is
undesirable because the effect of partly evaluating expressions with strategy s is undone (and cannot be
observed) if strategy t fails. The solution is to replace all occurrences of <?> (and derived combinators
such as sequence) in the generated strategy by a new combinator <?, which we define as:
(<?) :: Strategy a → Strategy a → Strategy
s <? t = s <?> (t . succeed)
4
Prototype
The prototype for practicing with evaluation steps is divided into a number of separate components.
Figure 7 shows the component model of the prototype: it consists of a front-end, a back-end, and a
strategy component. The strategy component contains all rewrite rules and rewrite strategies for a certain
evaluation strategy. The back-end component uses the external components I DEAS and Helium, and
reads the prelude file with function definitions that resides on the server. This paper focuses on the
back-end and strategy components.
In the future we hope to integrate the evaluator with the A SK -E LLE programming tutor [7] for learning Haskell. This tutor supports students in solving introductory programming exercises by providing
hints on how to continue, and by diagnosing intermediate programs that are submitted by a student.
Having a stepwise evaluator in the A SK -E LLE environment is promising because it allows students to
see their programs being evaluated, and to pinpoint mistakes in their definitions (e.g. by evaluating a
counter-example). A SK -E LLE is also build on top of the I DEAS framework and the Helium compiler.
Decomposing the evaluator into components, some of which are shared with A SK -E LLE, should make
future integration easier.
Front-end. The front-end is web-based and written in HTML and JavaScript. It uses JSON to communicate with the back-end. It provides an interface to inspect the evaluation of a Haskell expression or to
T. Olmer, B. Heeren & J. Jeuring
61
practice with the evaluation of a Haskell expression. The prototype front-end can easily be replaced by
another front-end and the purpose of this prototype front-end is to show what kind of I DEAS services [8]
can be used.
A user can select an example Haskell expression by clicking on the ‘Select’ button (see Figure 1)
or she can enter a Haskell expression. The prototype currently only supports a subset of the Haskell
syntax, so it is possible that this operation fails. After typing or selecting a Haskell expression the user
can choose between the innermost evaluation strategy and the outermost evaluation strategy. The user
can now call several standard I DEAS services such as the service for calculating the number of steps left,
getting information about the next rule that should be applied, or finding out what the result is of applying
the rule. The user can fill in the next evaluation step, possibly with the help of the services, and click on
the button ‘Diagnose’ to see if the provided next step is the correct step according to the strategy.
The string representation of a rule is used by the feedback service that gives information about the
next rule that should be applied. This string representation can be modified in a script file [8] where rule
identifiers are mapped to textual representations. Every rewrite rule has an identifier. For example, the
identifier of the foldl rewrite rule is "eval.foldl.rule" (Section 3.1). This identifier is mapped to
‘Apply the fold left rule to process a list using an operator that associates to the left’ in the script file.
This script file can be changed without recompiling the evaluator, which makes it possible to easily adapt
the information, for example to support feedback in another language.
Back-end. The back-end is developed in Haskell, and uses the Helium compiler [10] for parsing and
pretty-printing expressions. The back-end operates as a glue component that connects all other components. It receives a string from the front-end, uses the Helium compiler to parse the string, and converts
the abstract syntax tree produced by Helium to the expression datatype. The back-end receives expression
results from I DEAS, converts values from the expression datatype back to Helium’s expression datatype,
and uses the Helium compiler to convert this value to a string. This string is presented to the user using
the front-end.
To determine if a provided step follows the evaluation strategy, the I DEAS framework is instantiated
with functionality for determining if the provided expression is equal to the expected expression. The
evaluator therefore needs to implement two functions that are required by the framework: one function
to determine if two expressions are semantically equivalent, and one function to determine if two expressions are syntactically equivalent. Syntactic equivalence is obtained by deriving an instance of the Eq
type class for the Expr datatype. Semantic equivalence is more subtle because a student may submit an
expression that is syntactically different from the expression we expect to get from the step, but semantically the same. Semantic equivalence is defined by using a function that applies the rewrite strategy until
it cannot be applied any further and returns True if both results are the same. With these two functions
we can also spot mistakes in rewrite steps, even when an incorrect expression happens to produce the
same result.
Survey. We carried out a small survey to find out if the tool has potential in an educational setting. We
invited instructors of functional programming courses at several universities, primarily but not only in
the Netherlands, and students that follow or completed the functional programming course at the Open
Universiteit to experiment with the prototype, and to answer open-ended and closed-ended questions
about the tool. All participants were approached via email on April 30, 2014. By May 12, 7 instructors
(out of 8) and 9 students (out of 29) completed the survey. All questions and answers are included in the
first author’s master’s thesis [22]. We discuss the most important observations.
62
Evaluating Haskell expressions in a tutoring environment
All participants agree that it is useful for students following an introductory course in functional programming to inspect how an expression is evaluated. Eleven participants agree that it is also useful to
inspect multiple evaluation strategies. One student argued that the examples in Hutton’s textbook [12]
are sufficient to get a good understanding of the evaluation steps, and two participants argued that inspecting evaluation steps according to multiple evaluation strategies may be confusing for those who start
programming for the first time. There is less consensus about the question if more evaluation strategies
should be supported. Four instructors argue that lazy evaluation should be supported because it is very
subtle, but two other instructors argue that this is too much for the basic understanding of programming
concepts. Another instructor thinks it would be nice to add lazy evaluation, but is not sure if this is worth
the effort. Students find the addition of lazy evaluation useful.
Practicing with the evaluation steps was well received by all participants. Three participants argue
that it is only useful at the very start, and that students can quickly become bored by entering all evaluation steps. Another remark from an instructor is that studying evaluation steps is not the only way to
reason about a program. To ensure that students do not get bored quickly it might be useful to skip certain
evaluation steps. However, participants disagree about whether or not this feature should be supported.
Five instructors and one student argue that it should not be possible to skip steps to illustrate that each
step is necessary. The other participants would like to see this feature added to the prototype.
Nine participants find it possibly useful to add user-defined function definitions and to adjust function
definitions to be able to inspect the effect of these changes on the evaluation of expressions. Three
participants do not find this useful and think support for prelude functions is sufficient. Some participants
suggest to also show the function definitions. This could help students in understanding the evaluation
steps. Another suggestion is to show a derivation according to multiple evaluation strategies side by
side, so that the student can easily observe the differences. Some participants also suggest to detect if an
expression cannot be evaluated according to the innermost strategy and notify the user accordingly.
It is clear that instructors not always agree upon desirability of features. Therefore, it is important
that features are configurable for instructors, and that an instructor can adapt the tool to her own course.
5
Related work
There are roughly three approaches to inspect the evaluation steps of a Haskell expression: trace generation, observing intermediate data structures, and using rewrite rules. The central idea of the trace
generation approach, which is mainly used for debugging, is that every expression is transformed into
an expression that is supplemented with a description in the trace. The trace information is saved in a
datatype that can be viewed by a trace viewing component. There are two methods for trace generation.
The first method is to instrument Haskell source code. Pure Haskell functions are transformed to Haskell
functions that store the evaluation order in a certain datatype that can be printed to the user. This approach is used to show complete traces in so called redex trail format [27], and this is also the approach
used in the Hat debug library [5]. An advantage of this method is that it is completely separated from
the compiler, so it does not matter which compiler is used. A disadvantage of this method is that the
instrumentation of the original code can alter the execution of the program. The second method, which
is used for example by WinHIPE [24], is to instrument the interpreter. The advantage of this method
is that the execution of the program is exactly the same, but a disadvantage is that the interpreter (part
of the compiler) needs to be adjusted. The approach to observe intermediate data structures, which is
also mainly used for debugging, is used in the Hood debugger [27]. The approach to specify rewrite
rules to inspect the evaluation of expressions is used for example in the stepeval project where a subset
T. Olmer, B. Heeren & J. Jeuring
63
of Haskell expressions can be inspected [20]. With the above approaches it is possible to inspect the
evaluation steps of a Haskell expression, but it is not possible to practice with these evaluation steps.
Several intelligent tutoring systems have been developed that support students with learning a functional programming language. One of the main problems for novice programmers is to apply programming concepts in practice [16]. To keep students motivated to learn programming it is therefore important
to teach it incrementally, to practice with practical exercises, and to give them early and direct feedback
on their work [29]. The main advantage of an intelligent tutoring system is that a student can get feedback at any moment. An intelligent tutoring system consists of an inner loop and an outer loop. The
main responsibility of the outer loop is to select an appropriate task for the student; the main responsibility of the inner loop is to give hints and feedback on student steps. The Web-Based Haskell Adaptive
Tutor (WHAT) focuses more on the outer loop. It classifies each student into a group of students that
share some attributes and will behave differently based on the group of the student [19]. With WHAT, a
student can practice with three kinds of problems: evaluating expressions, typing functions, and solving
programming assignments. A disadvantage of this tutor is that it does not support the stepwise development of a program. A SK -E LLE [7] is a Haskell tutor system that focuses primarily on the inner loop. Its
goal is to help students learn functional programming by developing programs incrementally. Students
receive feedback about whether or not they are on the right track, can ask for a hint when they are stuck,
and see how a complete program is constructed stepwise [13].
6
Conclusions and future work
In this paper we have presented a prototype tool that can be used to show the evaluation steps of a
Haskell expression according to different evaluation strategies. We support the left-most innermost evaluation strategy and an outermost evaluation strategy based on pattern matching. The tool can also be
used to practice with evaluation steps. This prototype may help students to better understand important
programming concepts such as pattern matching, higher-order functions, and recursion. How effective
the tool is in practice needs to be further investigated by collecting more empirical data.
The evaluation process is driven by the definition of the rewrite rules and the evaluation strategy that
is used. The rewrite rules and evaluation strategies for the definitions can be generated by the prototype
from a set of function definitions. The extension of adding user-defined function definitions makes
it possible for users to easily get an understanding of how their function will be evaluated. Another
advantage is that alternative definitions for prelude functions can be tried, and that the results can be
inspected. For example, a student can use a strict version of foldl for sum, or define sum recursively. At
the moment, adding user-defined functions requires an instructor to change the content of the prelude file
on the server. Preferably, some support for this feature is added to the front-end.
The expression language in the current tool is limited. In the future, we hope to support other language constructs such as guards, conditional expressions, list comprehensions, algebraic datatypes, etc.
Other future work is to offer the possibility to change the step size of a function, and to add sharing to an
evaluation strategy:
• In the examples, we use a fixed step size of one. The step size is the number of steps that the
evaluator uses to rewrite a certain expression. For example, the expression 3 + (4 + 7) is evaluated
to 14 in two steps, although most students will typically combine these steps. More research must
be carried out to automatically derive or configure a certain step size that suits most students.
• The lazy evaluation strategy used by Haskell combines the outermost evaluation strategy with
sharing. Currently, sharing is not supported in the prototype. To help students learn about which
64
Evaluating Haskell expressions in a tutoring environment
computations are shared, we plan to extend the prototype along the lines of Launchbury’s natural
semantics for lazy evaluation [17], and by making the heap explicit.
The long-term goal of our work is to integrate the functionality of the prototype in the A SK -E LLE
programming tutor, which then results in a complete tutoring platform to help students learn programming. We are also considering to combine the evaluator with QuickCheck properties [6]: when QuickCheck finds a minimal counter-example that falsifies a function definition (e.g. for a simple programming
exercise), then we can use the evaluator to explain more precisely why the result was not as expected, or
use the evaluator as a debugging tool.
Acknowledgements The authors would like to thank the students and instructors that participated in
the survey for their feedback and suggestions. We also thank the anonymous reviewers for their detailed
comments.
References
[1] Martin Abadi, Luca Cardelli, Pierre-Louis Curien & Jean-Jacques Lvy (1990): Explicit Substitutions. In:
Proceedings of the 17th ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages,
POPL ’90, ACM, New York, NY, USA, pp. 31–46, doi:10.1145/96709.96712.
[2] Adam Bakewell & Colin Runciman (2000): The Space Usage Problem: An Evaluation Kit for Graph Reduction Semantics. In: Selected Papers from the 2nd Scottish Functional Programming Workshop (SFP00),
Intellect Books, Exeter, UK, pp. 115–128, doi:10.1.1.23.3087.
[3] Richard S. Bird & Philip. Wadler (1998): Introduction to functional programming using Haskell. PrenticeHall.
[4] Manuel M.T. Chakravarty & Gabriele Keller (2004): The risks and benefits of teaching purely
functional programming in first year.
Journal of Functional Programming 14(1), pp. 113–123,
doi:10.1017/S0956796803004805.
[5] Olaf Chitil, Colin Runciman & Malcolm Wallace (2003): Transforming Haskell for Tracing. In Ricardo
Peña & Thomas Arts, editors: Implementation of Functional Languages, Lecture Notes in Computer Science
2670, Springer Berlin Heidelberg, pp. 165–181, doi:10.1007/3-540-44854-3 11.
[6] Koen Claessen & John Hughes (2000): QuickCheck: A Lightweight Tool for Random Testing of Haskell Programs. In: Proceedings of the Fifth ACM SIGPLAN International Conference on Functional Programming,
ICFP ’00, ACM, New York, NY, USA, pp. 268–279, doi:10.1145/351240.351266.
[7] Alex Gerdes (2012): Ask-Elle: a Haskell Tutor. PhD thesis, Open Universiteit Nederland.
[8] Bastiaan Heeren & Johan Jeuring (2014): Feedback services for stepwise exercises. Science of Computer
Programming 88(0), pp. 110 – 129, doi:10.1016/j.scico.2014.02.021.
[9] Bastiaan Heeren, Johan Jeuring & Alex Gerdes (2010): Specifying Rewrite Strategies for Interactive Exercises. Mathematics in Computer Science 3(3), pp. 349–370, doi:10.1007/s11786-010-0027-4.
[10] Bastiaan Heeren, Daan Leijen & Arjan van IJzendoorn (2003): Helium, for learning Haskell. In: Proceedings
of the 2003 ACM SIGPLAN workshop on Haskell, Haskell ’03, ACM, New York, NY, USA, pp. 62–71,
doi:10.1145/871895.871902.
[11] Gérard Huet (1997): The Zipper.
doi:10.1017/S0956796897002864.
Journal of Functional Programming 7(5), pp. 549–554,
[12] Graham Hutton (2007):
Programming
doi:10.1017/CBO9780511813672.
in
Haskell.
Cambridge
University
Press,
T. Olmer, B. Heeren & J. Jeuring
65
[13] Johan Jeuring, Alex Gerdes & Bastiaan Heeren (2012): A Programming Tutor for Haskell. In Viktória Zsók,
Zoltán Horváth & Rinus Plasmeijer, editors: Central European Functional Programming School, Lecture
Notes in Computer Science 7241, Springer Berlin Heidelberg, pp. 1–45, doi:10.1007/978-3-642-31374-5.
[14] Aravind K. Krishna & Amruth N. Kumar (2001): A Problem Generator to Learn Expression: Evaluation in
CSI, and Its Effectiveness. J. Comput. Sci. Coll. 16(4), pp. 34–43.
[15] Amruth N. Kumar (2005): Results from the Evaluation of the Effectiveness of an Online Tutor on Expression
Evaluation. In: Proceedings of the 36th SIGCSE Technical Symposium on Computer Science Education,
SIGCSE ’05, ACM, New York, NY, USA, pp. 216–220, doi:10.1145/1047344.1047422.
[16] Essi Lahtinen, Kirsti Ala-Mutka & Hannu-Matti Järvinen (2005): A Study of the Difficulties of
Novice Programmers. In: Proceedings of the 10th Annual SIGCSE Conference on Innovation and
Technology in Computer Science Education, ITiCSE ’05, ACM, New York, NY, USA, pp. 14–18,
doi:10.1145/1151954.1067453.
[17] John Launchbury (1993): A Natural Semantics for Lazy Evaluation. In: Proceedings of the 20th ACM
SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL ’93, ACM, New York,
NY, USA, pp. 144–154, doi:10.1145/158511.158618.
[18] Miran Lipovaca (2011): Learn You a Haskell for Great Good!: A Beginner’s Guide, 1st edition. No Starch
Press, San Francisco, CA, USA.
[19] Natalia López, Manuel Núñez, Ismael Rodrı́guez & Fernando Rubio (2002): What: Web-Based Haskell
Adaptive Tutor. In Donia Scott, editor: Artificial Intelligence: Methodology, Systems, and Applications,
Lecture Notes in Computer Science 2443, Springer Berlin Heidelberg, pp. 71–80, doi:10.1007/3-540-461485 8.
[20] Ben Millwood (2011): stepeval library: Evaluating a Haskell expression step-by-step. Available at https:
//github.com/bmillwood/stepeval.
[21] Thomas van Noort, Alexey Rodriguez Yakushev, Stefan Holdermans, Johan Jeuring, Bastiaan Heeren &
José Pedro Magalhães (2010): A lightweight approach to datatype-generic rewriting. Journal of Functional
Programming 20, pp. 375–413, doi:10.1017/S0956796810000183.
[22] Tim Olmer (2014): Evaluation of Haskell expressions in a tutoring environment. Master’s thesis, Open
Universiteit Nederland. Available at http://hdl.handle.net/1820/5389.
[23] Bryan O’Sullivan, John Goerzen & Don Stewart (2008): Real World Haskell, 1st edition. O’Reilly Media,
Inc.
[24] Cristóbal Pareja-Flores, Jamie Urquiza-Fuentes & J. Ángel Velázquez-Iturbide (2007): WinHIPE: An IDE
for Functional Programming Based on Rewriting and Visualization. SIGPLAN Not. 42(3), pp. 14–23,
doi:10.1145/1273039.1273042.
[25] Jan Rochel (2010): The Very Lazy λ -calculus and the STEC Machine. In: Proceedings of the 21st International Conference on Implementation and Application of Functional Languages, IFL’09, Springer-Verlag,
Berlin, Heidelberg, pp. 198–217, doi:10.1007/978-3-642-16478-1 12.
[26] Peter Sestoft (2002): Demonstrating lambda calculus reduction. In: The Essence of Computation: Complexity, Analysis, Transformation. Essays Dedicated to Neil D. Jones, number 2566 in Lecture Notes in Computer
Science, Springer-Verlag, pp. 420–435, doi:10.1007/3-540-36377-7 19.
[27] Jan Sparud & Colin Runciman (1997): Tracing lazy functional computations using redex trails. In
Hugh Glaser, Pieter Hartel & Herbert Kuchen, editors: Programming Languages: Implementations, Logics, and Programs, Lecture Notes in Computer Science 1292, Springer Berlin Heidelberg, pp. 291–308,
doi:10.1007/BFb0033851.
[28] Simon Thompson (2011): Haskell: The Craft of Functional Programming, 3rd edition. Addison-Wesley
Longman Publishing Co., Inc.
[29] Arto Vihavainen, Matti Paksula & Matti Luukkainen (2011): Extreme Apprenticeship Method in Teaching
Programming for Beginners. In: Proceedings of the 42Nd ACM Technical Symposium on Computer Science
Education, SIGCSE ’11, ACM, New York, NY, USA, pp. 93–98, doi:10.1145/1953163.1953196.
66
Evaluating Haskell expressions in a tutoring environment
[30] Eelco Visser, Zine-el-Abidine Benaissa & Andrew Tolmach (1998): Building Program Optimizers with
Rewriting Strategies. In: ICFP 1998: International Conference on Functional Programming, pp. 13–26,
doi:10.1145/289423.289425.
| 6cs.PL
|
Self-paced Learning for Weakly Supervised Evidence
Discovery in Multimedia Event Search
Mengyi Liu1,2 , Lu Jiang2 , Shiguang Shan1 , Alexander G. Hauptmann2
1
Key Lab of Intelligent Information Processing of Chinese Academy of Sciences (CAS),
Institute of Computing Technology, CAS, Beijing, 100190, China
2
School of Computer Science, Carnegie Mellon University, 15213, USA
arXiv:1608.03748v3 [cs.CV] 23 Oct 2017
{mengyi.liu. shiguang.shan}@vipl.ict.ac.cn, {lujiang, alex}@cs.cmu.edu
ABSTRACT
Multimedia event detection has been receiving increasing
attention in recent years. Besides recognizing an event, the
discovery of evidences (which is refered to as “recounting”)
is also crucial for user to better understand the searching
result. Due to the difficulty of evidence annotation, only
limited supervision of event labels are available for training
a recounting model. To deal with the problem, we propose
a weakly supervised evidence discovery method based on
self-paced learning framework, which follows a learning process from easy “evidences” to gradually more complex ones,
and simultaneously exploit more and more positive evidence
samples from numerous weakly annotated video segments.
Moreover, to evaluate our method quantitatively, we also
propose two metrics, PctOverlap and F1-score, for measuring the performance of evidence localization specifically. The
experiments are conducted on a subset of TRECVID MED
dataset and demonstrate the promising results obtained by
our method.
CCS Concepts
•Computing methodologies → Computer vision; Visual content-based indexing and retrieval; Machine learning; Unsupervised learning;
Keywords
Multimedia event search, Event recounting, Evidence discovery, Weakly supervised learning, Self-paced learning
1.
INTRODUCTION
Nowadays multimedia contents have been produced and
shared ubiquitous in our daily life, which has encouraged
people to develop algorithms for multimedia search and analysis in various applications. As one of the most popular directions, multimedia event detection has been receiving increasing attention in recent years. Different from the atomic
object or action recognition, which focus retrieving simple
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected].
WOODSTOCK ’97 El Paso, Texas USA
c 2017 ACM. ISBN 123-4567-24-567/08/06. . . $15.00
DOI: 10.475/123 4
primitives [1], event detection aims to identify more complex scenario, for example, semantically meaningful human
activities, taking place within a specific environment, and
containing a number of necessary objects [2], which makes
it more suitable for the purpose of multimedia search.
Due to such complexity mentioned above, only concern
of recognizing an event is insufficient for user to understand
the searching result thoroughly. A question that “why is this
video classified as this event” is required to be answered, that
is, our system should provide the exact temporal locations of
several key-frames or key-shots from the whole video which
contain observed evidences that lead to our decisions. This
task is referred to as event recounting, where several efforts
have been contributed to this field. For example, [1] adopted
a semantic concept based event representation for learning
a discriminative event model, and generated recounting by
weighting the contribution of each individual concept to the
final event classification decision score. [3] proposed to identify event oriented discriminative video segments and their
descriptions with a linear SVM classifier and noise-filtered
concept detectors, then user friendly concepts including objects, scenes, and speech were extracted as recounting results
to generate descriptions. As event detection and recounting are two highly related task that could benefit with each
other, some recent work aimed to address these two problems simultaneously. [4] introduced an evidence localization model where evidence locations were modeled as latent
variables, and optimized the model via max-margin framework under the constraints on global video appearance, local evidence appearance, and the temporal structure of the
evidence. [5] proposed a joint framework to optimize both
event detection model and recounting model using improved
Alternating Direction Method of Multiplier (ADMM) algorithm. [6] proposed a flexible deep CNN architecture named
DevNet that detected pre-defined events and provided key
spatio-temporal evidences at the same time. Figure 1 shows
an illustration of event recounting results in an ideal multimedia search system.
Although these attempts have obtained promising results
by indicating plausible observations, the event recounting
task still remains a less addressed problem due to the challenge of evidence annotation, which leads to two limitations of the existing techniques. First, with only event labels of training videos, the evidential or non-evidential part
are confused with each other and distinguished all based
on category information, which omits some key evidences
shared among different events, or even background samples. Second, without the ground truth of evidence loca-
Video (length: 2’39”)
Evidence
...
Time duration
Score
Related concepts
00’14”~00’38”
0.87
Ring box; 2 or more people; going down on one knee
00’46”~01’01”
0.73
2 or more people; Holding ring; crying; hugging
01’22”~01’47”
0.59
2 and more people; hugging; kissing; cheering; clapping
...
...
...
Figure 1: An illustration of event recounting results in an ideal MED system. For each retrieved video
sample, the temporal locations of key segments which contain observed evidences are provided with their
predicted confidence score. The evidences are indicated by a list of semantic concepts detected in each key
segments.
tions, there could not be a substantial and quantitative comparison among different methods. The performance of a system can only be evaluated by making subjective and qualitative judgement that whether the recounted evidences or
semantic concepts are reasonable or not. In this paper, we
focus on the mentioned issues and make efforts in the following two aspects: (1) We propose a weakly supervised evidence discovery method based on self-paced learning framework [7], which follows a learning process from easy “evidences” to gradually more complex ones, and simultaneously exploit more and more positive evidence samples from
numerous weakly annotated video segments. (2) To evaluate our method quantitatively, we also propose two metrics, Percentage of Overlap (PctOverlap) and F1-score, for
measuring the performance of evidence localization according to a small group of ground truth annotated by humans
(The collection and generation of the ground truth are detailed in Section 4 below). The experiments are conducted
on a subset of TRECVID MED dataset and demonstrate
the promising results obtained by our method.
The rest of this paper is organized as follows. Section 2
presents the semantic concept based video representation,
which can provide high-level semantic information that should
benefit the evidence interpretation. Section 3 introduces the
essential technique in this paper, i.e. self-paced learning
framework, with its detailed formulation and optimization
process. In Section 4, we provide comprehensive evaluations
of the whole framework and comparisons with several highly
related methods. Finally, we conclude the work and discuss
possible future directions in Section 5.
2.
SEMANTIC CONCEPT FEATURE
Low-level feature based video representation, for example,
SIFT [8], STIP [9], and dense trajectory based features [10,
11], has been widely used in action recognition and event
detection. However, those low-level features hardly have semantic meanings, thus are not suitable for interpretation
purposes [5], such as the recounting task, which requires
some higher-level information for event or evidence description. Recently, semantic representations based on these kind
of attributes or concepts have been increasingly popular in
the field of event detection and recounting [1, 4, 5, 12, 13,
14]. With the same spirit, we also learn to generate a video
representation based on various semantic concepts, such as
objects, scenes, and activities.
Specifically, we pre-defined our concept collection C =
{C1 , C2 , ..., Cc } (The sources of these concepts is detailed in
Section 4. Table 1 provides some examples of the concepts
grouped by their types). For each concept, we collect training samples, i.e. video segments, from auxiliary datasets,
and employ improved dense trajectory features [11] for representation. Based on the low-level features, binary linear
SVM are used for training concept detector, and finally we
can generate c concept detectors totally. The next step is
to process testing event videos. For the purpose of identifying the key evidences temporally, we first segment each
video sample into a number of shots using well-established
shot boundary detection techniques [15]. For each shot, we
extract the same dense trajectory features and apply all the
concept detectors on this shot to obtain c confidence scores
as a representation (Note that the scores should be normalized from 0 to 1). Formally, we denote the concept repre(k)
sentation of the i-th shot from the k-th video as xi ∈ Rc .
Suppose there are nk shots in the k-th video, the collection
(k) nk
of all the shots can be represented as Xk = {xi }i=1
, where
k = 1, 2, ..., m, and m is the total number of videos.
3.
SELF-PACED LEARNING
Self-paced learning [7] is a lately proposed theory inspired
by the learning process of humans or animals. The idea is
to learn the model gradually from easy samples to complex
ones in a iteratively self-paced fashion. This theory has been
widely applied to various problems, including image classification [16], visual tracking [17], segmentation [18, 19], and
multimedia event detection [20, 21].
In the context of evidence recounting problem, the easy
samples are video shots with high confidence scores obtained
by a binary event-oriented detector. Based on these initialized training samples, our algorithm learns a gradually “mature” model by mining and appending more and more complex evidence samples iteratively according to their losses,
and also adaptively determines their weights in the next iteration. Now we start to introduce the detailed problem
formulation and optimization in this section.
3.1
Problem Formulation
Given m video candidates with only annotation of event
labels, continue to use notations in Section 2, the k-th sam(k)
(k) nk
ples can be represented as {Xk , Yk } = {{xi , yi }i=1
},
(k)
where xi denotes the representation of the i-th shot from
(k)
the k-th video, and yi ∈ {−1, +1} denotes its label whether
Table 1: An illustration of examples of semantic concepts grouped by their types, i.e. scene / site, object /
tool, person / animal, human activity, and others.
Scene / Site
Object / Tool
Person / Animal
Human Activity
Others
Dinning room
Refrigerator
Female news reporter
Speaking or singing
Moonlight
Birthday party
Helicopter
Man wearing a suit
Shooting
Religious
Daytime outdoor
Appliance
Laughing baby
Cheering
Bomber bombing
Sunny beach
Ring box
Politician
Bicycling
Airplane landing
Convention center
Musical instrument
Athlete
Handshaking
Background static
Election campaign
uniform
Seal balancing a ball
Rock climbing
Network logo
Hospital
vehicle
Domestic animal
Meeting
Earthquake
Laboratory
Machine gun
Dolphin
Sailing ship
Overlaid text
it can be regarded as an “evidence” or not. This formulation partially agree with the definition of Multiple Instance
Learning [22, 23], that we only know the label for each “bag”
but not the instances assigned to a certain “bag”. The same
point with MIL is, if Yk = +1, which indicates that this
video is categorized as a certain event, then at least one
(k)
(k)
instance xi ∈ Xk is a positive sample (i.e. yi = +1),
which means that there exists at least one evidence leading
to the decision. The different point with MIL is, if Yk = −1,
in most cases there are no evidence in this video, but this
cannot be guaranteed since there exists some complex and
confused evidences shared among different events or even
background videos. while in traditional MIL framework,
(k)
(k)
Yk = −1 leads to yi = −1 for all yi ∈ Yi .
Although we cannot employ the solution for MIL problem
directly, we can exploit the same idea of heuristic optimization proposed in [23], i.e. supposing all the instances have
their initialized pseudo labels and seeking for the optimal
hyperplane and labels alternatively. Here in our task, we
introduce all shots extracted from the background videos
as negative samples, and all shots from the videos labeled
as a certain event as positive samples. A linear SVM is
employed to train the initialized classifier, then the current
samples and model parameters are served as an initialization
for Self-paced Learning in the next step.
(k) nk m
(k)
For all the video shots {Xk , Yk }m
k=1 = {{xi , yi }i=1 }k=1 ,
(k)
where yi is kind of pseudo label which need to be optimized
(k)
(k)
during self-paced learning process. Let `(yi , f (xi ; w, b))
denote the loss function which calculates the cost between
(k)
(k)
the (pseudo) label yi and the predicted label f (xi ; w, b),
where w and b represents the model parameters in decision
function f . In SPL, the goal is to jointly learn the model
parameters {w, b}, the pseudo label y and the latent weight
variable v according to the objective function E:
min 1 ||w||22
w,b,y,v 2
+
Pm Pnk
k=1
i=1
(k)
(k)
(k)
vi `(yi , f (xi ; w, b)) + g(v, λ)
s.t. y ∈ {−1, +1}N , v ∈ [0, 1]N ,
(1)
P
where N = m
k=1 nk denotes the total number of instances
(1)
(1)
(m)
(m)
from m videos, y = [y1 , ..., yn1 , ..., y1 , ..., ynk ] denotes
(1)
(1)
the pseudo labels for all instances, and v = [v1 , ..., vn1 , ...,
(m)
(m)
v1 , ..., vnk ] denotes their weighting parameters which re-
flects the sample importance in training the model, ` is the
(k)
standard hinge loss of xi under classifier f (In this work,
we simply employ the linear SVM version), calculated from:
(k)
(k)
(k)
(k)
`(yi , f (xi ; w, b)) = max{0, 1 − yi (wT xi
+ b)}. (2)
More importantly, g(v, λ) is the regularization term called
self-paced function which specifies how the sample weights
are generated. Here λ (λ > 0) is a parameter for determining
the learning rate. g(v, λ) can be defined in various forms in
terms of the learning rate [21]. A conventional one proposed
in [7] is based on the l1 -norm of v as:
g(v, λ) = −λ||v||1 = −λ
nk
m X
X
(k)
vi .
(3)
k=1 i=1
This regularizer is very general and has been applied to various learning tasks with different loss functions [18, 13, 21].
Up to now, we can observe that the objective function is
subjected to two parts of constraints: one is the max-margin
constraints inherited from traditional SVM; another one is
self-paced term taking control of the pseudo labels and sample weights respectively. This objective is difficult to optimize directly due to its non-convexity. In the next subsection, we introduce the effective Cyclic Coordinate Method
(CCM) [24] to solve this problem as in [7, 25, 21, 19].
3.2
Optimization
Cyclic Coordinate Method (CCM) is a kind of iterative
method for non-convex optimization, in which the model
variables are divided into independent blocks (two blocks in
our case): (1) classifier parameters {w, b}; (2) pseudo labels
y and sample weights v. We switch between the two blocks
iteratively, that one block of variables can be optimized while
fixing the other block. Taking the input MIL-inspired initialization, in each iteration, the alternative optimization
process can be presented as follows:
Optimizing {w, b} while fixing y and v. In this step,
we fix the pseudo labels y and weight variables v as constant,
then the objective (1) is updated to only represent the sum
of weighted loss across all instances as E(w, b):
m
nk
X X (k) (k)
1
(k)
min ||w||22 +
vi `(yi , f (xi ; w, b)).
w,b 2
i=1
(4)
k=1
(k) (k)
Generally, vi `i is the discounted hinge loss of the shot
(k)
instance xi . To simplify the solution, in conventional SPL,
Table 2: Event names with their evidential description in TRECVID MED dataset.
Event ID & Name
Evidential Description
Event 221: Attempting a bike trick
Scene or site: outdoors (skate park, parking lot or street);
Objects or people: person, bike, ramps, helmet, concrete floor, audience
Event 222: Cleaning an appliance
Scene or site: indoors (kitchen, shop);
Objects or people: appliance (fridge, stove, dishwasher), sponge, cloths
Event 223: Dog show
Scene or site: indoors (exhibition hall) or outdoors (fairground);
Objects or people: judge(s), handler(s), dogs, name tags, leash, frisbee
Event 225: Marriage proposal
Scene or site: outdoors or indoors;
Objects or people: small box, ring, at least two people (kissing/hugging)
Event 226: Renovating a home
Scene or site: indoors (basement, garage) or outdoors (roof);
Objects or people: tools (hammer, wrench, drill, ladder, saw), wood, cement
Event 227: Rock climbing
Scene or site: outdoors (in nature) or indoors (gym);
Objects or people: carabiners, rope, helmet, rook formation, climbers
Event 228: Town hall meeting
Scene or site: indoors (stage, arena, auditorium, church fellowship hall);
Objects or people: microphone, security guards or bouncers, politicians
Event 229: Winning a race without a vehicle
Scene or site: outdoors (field, track, road, or stadium), indoors (gym, pool);
Objects or people: athletes, finish line, boundary markers, spectators
Event 230: Working on a metal crafts project
Scene or site: indoors or outdoors (workshop)
Objects or people: metal pieces, rivets, hammer, caliper, solder, drill
(k)
all the vi are forced to be binary value, i.e. 0 or 1. Thus
the objective (4) degenerates to a conventional linear SVM
which only considers the selected samples whose weight equals
1. However, on the other hand, this binary setting of v
has limited ability for balancing the positive and negative
costs, since in our task there exists only few positive evidence (event) examples while a large number of negative
(background) samples. To address this problem, we employ
the similar idea in Exemplar-SVM [26] which introduces two
regularization parameters (i.e. C+ and C− ) to balance the
effects of these two types of costs. Differently, in our formulation, there is a small set of positive samples rather than
a single “exemplar”. Accordingly, we can rewrite (4) as an
ESVM-like form as follows:
P (k)
(k)
min 21 ||w||22 + C+ vi `(+1, f (xi ; w, b))
w,b
(5)
P (k)
(k)
+C− vi `(−1, f (xi ; w, b))
By solving (5), we can obtain w as the classification hyperplane, which is going to be fixed for the next step optimization.
Optimizing y and v while fixing {w, b}. With the
fixed classifier parameters, we can omit the ||w||22 term and
the objective (1) becomes E(y, v, λ):
P
Pnk (k) (k)
(k)
min m
k=1
i=1 vi `(yi , f (xi ; w, b)) + g(v, λ)
y,v
(6)
s.t. y ∈ {−1, +1}N , v ∈ [0, 1]N .
Based on (6), learning y is independent of v. and also, all
the pseudo labels are independent with each other in the
loss function. As each label can only take two integer values
+1 and −1, the global optimal solution can be achieved by
enumerating 2N times.
After obtaining the optimal y, the final task for us is to
(k)
optimize v. Following the solution in [7], the weight vi for
(k)
sample xi can be calculated by:
(
(k)
(k)
1, `(yi , f (xi ; w, b)) < λ
(k)
(7)
vi =
(k)
(k)
0, `(yi , f (xi ; w, b)) ≥ λ.
The criterion in (7) indicates that if the loss of an instance
is less than the current threshold λ, which means “easy”, it
will be selected for training in the next iteration, or otherwise unselected. Here λ controls the learning pace that how
many training samples should be selected at this time. As λ
increases, the tolerance of sample loss becomes larger, and
more “hard” samples will be appended to the training set to
learn a stronger model. Formally, we summarize the whole
optimization procedure in Algorithm 1.
Algorithm 1 : Iterative optimization of Self-paced
Learning
1: t = 0
2: Initialize pseudo labels y and weights v;
3: while t < max iter do
4:
{w(t+1) , b(t+1) } = arg max E(w(t) , b(t) )
5:
y (t+1) , v (t+1) = arg max E(y (t) , v (t) ; λ)
6:
λ(t+1) = λ(t) + ∆λ
y,v
w,b
7: end while
8: return y (max iter) , v (max iter)
4.
EXPERIMENTS
4.1
Dataset and Protocol
We conduct our experiments on TRECVID Multimedia
Event Detection dataset [27, 28]. Specifically, there are two
sets MED13 and MED14 collected by National Institute of
Standards and Technology (NIST) for the TRECVID competition. Each dataset includes 20 complex events with 10
events in common from E221 to E230. In this paper, we
only take the common part for evaluation. A detailed list of
event names with their evidential description is provided in
Table 2.
(a) SIN (PctOverlap)
(b) SIN (F1-score)
indications of evidences.
In order to conduct quantitative evaluation, we also propose two metrics, Percentage of Overlap (PctOverlap) and
F1-score, for measuring the performance of evidence localization. For better explanation, we first define the following
notations: (1) p (prediction): all temporal regions with predicted scores higher than a certain threshold (0.5 in this
paper); (2) g (groundtruth): all temporal regions with annotated scores 1 (or higher than a certain threshold); (3)
o (overlap): intersection regions of “prediction” and “ground
truth”. Based on these notations, P ctOverlap = o/(p +
g − o). As precision = o/p, recall = o/g, we can have
F1-score= 2 ∗ (precision ∗ recall)/(precision + recall) =
(2 ∗ o)/(p + g).
4.2
(c) Sports (PctOverlap)
(e) YFCC (PctOverlap)
(d) Sports (F1-score)
(f) YFCC (F1-score)
Parameter Settings
For semantic concept representation, we pre-train the concept detector on three auxiliary datasets: TRECVID SIN
dataset (346 concepts), Google Sports dataset [29] (478 concepts), and Yahoo Flickr Creative Commons (YFCC) dataset
[30] (609 concepts), and the prediction scores of these detectors are served as a feature vector of each video shot. The
SPL framework is based on these three kinds of features corresponding to different concept sets. In the learning process
according to Algorithm 1, we set max iter = 50, and in
each iteration, the learning pace controller ∆λ = 0.02. The
regularization parameters in (5) are set as C+ = 0.5 and
C− = 0.01, which follow the default settings in [26], and
proved to be insensitive in our experiments. Table 3 demonstrates the performance in different SPL iterations based on
three concept sets.
Table 4: Performance comparison based on different
learning methods.
(a) Metric: PctOverlap
Method
RF
AB
MIL
SPL
TRECVID SIN
10Ex
100Ex
0.3723
0.4113
0.3531
0.4617
0.4446
0.4571
0.4306
0.5028
Google Sports
10Ex
100Ex
0.3513
0.4186
0.3396
0.4724
0.4373
0.4137
0.4544
0.4807
YFCC
10Ex
100Ex
0.3513
0.3837
0.3102
0.4594
0.4603
0.4585
0.4258
0.4868
(b) Metric: F1-score
Table 3: Performance in different SPL iterations
based on three concept sets. (a)(b) TRECVID SIN.
(c)(d) Google Sports. (e)(f ) YFCC.
According to the evaluation procedure outlined by TRECVID MED task, the dataset can be divided into 3 partitions: (1) Background, which contains 4, 992 background
videos not belonging to any of the target events; (2) 10Ex
(or 100Ex ), which contains 10 (or 100) positive video examples for each event as the training samples; (3) MEDTest,
which contains about 25, 000 videos for testing. Here in our
recounting task, we select a small number samples from each
event (about 20 for each in average), and annotate the temporal evidence locations manually. To alleviate the bias of
different annotators, we average the results from 8 persons
to obtain the final ground truth. Note that, this annotation is only performed on test data for evaluation purpose.
For training data, only event labels are available without
Method
RF
AB
MIL
SPL
TRECVID SIN
10Ex
100Ex
0.4984
0.5306
0.4829
0.5826
0.5651
0.5738
0.5538
0.6136
Google Sports
10Ex
100Ex
0.4537
0.5413
0.4630
0.5932
0.5592
0.5367
0.5519
0.5894
YFCC
10Ex
100Ex
0.4712
0.5119
0.4522
0.5808
0.5817
0.5751
0.5502
0.6050
According to Table 3, we can observe an approximately
rising trend of the performance for both PctOverlap and F1score, as the number of iteration increases. Specifically, for
TRECVID SIN, the SPL converges really fast and achieves
the peak at #iteration = 14. This phenomenon of fast
convergence also appears in the 10Ex setting for YFCC (at
#iteration = 10). Another observation is about the improvements compared to BasicMIL (the details for are presented in the next subsection). For all the three concept
sets, the relative improvements for 10Ex setting are much
more significant than that for 100Ex, which indicates that
our method possesses strong superiority in weakly super-
results due to its straight-forward manner of data usage, i.e.
no sampling and equal weights. RandomForest performs
better since it considers sample weights implicitly by random sampling. AdaBoost is shown to be the best baseline
method, because it performs much more similar mechanism
that gradually selects “hard” samples out according to the
“error”, where in SPL the criterion is “loss”.
HVC147344
HVC852674
(a) Metric: PctOverlap
(b) Metric: F1-score
Figure 2: Performance comparison based on different learning methods.
0.79
vised learning especially for extremely few samples.
4.3
Performance Comparison
We compare the propose framework to three classical methods: RandomForest (RF) [31], AdaBoost (AB) [32],
and BasicMIL (MIL) as in [20]. Random forest and AdaBoost are both classical approaches which introducing sample weights implicitly in training process by random or rulebased sampling, which share the similar spirit as SPL. While
in BasicMIL manner, the model is trained using all samples
simultaneously with equal weights. Here we perform BasicMIL using SVM for fair comparison with SPL. All of the
results are shown in Table 4 and Figure 2.
0.74
Running , Stadium, Athlete
HVC774005
HVC764035
0.92
HVC407348
HVC525265
0.83
0.88 0.77
0.63
Sport climbing, Climbing, Indoor
Rock climbing, Mountaineer, Person
HVC934853
HVC732600
0.96
0.68 0.81
0.82
0.94
Climbing, Person, Rope Climbing
Nature, Vegetation, Primate
HVC620324
HVC316737
0.69 0.74 0.82
Climbing, Bouldering, Vegetation
0.54
Swimming, Waterfront, Hurdles
Running, Sport, Person
0.71
0.94
Walking, 3 or more people, Grass
HVC419799
HVC864271
0.97
0.89
Swimming, 3 or more people, Face
0.77 0.80
0.59
0.78
0.89
Field, Jumping, Children
Figure 4: Examples of evidence localization results
obtained by our method (E229 Winning a race without a vehicle). The colors indicate the strength of
decision confidence (red: 1-0.85; yellow: 0.84-0.7;
green: 0.69-0.5).
We also conduct a late fusion among different concept
features, and obtain the comparison results in Table 5, in
which we also demonstrate the performance for each individual event respectively. Figure 5 provides the corresponding results for a more intuitive visualization. Moreover, to
justify our results qualitatively, we also illustrate some video
examples with predicted evidence locations as well as their
recounting concepts. Figure 3 shows the event “E227 Rock
Climbing”, in which the concepts such as Climbing, Mountaineer, Person, are seemed to be appear in high frequency.
Figure 4 shows the event “E229 Winning a race without a
vehicle”, and we can observe that concepts Sport, Running,
Athlete are most likely to appear with high confidences.
Person, Daytime outdoor, Canyoning
5.
Figure 3: Examples of evidence localization results
obtained by our method (E227 Rock climbing). The
colors indicate the strength of decision confidence
(red: 1-0.85; yellow: 0.84-0.7; green: 0.69-0.5).
According to Table 4, BasicMIL always shows the worst
CONCLUSIONS
In this paper, we propose a weakly supervised evidence
discovery method based on self-paced learning framework,
which follows a learning process from easy “evidences” to
gradually more complex ones, and simultaneously exploit
more and more positive evidence samples from numerous
weakly annotated video segments. Our method is evaluated
Table 5: Performance comparison of different methods after late fusion.
(a) MED10Ex
Event ID
E221
E222
E223
E225
E226
E227
E228
E229
E230
Average
RF
0.1891
0.3284
0.3931
0.1793
0.4486
0.3741
0.5095
0.2477
0.4046
0.3414
PctOverlap
AB
MIL
0.3163
0.2368
0.4185
0.2947
0.7537
0.5150
0.1622
0.1422
0.4147
0.3322
0.5121
0.4939
0.5243
0.2615
0.4816
0.3696
0.4207
0.3496
0.4449
0.3328
RF
0.2947
0.3990
0.5690
0.2428
0.6039
0.4618
0.6875
0.4213
0.5115
0.4657
PctOverlap
AB
MIL
0.2925
0.3059
0.3919
0.3444
0.6732
0.7319
0.1980
0.1844
0.5359
0.3957
0.4909
0.5670
0.6305
0.5801
0.4396
0.4916
0.5770
0.4818
0.4700
0.4536
SPL
0.2760
0.3470
0.7750
0.2080
0.4676
0.4970
0.7622
0.5623
0.5461
0.4935
RF
0.2768
0.4658
0.5324
0.2771
0.5860
0.5159
0.6258
0.3600
0.5373
0.4641
F1-score
AB
MIL
0.4300
0.3602
0.5542
0.4208
0.8516
0.6605
0.2560
0.2229
0.5571
0.4666
0.6505
0.6351
0.6402
0.3543
0.6213
0.5082
0.5490
0.4805
0.5678
0.4566
SPL
0.3907
0.4888
0.8426
0.3085
0.6103
0.6421
0.8390
0.6841
0.6621
0.6076
RF
0.3968
0.5322
0.6829
0.3520
0.7228
0.6074
0.7876
0.5466
0.6311
0.5844
F1-score
AB
MIL
0.3948
0.4308
0.5261
0.4615
0.7749
0.8377
0.3093
0.2804
0.6636
0.5259
0.6279
0.6882
0.7278
0.7125
0.5553
0.6236
0.6809
0.6032
0.5845
0.5737
SPL
0.4030
0.4977
0.8349
0.3127
0.6121
0.6790
0.9033
0.6777
0.6478
0.6187
(b) MED100Ex
Event ID
E221
E222
E223
E225
E226
E227
E228
E229
E230
Average
on TRECVID MED dataset and shows promising results
both quantitatively and qualitatively. For future work, we
will attempt to investigate various forms of self-paced learning function which can be effectively adapted to our specific
task for further improvement.
6.
REFERENCES
[1] Yu, Q., Liu, J., Cheng, H., Divakaran, A., Sawhney,
H.: Multimedia event recounting with concept based
representation. In: ACM Multimedia. (2012)
[2] Li, L.J., Fei-Fei, L.: What, where and who? classifying
events by scene and object recognition. In:
International Conference on Computer Vision. (2007)
[3] Sun, C., Burns, B., Nevatia, R., Snoek, C., Bolles, B.,
Myers, G., Wang, W., Yeh, E.: Isomer: Informative
segment observations for multimedia event recounting.
In: ACM International Conference on Multimedia
Retrieval. (2014)
[4] Sun, C., Nevatia, R.: Discover: Discovering important
segments for classification of video events and
recounting. In: IEEE Conference on Computer Vision
and Pattern Recognition. (2014)
[5] Chang, X., Yu, Y.L., Yang, Y., Hauptmann, A.G.:
Searching persuasively: Joint event detection and
evidence recounting with limited supervision. In:
ACM Multimedia. (2015)
[6] Gan, C., Wang, N., Yang, Y., Yeung, D.Y.,
Hauptmann, A.G.: Devnet: A deep event network for
multimedia event detection and evidence recounting.
In: IEEE Conference on Computer Vision and Pattern
SPL
0.3007
0.3542
0.7574
0.2118
0.4710
0.5579
0.8593
0.5547
0.5252
0.5102
Recognition. (2015)
[7] Kumar, M.P., Packer, B., Koller, D.: Self-paced
learning for latent variable models. In: The Conference
on Neural Information Processing Systems. (2010)
[8] Lowe, D.G.: Distinctive image features from
scale-invariant keypoints. International Journal of
Computer Vision 60(2) (2004) 91–110
[9] Laptev, I., Marszalek, M., Schmid, C., Rozenfeld, B.:
Learning realistic human actions from movies. In:
IEEE Conference on Computer Vision and Pattern
Recognition. (2008)
[10] Wang, H., Kläser, A., Schmid, C., Liu, C.L.: Action
recognition by dense trajectories. In: IEEE Conference
on Computer Vision and Pattern Recognition. (2011)
[11] Wang, H., Schmid, C.: Action recognition with
improved trajectories. In: International Conference on
Computer Vision. (2013)
[12] Merler, M., Huang, B., Xie, L., Hua, G., Natsev, A.:
Semantic model vectors for complex video event
recognition. IEEE Transaction on Multimedia 14(1)
(2012) 88–101
[13] Liu, J., Yu, Q., Javed, O., Ali, S., Tamrakar, A.,
Divakaran, A., Cheng, H., Sawhney, H.: Video event
recognition using concept attributes. In: IEEE Winter
Conference on Applications of Computer Vision.
(2013)
[14] Mazloom, M., Habibian, A., Snoek, C.G.: Querying
for video events by semantic signatures from few
examples. In: ACM Multimedia. (2013)
[15] Boreczky, J.S., Rowe, L.A.: Comparison of video shot
(a) Metric: PctOverlap
(b) Metric: F1-score
Figure 5: Performance comparison of different methods after late fusion.
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
boundary detection techniques. Journal of Electronic
Imaging 5(2) (1996) 122–128
Tang, Y., Yang, Y.B., Gao, Y.: Self-paced dictionary
learning for image classification. In: ACM Multimedia.
(2012)
Supancic, J., Ramanan, D.: Self-paced learning for
long-term tracking. In: IEEE Conference on Computer
Vision and Pattern Recognition. (2013)
Kumar, M.P., Turki, H., Preston, D., Koller, D.:
Learning specific-class segmentation from diverse data.
In: International Conference on Computer Vision.
(2011)
Zhang, D., Meng, D., Li, C., Jiang, L., Zhao, Q., Han,
J.: A self-paced multiple-instance learning framework
for co-saliency detection. In: International Conference
on Computer Vision. (2015)
Jiang, L., Meng, D., Yu, S.I., Lan, Z., Shan, S.,
Hauptmann, A.: Self-paced learning with diversity. In:
The Conference on Neural Information Processing
Systems. (2014)
Jiang, L., Meng, D., Mitamura, T., Hauptmann, A.G.:
Easy samples first: Self-paced reranking for
zero-example multimedia search. In: ACM
Multimedia. (2014)
Dietterich, T.G., Lathrop, R.H., Lozano-Pérez, T.:
Solving the multiple instance problem with
axis-parallel rectangles. Artificial Intelligence 89(1)
(1997) 31–71
Andrews, S., Tsochantaridis, I., Hofmann, T.: Support
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
vector machines for multiple-instance learning. In:
The Conference on Neural Information Processing
Systems. (2002)
Gorski, J., Pfeuffer, F., Klamroth, K.: Biconvex sets
and optimization with biconvex functions: a survey
and extensions. Mathematical Methods of Operations
Research 66(3) (2007) 373–407
Tang, K., Ramanathan, V., Fei-Fei, L., Koller, D.:
Shifting weights: Adapting object detectors from
image to video. In: The Conference on Neural
Information Processing Systems. (2012)
Malisiewicz, T., Gupta, A., Efros, A.A.: Ensemble of
exemplar-svms for object detection and beyond. In:
International Conference on Computer Vision. (2011)
NIST. http://nist.gov/itl/iad/mig/med13.cfm (2013)
NIST. http://nist.gov/itl/iad/mig/med14.cfm (2014)
Karpathy, A., Toderici, G., Shetty, S., Leung, T.,
Sukthankar, R., Fei-Fei, L.: Large-scale video
classification with convolutional neural networks. In:
IEEE Conference on Computer Vision and Pattern
Recognition. (2014)
Thomee, B., Shamma, D.A., Friedland, G., Elizalde,
B., Ni, K., Poland, D., Borth, D., Li, L.J.: The new
data and new challenges in multimedia research. arXiv
(2015)
Breiman, L.: Random forests. Machine Learning 45(1)
(2001) 5–32
Friedman, J.H.: Stochastic gradient boosting.
Computational Statistics & Data Analysis 38(4)
(2002) 367–378
| 1cs.CV
|
Stabilizing Transmission Intervals for Nonlinear Delayed
Networked Control Systems
[E XTENDED V ERSION ]
arXiv:1604.04421v1 [cs.SY] 15 Apr 2016
Domagoj Tolić and Sandra Hirche ∗†‡
Abstract
less) communication network [12]. NCSs offer several advantages, such as reduced installation and maintenance costs
as well as greater flexibility, over conventional control systems in which parts of control loops exchange information
via dedicated point-to-point connections. At the same time,
NCSs generate imperfections (such as sampled, corrupted, delayed and lossy data) that impair the control system performance and can even lead to instability. In order to reduce data
loss (i.e., packet collisions) among uncoordinated NCS links,
scheduling protocols are employed to govern the communication medium access. Since the aforementioned networkinduced phenomena occur simultaneously, the investigation
of their cumulative adverse effects on the NCS performance is
of particular interest. This investigation opens the door to various trade-offs while designing NCSs. For instance, dynamic
scheduling protocols (refer to [11] and [17]), model-based estimators [8] or smaller transmission intervals can compensate
for greater delays at the expense of increased implementation
complexity/costs [5].
In this article, we consider a nonlinear delayed system to be controlled by a nonlinear delayed dynamic controller over a communication network in the presence of exogenous/modeling disturbances, scheduling protocols among
lossy NCS links, time-varying signal delays, time-varying
transmission intervals and distorted data. Notice that networked control is not the only source of delays and that delays might be present in the plant and controller dynamics as
well. Therefore, we use the term delayed NCSs. The present
article takes up the emulation-based approach from [27] for
investigating the cumulative adverse effects in NCSs and extends it towards plants and controllers with delayed dynamics as well as towards nonuniform time-varying NCS link delays. In other words, different NCS links induce different and
nonconstant delays. It is worth mentioning that [27] generalizes [11] towards corrupted data and the so-called large delays. Basically, we allow communication delays to be larger
than the transmission intervals. To the best of our knowledge,
the work presented herein is the most comprehensive study
of the aforementioned cumulative effects as far as the actual
plant-controller dynamics (i.e., time-varying, nonlinear, delayed and with disturbances) and interconnection (i.e., output
feedback) as well as the variety of scheduling protocols (i.e.,
UGES protocols) and other network-induced phenomena are
concerned (i.e., variable delays, lossy communication chan-
In this article, we consider a nonlinear process with delayed dynamics to be controlled over a communication network in the
presence of disturbances and study robustness of the resulting
closed-loop system with respect to network-induced phenomena such as sampled, distorted, delayed and lossy data as well
as scheduling protocols. For given plant-controller dynamics
and communication network properties (e.g., propagation delays
and scheduling protocols), we quantify the control performance
level (in terms of L p -gains) as the transmission interval varies.
Maximally Allowable Transfer Interval (MATI) labels the greatest transmission interval for which a prescribed L p -gain is attained. The proposed methodology combines impulsive delayed
system modeling with Lyapunov-Razumikhin techniques to allow for MATIs that are smaller than the communication delays.
Other salient features of our methodology are the consideration
of variable delays, corrupted data and employment of modelbased estimators to prolong MATIs. The present stability results
are provided for the class of Uniformly Globally Exponentially
Stable (UGES) scheduling protocols. The well-known Round
Robin (RR) and Try-Once-Discard (TOD) protocols are examples of UGES protocols. Finally, two numerical examples are
provided to demonstrate the benefits of the proposed approach.
1
Introduction
Networked Control Systems (NCSs) are spatially distributed
systems for which the communication between sensors, actuators and controllers is realized by a shared (wired or wire∗ D. Tolić is with Faculty of Electrical Engineering and Computing, University of Zagreb, Unska 3, 10000 Zagreb, Croatia. S. Hirche is with the
Chair of Information-oriented Control, Technical University of Munich, Arcisstraße 21, D-80290 Munich, Germany. (e-mail: [email protected],
[email protected]).
† This work has been supported by the European Community Seventh
Framework Programme under grant No. 285939 (ACROSS) and by the Institute for Advanced Study of Technical University of Munich, Germany.
D. Tolić is partially supported by the Air Force Research Laboratory under
agreement number FA8655-13-1-3055.
‡ The U.S. Government is authorized to reproduce and distribute reprints
for Governmental purposes notwithstanding any copyright notation thereon.
The views and conclusions contained herein are those of the authors and
should not be interpreted as necessarily representing the official policies or
endorsements, either expressed or implied, of the Air Force Research Laboratory or the U.S. Government.
1
nels with distortions). For instance, [19] focuses on timevarying nonlinear control affine plants (i.e., no delayed dynamics in the plant nor controller) and state feedback with
a constant delay whilst neither exogenous/modeling disturbances, distorted data nor scheduling protocols are taken into
account. The authors in [15] and [33] consider linear control systems, impose Zero-Order-Hold (ZOH) sampling and
do not consider noisy data nor scheduling protocols. In addition, [15] does not take into account disturbances. Similar
comparisons can be drawn with respect to other related works
(see [11, 12, 19, 27, 32] and the references therein).
future challenges are in Section 6. The proofs are provided in
the Appendix.
2
2.1
Preliminaries
Notation
To simplify notation, we use (x, y) := [x> y> ]> . The dimension of a vector x is denoted nx . Next, let f : R → Rn be a
Lebesgue measurable function on [a, b] ⊂ R. We use
In order to account for large delays, our methodology employs impulsive delayed system modeling and LyapunovRazumikhin techniques when computing Maximally Allowable Transmission Intervals (MATIs) that provably stabilize
NCSs for the class of Uniformly Globally Exponentially Stable (UGES) scheduling protocols (to be defined later on). Besides MATIs that merely stabilize NCSs, our methodology is
also capable to design MATIs that yield a prespecified level
of control system performance. As in [11], the performance
level is quantified by means of L p -gains. According to the
batch reactor case study provided in [27], MATI conservativeness repercussions of our approach for the small delay case
appear to be modest in comparison with [11]. This conservativeness emanates from the complexity of the tools for computing L p -gains of delayed (impulsive) systems as pointed out
in Section 5 and, among others, [3]. On the other hand, delayed system modeling (rather than ODE modeling as in [11])
allows for the employment of model-based estimators, which
in turn increases MATIs (see Section 5 for more). In addition, real-life applications are characterized by corrupted data
due to, among others, measurement noise and communication channel distortions. In order to include distorted information (in addition to exogenous/modeling disturbances) into
the stability analyses, we propose the notion of L p -stability
with bias.
k f [a, b]k p :=
Z
1
p
k f (s)k ds
p
[a,b]
to denote the L p -norm of f when restricted to the interval
[a, b]. If the corresponding norm is finite, we write f ∈
L p [a, b]. In the above expression, k · k refers to the Euclidean
norm of a vector. If the argument of k · k is a matrix A, then
it denotes the induced 2-norm of A. Furthermore, | · | denotes
the (scalar) absolute value function. The n-dimensional vector with all zero entries is denoted 0n . Likewise, the n by m
matrix with all zero entries is denoted 0n×m . The identity matrix of dimension n is denoted In . In addition, Rn+ denotes the
nonnegative orthant. The natural numbers are denoted N or
N0 when zero is included.
Left-hand and right-hand limits are denoted x(t − ) =
limt 0 %t x(t 0 ) and x(t + ) = limt 0 &tx(t 0 ), respectively. Next, for
a set S ⊆ Rn , let PC([a, b], S ) = φ : [a, b] → S φ(t) = φ(t + )
for every t ∈ [a, b), φ(t − ) exists in S for all t ∈ (a, b] and
φ(t − ) = φ(t) for all but at most a finite number of points
t ∈ (a, b] . Observe that PC([a, b], S ) denotes the family of
right-continuous functions on [a, b) with finite left-hand limits on (a, b] contained in S and whose discontinuities do not
accumulate in finite time. Finally, let 0̃nx denote the zero element of PC([−d, 0], Rnx ).
The main contributions of this article are fourfold: a) the
design of MATIs in nonlinear delayed NCSs with UGES protocols even for the so-called large delays; b) the LyapunovRazumikhin-based procedure for rendering L p -stability of
nonlinear impulsive delayed systems and computing the associated L p -gains; c) the consideration of NCS links with
nonidentical time-dependent delays; and d) the inclusion of
model-based estimators. In contrast to our conference paper [27], this article incorporates variable delays, contains
proofs, provides a nonlinear numerical example with delayed
plant dynamics and designs model-based estimation that prolongs MATIs [8]. Furthermore, this article accompanies [28].
2.2
Impulsive Delayed Systems
In this article, we consider nonlinear impulsive delayed systems
+
χ(t ) = hχ (t, χt )
t ∈T
)
Σ
(1)
χ̇(t) = fχ (t, χt , ω)
otherwise ,
y = `χ (t, χt , ω)
where χ ∈ Rnχ is the state, ω ∈ Rnω is the input and y ∈ Rny
is the output. The functions fχ and hχ are regular enough to
guarantee forward completeness of solutions which, given initial time t0 and initial condition χt0 ∈ PC([−d, 0], Rnχ ), where
d ≥ 0 is the maximum value of all time-varying delay phenomena, are given by right-continuous functions t 7→ χ(t) ∈
PC([t0 − d, ∞], Rnχ ). Furthermore, χt denotes the translation
operator acting on the trajectory χ(·) defined by χt (θ) :=
χ(t + θ) for −d ≤ θ ≤ 0. In other words, χt is the restriction of trajectory χ(·) to the interval [t − d,t] and translated to
[−d, 0]. For χt ∈ PC([−d, 0], Rnχ ), the norm of χt is defined
The remainder of this article is organized as follows. Section 2 presents the utilized notation and stability notions regarding impulsive delayed systems. Section 3 states the problem of finding MATIs for nonlinear delayed NCSs with UGES
protocols in the presence of nonuniform communication delays and exogenous/modeling disturbances. A methodology
to solve the problem is presented in Section 4. Detailed numerical examples are provided in Section 5. Conclusions and
2
by kχt k = sup−d≤θ≤0 kχt (θ)k. Jumps of the state are denoted
χ(t + ) and occur at time instants t ∈ T := {t1 ,t2 , . . .}, where
ti < ti+1 , i ∈ N0 . The value of the state after a jump is given by
χ(t + ) for each t ∈ T . For a comprehensive discussion regarding the solutions to (1) considered herein, refer to [2, Chapter
2 & 3]. Even though the considered solutions to (1) allow
for jumps at t0 , we exclude such jumps in favor of notational
convenience.
and a controller with delayed dynamics
ẋc = fc (t, xct , y, ωc ),
u = gc (t, xct ),
where x p ∈ Rn p and xc ∈ Rnc are the states, y ∈ Rny and u ∈
Rnu are the outputs, and (u, ω p ) ∈ Rnu × Rnω p and (y, ωc ) ∈
Rny × Rnωc are the inputs of the plant and controller, respectively, where ω p and ωc are external disturbances to (and/or
modeling uncertainties of) the plant and controller, respectively. The translation operators x pt and xct are defined in
Section 2.2 while the corresponding plant and controller delays are d p ≥ 0 and dc ≥ 0, respectively. For notational convenience, constant plant and controller delays are considered.
Let us now model the communication network between the
plant and controller over which intermittent and realistic exchange of information takes place (see Figure 1). The value
of u computed by the controller that arrives at the plant is denoted û. Similarly, the values of y that the controller actually
receives are denoted ŷ. Consequently, we have
Definition 1 (Uniform Global Stability). For ω ≡ 0nω , the
system Σ is said to be Uniformly Globally Stable (UGS) if
for any ε > 0 there exists δ(ε) > 0 such that, for each t0 ∈
R and each χt0 ∈ PC([−d, 0], Rnχ ) satisfying kχt0 k < δ(ε),
each solution t 7→ χ(t) ∈ PC([t0 − d, ∞], Rnχ ) to Σ satisfies
kχ(t)k < ε for all t ≥ t0 and δ(ε) can be chosen such that
limε→∞ δ(ε) = ∞.
Definition 2 (Uniform Global Asymptotic Stability). For ω ≡
0nω , the system Σ is said to be Uniformly Globally Asymptotically Stable (UGAS) if it is UGS and uniformly globally
attractive, i.e., for each η, ζ > 0 there exists T (η, ζ) > 0 such
that kχ(t)k < η for every t ≥ t0 + T (η, ζ) and every kχt0 k < ζ.
u = û,
Definition 4 (L p -Stability with Bias b). Let p ∈ [1, ∞]. The
system Σ is L p -stable with bias b(t) ≡ b ≥ 0 from ω to y with
(linear) gain γ ≥ 0 if there exists K ≥ 0 such that, for each t0 ∈
R and each χt0 ∈ PC([−d, 0], Rnχ ), each solution to Σ from
χt0 satisfies ky[t0 ,t]k p ≤ Kkχt0 k + γkω[t0 ,t]k p + kb[t0 ,t]k p for
each t ≥ t0 .
(4)
where yt and ut are translation operators and the maximal network-induced delay d ≥ 0 (e.g., propagation delays
and/or delays arising from protocol arbitration). The operator (yt , ut ) in (5) delays each component of (y, u) for
the respective delay. Essentially, if the ith component of
(y(t), u(t)), that is (y(t), u(t))i , is transmitted with delay
di : R → R+ , then the ith component of (yt , ut ), that is
(yt , ut )i , is in fact (y(t − di (t)), u(t − di (t)))i . Accordingly,
d := max{supt∈R d1 (t), . . . , supt∈R dny +nu (t)}.
Due to intermittent transmissions of the components of y
and u, the respective components of ŷ and û are updated at
time instants t1 ,t2 , . . . ,ti , . . . ∈ T , i.e.,
)
ŷ(ti+ ) = yt + hy (ti , e(ti ))
ti ∈ T ,
(6)
û(ti+ ) = ut + hu (ti , e(ti ))
Definition 5 (L p -Detectability). Let p ∈ [1, ∞]. The state χ
of Σ is L p -detectable from (y, ω) with (linear) gain γ ≥ 0
if there exists K ≥ 0 such that, for each t0 ∈ R and each
χt0 ∈ PC([−d, 0], Rnχ ), each solution to Σ from χt0 satisfies kχ[t0 ,t]k p ≤ Kkχt0 k + γky[t0 ,t]k p + γkω[t0 ,t]k p for each
t ≥ t0 .
Definitions 1, 2 and 3 are motivated by [14], while Definition 5 is inspired by [20]. Definition 4 is motivated by [20]
and [13]. When b = 0, we say “L p -stability” instead of “L p stability with bias 0”.
where hy : R × Rne → Rny and hu : R × Rne → Rnu model
measurement noise, channel distortion and the underlying
scheduling protocol. The role of hy and hu is as follows. Suppose that the NCS has l links. Accordingly, the error vector e
can be partitioned as e := (e1 , . . . , el ). In order to avoid cumbersome indices, let us assume that each NCS link is characterized by its own delay. Hence, there are merely l (rather
than ny + nu ) different delays di : R → R+ in (5). Besides the
already introduced upper bound d on di (t)’s, we assume that
di (t)’s are differentiable with bounded |d˙i (t)|. As orchestrated
Problem Formulation
Consider a nonlinear control system consisting of a plant with
delayed dynamics
ẋ p = f p (t, x pt , u, ω p ),
y = g p (t, x pt ),
y = ŷ,
on the right hand sides of (2) and (3). In our setting, the quantity û is the delayed and distorted input u fed to the plant (2)
while the quantity ŷ is the delayed and distorted version of y
received by the controller (3). We proceed further by defining
the error vector
ey (t)
ŷ(t) − yt
e=
:=
,
(5)
eu (t)
û(t) − ut
Definition 3 (Uniform Global Exponential Stability). For
ω ≡ 0nω , the system Σ is said to be Uniformly Globally Exponentially Stable (UGES) if there exist positive constants λ and
M such that, for each t0 ∈ R and each χt0 ∈ PC([−d, 0], Rnχ ),
each solution t 7→ χ(t) ∈ PC([t0 − d, ∞], Rnχ ) to Σ satisfies
kχ(t)k ≤ Mkχt0 ke−λ(t−t0 ) for each t ≥ t0 .
3
(3)
(2)
3
(i) akek ≤ W (i, e) ≤ akek, and
(ii) W (i + 1, h(ti , e)) ≤ ρW (i, e),
for all (i, e) ∈ N0 × Rne .
Notice that, even though the delays could result from protocol
arbitration, the delays are not a part of the UGES protocol definition [11,20]. In addition, T is not a part of the protocol, but
rather a consequence, as it is yet to be designed. Commonly
used UGES protocols are the Round Robin (RR) and TryOnce-Discard protocol (TOD) (consult √
[5,11,20]).p
The correl, ρRR = (l − 1)/l
sponding constants are aRR = 1, aRR = p
for RR and aT OD = aT OD = 1, ρT OD = (l − 1)/l for TOD.
Explicit expressions of the noise-free h(t, e) for RR and TOD
are provided in [20], but are not needed in the context of this
article.
The properties imposed on the NCS in Figure 1 are summarized in the following standing assumption.
Figure 1: A diagram of a control system with the plant and controller interacting over a communication network with intermittent
information updates. The two switches indicate that the information between the plant and controller are exchanged (complying with
some scheduling protocol among the NCS links) at discrete time instants belonging to a set T . The communication delays in each NCS
link are time varying and, in general, different.
by (6), if the jth NCS link is granted access to the communication medium at some ti ∈ T , the corresponding components of
(ŷ(ti ), û(ti )) jump to the received values. It is to be noted that
all other components of (ŷ(ti ), û(ti )) remain unaltered. Consequently, the related components of e(ti ) reset to the noise
ν j (ti ) present in the received data, i.e.,
e j (ti+ ) = ν j (ti ),
Assumption 1. The jump times of the NCS links at the controller and plant end obey the underlying UGES scheduling protocol (characterized through h) and occur at transmission instants belonging to T := {t1 ,t2 , . . . ,ti , . . .}, where
ε ≤ ti+1 − ti ≤ τ for each i ∈ N0 with ε > 0 arbitrarily small.
The received data is corrupted by measurement noise and/or
channel distortion (characterized through h as well). In addition, each NCS link is characterized by the network-induced
delay di (t), i ∈ {1, . . . , l}.
(7)
and we assume that
sup
kν j (t)k = Kν .
t∈R, j∈{1,...,l}
The existence of a strictly positive τ, and therefore the existence of ε > 0, is demonstrated in Remark 3.
A typical closed-loop system (2)-(8) with continuous (yet
delayed) information flows in all NCS links might be robustly
stable (in the L p sense according to (15)) only for some sets
of di (t), i ∈ {1, . . . , l}. We refer to the family of such delay
sets as the family of admissible delays and denote it D . Next,
given some admissible delays di (t), i ∈ {1, . . . , l}, the maximal τ which renders L p -stability (with a desired gain) of the
closed-loop system (2)-(8) is called MATI and is denoted τ.
We are now ready to state the main problem studied herein.
Noise ν j (ti ), which is embedded in hy and hu , models any
discrepancy between the received values and their actual values at time ti − d j (t) (when the jth NCS link of (y(t), u(t))
was sampled). As already indicated, this discrepancy can be
a consequence of measurement noise and channel distortion.
We point out that ν j has nothing to do with ω p nor ωc . Observe that out-of-order packet arrivals, as a consequence of
the time-varying delays, are allowed for.
In between transmissions, the values of ŷ and û need not to
be constant as in [11], but can be estimated in order to extend
transmission intervals (consult [8] for more). In other words,
for each t ∈ [t0 , ∞) \ T we have
ŷ˙ = fˆp t, x pt , xct , ŷt , ût , ω p , ωc ,
û˙ = fˆc t, x pt , xct , ŷt , ût , ω p , ωc ,
(8)
Problem 1. Given admissible delays di (t), i ∈ {1, . . . , l}, estimator (8) and the UGES protocol of interest, determine the
MATI τ to update components of (ŷ,û) such that the NCS (2)(8) is L p -stable with bias and a prespecified L p -gain for some
p ∈ [1, ∞].
where the translation operators ŷt and ût are with delay d. The
commonly used ZOH strategy is characterized by ŷ˙ ≡ 0ny and
û˙ ≡ 0nu .
The following definition of UGES scheduling protocols is
extracted from [20] and [11].
Remark 1. Even though our intuition (together with the case
studies provided herein and in [27]) suggests that merely
“small enough" delays (including the zero delay) are admissible because the control performance impairs (i.e., the corresponding L p -gain increases) with increasing delays, this obDefinition 6. Consider the noise-free setting, i.e., Kν = 0. The servation does not hold in general [10], [21, Chapter 1.],
protocol given by h := (hy , hu ) is UGES if there exists a func- [22]. In fact, “small" delays may destabilize some systems
tion W : N0 × Rne → R+ such that W (i, ·) : Rne → R+ is lo- while “large" delays might destabilize others. In addition,
cally Lipschitz (and hence almost everywhere differentiable) even a second order system with a single discrete delay might
for every i ∈ N0 , and if there exist positive constants a, a and toggle between stability and instability as this delay is being
0 ≤ ρ < 1 such that
decreased. Clearly, the family D needs to be specified on a
4
4.1
case-by-case basis. Hence, despite the fact that the case studies presented herein and in [27] yield MATIs that hold for all
smaller time-invariant delays (including the zero delay) than
the delays for which these MATIs are computed for, it would
be erroneous to infer that this property holds in general.
4
L p -Stability with Bias of Impulsive Delayed LTI Systems
Before invoking the small-gain theorem in the upcoming subsection, let us establish conditions on the transmission interval
τ and delay d(t) that yield L p -stability with bias for a class of
impulsive delayed LTI systems. Clearly, the results of this
subsection are later on applied towards achieving L p -stability
with bias and an appropriate L p -gain of Σe .
Consider the following class of impulsive delayed LTI system
Methodology
Along the lines of [20], we rewrite the closed-loop system (2)(8) in the following form amenable for small-gain theorem
(see [14, Chapter 5]) analyses:
)
x(t + ) = x(t)
t ∈T
(9a)
e(t + ) = h(t, e(t))
)
ẋ = f (t, xt , e, ω)
otherwise,
(9b)
ė = g(t, xt , et , ωt )
˙ = aξ(t − d(t)) + ũ(t),
ξ(t)
+
ξ(t ) = cξ(t) + ν̃(t),
t∈
/T
(14a)
t ∈T,
(14b)
where a ∈ R and c ∈ (−1, 1), initialized with some ξt0 ∈
˘ 0], R). In addition, d(t) is a continuous function upPC([−d,
per bounded by d˘ while ũ, ν̃ : R → R denote external inputs
and ν̃ ∈ L∞ .
Lemma 1. Assume ũ ≡ 0, ν̃ ≡ 0 and consider a positive con2
stant r. In addition, let λ1 := ar , and λ2 := c2 for c 6= 0 or
merely λ2 ∈ (0, 1) for c = 0. If there exist constants λ > 0,
M > 1 such that the conditions
(I) τ λ + r + λ1 Me−λτ < ln M, and
˘
(II) τ λ + r + λλ1 eλd < − ln λ2
where x := (x p , xc ), ω := (ω p , ωc ), and functions f , g and h
are given by (10) and (11). We assume enough regularity on f
and g to guarantee existence of the solutions on the interval of
interest [2, Chapter 3]. Observe that differentiability of di (t)’s
and boundedness of |d˙i (t)| play an important role in attaining
regularity of g. For the sake of simplicity, our notation does
not explicitly distinguish between translation operators with
delays d p , dc , d or 2d in (10), (11) and in what follows. In
this regard, we point out that the operators x pt and xct are with
delays d p and dc , respectively, the operators g pt and gct within
fˆp and fˆc are with delay 2d while all other operators are with
delay d. In what follows we also use d := 2d + max{d p , dc },
which is the maximum value of all delay phenomena in (11).
For future reference, the delayed dynamics
2
hold, then the system (14) is UGES and kξ(t)k ≤
√
λ
Mkξt0 ke− 2 (t−t0 ) for all t ≥ t0 .
The previous lemma, combined with the work presented in
[1], results in the following theorem.
Theorem 1. Suppose that the system given by (14) is UGES
with constants λ > 0 and M > 1 and that supt∈R
kν̃(t)k ≤ K̃ν .
√
x(t + ) = x(t)
t ∈T
(12a)
K̃ν M
Then, the system (14) is L p -stable with bias λε from ũ to ξ
e 2 −1
ẋ = f (t, xt , e, ω)
otherwise,
(12b)
√
and with gain λ2 M for each p ∈ [1, ∞].
are termed the nominal system Σn , and the impulsive delayed
dynamics
4.2 Obtaining MATIs via the Small-Gain Thee(t + ) = h(t, e(t))
t ∈T
orem
(13a)
We are now ready to state and prove the main result of this
article. Essentially, we interconnect Σn and Σe via suitable
are termed the error system Σe . Observe that Σn contains outputs (i.e., H(xt , ωt ) and W (i, e), respectively), impose the
delays, but does not depend on h nor T as seen from (12). small-gain condition and invoke the small-gain theorem.
Instead, h and T constitute the error subsystem Σe as seen Theorem 2. Suppose the underlying UGES protocol,
from (13).
d1 (t), . . . , dl (t) and Kν ≥ 0 are given. In addition, assume
The remainder of our methodology interconnects Σn and Σe that
using appropriate outputs. Basically, W (i, e) from Definition
function H : PC([−d, 0], Rnx ) ×
6 is the output of Σe while the output of Σn , denoted H(xt , ωt ), (a) there exists a ncontinuous
m
PC([−d, 0], R ω ) → R such that the system Σn given by
is obtained from g(t, xt , et , ωt ) and W (i, e) as specified in Sec(12)
is L p -stable from (W, ω) to H(xt , ωt ) for some p ∈
tion 4.2. Notice that the outputs H(xt , ωt ) and W (i, e) are
[1,
∞],
i.e., there exist KH , γH ≥ 0 such that
auxiliary signals used to interconnect Σn and Σe and solve
Problem 1, but do not exist physically. Subsequently, the
kH[t,t0 ]k p ≤KH kxt0 k + γH k(W, ω)[t,t0 ]k p ,
(15)
small-gain theorem is employed to infer L p -stability with
bias. Proofs of the upcoming results are in the Appendix.
for all t ≥ t0 , and
ė = g(t, xt , et , ωt )
otherwise,
(13b)
5
f p t, x pt , gct (t, xct ) + eu (t) , ω p (t)
{z
}
|
(2),(3)
f1 (t, xt , e, ω)
=û(t) using (3) and (5)
;
f (t, xt , e, ω) :=
=:
f2 (t, xt , e, ω)
fc t, xct , g pt (t, x pt ) + ey (t), ωc (t)
{z
}
|
hy t, e(t)
hu t, e(t)
h(t, e(t)) :=
(10)
=ŷ(t) using (2) and (5)
fˆp t,x pt ,xct ,g pt (t,x pt )+eyt ,gct (t,xct )+eut ,ω(t)
|
(5)
g(t, xt , et , ωt ) := z
fˆ
c
−
{z
}|
}|
{
model-based estimator (8)
t,x pt ,xct ,g pt (t,x pt )+eyt ,gct (t,xct )+eut ,ω(t)
c
− ∂g
∂t
|
(b) there exists L ≥ 0 and d : R → R+ , supt∈R d(t) = d,˘ such
that for almost all t ≥ t0 , almost all e ∈ Rne and for all
(i, xt , ωt ) ∈ N0 × PC([−d, 0], Rnx ) × PC([−d, 0], Rnω ) it
holds that
E
D ∂W (i, e)
, g(t, xt , et , ωt ) ≤LW (i, e(t − d(t)))+
∂e
+ kH(xt , ωt )k.
(16)
∂g p
∂t
(t,x pt )− ∂x p (t,x pt ) f1t (t,xt ,e,ω)
p t
∂g
t
}
∂gc
(t,xct ) f2t (t,xt ,e,ω)
∂xc t
{z
}
{z
=−ẏt using (2) and (10)
t
(t,xct )−
(11)
=−u̇t using (3) and (10)
Remark 4. The right hand side of (16) might not be descriptive enough for many problems
D of interest. In general,
E
(16) should be sought in the form ∂W∂e(i,e) , g(t, xt , et , ωt ) ≤
`
`
∑m
k=1 LkW (i, e(t − dk (t))) + kH(xt , ωt )k, where dk : R → R+
and m ≥ 1. As this general form leads to tedious computations
(as evident from the proof of Lemma 1 in the Appendix), we
postpone its consideration for the future. For the time being,
one can intentionally delay the communicated signals in orThen, the NCS (9) is L p -stable with bias from ω to (H, e) for der to achieve a single discrete delay d(t) in (16). This idea is
each τ for√which there exist M > 1 and λ > 0 satisfying (I), often found in the literature and can be accomplished via the
Controller Area Network (CAN) protocol, time-stamping of
(II) and λ2 MγH < 1 with parameters a = aa L and c = ρ.
data and introduction of buffers at receiver ends (refer to [12]
Remark 2. According to Problem 1, condition (a) re- and references therein).
quires the underlying delays to be admissible, i.e., Remark 5. Noisy measurements can be a consequence of
{d1 (t), . . . , dl (t)} ∈ D . Condition (a) implies that the nom- quantization errors. According to [18], feedback control
inal system (i.e., the closed-loop system) is robust with re- prone to quantization errors cannot yield closed-loop systems
spect to intermittent information and disturbances. Besides with linear L -gains. Hence, the bias term in the linear gain
p
L p -stability, typical robustness requirements encountered in L -stability with bias result of Theorem 2 cannot be removed
p
the literature include Input-to-State Stability (ISS) and pas- without contradicting the points in [18]. Further investigasivity [30]. Condition (b) relates the current growth rate of tions of quantized feedback are high on our future research
W (i, e) with its past values. As shown in Section 5, all rec- agenda.
ommendations and suggestions from [20] and [11] regarding how to obtain a suitable W (i, e) readily apply because Remark 6. Let us consider the case of lossy communication
W (i, e) characterizes the underlying UGES protocol (and not channels. If there is an upper bound on the maximum number of successive dropouts, say Nd ∈ N, simply use Nτd as the
the plant-controller dynamics).
transmission interval in order for Theorem 2 to hold. MoreRemark 3 (Zeno-freeness). The left-hand sides of conditions over, the transmission instants among NCS links need not to
(I) and (II) from Lemma 1 are nonnegative continuous func- be (and often cannot be) synchronized. In this case, each NCS
tions of τ ≥ 0 and approach ∞ as τ → ∞. Also, these left-hand must transmit at a rate smaller than τRR (instead of τRR l),
sides equal zero for τ = 0. Note that both sides of (I) and (II) where τRR is the MATI obtained for the RR protocol, in order
are continuous in λ, M, λ1 , λ2 and d.˘ Hence, for every λ > 0, to meet the prespecified performance requirements. Observe
λ1 ≥ 0, M > 1, λ2 ∈ (0, 1) and d˘ ≥ 0 there exists
√ τ > 0 such that this leads to asynchronous transmission protocols, which
that (I) and (II) are satisfied. Finally, since λ2 M is contin- in turn increases the likelihood of packet collisions [17].
uous in λ and M, we infer
√ that for every finite γH > 0 there Corollary 1. Assume that the conditions of Theorem 2 hold
exists τ > 0 such that λ2 MγH < 1. In other words, for each
and that x is L p -detectable from (W, ω, H). Then the NCS (9)
admissible di (t), i ∈ {1, . . . , l}, the unwanted Zeno behavior
is L p -stable with bias from ω to (x, e).
is avoided and the proposed methodology does not yield continuous feedback that might√be impossible to implement. No- In the following proposition, we provide conditions that yield
tice that each τ yielding λ2 MγH < 1 is a candidate for τ. UGS and GAS of the interconnection Σn and Σe . Recall that
Depending on r, λ2 , λ and M, the maximal such τ is in fact ω ≡ 0nω and Kν = 0 are the disturbance and noise settings,
MATI τ.
respectively, corresponding to UGS and GAS.
6
Proposition 1. Assume that the interconnection of systems Σn
and Σe , given by (12) and (13), is L p -stable from ω to (x, e). If
p = ∞, then this interconnection is UGS. When p ∈ [1, ∞), assume that f (t, xt , e, 0nω ) and g(t, xt , et , 0̃nω ) are (locally) Lipschitz uniformly in t as well as that kH(xt , 0̃nω )k → 0 as
kxt k → 0. Then, this interconnection is GAS.
ė(t) = ŷ˙ − Be(t − d) +
0.5 −1
x(t)+
0
0
| {z }
C1
0 0
0 0
+
x(t − d) +
x(t − 2d)+
2 2
0 0.3
| {z }
| {z }
C2
5
5.1
Numerical Examples
(19)
C3
0.25 0
0 0
+
x(t)N(xt , e) +
x(t − d)N(xt , et )+
0
0
−1 0
| {z }
| {z }
Constant Delays
C4
The following example is motivated by [31, Example 2.2.]
and all the results are provided for p = 2. Consider the following nonlinear delayed plant (compare with (2))
ẋ p1 (t)
=
ẋ p2 (t)
−0.5x p1 (t)+x p2 (t)−0.25x p1 (t) sin u(t)x p2 (t−d p1 )
+
x p1 (t) sin u(t)x p2 (t−d p1 ) +1.7x p2 (t−d p2 )+u(t)−x p2 (t)
h
i
ω (t)
+ ω1 (t)
C5
0 0
−1 0
ω(t) +
ω(t − d),
+
0 −1
0 0
| {z }
| {z }
C6
C7
h
− 2 x p1 (t) + e1 (t) − 2 x p2 (t − d) +
where N(xt , e) := sin
!
h
i
e2 (t) x p2 (t − d p1 ) and N(xt , et ) := sin − 2 x p1 (t − d) +
!
i
e1 (t − d) − 2 x p2 (t − 2d) + e2 (t − d) x p2 (t − d p1 − d) .
2
controlled with (compare with (3))
u(t) = −2x p1 (t) − 2x p2 (t).
According to [20] and [11], we select WRR (i, e) := kD(i)ek
and WT OD (t, e) := kek, where D(i) is a diagonal matrix whose
diagonal
√ elements are lower bounded by 1 and upper bounded
by l. Next, we determine LRR , HRR (x, ω, d), LT OD and
HT OD (x, ω, d) from Theorem 2 for the ZOH strategy (i.e.,
ŷ˙ ≡ 0ny ) obtaining (20) and (21). In order to estimate γH ,
we utilize Lyapunov-Krasovskii functionals according to [4,
Chapter 6] and [6]. Basically, if there exist γ ≥ 0 and a
Lyapunov-Krasovskii functional V (xt ) for the nominal system (12), that is (18), with the input (W, ω) and the output H
such that its time-derivative along the solution of (12) with a
zero initial condition satisfies:
As this controller is without internal dynamics, therefore x(t) := x p (t) = (x p1 (t), x p2 (t)). Additionally, ω(t) :=
(ω1 (t), ω2 (t)).
Let us consider the NCS setting in which noisy information
regarding x p1 and x p2 are transmitted over a communication
network while the control signal is not transmitted over a communication network nor distorted (i.e., û = u). In addition,
consider that the information regarding x p2 arrives at the controller with delay d while information regarding x p1 arrives in
timely manner. For the sake of simplicity, let us take d = d p2 .
Apparently, the output of the plant is y(t) = x p (t) = x(t) and
there are two NCS links so that l = 2. Namely, x p1 is transmitV̇ (xt ) + H > H − γ2 (W,ω)> (W, ω) ≤ 0,
ted through one NCS link while x p2 is transmitted through the
(22)
∀xt ∈ C([−d, 0], Rnx ),
second NCS link. The repercussions of these two NCS links
are modeled via the following error vector (compare with (5)) than the corresponding L -gain γ is less than γ. The func2
H
tional used herein is
!
e
x p1 (t)
0
Z 0
.
e = 1 = ŷ −
+
e2
0
x p2 (t − d)
x(t + θ)> Ex(t + θ)dθ,
(23)
V (xt ) = x(t)>Cx(t) +
{z
}
|
−d
yt
where C and E are positive-definite symmetric matrices.
The expressions (10) and (11) for this example become:
Next, we illustrate the steps behind employing (23). Let
us
focus on TOD (i.e., the input is (e, ω)) and the output
−0.5 1
0
0
ẋ(t) =
x(t) +
x(t − d)+
(17) C1 x(t) +C2 x(t − d) +C6 ω(t). The same procedure is repeated
−2 −1
0 −0.3
|
{z
}
| {z }
for the remaining terms of HRR (xt , ωt ) and HT OD (xt , ωt ). For
A1
A2
the Lyapunov-Krasovskii functional in (23), the expression
(22) boils down to the Linear Matrix Inequality (LMI) (see [4]
for more) given by (24). Notice that the above LMI has to hold
−0.25 0
0
0
+
x(t)N(xt , e) +
e(t) + ω(t), (18) for all N(xt , e) ∈ [−1, 1]. Using the LMI Toolbox in MAT1
0
−2 −2
LAB, we find that the minimal γ for which (24) holds is in
|
{z
}
| {z }
B1
B
fact γH . For our TOD example and the specified output, we
7
D ∂W (i, e) E
√
RR
, ė ≤ kD(i)ėk ≤ lkBk kD(i)e(t − d)k +
| {z } |
{z
}
∂e
LRR
WRR (i,e(t−d))
√
+ l kC1 x(t) +C2 x(t − d) +C3 x(t − 2d) +C6 ω(t) +C7 ω(t − d)k + kC4 x(t)k + kC5 x(t − d)k ,
|
{z
}
(20)
HRR (xt ,ωt )
E
T OD (i, e)
, ė ≤ kBk ke(t − d)k +
|{z} | {z }
∂e
D ∂W
LT OD WT OD (i,e(t−d))
+ kC1 x(t) +C2 x(t − d) +C3 x(t − 2d) +C6 ω(t) +C7 ω(t − d)k + kC4 x(t)k + kC5 x(t − d)k ,
|
{z
}
(21)
HT OD (xt ,ωt )
>
>
>
A>
1 C+CA1 +E+N(xt ,e)(B1 C+CB1 )+C1 C1 CA2 +C1 C2
>C
>C
A>
C+C
−E+C
1
2
2
2 2
B> C
02×2
C+C6>C1
C6>C2
CB
C+C1>C6
02×2
C2>C6
≤ 0.
2
−γ I2
02×2
02×2 −γ2 I2 +C6>C6
(24)
estimator (such as ŷ˙ = kBŷ(t − d) for some k ∈ R) can be employed as d approaches zero (because the estimator slightly
decreases the MATIs as seen in Figure 2) to render greater
MATIs in comparison with the scenarios without estimation.
In addition, notice that the case d = 0 boils down to ODE
modeling so that we can employ less conservative tools for
computing L2 -gains. Accordingly, the 4 × 4 LMI given by
(24) becomes a 3 × 3 LMI resulting in a smaller γH . Furthermore, the constant a in Theorem 2 becomes L, rather than aa L,
which in turn decreases γW for the same τ. Apparently, MAγd
TIs pertaining to UGAS are greater than the MATIs pertaining
γW γH < 1 −
γdes
to L p -stability from ω to (x, e) and these are greater than the
that needs to be satisfied (by changing γW through changing MATIs pertaining to L p -stability from ω to x with γdes = 50.
For completeness, we provide the gains used to obtain
MATIs) in order to achieve the desired gain γdes . In addiFigure
2: γH,T OD = 9.6598 for UGAS with ZOH and d =
tion, observe that the conditions of Proposition 1 hold (and
0;
γ
H,T
OD = 4.3344 for UGAS with estimation and d =
the closed-loop system is an autonomous system) so that we
0;
γ
H,T
OD = 22.3631 for UGAS with ZOH and d > 0;
can infer UGAS when ω ≡ 0nω and Kν = 0.
γ
=
27.3659 for UGAS with estimation and d > 0;
H,T
OD
Let us now introduce the following estimator (compare
γ
=
10.8958 for L p -stability with ZOH and d = 0;
H,T
OD
with (8))
γH,T OD = 5.3258 for L p -stability with estimation and d = 0;
γH,T OD = 26.4601 for L p -stability with ZOH and d > 0;
1 0
ŷ˙ = Bŷ(t − d) = B e(t − d) +
x(t − d)+
γH,T OD = 31.7892 for L p -stability with estimation and d > 0;
0 0
γd = 3.5884
!
√ for d = 0; and, γd = 7.9597 for d > 0. Recall that
γH,RR = 2γH,T OD .
0 0
+
x(t − 2d) ,
(25)
0 1
obtain γH = 18.7051. This γH holds for all d ≥ 0. In other
words, any d ≥ 0 is an admissible delay √
and belongs to the
family D . For RR, simply multiply γH by 2.
Detectability of x from (W, x, H), which is a condition of
Corollary 1, is easily inferred by taking x(t) to be the output of
the nominal system and computing the respective L2 -gain γd .
Next, let us take the output of interest to be x and find MATIs
that yield the desired L p -gain from ω to x to be γdes = 50.
Combining (57) with γd leads to the following condition
5.2
Time-Varying Delays
which can be employed when one is interested in any of the
three performance objectives (i.e., UGAS, L p -stability or L p - The following example is taken from [25, 29] and the results
are provided for p = 2. Consider the inverted pendulum (comstability with a desired gain).
Figure 2 provides evidence that the TOD protocol results in pare with (2)) given by
greater MATIs (at the expense of additional implementation
ẋ p1 = x p2 + ω1
complexity/costs) and that the model-based estimators signif1
icantly prolong MATIs, when compared with the ZOH stratẋ p2 = (−g cos(x p1 ) + u) + ω2 ,
egy, especially as d increases. We point out that a different
L
8
Try−Once−Discard (TOD)
Round Robin (RR)
60
40
UGAS/no estim.
Lp−stab./no estim.
des. gain/no estim.
UGAS/with estim.
Lp−stab./with estim.
des. gain/with estim.
UGAS/no estim.
Lp−stab./no estim.
des. gain/no estim.
UGAS/with estim.
Lp−stab./with estim.
des. gain/with estim.
30
MATI [ms]
25
20
50
40
MATI [ms]
35
UGAS/no estim.
Lp−stab./no estim.
des. gain/no estim.
UGAS/with estim.
Lp−stab./with estim.
des. gain/with estim.
UGAS/no estim.
Lp−stab./no estim.
des. gain/no estim.
UGAS/with estim.
Lp−stab./with estim.
des. gain/with estim.
30
15
20
10
10
5
0
0
0
50
100
delay d [ms]
150
200
0
50
100
delay d [ms]
(a)
150
200
(b)
Figure 2: Numerically obtained MATIs for different delay values d ≥ 0 in scenarios with and without estimation: (a) RR; and, (b) TOD.
0
0 −1
˙
ω(t)+
ė(t) = ŷ +
x(t) + −1
0 0
| 0{z 0}
| {z }
C
where g = 9.8 and L = 2, controlled with
u = −Lλx p2 + g cos(x p1 ) − K(x p2 + λx p1 ),
where K = 50 and λ = 1. Clearly, the control system goal
is to keep the pendulum at rest in the upright position. As
this controller is without internal dynamics, therefore x(t) :=
x p (t) = (x p1 (t), x p2 (t)). Additionally, ω(t) := (ω1 (t), ω2 (t)).
Consider the NCS setting in which noisy information regarding x p1 and x p2 are transmitted over a communication network while the control signal is not transmitted over a communication network nor distorted (i.e., û = u). In addition,
consider that the information regarding x p2 arrives at the con˙
troller with delay d(t) ≤ d˘ and |d(t)|
≤ d˘1 while information
regarding x p1 arrives instantaneously. Apparently, the output
of the plant is y(t) = x p (t) = x(t) and there are two NCS links
so that l = 2. Namely, x p1 is transmitted through one NCS
link while x p2 is transmitted through the second NCS link.
The repercussions of these two NCS links are modeled via
the following error vector (compare with (5))
!
e1
x p1 (t)
0
.
e=
= ŷ −
+
e2
0
x p2 (t − d(t))
|
{z
}
−
0
+
+
n(x1 (t), e1 (t))
0
−Kλ
L
|
0
0
ω(t − d(t))+
+ 00 −1
n(x1 (t − d(t)), e1 (t − d(t)))
| {z }
C2
0
+
0
|
!
0
˙
x(t − 2d(t)) (1 − d(t)),
+ λL
{z
}
K
L
B3
where n(x1 (t), e1 (t)) =
−2g
L
sin
e1 (t)+2x1 (t)
2
sin
e1 (t)
2
.
According to [20] and [11], we select WRR (i, e) := kD(i)ek
and WT OD (t, e) := kek, where D(i) is a diagonal matrix whose
diagonal
√ elements are lower bounded by 1 and upper bounded
by l. Next, we determine LRR , HRR (x, ω, d), LT OD and
HT OD (x, ω, d) from Theorem 2 for the ZOH strategy (i.e.,
ŷ˙ ≡ 0ny ) obtaining (26) and (27).
The Lyapunov-Krasovskii functional used for the pendulum example is
V (t, xt , ẋt ) = x(t)> Px(t) +
A2
0
x(t − d(t))−
Kλ
0
L
| {z }
0
B2
The expressions (10) and (11) for this example become:
0
1
0
0
ẋ(t) = −Kλ
x(t) +
x(t − d(t))+
0 −K
−1
L − λL
L
|
{z
}
|
{z
}
− Be(t − d(t)) +
+
yt
A1
1
B1
+ d˘
0
e(t) + ω(t),
−K
−
λL
L
{z
}
Z 0 Z t
−d˘ t+θ
Z t
t−d˘
x(s)> Sx(s)ds+
ẋ(s)> Rẋ(s)dsdθ +
Z t
x(s)> Qx(s)ds,
t−d(t)
where P is a positive-definite symmetric matrix while S, R
and Q are positive-semidefinite symmetric matrices. Next,
B
9
D ∂W (i, e) E
√
g
RR
,ė ≤ kD(i)ėk ≤ l 1 + d˘1 kBk +
kD(i)e(t − d(t))k +
{z
}
∂e
L }|
|
{z
WRR (i,e(t−d(t)))
LRR
√
˙
B2 x(t − d(t)) + B3 x(t − 2d(t)) +C2 ω(t − d(t))
,
+ l B1 x(t) +C1 ω(t) + 1 − d(t)
|
{z
}
(26)
HRR (xt ,ωt )
E
g
T OD (i, e)
,ė ≤ 1 + d˘1 kBk +
ke(t − d(t))k +
{z
}
∂e
L} |
|
{z
D ∂W
LT OD
+
WT OD (i,e(t−d(t)))
˙
B1 x(t) +C1 ω(t) + 1 − d(t)
B2 x(t − d(t)) + B3 x(t − 2d(t)) +C2 ω(t − d(t))
,
|
{z
}
(27)
HT OD (xt ,ωt )
we take the output of interest to be x and find MATIs that
yield the desired L p -gain from ω to x to be γdes = 15. In addition, observe that the conditions of Proposition 1 hold (and
the closed-loop system is an autonomous system) so that we
can infer UGAS when ω ≡ 0nω and Kν = 0.
We use the following estimator (compare with (8))
of MATIs) such that the performance objectives of interest
are not compromised. Depending on the noise and disturbance setting, the performance objective can be UGAS or
L p -stability (with a prespecified gain and towards the output of interest). Our framework incorporates time-varying
delays and transmission intervals that can be smaller than
the delays, plants/controllers with delayed dynamics, external disturbances (or modeling uncertainties), UGES schedul˙
ŷ˙ = Bŷ(t − d(t))(1 − d(t))
= B e(t − d(t))+
ing protocols (e.g., RR and TOD protocols), distorted data
!
and model-based estimators. As expected, the TOD proto1 0
0 0
˙
col results in greater MATIs than the RR protocol. Likewise,
+
x(t − d(t)) +
x(t − 2d(t)) (1 − d(t)),
0 0
0 1
estimation (rather than the ZOH strategy) in between two conwhich can be employed in any of the three performance objec- secutive transmission instants extends the MATIs.
The primary goal of our future work is to devise conditions
tives (i.e., UGAS, L p -stability or L p -stability with a desired
gain) provided d(t) is known. One can use the ideas from rendering L p -stability of the error dynamics involving several
time-varying delays (see Remark 4). In addition, in light of
Remark 4 towards obtaining known delays.
Figures 3 and 4 provide evidence that the TOD protocol [7], we plan to design event- and self-triggered realizations of
results in greater MATIs (at the expense of additional imple- our approach.
mentation complexity/costs) and that the model-based estimators significantly prolong MATIs, when compared with the
ZOH strategy. In addition, notice that the case d˘ = 0 boils Appendix
down to ODE modeling so that we can employ less conservative tools for computing L2 -gains. Apparently, MATIs per- 6.1 Proof of Lemma 1
taining to UGAS are greater than the MATIs pertaining to L p stability from ω to (x, e) and these are greater than the MATIs This proof follows the exposition in [34]. The following two
pertaining to L p -stability from ω to x with γdes = 15. As ex- definitions regarding (1) are utilized in this proof and are
pected, time-varying delays upper bounded with some d˘ lead taken from [34].
to smaller MATIs when compared to constant delays d.˘ It is
Definition 7 (Lyapunov Function). The function V : [t0 , ∞) ×
worth mentioning that d˘ = 33 ms is the maximal value for
Rnξ → R+ is said to belong to the class ν0 if we have the
which we are able to establish condition (a) of Theorem 2.
following:
Consequently, the delays from Figures 3 and 4 are instances
of admissible delays. The exhaustive search for admissible
1. V is continuous in each of the sets [tk−1 ,tk ) × Rnξ , and
delays is an open problem that is out of scope of this article.
for each ξ ∈ Rnξ and each t ∈ [t ,t ), where k ∈ N, the
k−1 k
limit lim(t,y)→(t − ,x) V (t, y) = V (tk− , x) exists;
6
k
Conclusion
2. V (t, ξ) is locally Lipschitz in all ξ ∈ Rnξ ; and
In this article, we study how much information exchange between a plant and controller can become intermittent (in terms
3. V (t, 0) ≡ 0 for all t ≥ t0 .
10
25
MATI [ms]
20
15
UGAS/no estim.
Lp-stab./no estim.
des. gain/no estim.
UGAS/with estim.
Lp-stab./with estim.
des. gain/with estim.
UGAS/no estim.
Lp-stab./no estim.
des. gain/no estim.
UGAS/with estim.
Lp-stab./with estim.
des. gain/with estim.
Try-Once-Discard (TOD)
40
35
30
25
MATI [ms]
UGAS/no estim.
Lp-stab./no estim.
des. gain/no estim.
UGAS/with estim.
Lp-stab./with estim.
des. gain/with estim.
UGAS/no estim.
Lp-stab./no estim.
des. gain/no estim.
UGAS/with estim.
Lp-stab./with estim.
des. gain/with estim.
Round Robin (RR)
30
20
15
10
10
5
5
0
0
0
5
10
15
20
25
0
30
5
10
15
20
25
30
delay d [ms]
delay d [ms]
(a)
(b)
Figure 3: Numerically obtained MATIs for various constant delay values d ≥ 0 in scenarios with and without estimation: (a) RR; and, (b)
TOD.
˙ ≤ 0.5
Round Robin (RR), |d|
˙ ≤ 0.5
Try-Once-Discard (TOD), |d|
16
UGAS/no estim.
Lp-stab./no estim.
des. gain/no estim.
UGAS/with estim.
Lp-stab./with estim.
des. gain/with estim.
14
12
20
15
MATI [ms]
10
MATI [ms]
UGAS/no estim.
Lp-stab./no estim.
des. gain/no estim.
UGAS/with estim.
Lp-stab./with estim.
des. gain/with estim.
8
6
10
4
5
2
0
0
5
10
15
20
25
0
30
0
delay upper bound [ms]
5
10
15
20
25
30
delay upper bound [ms]
(a)
(b)
˙
Figure 4: Numerically obtained MATIs for various time-varying delays d(t) such that d(t) ≤ d˘ and |d(t)|
≤ d˘1 = 0.5 in scenarios with and
without estimation: (a) RR; and, (b) TOD.
11
Definition 8 (Upper Dini Derivative). Given a function V : Let us now take s to be a function of time, that is, s := s(t).
˘ 0] for all
[t0 , ∞) × Rnξ → R+ , the upper right-hand derivative of V From (29) and (36) with s := s(t) = −d(t) ∈ [−d,
with respect to system (1) is defined by D+V (t, ξ(t)) = t ∈ [t ∗∗ ,t ∗ ], we obtain
lim supδ&0 1δ [V (t + δ, ξ(t + δ)) −V (t, ξ(t))].
D+V (t, ξ(t)) ≤ r + λ1 Me−λ(t1 −t0 ) V (t, ξ(t)),
Proof. We prove this theorem employing mathematical induction. Consider the following Lyapunov function for (14)
for all t ∈ [t ∗∗ ,t ∗ ]. Recall that 0 < t ∗ −t ∗∗ ≤ t1 −t0 ≤ τ. Having
with ũ ≡ 0, ν̃ ≡ 0:
that said, it follows from (31), (34) and (35) that
V (t, ξ(t)) = rξ(t)2 .
(28)
−λ(t1 −t0 ) )(t ∗ −t ∗∗ )
V (t ∗ , ξ(t ∗ )) ≤ V (t ∗∗ , ξ(t ∗∗ ))e(r+λ1 Me
Using 2ab ≤ a2 + b2 , a, b ∈ R in what follows, we obtain
−λ(t1 −t0 ) )(t ∗ −t ∗∗ )
= rkξt0 k2 e(r+λ1 Me
D+V (t,ξ(t)) ≤ 2rξ(t)aξ(t − d(t))
2
2
2
−λτ )τ
≤ rkξt0 k2 e(r+λ1 Me
2
≤r ξ(t) + a ξ(t − d(t))
≤rV (t, ξ(t)) + λ1V (t − d(t), ξ(t − d(t))),
≤ rMkξt0 k2 e−λ(t1 −t0 ) = V (t ∗ , ξ(t ∗ )),
(29)
which is a contradiction. Hence, (32) holds, i.e., (30) holds
over [t0 ,t1 ).
It is now left to show that (30) holds over [tk−1 ,tk ) for each
k ∈ N, k ≥ 2. To that end, assume that (30) holds for each
k ∈ {1, 2, . . . , m}, where m ∈ N, i.e.,
along the solutions of (14) with ũ, ν̃ ≡ 0, for each t 6∈ T . In
what follows, we are going to show that
V (t, ξ(t)) ≤ rMkξt0 k2 e−λ(t−t0 ) ,
∀t ≥ t0 ,
(30)
where
M > eλτ ≥ eλ(t1 −t0 ) .
V (t, ξ(t)) ≤ rMkξt0 k2 e−λ(t−t0 ) ,
(31)
∀t ∈ [t0 ,t1 ),
V (t, ξ(t)) ≤ rMkξt0 k2 e−λ(t−t0 ) , ∀t ∈ [tm ,tm+1 ).
(32)
∀t ∈ [t0 ,t1 ),
(37)
(38)
For the sake of contradiction, suppose that (38) does not hold.
Then, we can define
n
o
t := inf t ∈ [tm ,tm+1 ) V (t, ξ(t)) > rMkξt0 k2 e−λ(t−t0 ) .
holds by showing that the basis of mathematical induction
V (t, ξ(t)) ≤ rMkξt0 k2 e−λ(t1 −t0 ) ,
∀t ∈ [t0 ,tk ),
for every k ∈ {1, 2, . . . , m}. Let us now show that (30) holds
over [tm ,tm+1 ) as well, i.e.,
One can easily verify that (I) implies (31). Notice that this
choice of M yields kξt0 k2 < Mkξt0 k2 e−λ(t1 −t0 ) .
According to the principle of mathematical induction, we
start showing that
V (t, ξ(t)) ≤ rMkξt0 k2 e−λ(t−t0 ) ,
(I)
< rMkξt0 k2 e−λτ
(33)
holds. For the sake of contradiction, suppose that (33) does From (14b) with ν̃ ≡ 0 and (37), we know that
not hold. From (28) and (31), we infer that there exists t ∈
(t0 ,t1 ) such that
V (tm+ , ξ(tm+ )) = rξ(tm+ )2 = rc2 ξ(tm )2 = λ2V (tm− , ξ(tm− ))
≤ λ2 rMkξt0 k2 e−λ(tm −t0 )
V (t, ξ(t)) > rMkξt0 k2 e−λ(t1 −t0 )
= λ2 rMkξt0 k2 eλ(t−tm ) e−λ(t−t0 )
˘ 0],
> rkξt0 k2 ≥ V (t0 + s, ξ(t0 + s)), ∀s ∈ [−d,
< λ2 reλ(tm+1 −tm ) Mkξt0 k2 e−λ(t−t0 )
which implies that there exists t ∗ ∈ (t0 ,t) such that
< rMkξt0 k2 e−λ(t−t0 ) ,
V (t ∗ , ξ(t ∗ )) = rMkξt0 k2 e−λ(t1 −t0 ) ,
V (t, ξ(t)) ≤ V (t ∗ , ξ(t ∗ )),
˘ ∗ ],
∀t ∈ [t0 − d,t
and there exists t ∗∗ ∈ [t0 ,t ∗ ) such that
V (t ∗∗ , ξ(t ∗∗ )) = rkξt0 k2 ,
V (t ∗∗ , ξ(t ∗∗ )) ≤ V (t, ξ(t)),
∀t ∈ [t ∗∗ ,t ∗ ].
λ(tm+1 −tm ) ≤ λ eλτ < 1. One
2
(34) where λ2 ∈ (0, 1) is such that λ2 e
can easily verify that (II) implies λ2 eλτ < 1. Another fact to
notice is that t 6= tm . Employing the continuity of V (t, ξ(t))
over the interval [tm ,tm+1 ), we infer
V (t, ξ(t)) = rMkξt0 k2 e−λ(t−t0 ) ,
(35)
V (t, ξ(t)) ≤ V (t, ξ(t)),
˘ 0] we have
Using (34) and (35), for any s ∈ [−d,
= Me−λ(t1 −t0 )V (t ∗∗ , ξ(t ∗∗ ))
≤ Me
V (t, ξ(t)),
(39)
In addition, we know that there exists t ∗ ∈ (tm ,t) such that
V (t + s, ξ(t + s)) ≤ rMkξt0 k2 e−λ(t1 −t0 )
−λ(t1 −t0 )
∀t ∈ [tm ,t].
V (t ∗ , ξ(t ∗ )) = λ2 reλ(tm+1 −tm ) Mkξt0 k2 e−λ(t−t0 ) ,
∗∗
∗
∀t ∈ [t ,t ].
V (t ∗ , ξ(t ∗ )) ≤ V (t, ξ(t)) ≤ V (t, ξ(t)),
(36)
12
∀t ∈ [t ∗ ,t].
(40)
˘ 0], Theorem 3 (Riesz-Thorin Interpolation Theorem). Let
We proceed as follows, for any t ∈ [t ∗ ,t] and any s ∈ [−d,
˘ m ) or t + s ∈ [tm ,t]. If t + s ∈ [t0 − F : Rn → Rm be a linear operator and suppose that
then either t + s ∈ [t0 − d,t
˘ m ), then from (37) we have
d,t
p0 , p1 , q0 , q1 ∈ [1, ∞] satisfy p0 < p1 and q0 < q1 . For any θ ∈
[0, 1] define pθ , qθ by 1/pθ = (1 − θ)/p0 + θ/p1 and 1/qθ =
θ
V (t + s,ξ(t + s)) ≤ rMkξt0 k2 e−λ(t+s−t0 )
(1 − θ/q0 ) + θ/q1 . Then, kFk pθ →qθ ≤ kFk1−θ
p0 →q0 kFk p1 →q1 ,
where kFk p·→q· denotes the norm of the mapping F between
= rMkξt0 k2 e−λ(t−t0 ) e−λs
the
L p· and Lq· space. In particular, if kFk p0 →q0 ≤ M0 and
˘
≤ rMkξt0 k2 e−λ(t−t0 ) eλ(t−t) eλd
kFk p1 →q1 ≤ M1 , then kFk pθ →qθ ≤ M01−θ M1θ .
˘
≤ reλd eλ(tm+1 −tm ) Mkξt0 k2 e−λ(t−t0 ) .
If t + s ∈ [tm ,t], then from (39) we have
(41) Proof. From the UGES assumption of the theorem, we infer
that the fundamental matrix Φ(t,t0 ) satisfies:
kΦ(t,t0 )k ≤
V (t + s, ξ(t + s)) ≤ rMkξt0 k2 e−λ(t−t0 )
√ − λ (t−t )
Me 2 0 ,
∀t ≥ t0 ,
˘
≤ reλd eλ(tm+1 −tm ) Mkξt0 k2 e−λ(t−t0 ) .
(42) uniformly in t0 . Refer to [1, Definition 3.] for the exact definition of a fundamental matrix. Now, [1, Theorem 3.1.] proApparently, the upper bounds (41) and (42) are the same; vides
˘ m ) or t + s ∈
Z t
hence, it does not matter whether t + s ∈ [t0 − d,t
ξ(t) = Φ(t,t0 )ξ(t0 ) + Φ(t, s)ũ(s)ds
[tm ,t]. Therefore, from (40) and this upper bound, we have for
t0
˘ 0]
any s ∈ [−d,
Z t
+ Φ(t, s)aξs (−d(t))ds
˘
˘
t0
eλd
eλd
∗
∗
V (t + s, ξ(t + s)) ≤
V (t , ξ(t )) ≤
V (t, ξ(t)), (43)
+ ∑ Φ(t,ti )ν̃(ti ),
∀t ≥ t0 ,
(44)
λ2
λ2
t0 <ti ≤t
for all t ∈ [t ∗ ,t]. Once more, let us take s to be a function of
˘ 0] for all t ∈ [t ∗ ,t]. Now, where ξs (−d(t)) = 0 when s − d(t) ≥ t0 . The above equality
time, that is, s := s(t) = −d(t) ∈ [−d,
along with
from (29) and (43), we have
D+V (t, ξ(t)) ≤ r +
λ1 λd˘
e V (t, ξ(t)),
λ2
∗
∗
∑
t0 <ti ≤t
Recall that 0 < t − t ∗ ≤ tm+1 − tm ≤ τ. Accordingly, we reach
(r+ λ1 eλd )(t−t ∗ )
V (t, ξ(t)) ≤ V (t , ξ(t ))e
2
˘
λ
(r+ λ1 eλd )(t−t ∗ )
λ
˘
1 λd
λτ (r+ λ2 e )τ
≤ λ2 e e
(II)
< rMkξt0 k e
Z t
2
+
t0
rMkξt0 k2 e−λ(t−t0 )
2 −λ(t−t0 )
+
≤
−1
Z t
kΦ(t, s)kkũ(s)kds
t0
kΦ(t, s)k|a|kξs (−d(t))kds
∑
kΦ(t,ti )kkν̃(ti )k
Z
√
+ |a| Mkξt0 k
The following two well-known results can be found in, for
example, [24].
t0 +d(t)
λ
e− 2 (t−s) ds+
t0
√
+ K̃ν M
1
e
∀t ≥ t0 .
Proof of Theorem 1
√ − λ (t−t )
√ Zt λ
Me 2 0 kξt0 k + M e− 2 (t−s) kũ(s)kds
t0
which is a contradiction; hence, (30) holds over [tm ,tm+1 ).
Employing mathematical induction, one immediately infers
that (30) holds over [tk−1 ,tk ) for each k ∈ N. From (28) and
(30), it follows that
6.2
e
λε
2
t0 <ti ≤t
= V (t, ξ(t)),
√
λ
kξ(t)k ≤ Mkξt0 ke− 2 (t−t0 ) ,
i=1
kξ(t)k ≤ kΦ(t,t0 )kkξt0 k +
= λ2 reλ(tm+1 −tm ) Mkξt0 k2 e−λ(t−t0 ) e
1
λ
e− 2 (t−ti ) ≤ ∑ e− 2 εi =
immediately yields
˘
λ
∞
λ
∀t ∈ [t ∗ ,t].
λε
2
∀t ≥ t0 ,
,
−1
(45)
uniformly in t0 .
Let us now estimate the contribution of the initial condition
ξt0 towards kξ[t0 ,t]k p by setting ũ ≡ 0 and K̃ν = 0. In other
words, we have
√
λ
kξ(t)k ≤ Me− 2 (t−t0 ) kξt0 k
Z
√
+ |a| Mkξt0 k
t0 +d˘
λ
(46)
e− 2 (t−s) ds, ∀t ≥ t0 .
Lemma 2 (Young’s Inequality). Let ∗ denote convolution
t0
over an interval I, f ∈ L p [I] and g ∈ Lq [I]. The Young’s
inequality is k f ∗ gkr ≤ k f k p kgkq for 1r = 1p + 1q − 1 where In what follows, we use (a + b) p ≤ 2 p−1 a p + 2 p−1 b p and
1
1
1
p, q, r > 0.
(a + b) p ≤ a p + b p , where a, b ≥ 0 and p ∈ [1, ∞) (see, for
13
example, [9, Lemma 1 & 2]). Raising (46) to the pth ∈ [1, ∞) for all p ∈ [1, ∞].
power, integrating over [t0 ,t] and taking the pth root yields
Finally, summing up the contributions of ξt0 , ũ(t) and ν̃(t)
!
produces
√
√ 2 dλ˘
p−1
!
kξ[t0 ,t]k p ≤
M + |a| M e 2 − 1 2 p ×
√
1 1p
λ
2 dλ˘
kξt0 k
kξ[t0 ,t]k p ≤ 2 M 1 + |a| e 2 − 1
2 1
λ
pλ
p
×
kξt0 k,
∀t ≥ t0 ,
2√
pλ
+
Mkũ[t0 ,t]k p + kb[t0 ,t]k p ,
∀t ≥ t0 ,
λ
where we used
for any p ∈ [1, ∞].
!1
Z ∞ Z t0 +d˘
p
p
2 1p
2 dλ˘
− λ2 (t−s)
= e 2 −1
.
e
ds dt
λ
pλ
t0
t0
6.3
(47)
Proof of Theorem 2
Proof. Combining (i) of UGES protocols and (16), one obtains:
1
D ∂W (i, e)
E a
p−1 2 p
lim 2 p
= 2.
,
g(t,
x
,
e
,
ω
)
≤ LW ( j, e(t − d(t)))
t
t
t
p→∞
pλ
∂e
a
+ kH(xt , ωt )k.
(52)
Let us now estimate the contribution of the input ũ(t) towards kξ[t0 ,t]k p by setting kξt0 k = 0 and K̃ν = 0. In other
for any i, j ∈ N. Hence, the index i in W (i, e) can be omitted
words, we have
in what follows. Now, we define Z(t) := W (e(t)) and reach
√ Z t − λ (t−s)
2
kũ(s)kds,
∀t ≥ t0 .
(48)
kξ(t)k ≤ M e
dZ(t) a
t0
≤ LZ(t − d(t)) + kH(xt , ωt )k,
(53)
dt
a
R ∞ −λs
2
Note that 0 e 2 ds = λ . Now, integrating the previous in/ T . For a justification of the transition from
equality over [t0 ,t] and using Lemma 2 with p = q = r = 1 for almost all t ∈
(52) to (53), refer to [20, Footnote 8]. Likewise, property (ii)
yields the L1 -norm estimate:
of UGES protocols yields
2√
kξ[t0 ,t]k1 ≤
Mkũ[t0 ,t]k1 ,
∀t ≥ t0 .
(49)
λ
Z(t + ) ≤ ρZ(t) + aν j (t),
(54)
Taking the max over [t0 ,t] in (48) and using Lemma 2 with
for all t ∈ T , where ν j (t), j ∈ {1, . . . , l}, is the jth NCS
q = r = ∞ and p = 1 yields the L∞ -norm estimate:
link noise given by (7) and upper bounded with Kν . Notice
2√
Mkũ[t0 ,t]k∞ ,
∀t ≥ t0 .
(50) that |Z(t)| = |W (e(t))|. Next, we use the comparison lemma
kξ[t0 ,t]k∞ ≤
λ
for impulsive delayed systems [16, Lemma 2.2]. Basically,
From (44), one infers that we are dealing with a linear op- the fundamental matrix of (53)-(54) is upper bounded with
a
erator, say F, that maps ũ to e with bounds for the norms the fundamental matrix of (14) with parameters a := a L and
kFk1 ≤ kFk∗1 and kFk∞ ≤ kFk∗∞ , where kFk∗1 and kFk∗∞ are c := ρ. Refer to [1, Definition 3.] for the exact definition of a
given by (49) and (50), respectively. Because kFk∗1 = kFk∗∞ , fundamental matrix. Of course, the corresponding transmisTheorem 3 gives that kFk p ≤ kFk∗1 = kFk∗∞ for all p ∈ [1, ∞]. sion interval τ in (14b), and therefore in (54), has
√ to allow
2
for
M
>
1
and
λ
>
0
that
satisfy
(I),
(II)
and
This yields
λ MγH < 1
(as stated in Theorem 2). Essentially, (I) √
and (II) yield L p 2√
2
kξ[t0 ,t]k p ≤
Mkũ[t0 ,t]k p ,
∀t ≥ t0 ,
(51) stability with bias from H to W , while λ MγH < 1 allows
λ
us to invoke the small-gain theorem. Following the proof of
Theorem 1, one readily establishes L p -stability from H to W
for any p ∈ [1, ∞].
with
bias, i.e.,
Let us now estimate the contribution of the noise ν̃(t) towards kξ[t0 ,t]k p by setting kξt0 k = 0 and ũ ≡ 0. In other
kW [t0 ,t]k p ≤ KW kWt0 k + γW kH[t0 ,t]k p + kb[t0 ,t]k p , (55)
words, we have
When p = ∞, simply take the limit
√
kξ(t)k ≤ K̃ν M
1
,
−1
√
By identifying b(t) ≡ b := K̃ν M
e
for any t ≥ t0 any p ∈ [1, ∞], where
∀t ≥ t0 .
λε
2
1
λε
, we immediately ob-
e 2 −1
tain
kξ[t0 ,t]k p ≤ kb[t0 ,t]k p ,
∀t ≥ t0 ,
14
!
√
1 1p
a 2 dλ˘
KW := 2 M 1 + L e 2 − 1
,
a λ
pλ
√
2√
aKν M
γW :=
M, b := λε
.
λ
e 2 −1
where K1 := K p .
First, we need to establish UGS of the interconnection Σn
and Σe when ω ≡ 0nω . Before we continue, note that the jumps
kH[t,t0 ]k p ≤KH kxt0 k + γH kW [t,t0 ]k p + γH kω[t0 ,t]k p , (56) in (12) and (13) are such that k(x, e)(t + )k ≤ k(x, e)(t)k for
for all t ≥ t0 . Combining the above with (55) and property (i) each t ∈ T . Apparently, jumps are not destabilizing and can
be disregarded in what follows. Along the lines of the proof
of UGES protocols yields
for [23, Theorem 1], we pick any ε > 0. Let K be the set
aKW /a
γW KH /a
{(x, e) 2ε ≤ k(x, e)k ≤ ε}, K1 be the set {(x, e) k(x, e)k ≤ ε},
ke[t0 ,t]k p ≤
ket0 k +
kxt0 k
(1 − γW γH )
(1 − γW γH )
and take
1/a
γW γH /a
a :=
sup
k f (t, xt , e, 0nω ), g(t, xt , et , 0̃nω ) k,
kω[t0 ,t]k p +
kb[t0 ,t]k p , (57)
+
(1 − γW γH )
(1 − γW γH )
t∈R
Let us now infer L p -stability with bias from ω to (H, e) via
the small-gain theorem. Inequality (15) implies
(x,e)(t)∈K
(x,e)t ∈PC([−d,0],K1 )
KH
aKH KW
kxt0 k +
ket k
1 − γW γH
1 − γW γH 0
γH
γH
kω[t0 ,tk p +
kb[t0 ,tk p .
+
1 − γW γH
1 − γW γH
kH[t0 ,t]k p ≤
where f (t, xt , e, 0nω ) and g(t, xt , et , 0̃nω ) are given by (10) and
(11), respectively. This supremum exists because the underlying dynamics are Lipschitz uniformly in t. Next, choose
0 < δ < 2ε such that r < δ implies K1 k(xt0 , et0 )k p < s0 , where
From the above two inequalities, L p -stability from ω to (H, e) s := ε(ε/2) p . Let k(x , e )k < δ. Then k(x, e)(t)k < ε
t0 t0 d
0
γ
√
2a
γH (1+ W )
γH + 1
with bias 1−γW γaH aKλεν M and gain 1−γW γaH is immediately ob- for all t ≥ t0 . Indeed, suppose that there exists some t > t0
e 2 −1
such that k(x, e)(t)k ≥ ε. Then, there is an interval [t1 ,t2 ]
tained.
such that k(x, e)(t1 )k = 2ε , k(x, e)(t2 )k = ε, (x, e)(t) ∈ K ,
(x, e)t ∈ PC([−d, 0], K1 ), for all t ∈ [t1 ,t2 ]. Hence,
6.4 Proof of Corollary 1
Z ∞
Z t2
ε p
p
p
.
k(x,
e)(s)k
ds
≥
k(x,
e)(s)k
ds
≥
(t
−
t
)
2
1
Proof. The L p -detectability of x from (W, ω, H) implies that
2
t0
t1
there exist Kd , γd ≥ 0 such that
On the other hand,
kx[t0 ,t]k p ≤Kd kxt0 k + γd kH[t0 ,t]k p + γd k(W, ω)[t0 ,t]k p
ε
≤ k(x, e)(t2 ) − (x, e)(t1 )k
≤Kd kxt0 k + γd kH[t0 ,t]k p + γd kW [t0 ,t]k p
2
Z t2
+γd kω[t0 ,t]k p
(58)
≤
k( f (t, x , e, 0 ), g(t, x , e , 0̃ ))kds
t
t1
for all t ≥ t0 . Plugging (56) into (58) leads to
nω
t
t
nω
≤ a(t2 − t1 ),
kx[t0 ,t]k p ≤ Kd kxt0 k + γd KH kxt0 k + γd γH kW [t,t0 ]k p
and, combining the above with (59), we conclude that
+ γd γH kω[t0 ,t]k p + γd kW [t0 ,t]k p + γd kω[t0 ,t]k p
s0 ≤
≤ Kd kxt0 k + γd KH kxt0 k + (aγd γH + aγd )ke[t,t0 ]k p
Z ∞
t0
k(x, e)(s)k p ds ≤ K1 k(xt0 , et0 )k p ,
+ (γd γH + γd )kω[t0 ,t]k p
which is a contradiction.
Second, let us show asymptotic convergence of k(x, e)(t)k
for all t ≥ t0 . Finally, we include (57) into (58) and add the
obtained inequality to (57) which establishes L p -stability with to zero. Using (59), we infer that (x, e)(t) ∈ L p . Owing to
the Lipschitz dynamics of the corresponding system, we conbias from ω to (x, e).
clude that (ẋ, ė)(t) ∈ L p as well. Now, one readily establishes
asymptotic convergence of kxk to zero using [26, Facts 1-4].
6.5 Proof of Proposition 1
Consequently, asymptotic convergence of kek to zero follows
Proof. For the case p = ∞, UGS of the interconnection Σn and from (45) by observing that ξ(t) and ũ(t) in (45) correspond to
Σe is immediately obtained using the definition of L∞ -norm. W (t) and H(t), respectively, and using (i) of Definition 6.
Therefore, the case p ∈ [1, ∞) is more interesting. From the
conditions of the proposition, we know that there exist K ≥ 0
References
and γ ≥ 0 such that
k(x, e)[t0 ,t]k p ≤ Kk(xt0 , et0 )k + γkω[t0 ,t]k p , ∀t ≥ t0 .
Recall that ω ≡ 0nω when one is interested in asymptotic stability. By raising both sides of the above inequality to the pth
power, we obtain:
Z t
t0
p
p
k(x, e)(s)k ds ≤ K1 k(xt0 , et0 )k ,
∀t ≥ t0 ,
[1] A. Anokhin, L. Berezansky, and E. Braverman. Exponential stability of linear delay impulsive differential
equations. Journal of Mathematical Analysis and Applications, 193(3):923–941, 1995.
[2] G. H. Ballinger. Qualitative Theory of Impulsive Delay
Differential Equations. PhD thesis, Univ. of Waterloo,
Canada, 1999.
(59)
15
[3] D. P. Borgers and W. P. M. H. Heemels. Stability analysis of large-scale networked control systems with local networks: A hybrid small-gain approach. In Hybrid
Systems: Computation and Control (HSCC), pages 103–
[16]
112, 2014.
[4] S. Boyd, L. El-Ghaoui, E. Feron, and V. Balakrishnan.
Linear matrix inequalities in systems and control theory.
SIAM, Philadelphia, 1994.
exponential stabilization through communication network. IEEE Trans. on Automatic Control, 20(4):887–
900, 2012.
X. Liu, X. Shen, and Y. Zhang. A comparison principle and stability for large-scale impulsive delay differential systems. Anziam journal: The Australian &
New Zealand industrial and applied mathematics journal, 47(2):203–235, 2005.
[5] D. Christmann, R. Gotzhein, S. Siegmund, and F. Wirth. [17] M. H. Mamduhi, D. Tolić, A. Molin, and S. Hirche.
Event-triggered scheduling for stochastic multi-loop
Realization of Try-Once-Discard in wireless multihop
networked control systems with packet dropouts. In
networks. IEEE Transactions on Industrial Informatics,
Proceedings of the IEEE Conference on Decision and
10(1):17–26, Feb 2014.
Control, pages 2776–2782, Los Angeles, CA, Decem[6] D. F. Coutinho and C. E. de Souza. Delay-dependent rober 2014.
bust stability and L2 -gain analysis of a class of nonlinear time-delay systems. Automatica, 44(8):2006–2018, [18] N. C. Martins. Finite gain l p stability requires analog
control. Systems and Control Letters, 55(11):949–954,
2008.
2006.
[7] V. S. Dolk, D. P. Borgers, and W. P. M. H. Heemels. Dynamic event-triggered control: Tradeoffs between trans- [19] F. Mazenc, M. Malisoff, and T. N. Dinh. Robustness of
nonlinear systems with respect to delay and sampling of
mission intervals and performance. In Proceedings of
the controls. Automatica, 49(6):1925–1931, 2013.
the IEEE Conference on Decision and Control, pages
2764–2769, Los Angeles, CA, December 2014.
[20] D. Nešić and A. R. Teel. Input-output stability properties
of Networked Control Systems. IEEE Transactions on
[8] T. Estrada and P. J. Antsaklis. Model-based control with
Automatic Control, 49(10):1650–1667, October 2004.
intermittent feedback: Bridging the gap between continuous and instantaneous feedback. Int. J. of Control, [21] S.-I. Niculescu. Delay Effects on Stability: A Ro83(12):2588–2605, 2010.
bust Control Approach. Number 269 in Lecture Notes
in Control and Information Sciences. Springer-Verlag,
∞
[9] R. A. Freeman. On the relationship between induced l London, 2001.
gain and various notions of dissipativity. In Proceedings
of the IEEE Conference on Decision and Control, pages [22] H. Smith. An Introduction to Delay Differential Equa3430–3434, 2004.
tions with Applications to the Life Sciences, volume 57
of Texts in Applied Mathematics. Springer, New York,
[10] J. K. Hale. Theory of functional differential equations,
2011.
volume 3 of Applied Mathematical Sciences. SpringerVerlag, New York, 1977.
[23] E. D. Sontag. Comments on integral variants of ISS.
Systems & Control Letters, 34(1âĂŞ2):93 – 100, 1998.
[11] W. P. M. H. Heemels, A. R. Teel, N. V. de Wouw, and
D. Nešić. Networked Control Systems with communica- [24] M. Tabbara, D. Nešić, and A. R. Teel. Stability of
wireless and wireline networked control systems. IEEE
tion constraints: Tradeoffs between transmission interTransactions on Automatic Control, 52(9):1615–1630,
vals, delays and performance. IEEE Tran. on Automatic
September 2007.
Control, 55(8):1781–1796, 2010.
[12] J. P. Hespanha, P. Naghshtabrizi, and X. Yonggang. [25] P. Tallapragada and N. Chopra. On event triggered trajectory tracking for control affine nonlinear systems. In
A survey of recent results in Networked Control SysProceedings of the IEEE Conference on Decision and
tems. Proceedings of the IEEE, 95(1):138 – 162, JanControl, pages 5377–5382, December 2011.
uary 2007.
[13] Z. P. Jiang, A. R. Teel, and L. Praly. Small-gain theorem for ISS systems and applications. Mathematics of
Control, Signals and Systems, 7(2):95–120, 1994.
[26] A. R. Teel. Asymptotic convergence from L p stability.
IEEE Transactions on Automatic Control, 44(11):2169–
2170, 1999.
[14] H. Khalil. Nonlinear Systems. Prentice Hall, 3rd edition, [27] D. Tolić and S. Hirche. Stabilizing transmission intervals and delays for nonlinear networked control systems:
2002.
The large delay case. In Proceedings of the IEEE Con[15] A. Kruszewski, W. J. Jiang, E. Fridman, J. P. Richard,
ference on Decision and Control, pages 1203–1208, Los
and A. Toguyeni. A switched system approach to
Angeles, CA, December 2014.
16
[28] D. Tolić and S. Hirche. Stabilizing transmission intervals for nonlinear delayed networked control systems.
IEEE Trans. on Automatic Control, April 2017. to appear.
[29] D. Tolić, R. Sanfelice, and R. Fierro. Self-triggering in
nonlinear systems: A small-gain theorem approach. In
IEEE Mediterranean Conf. Contr. and Automat., pages
935–941, Barcelona, Spain, 2012.
[30] D. Tolić, R. G. Sanfelice, and R. Fierro. Input-output
triggered control using Lp-stability over finite horizons.
Int. J. Robust and Nonlinear Control, 25(14):2299–
2327, 2015.
[31] J. Tsinias. A theorem on global stabilization of nonlinear
systems by linear feedback. Systems & Control Letters,
17(5):357–362, 1991.
[32] G. C. Walsh, H. Ye, and L. G. Bushnell. Stability analysis of Networked Control Systems. IEEE Transactions
on Control Systems Technology, 10(3):438–446, 2002.
[33] B. Zhang, A. Kruszewski, and J. P. Richard. A novel
control design for delayed teleoperation based on delayscheduled Lyapunov-Krasovskii functionals. International Journal of Control, 87(8):1694–1706, 2014.
[34] J. Zhou and Q. Wu. Exponential stability of impulsive
delayed linear differential equations. IEEE Transactions
on Circuits and Systems – II: Express Briefs, 56(9):744–
748, September 2009.
17
| 3cs.SY
|
Near-linear Time Algorithms for Approximate Minimum Degree
Spanning Trees
Ran Duan
∗
Haoqing He
†
Tianyi Zhang
‡
arXiv:1712.09166v1 [cs.DS] 26 Dec 2017
Abstract
Given a graph G = (V, E), n = |V |, m = |E|, we wish to compute a spanning tree whose
maximum vertex degree is as small as possible. Computing the exact optimal solution is known
to be NP-hard, since it generalizes the Hamiltonian path problem. For the approximation version
of this problem, a Õ(mn) time algorithm that computes a spanning tree of degree at most ∆∗ +1
is previously known [Fürer, Raghavachari 1994]; here ∆∗ denotes the optimal tree degree. In
this paper we give the first near-linear time algorithm for this problem. Specifically speaking,
we first propose a simple Õ(m) time algorithm that achieves an O(∆∗ log n) approximation;
then we further improve this algorithm to obtain a (1 + δ)∆∗ + O( δ12 log n) approximation in
Õ( δ16 m) time.
∗
Institute for Interdisciplinary Information Sciences, Tsinghua University, [email protected]
Institute for Interdisciplinary Information Sciences, Tsinghua University, [email protected]
‡
Institute for Interdisciplinary Information Sciences, Tsinghua University, [email protected]
†
1
Introduction
Computing minimum degree spanning trees is a fundamental problem that has inspired a long line
of research. Let G = (V, E) be an undirected graph, and we wish to compute a spanning tree of G
whose tree degree, or maximum vertex degree in the tree, is the smallest. Clearly this problem is
NP-hard as the Hamiltonian path problem can be reduced to it, and so we could only hope for a
good approximation in polynomial time. The optimal approximation of this problem was achieved
in [6] where the authors proposed an 1 Õ(mn) time algorithm that computes a spanning tree of
tree degree ≤ ∆∗ + 1; conventionally n = |V |, m = |E| and ∆∗ denotes the optimal tree degree.
However, polynomial time algorithms does not always mean efficient on large data sets, so finding
approximation algorithms of almost linear time is a very popular and important topic nowadays.
1.1
Our results
The major results of this paper are two near-linear time algorithms for minimum degree spanning trees in undirected graphs. These are the first near-linear time algorithms for this problem.
Formally we propose the following two theorems.
Theorem 1. There is an 2 O(mα(n) log2 n + n log3 n) time algorithm that computes a spanning tree
with tree degree O(∆∗ log n).
As in many algorithms of this problem such as [6], this algorithm iteratively improves the
spanning tree T by finding replacement edge connecting two low-degree vertices. To achieve almost
linear time, we fix a degree threshold k ≥ 3, and repeatedly search for edges connecting two vertices
of tree degree ≤ k − 2 such that the tree path between its two endpoints contains a vertex of tree
degree ≥ k. We can efficiently maintain the spanning tree by the link-cut tree structure [16]. When
there are not many vertices of tree degree k − 1, we can argue a lower bound on ∆∗ in terms of k.
However, the algorithm may generate a large number of (k − 1)-degree vertices which undermines
the lower bound on ∆∗ . To circumvent such difficulties, we iteratively perform this procedure on
larger and larger k’s. If the number of vertices of degree ≥ k becomes smaller and smaller, we can
finally bound the number of (k −1)-degree vertices. The crucial observation is that if a vertex which
starts out as a low-degree vertex for previous k now becomes (k − 1)-degree, lots of high-degree
vertices must have lost some tree neighbours. By carefully selecting a series of threshold k’s, finally
we can argue a lower bound on ∆∗ or decrease the degree of T by a constant factor.
Theorem 2. For any constant δ ∈ (0, 16 ), there is an algorithm that runs in time O( δ16 m log7 n)
5
which computes a spanning tree with tree degree at most (1 + δ)∆∗ + 32δ
2 log n.
Theorem 2 refines Theorem 1’s approach by an augmenting path approach. In each iteration,
the algorithm conducts a series of tree modifications to remove all augmenting paths of the shortest
length, and so in the next iteration the shortest length of augmenting paths would increase. To
facilitate our search for shortest augmenting paths, we divide the graph into O( 1δ log n) layers and
then look for edges that connect two different tree components on the bottom layer. If such an
edge is successfully detected, then we add this edge to the tree and propagate a sequence of tree
edge insertions and deletions upwards to higher layers. When no such edges can be found, we argue
1
2
Õ(·) hides poly-logarithmic factors.
α(·) refers to the inverse Ackermann function.
1
that every layer can yield a lower bound on ∆∗ which jointly proves a lower bound with a constant
multiplicative error.
1.2
Related work
There is a line of works that are concerned with low-degree trees in weighted undirected graphs. In
this scenario, the target low-degree that we wish to compute is constrained by two parameters: an
upper bound B on tree degree, an upper bound C on the total weight summed over all tree edges.
The problem was originally formulated in [3]. Two subsequent papers [10, 11] proposed polynomial
w
bB + logb n, ∀b, w > 1. The
time algorithms that compute a tree with cost ≤ wC and degree ≤ w−1
cost was improved from wC to C in [2] while degree upper bound becomes bB + 2(b + 1) log b n; the
authors also proposed an quasi-polynomial algorithm that finds a tree with cost ≤ C and degree
B + O(log n/ log log n). [2]’s result was improved by [8] where for all k, a spanning tree of degree
≤ k + 2 and of cost at most the cost of the optimum spanning tree of maximum degree at most k
can be computed in polynomial time. The degree bound was later further improved from k + 2 to
the optimal k + 1 in [15].
Another variant is minimum degree Steiner trees which is related to network broadcasting
[13, 14, 4]. For undirected graphs, authors of [6] showed that the same approximation guarantee
and running time can be achieved as with minimum degree spanning trees in undirected graphs, i.e.,
a solution of tree degree ∆∗ +1 and a running time of Õ(mn). For the directed case, [4] showed that
directed minimum degree Steiner trees problem cannot be approximated within (1−ǫ) log |D|, ∀ǫ > 0
unless NP ⊆ DTIME(nlog log n ), where D is the set of terminals.
The minimum degree tree problem can also be formulated in directed graphs. This problem was
first studied in [5] where the authors proposed a polynomial time algorithm that finds a directed
spanning tree of degree at most O(∆∗ log n). The approximation guarantee was improved to roughly
O(∆∗ + log n) in [12, 9] while the time complexity became nO(log n) . The problem becomes much
easier when G is acyclic, as shown in [17], where a directed spanning tree of degree ≤ ∆∗ + 1 is
computable in polynomial time. The approximation was greatly advanced to ∆∗ + 2 in [1] by an
LP-based polynomial time algorithm, and this problem has become more-or-less closed since then.
2
Definitions
Logarithms are taken at base 2. Assume G = (V, E) is a connected graph, n = |V |, m = |E|, and
we assume n ≥ 4. During the execution of our algorithm, a spanning tree T will be maintained
and our algorithm will repeatedly modify T to reduce its degree ∆ = deg(T). For every u ∈ V , let
deg(u) be the tree degree of u. For each pair u, v ∈ V , let ρu,v be the unique tree path on T that
connects P
u and v. For each 1 ≤ k ≤ n, define Sk = {u | deg(u) ≥ k}, Nk = {u | deg(u) = k}, and
let dk = u∈Sk deg(u), that is, the total degree of all vertices of degree at least k.
3
3.1
An O(∆∗ log n) Approximation
Main algorithm
Starting from an arbitrary spanning tree T with degree ∆, the core of the main algorithm is a
near-linear time subroutine that, as long as ∆ ≥ 20 log n, either reduces ∆ to (1 − Ω(1)) · ∆ or
2
terminates with the guarantee that ∆ = O(∆∗ log n); the main algorithm simply repeatedly apply
this subroutine until ∆ < 20 log n or ∆ = O(∆∗ log n). This subroutine consists of two parts:
(1) a low-level fast degree reduction algorithm that, given any degree threshold k, modifies T to
reduce the total number of high-degree vertices; (2) a high-level scheduling algorithm that selects a
sequence of degree thresholds and feed them to the low-level degree reduction algorithm as inputs.
For every 0 ≤ i ≤ log n + 1, let us define a sequence of degree thresholds:
1
1 i
3
) )⌋
ki = ⌊ ∆ + ∆ · (1 − (1 −
4
4
log n
Clearly ki ’s are increasing as ∀0 ≤ i ≤ log n,
1
1 i 1
1 i+1
1
ki+1 − ki ≥ ∆ · (1 −
) − ∆ · (1 −
)
−1≥
·∆−1>0
4
log n
4
log n
16 · log n
The last two inequality holds as (1 − log1 n )i ≥ 1/4 and ∆ ≥ 20 log n.
The high-level scheduling algorithm (2) is described in Scheduling shown in Algorithm 1. If it
returns false, an upper bound ∆ = O(∆∗ log n) would be established; otherwise when it returns
true, it means ∆ would be reduced to (1 − Ω(1)) · ∆. The low-level degree reduction algorithm (1)
is described in FastDegreeReduction in Algorithm 2. The rough idea is that we repeatedly looks for
edges that connect two vertices of tree degree ≤ k − 2 from different components of T \ Sk and add
these edges to T, while at the same time we delete some edges incident on Sk so T stays a tree. In
order to implement this idea in near-linear time, we have to neglect those (k − 2)-degree vertices
that have once become (k − 1)-degree. A key operation of our algorithm is marking. During one
execution of FastDegreeReduction, a vertex gets marked whenever its tree degree becomes k − 1
and it stays so even if its degree goes smaller, and instead of searching for edges between two
(k − 2)-degree vertices, we only care about edges between two unmarked vertices.
Algorithm 1: Scheduling
7
input params: T with maximum tree degree ∆ ≥ 20 log n;
for i = 1, 2, · · · , 1 + log n do
count ← |Ski−1 |;
if |Ski | > 12 · count then
invoke FastDegreeReduction with input k = ki ;
if |Ski | > 12 · count then
return false;
8
return true;
1
2
3
4
5
6
3.2
Implementation and running time
We specify some implementation details of FastDegreeReduction.
(1) To efficiently implement line-4, we enumerate all edges (u, v) ∈ E one by one. Using the unionfind data structure [7], we check if one of u, v is marked or both of u, v belong to the same
component, we move on to the next edge; otherwise we execute line-5 through line-13. The
total running time of this part would be O(mα(n)).
3
Algorithm 2: FastDegreeReduction
1
2
3
4
5
6
7
8
9
10
11
12
13
14
input params: k;
for every u ∈
/ Sk , let Cu be the connected component containing u of T \ Sk ;
mark all vertices in Sk−1 and unmark all other vertices;
while ∃(u, v) ∈ E, both u, v are unmarked and belong to different components in T \ Sk do
find a vertex w ∈ Sk ∩ ρu,v , and (w, z) be an edge ∈ ρu,v which is incident on w;
T ← T ∪ {(u, v)} \ {(w, z)};
merge components Cu and Cv ;
for x ∈ {u, v} do
if deg(x) = k − 1 then
mark x;
for x ∈ {w, z} ∩ Sk do
if deg(x) = k − 1 then
remove x from Sk and merge tree components of T \ Sk accordingly;
return;
(2) For line-5, to efficiently retrieve a vertex w ∈ Sk ∩ ρu,v given u, v, we maintain T using the
link-cut tree data structure [16]. We set the weight of each u ∈ Sk to be 1, and weight of each
u∈
/ Sk equal to 0. Then w ∈ Sk ∩ ρu,v can be found in O(log n) amortized time by querying
the maximum weight vertex on the tree path ρu,v using the link-cut tree data structure. Note
that such Sk ∩ ρu,v is always non-empty because u, v belong to different connected components
of T \ Sk . For line-6, edge updates to T can be handled using the link-cut tree as well. Since
there are less than n components in T \ B, the total time would be O(n log n).
(3) On line-7, merging components Cu and Cv can be done in O(α(n)) time using the union-find
data structure.
(4) On line-13, when a vertex x is removed from Sk , we need to add x to T \ Sk and possibly merge
some connected components. This can be done by enumerating x’s incident tree edges and
using the union-find data structure. The total cost of such operations would be O(nα(n)).
To conclude, the overall running time complexity of FastDegreeReduction is by O(mα(n) +
n log n). The running time of Scheduling then becomes O(mα(n) log n+n log2 n) since i can increase
to at most O(log n).
To upper bound the running time of the main algorithm, we need the following lemma that
characterize the performance of Scheduling.
Lemma 3. If Scheduling returns true, then the degree ∆ of T has at least dropped by a constant
31
.
factor of 32
Proof. When Scheduling returns true, we claim that count declines by a factor of 21 after each
iteration of the while-loop. In fact, on the one hand, if the condition of line-5 does not hold, i.e.
Ski−1 ≥ 2|Ski |, then as count was previous set to |Ski−1 | and now equal to |Ski |, count declines
by a factor of 12 . On the other hand, if the condition of line-5 holds, then because Scheduling
4
returns true, the condition on line-7 always fails, i.e. |Ski | ≤ 12 · count, and therefore when we
set count ← |Ski | the value of count would decrease by a factor of 21 . Hence, the while-loop of
Scheduling can iterate for at most log n + 1 times before count becomes 0; that is to say, |Ski | = 0
for some i ≤ log n + 1. By definition,
3
ki ≤ ∆ +
4
3
≤ ∆+
4
3
≤ ∆+
4
Here we use the fact that (1 −
31
32 · ∆.
1
1 i
∆ · (1 − (1 −
))
4
log n
1
1 log n+1
∆ · (1 − (1 −
)
)
4
log n
1
1
31
∆ · (1 − ) =
·∆
4
8
32
1
log n+1
log n )
≥ 81 . As |Ski | = 0, deg(T) must now be smaller than
Now we can upper bound the running time of the main algorithm as stated in the following
lemma.
Lemma 4. The running time of the main algorithm is O(mα(n) log 2 n + n log3 n).
Proof. By Lemma 3, every invocation of Scheduling that returns true decreases ∆ by a factor of
31
32 . Therefore, such kind of invocations can be at most O(log n) many. Also, there can be at most 1
instance of Scheduling that returns false because the main algorithm terminates immediately after
that. Overall, the total running time of Scheduling would be O(log n · (mα(n) log n + n log2 n)) =
O(mα(n) log2 n + n log3 n).
3.3
Approximation guarantee
To prove approximation guarantee, we will utilize the following lemmas.
Lemma 5. Let V1 , V2 , · · · , Vl ⊆ V be disjoint vertex
S subsets. A set W is called “boundary” (with respect to V1 , V2 , · · · , Vl ), if any edge incident on li=1 Vi whose both endpoints are not simultaneously
l−1
contained in any single Vi , is incident on at least one vertex from W . Then, ∆∗ ≥ |W
|.
S
Proof. For any spanning tree, there are at least l−1 edges incident on li=1 Vi whose both endpoints
are not simultaneously contained in any Vi , 1 ≤ i ≤ l. Then by definition of W , any one of these
l − 1 edges is incident on at least on vertex of W , and thus by the pigeon-hole principle, there exists
l−1
a u ∈ W whose tree degree is ≥ |W
|.
Lemma 6. For any vertex subset B, the number of connected components in T \ B is at least
P
u∈B deg(u) − 2|B| + 2.
P
Proof. Note that there are at least u∈BP
deg(u)−|B|+1 tree edges incident on B, and so removing
all of these edges would break T into ≥ u∈B deg(u)
P − |B| + 2 components. Therefore, excluding
singleton components formed by B, there are ≥
u∈B deg(u) − 2|B| + 2 components are from
T \ B.
5
Now we prove when the main algorithm terminates, ∆ = O(∆∗ log n). If the main algorithm
terminates with ∆ < 20 log n, then automatically we have ∆ < 20 log n = O(log n) = O(∆∗ log n).
Next we focus on the case when the main algorithm terminates on a false returned by Scheduling. In
this case, there was an execution of Scheduling that returned on line-7. By the branching condition
of line-6, we know that |Ski | > 21 · count by the end of this execution of Scheduling.
Consider the most recent invocation of FastDegreeReduction. By the end of this invocation, let
V1 , V2 , · · · , Vl be the sequence of all different connected components spanned by T \ Sk . Let M
be the set of all marked vertices. To apply Lemma 5, we claim M is a boundary set with respect
to V1 , V2 , · · · , Vl ; this is because, for any edge (u, v) such that u, v belong to different connected
components of T \ Sk , one of u, v must be marked since otherwise FastDegreeReduction would
continue to merge Cu and Cv instead of terminating. Therefore, Lemma 5 immediately yields
l−1
∆∗ ≥ |M
|.
One last thing is to lower-bound l and upper-bound |M |.
(1) Lower bounding l.
Let Sk′ i−1 and Sk′ i be the snapshots of Ski−1 and Ski before this instance of FastDegreeReduction
began. So by the algorithm we have count = |Sk′ i−1 | and thus |Ski | ≥ 21 · count = 21 |Sk′ i−1 | ≥
1 ′
2 |Ski |. Then clearly, the number of connected components of T \ Sk is at least
X
l≥
deg(u) − 2|Sk | + 2
u∈Sk
≥
X
deg(u) − 2|Ski | + 2 ≥ ki |Ski | − 2|Ski | + 2
u∈Ski
> ki |Ski | − 2|Ski | ≥ (
ki
3
3
∆
3
− 1) · |Sk′ i | ≥ ( ∆ +
− ) · |Sk′ i | > ∆|Sk′ i |
2
8
8 log n 2
8
The first inequality holds by Lemma 6; the last two inequalities holds by ki = ⌊ 43 ∆ + 41 ∆(1 −
∆
(1 − log1 n )i )⌋ > 34 ∆ + 4 log
n − 1 and ∆ ≥ 20 log n.
(2) Upper bounding |M |.
There are two kinds of marked vertices.
(a) Either u ∈ Sk′ i −1 was marked at the beginning, or u ∈ Sk′ i−1 \ Sk′ i −1 whose tree degree later
got increased to ki − 1 at some point while the algorithm kept modifying T. The total
number of such vertices is at most |Sk′ i−1 | < 2|Sk′ i |.
(b) u is a marked vertex and u ∈
/ Sk′ i−1 . In this case, deg(u) < ki−1 before this instance of
FastDegreeReduction began. Since u is marked, deg(u) increases to ki − 1 at some point.
Every time we modify T, at least one vertex in Ski loses one degree and at most two
unmarked vertices get one degree separately. So for a vertex u ∈
/ Sk′ i−1 to be marked, the
vertices in Ski loses at least 12 (ki − ki−1 ) degrees on average. As each vertex will be removed
from Ski after it loses at most ∆ − ki + 1 tree degree, the total number of such vertex u
can be at most:
2(∆ − ki + 1)|Sk′ i |
<
ki − ki−1
1
i
′
log n ) + 2)|Ski |
1
i−1 − (1 − 1 )i )
log n )
log n
2( 14 ∆(1 −
1
4 ∆((1
−
6
−1
< 3|Sk′ i | log n
Overall, |M | < 2|Sk′ i | + 3|Sk′ i | log n = (2 + 3 log n)|Sk′ i |.
Summing up (1) and (2), we have
∆∗ ≥
3
′
l−1
8 ∆|Ski |
≥
= Ω(∆/ log n)
|M |
(2 + 3 log n)|Sk′ i |
or equivalently, ∆ = O(∆∗ log n).
4
A (1 + δ)∆∗ + O( δ12 log n) Approximation
n
In this section we prove Theorem 2. To obtain an improved approximation of (1 + δ)∆∗ + O( log
),
δ2
the rough idea is that we refine the fast degree reduction algorithm in the previous section using
an augmenting path technique.
1
) be a fixed parameter. The basic framework stays the same as in the previous
Let ǫ ∈ (0, 48
section. One difference is that the new main algorithm consists of two phases. In the large-step
2
n
, we repeatedly apply a near-linear time algorithm LargeStepScheduling
phase, as long as ∆ ≥ 10 log
ǫ3
that either reduces ∆ to ≤ (1 − ǫ) · ∆ or terminates with ∆ ≤ (1 + O(ǫ))∆∗ . In the small-step
2
n
n
≤ ∆ < 10 log
; in this case we repeatedly
phase we need to deal with the situation where 20 log
ǫ2
ǫ3
run a weaker near-linear time algorithm SmallStepScheduling that either reduces ∆ by 1 or provides
n
).
evidence that ∆ ≤ (1 + O(ǫ))∆∗ + O( log
ǫ2
Both algorithms LargeStepScheduling and SmallStepScheduling rely on a building block algorithm
AugPathDegRed; similar to Scheduling , both scheduling algorithms run a while-loop and repeatedly
feed inputs to AugPathDegRed. Algorithm AugPathDegRed efficiently reduces the total number of
vertices of high tree degree using an augmenting path technique, which is a significant improvement
over FastDegreeReduction.
For the rest of this section, we first propose and analyse the building block algorithm AugPathDegRed which underlies the core of our main algorithm. After that we specify how the two
phases large-step phase and small-step phase work. Finally, we prove Theorem 2.
4.1
4.1.1
Degree reduction via augmenting paths
Algorithm description
Let k ≤ ∆ be a fixed threshold. This algorithm is, in some way, an extension of the previous
algorithm FastDegreeReduction. As before, due to concerns of efficiency, a vertex gets marked if its
tree degree is ≥ k − 1, and it stays marked (throughout one execution of AugPathDegRed) even if
its tree degree decreases afterwards. Previously, we only look for a non-tree edge whose inclusion
could directly reduce some tree degrees of vertices in Sk , and when such edges no longer exist
the procedure terminates. In this case, AugPathDegRed would continue to explore possibilities of
improving the tree structure using the idea of augmenting paths. Intuitively, an augmenting path
consists of a sequence of non-tree edges that can jointly reduce tree degrees of Sk . Formally we
give its definition below.
Definition 7 (augmenting paths). An h-length augmenting path consists of a sequence of distinct
non-tree edges (w1 , z1 ), (w2 , z2 ), · · · , (wh , zh ) ∈ E with the following properties.
7
S
(i) ∃w0 ∈ ρw1 ,z1 ∩ Sk , wi ∈ ρwi+1 ,zi+1 \ ( hj=i+2 ρwj ,zj ), ∀0 ≤ i < h.
(ii) All zi ’s are unmarked, ∀1 ≤ i ≤ h; wi ’s are marked for 1 ≤ i < h and wh is unmarked.
Lemma 8 (tree modification). Given an augmenting path (w1 , z1 ), (w2 , z2 ), · · · , (wh , zh ) ∈ E, one
can modify T such that dk decreases and no vertices are added to Sk .
Proof. We modify T in an inductive way. For i = h−1, h−2, · · · , 0, as wi ∈ ρwi+1 ,zi+1 , we can take an
arbitrary tree edge (wi , x) ∈ ρwi+1 ,zi+1 , and then perform an update T ← T∪{(wi+1 , zi+1 )}\{(wi , x)}
which guarantees that S
T is still a spanning tree. Note that this update also preserves the property
that wj ∈ ρwj+1 ,zj+1 \ ( hl=j+2 ρwl ,zl ), ∀0 ≤ j < i; this is because, when wj ∈
/ ρwi+1 ,zi+1 , tree update
T ← T ∪ {(wi+1 , zi+1 )} \ {(wi , x)} does not change the connected components of T \ {wj }, and thus
S
the condition wj ∈ ρwj+1 ,zj+1 \ ( hl=j+2 ρwl ,zl ) stays intact.
During the process, if any deg(zi ), 1 ≤ i ≤ h becomes k − 1 during the process, mark zi . By
definition, dk decreases as w0 loses a tree neighbour; plus, because all deg(wi ), 1 ≤ i < h are
unchanged, and no vertices are newly added to Sk because deg(zi ) ≤ k − 2, ∀1 ≤ i ≤ h.
It is easy to notice that what FastDegreeRedection does is repeatedly looking for augmenting
paths of length 1 and then apply Lemma 8. To extend this algorithm, when we can no longer find
any augmenting paths of length 1, we turn to search for augmenting paths of length 2, and so on.
Generally speaking, when the currently shortest augmenting paths have length h, we apply Lemma
8 to decrease the total number of shortest augmenting paths, and when no further progress of such
kind can be made we argue the shortest length of augmenting paths must now increase. Finally
our algorithm terminates when h grows to ≥ 1 + log1+ǫ n, and then we prove a lower bound on ∆∗
based on the structure of T.
The algorithm for finding the shortest length of augmenting paths works as follows; actually the
algorithm computes an auxiliary layering of the graph that will also help tree modification later.
Initially we set B0 ← Sk . Inductively, supposeSwe have already computed B0 , B1 , · · · , Bh , h ≥ 0,
then we compute S
the forest spanned by T \ ( hi=0 Bi ). Here is an extra
Si notation: ∀0 ≤ i ≤ h,
i
i
for each u ∈ V \ ( j=0 Bj ), let Cu be the connected component of T \ ( j=0 Bj ) that contains u.
If there exists an edge (u, v) ∈ E such that both u, v are unmarked vertices, and that Cuh 6= Cvh ,
then the algorithm terminates and reports that the shortest length of augmenting paths
is equal
S
to h + 1; otherwise, we compute Bh+1 to be the set of all marked vertices u ∈ V \ ( hi=0 Bi ) such
that there exists an unmarked adjacent vertex v with Cuh 6= Cvh , and then continue. The above
procedure is summarised as the following pseudo-code Layering shown in Algorithm 3. Note that
once Bh+1 = ∅, the algorithm would continue to compute Bh+2 = Bh+3 = · · · = B1+log1+ǫ n = ∅.
After we have invoked Layering and computed a sequence of vertex subsets B0 , B1 , · · · , Bh which
naturally divides the graph into h+2 layers, we should start to apply tree modifications of Lemma 8
to decrease the total number of shortest augmenting paths. The difficulty in searching for shortest
augmenting paths is that, for a search that starts from a pair of adjacent and unmarked vertices
u, v satisfying Cuh 6= Cvh and goes up the layers Bh , Bh−1 , · · · , B1 , B0 , not every route can reach the
top layer B0 because the augmentations of some previous (h + 1)-length augmenting paths might
have already blocked the road. Therefore, a depth-first search needs to performed. To save running
time, some tricks are needed: if a certain vertex has been searched and failed to lead a way upwards
to B0 , then we tag this vertex so that future depth-first searches may avoid this tagged vertex; if
a certain edge has been searched before, then we tag this edge whatsoever. The following pseudocode AugDFS shown in Algorithm 4 may be a better illustration of this algorithm; the recursive
8
Algorithm 3: Layering
1
2
3
4
5
6
7
8
9
10
B0 ← S k ;
h ← 0;
while h ≤ 1 + log1+ǫ n do
S
compute the forest {Cuh } spanned by T \ ( hi=0 Bi );
if exists unmarked u, v such that (u, v) ∈ E, Cuh 6= Cvh then
break;
else
S
compute Bh+1 to be the set of all marked vertices in u ∈ V \ ( hi=0 Bi ) such that
there exists an unmarked adjacent vertex v with Cuh 6= Cvh ;
h ← h + 1;
return B0 , B1 , · · · , Bh ;
algorithm AugDFS searches for an (h+1)-length augmenting path (w1 , z1 ), (w2 , z2 ), · · · , (wh+1 , zh+1 )
given input (wh+1 , zh+1 ) ∈ E. Later we will prove, if AugDFS returns true, then the sequence
(w1 , z1 ), (w2 , z2 ), · · · , (wh+1 , zh+1 ) is indeed an augmenting path.
Now we come to describe the upper-level AugPathDegRed: basically, it repeatedly apply Layering
followed by several rounds of AugDFS until h ≥ 1 + log1+ǫ n. Here is the pseudo-code AugPathDegRed as shown in Algorithm 5.
Before proving termination of AugPathDegRed, we first need to argue some properties of Layering. The following lemma will serve as the basis for our future lower bounds on ∆∗ .
Lemma 9 (the blocking property). Throughout each iteration
Sof the repeat-loop in AugPathDegRed,
for any 1 ≤ i < h and any two adjacent vertices u, v ∈ V \ ( ij=0 Bj ) such that u is unmarked and
Cui 6= Cvi , then v ∈ Bi+1 .
Proof. By rules of Layering, this blocking property holds right after Layering outputs them. This
claim continuous to hold afterwards because tree modifications only merge components Cui ’s and
never splits any Cui ’s.
Here is an important corollary of this Lemma 9.
Corollary 10. Throughout each iteration of the repeat-loop, for any w ∈ Bi , 1 ≤ i ≤ h, suppose
w is S
adjacent to an unmarked z such that Cwi−1 6= Czi−1 . Then ρw,z only contains vertices from
V \ ( i−2
j=0 Bj ).
Proof. Suppose otherwise, then there would be a vertex x ∈ ρw,z ∩ Bj , j ≤ i − 2, then in this case
Cwj 6= Czj , and thus by Lemma 9 w ∈ Bj+1 which is a contradiction as j + 1 < i.
Now we can argue correctness of AugDFS.
Lemma 11. If AugDFS returns true, (w1 , z1 ), (w2 , z2 ), · · · , (wh+1 , zh+1 ) is an augmenting path.
Proof. Property (ii) in Definition 7 holds by rules of this algorithm. Now let us focus on property
(i). Take an arbitrary w0 ∈ ρw1 ,z1 ∩ B0 . By the algorithm it must be wi ∈ Bi , ∀0 ≤ i ≤ h, then
using Corollary 10 we know ρwi ,zi does not contain any wj , 0 ≤ j ≤ i − 2, so property (i) holds.
9
Algorithm 4: AugDFS
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
global variables: (w1 , z1 ), (w2 , z2 ), · · · , (wh+1 , zh+1 );
input params: i, 1 ≤ i ≤ h + 1;
if i = 1 then
return true;
for untagged w ∈ ρwi ,zi ∩ Bi−1 do
wi−1 ← w;
for unmarked z such that (w, z) is untagged and Czi−2 6= Cwi−2 do
zi−1 ← z;
run AugDFS with input i − 1;
tag (w, z);
if AugDFS has returned false then
continue;
else
if i = h + 1 then
modify T by augmenting path (w1 , z1 ), (w2 , z2 ), · · · , (wh+1 , zh+1 ) via Lemma 8
and mark new degree k − 1 vertices;
return true;
16
17
tag w;
18
return false;
Algorithm 5: AugPathDegRed
1
2
3
4
5
6
7
8
9
input params: threshold k;
mark all vertices from Sk−1 and unmark the rest;
repeat
run Layering which computes B0 , B1 , · · · , Bh ;
untag all vertices and edges;
for (u, v) ∈ E such that u, v are unmarked and adjacent, and that Cuh 6= Cvh do
run AugDFS with input h + 1 and global params (wh+1 , zh+1 ) ← (u, v);
until h > 1 + log1+ǫ n;
return T;
10
Finally we conclude this subsection with the lemma below, from which termination of AugPathDegRed immediately follows.
Lemma 12. Every iteration of the repeat-loop, if not the last, increases h by at least one.
Proof. By the rules of Layering, it is easy to see that at the beginning when Layering outputs
B0 , B1 , · · · , Bh , the shortest length of augmenting path is equal to h + 1. So it suffices to prove
that by the end of this iteration the shortest augmenting path has length > h + 1.
First we need to characterize all augmenting paths using B0 , B1 , · · · , Bh . Let the sequence
(w1 , z1 ), (w2 , z2 ), · · · , (wl , zl ) be an arbitrary augmenting path. We argue l ≥ h + 1, and
Si more
importantly, if l = h + 1, it must be wi ∈ Bi , ∀1 ≤ i ≤ h. We inductively prove that wi ∈ j=0 Bj
for i = 0, 1, · · · , l − 1. The basis is obvious as is required by property (i) in Definition 7. Now
assume
∈ Br for some r ≤ i. Then, by Corollary 10, it would
not be hard to see wi+1 ∈
Sr+1 wi S
Sh−1
i+1
B
⊆
B
.
Now,
on
the
one
hand
by
Corollary
10
ρ
∩
j
j
w
,z
l l
j=0
j=0
j=0 Bj = ∅, and on the other
Sl−1
hand wl−1 ∈ ρwl ,zl ∩ j=0 Bj , so l ≥ h + 1. Plus, we can see from the induction that, when l = h + 1
it must be wi ∈ Bi , ∀0 ≤ i ≤ h.
For any unmarked and adjacent vertices u, v such that Cuh 6= Cvh , consider the instance of
AugDFS with input (wh+1 , zh+1 ) ← (u, v). We make two claims.
(1) If there is an (h + 1)-length augmenting path ending with (wh+1 , zh+1 ) = (u, v), AugDFS would
succeed in finding one.
(2) If it has returned false, then there would be no (h+1)-augmenting path ending with (wh+1 , zh+1 ) =
(u, v) throughout the entire repeat-loop iteration.
If (1)(2) can be proved, then by the end of this repeat-loop iteration, there would be no (h + 1)length augmenting paths because at such augmenting path should end with a pair of adjacent
unmarked vertices. Next we come to prove (1)(2).
(1) The depth-first search of AugDFS exactly coincides with the conditions that wi ∈ Bi , except
that it skips all tagged vertices and edges. Now we prove that omitting tagged vertices and
edges does not miss any (h+ 1)-length augmenting paths. For a vertex w to be tagged, we must
have enumerated all of its untagged edges (w, z) but failed to find any augmenting paths, and
therefore any future depth-first searches on w would still end up in vain. For an edge (w, z)
to be tagged, either a further recursion AugDFS on line-9 has succeeded or failed in finding an
augmenting paths; in the former case, Cwi−2 and Czi−2 has been merged, and so the condition
Cwi−2 6= Czi−2 would be violated afterwards; in the latter case, we would not need to recur on
(w, z) anyway.
(2) If AugDFS has once failed to find any augmenting paths starting with (u, v), then all vertices
w ∈ ρu,v ∩ Bh visited by this instance of AugDFS should be tagged and they would be omitted
by all succeeding instances of AugDFS. Therefore ρu,v ∩ Bh would stay unchanged since then
(although ρu,v itself might change). Hence, image if we re-run AugDFS with (wh+1 , zh+1 ) ←
(u, v), it may return false without any recursion because all vertices in ρwh+1 ,zh+1 ∩ Bh are
tagged.
11
4.1.2
Lower bound on ∆∗
Suppose AugPathDegRed has terminated with B0 , B1 , · · · , Blog1+ǫ n+1 . Let us see it yields lower
bounds on ∆∗ . To apply Lemma 5, we first need to specify a sequence of disjoint vertex subsets,
which is what the following definition is about.
Definition 13. After an instance
S of AugPathDegRed has been executed, for an arbitrary component
Cuh , 0 ≤ h ≤ log1+ǫ n, u ∈ V \ ( hi=0 Bi ), it is called clean if all vertices in Cuh are unmarked.
S
Lemma 14. For any 0 ≤ h ≤ log1+ǫ n, suppose T \ ( hi=0 Bi ) has l clean components, then a lower
l−1
.
bound holds that ∆∗ ≥ Ph+1
i=0
|Bi |
S
Proof. By Lemma 9, any edge that connects a clean components of T \ ( hi=0 Bi ) outwards must
Sh+1
l−1
l−1
be incident on a vertex in i=0 Bi . Therefore by Lemma 5 we have ∆∗ ≥ Sh+1
= Ph+1
|
i=0
Bi |
i=0
|Bi |
From Lemma 14, it suffices to lower bound the total number of clean components. The next
lemma describes a scenario in which l must be large.
′
and
Lemma 15. Suppose an instance of AugPathDegRed has been executed. Let d′k , d′k−1 , Sk−1
′
Sk be snapshots of dk , dk−1 , Sk−1 and Sk right before this instance of AugPathDegRed started;
recall that dk , dk−1 , Sk−1 and Sk always refer to statistics of the current T after this instance of
AugPathDegRed has finished.
Assume the following three conditions:
(i) k ≥
4
ǫ
(ii) d′k ≥
1
ǫ(k−1)
(iii) dk ≥ (1 −
· d′k−1
ǫ2
′
2 log n )dk
S
Then, for each 0 ≤ h ≤ log1+ǫ n, the number of clean components in T \ ( hi=0 Bi ) is more than
P
k · (1 − 4ǫ) hi=0 |Bi |.
S
Proof. By Lemma 6, the number of tree components in T \ ( hi=0 Bi ) is at least
X
S
u∈ h
i=0 Bi
deg(u) − 2|
h
[
Bi | + 2
i=0
′
Let M be the set of all marked vertices ∈
/ Sk−1
(i.e., vertices that are initially unmarked) by the
S
end of AugPathDegRed. Then, the number of clean components in T \ ( hi=0 Bi ) is at least
X
S
u∈ h
i=0 Bi
deg(u) − 2
The argument consists of a lower bound on
h
X
′
|
|Bi | + 2 − |M ∪ Sk−1
i=0
P
S
u∈ h
i=0 Bi
12
′
|.
deg(u) and an upper bound on |M ∪Sk−1
(1) Lower bound on
P
S
u∈ h
i=0 Bi
deg(u).
P
By the algorithm B0 = Sk , then we have u∈B0 deg(u) = dk .
S
For any vertex u ∈ hi=1 Bi , deg(u) = k − 1 by the time u was first added to some Bi . After
that, deg(u) could only decrease when we modify T by an augmenting path (w1 , z1 ), · · · , (wt , zt )
where u = wj for some 1 ≤ j ≤ t. Since t ≤ 1 + log1+ǫ n, during a tree modification, at least
S
one vertex in Sk loses one degree and at most 1 + log1+ǫ n vertices in hi=1 Bi lose one degree
separately. As the total number of the degree loss in Sk is (d′k − dk ), we have
X
S
u∈ h
i=1 Bi
deg(u) ≥ (k − 1)
S
u∈ h
i=0 Bi
|Bi | − (d′k − dk )(1 + log1+ǫ n)
i=1
From above, we get a lower bound on
X
h
X
P
deg(u) ≥ dk + (k − 1)
S
u∈ h
i=0 Bi
h
X
deg(u),
|Bi | − (d′k − dk )(1 + log1+ǫ n)
i=1
≥ (k − 1)
h
X
|Bi | + (1 −
≥ (k − 1)
i=1
h
X
ǫ2
ǫ2
)d′k −
d′ · (1 + log1+ǫ n)
2 log n
2 log n k
|Bi | + (1 −
ǫ2
− ǫ)d′k
2 log n
i=1
(2) Upper bound on |M |.
The argument is similar to (1). An unmarked vertex u is marked only when we modify T by
an augmenting path (w1 , z1 ), · · · , (wt , zt ) where u = zj for some 1 ≤ j ≤ t or u = wt . Since
t ≤ 1 + log1+ǫ n, during a tree modification, at least one vertex in Sk loses one degree and at
most 2 + log1+ǫ n unmarked vertices are marked. Then we get a upper bound on |M |.
|M | ≤ (d′k − dk )(2 + log1+ǫ n) ≤ ǫ · d′k
′
|.
(3) Upper bound on |Sk−1
By easy calculations, d′k ≥
1
′
ǫ(k−1) dk−1
′
′
≥ 1ǫ |Sk−1
| ≤ ǫ · d′k .
|, and so |Sk−1
13
Summing up (1)(2)(3), we can conclude the proof:
X
S
u∈ h
i=0 Bi
deg(u) − 2
h
X
′
| > dk + (k − 1)
|Bi | + 2 − |M ∪ Sk−1
h
X
|Bi | − 2
i=1
i=0
h
X
|Bi | − 3ǫ · d′k
i=0
X
ǫ2
|Bi | − 3ǫ · d′k
≥ (1 −
) · d′k − 2|B0 | + (k − 3)
2 log n
h
i=1
X
ǫ2
|Bi |
− 3ǫ) · d′k − 2|B0 | + (k − 3)
2 log n
h
≥ (1 −
i=1
> k(1 − 3.5ǫ) · |B0 | − 2|B0 | + (k − 3)
h
X
|Bi | ≥ k(1 − 4ǫ) ·
|Bi |
i=0
i=1
The last inequality holds by k ≥ 4ǫ .
4.1.3
h
X
Implementation and running time
We present an implementation of AugPathDegRed that runs in O( ǫ12 m log2 n) time. We discuss some
implementation details of Layering, AugDFS and AugPathDegRed, and analyse their contributions
to the total running time in a single run of AugPathDegRed.
(1) Layering.
For every instance of Layering, computing the forest {Cuh }u∈V \(Sh Bi ) can be done in a single
i=0
pass of breath-first search which takes O(m) time. Computing Bh+1 , if necessary, is easily done
by scanning the edge set E which also takes O(m) time. As the while-loop iterates for at most
1 + log1+ǫ n times, and due to Lemma 12 Layering is invoked for at most 1 + log 1+ǫ n times, the
overall contribution of Layering is O( ǫ12 m log2 n).
(2) AugPathDegRed.
Excluding the contributions of AugDFS and Layering, all AugPathDegRed does is simply untagging all vertices and edges, as well as scanning the edge set (u, v) ∈ E and deciding if
Cuh 6= Cvh . As tree components only get merged and never split, we can use the union-find data
structure to support querying whether Cuh 6= Cvh in O(α(n)) time. Hence, AugPathDegRed’s
exclusive contributions to the total running time would be O( 1ǫ mα(n) log n).
(3) AugDFS.
Now we analyse the overall time complexity induced by AugDFS invoked on line-7 of AugPathDegRed. There are three technical issues to be resolved.
(a) How to enumerate untagged vertices ∈ ρwi ,zi ∩ Bi−1 (line-5)?
For each u ∈ Bi , ∀0 ≤ i ≤ h, assign u a weight of i; vertices that do not belong to any
Bi have weight h + 1. By Corollary 10, to enumerate vertices ∈ ρwi ,zi ∩ Bi−1 , it suffices
to enumerate the lightest vertices on ρwi ,zi , which can be done using a link-cut tree data
structure [16] built on T, each enumeration taking O(log n) amortized time. When a vertex
gets tagged, we change its weight to h + 1, and so future enumerations on ρwi ,zi ∩ Bi−1 may
skip this tagged vertex.
14
(b) How to enumerate unmarked z connected by an untagged edge (w, z) such that Czi−2 6= Cwi−2
(line-7)?
Each w decrementally maintains a list of all its neighbours. While we scan the list, if the
next edge (w, z) satisfies both conditions that Czi−2 6= Cwi−2 and z is unmarked, then the
algorithm starts a new iteration and recur; either way we cross the edge (w, z) off the list.
In this way, every edge appears on line-7 for at most once. Thus the total time of this part
is O(mα(n)); the additional α(n) factor comes from the union-find data structure that
helps deciding if Czi−2 6= Cwi−2 .
(c) How to implement the tree modification from Lemma 8?
Every tree modification involves insertions and deletions of O( 1ǫ log n) edges, as well as
merging O( 1ǫ log n) pairs of some tree components Cui . Using the link-cut tree, every edge
insertion and deletion takes update time O(log n), and every component-merging takes
time O(α(n)). Since every tree modification merges two components in T \ Sk (i.e., T \ B0 ),
there can be at most O(n) tree modifications throughout AugPathDegRed. Therefore, the
overall contribution of tree modifications is O( 1ǫ n log2 n).
The total time of (a)(b) is O(m log n) because every time a vertex is enumerated on line-5,
either it gets tagged or one of its edge gets tagged. Thus, the overall complexity of AugDFS is
O(m log n + 1ǫ n log2 n).
Summing up (1)(2)(3), the total running time of is dominated by time complexity of Layering
which is O( ǫ12 m log2 n).
4.2
4.2.1
Large-step phase
Algorithm description and running time
n
: it iterates over k = (1 − 2ǫ)∆ +
Algorithm largeStepScheduling deals with the case ∆ ≥ 10 ǫlog
2
1, (1 − 2ǫ)∆ + 2, · · · , (1 − ǫ)∆ and if dk−1 ≤ 2dk it invoke AugPathDegRed with input k. During the
course, if one instance fails to reduce dk significantly, then the algorithm reports a lower bound on
∆∗ and terminates. Here is its pseudo-code 6.
Algorithm 6: LargeStepScheduling
1
2
3
4
5
6
7
8
9
10
input params: T with maximum tree degree ∆;
for k = (1 − 2ǫ)∆ + 1, (1 − 2ǫ)∆ + 2, · · · , (1 − ǫ)∆ do
if dk = 0 then
break;
else if dk−1 ≤ 2dk then
d′k ← dk ;
run AugPathDegRed with input k;
ǫ2
′
if dk > (1 − 2 log
n ) · dk then
return false;
return true;
15
From the previous subsection we already know AugPathDegRed runs in near-linear time, so
here we only need to upper bound the total number of times AugPathDegRed gets invoked before
the algorithm returns either true or false. We consider in each iteration, by how much would dk
decrease. There are two cases.
(1) If the branching condition on line-5 does not hold, then dk−1 > 2dk , and so dk has shrunk by
a factor of at most 21 .
(2) In the case where the branching condition holds, if the algorithm does not return false within
ǫ2
ǫ2
′
this iteration, then dk ≤ (1 − 2 log
n ) · dk ; that is, dk has declined by a factor of at most 1 − 2 log n
during this iteration.
2
ǫ
Summing up (1)(2), the value of dk would decrease by a factor of at most 1− 2 log
n in each iteration.
2
n
because by the time dk would decrease
Then the for-loop would break when k ≥ (1 − 2ǫ)∆ + 2 log
ǫ2
to 0. Hence, the running time of LargeStepScheduling is bounded by O( ǫ13 m log3 n).
10 log2 n
,
ǫ3
A byproduct of the above observation is that, when ∆ ≥
10 log2 n
,
ǫ2
we have (1 − ǫ)∆ ≥ (1 −
2ǫ)∆ +
and thus if LargeStepScheduling eventually returns true, it must be d(1−ǫ)∆ = 0; in
other words, ∆ has been reduced by a factor of at most 1 − ǫ.
4.2.2
Approximation guarantee
We prove when LargeStepScheduling returns false, it must be ∆ ≤ (1 + 8ǫ) · ∆∗ . Consider the most
recent execution of AugPathDegRed before returning. By the previous subsection, this instance
of AugPathDegRed has created a sequence of disjoint vertex subsets B0 , B1 , · · · , B1+log1+ǫ n that
satisfies the blocking property.
′
Let d′k−1 , Sk−1
and Sk′ be snapshots of dk−1 , Sk−1 and Sk right before this instance of Aug1
d′k−1 by the branching condition on line-5. Using
PathDegRed started. Then d′k ≥ 21 d′k−1 ≥ ǫ(k−1)
Lemma 15 and Lemma 14, we can derive a lower bound
Ph
∗
i=0 |Bi |
∆ ≥ k(1 − 4ǫ) · Ph+1
, ∀0 ≤ h ≤ log1+ǫ n
i=0 |Bi |
By the pigeon-hole principle, there exists an h such that
∆∗ ≥ k(1 − 4ǫ) ·
or equivalently, ∆ ≤
4.3
4.3.1
1+ǫ
∆∗
1−6ǫ+8ǫ2
Ph
i=0 |Bi |
Ph+1
i=0 |Bi |
≥
1
1+ǫ ,
and hence
1
1 − 6ǫ + 8ǫ2
>
∆
1+ǫ
1+ǫ
1
< (1 + 8ǫ)∆∗ when ǫ ∈ (0, 48
).
Small-step phase
Algorithm description and running time
Algorithm SmallStepScheduling only deals with
define a potential function:
φ(T) =
20 log n
ǫ2
∆
X
i=∆+1−log n
16
≤∆<
ci · |Ni |
10 log2 n
.
ǫ3
Set c = 12 + 6 log1+ǫ n and
SmallStepScheduling works by repetitively selecting a degree k that maximizes ck · |Nk | and then
n
.
feed input k to AugPathDegRed until ∆ decreases; clearly k must be larger than ∆ − log n > 19 log
ǫ2
We formulate this procedure as pseudo-code SmallStepScheduling as shown in Algorithm 7 below.
Algorithm 7: SmallStepScheduling
1 input params: T with maximum tree degree ∆;
2 while ∆ has not changed do
3
pick a k ∈ arg maxi∈[∆+1−log n,∆] {ci · |Ni |};
4
d′k ← dk ;
5
run AugPathDegRed with input k;
ǫ2
′
6
if dk > (1 − 2 log
n ) · dk then
7
return false;
8
return true;
Now let us bound the running time of SmallStepScheduling. We study how many rounds of
AugPathDegRed could be invoked before branching conditions on line-2 or line-6 can be triggered.
In fact, an instance of AugPathDegRed has not triggered the condition on line-6, then dk has
ǫ2
decreased by a factor of at most 1 − 2 log
n . Let us analyse how φ(T) has decreased.
′
Let Nk , k ∈ [∆ + 1 − log n, ∆] be snapshots of Nk right before we execute AugPathDegRed,
P
i
′
and denote the potential of T at that time by φ′ (T) = ∆
i=∆+1−log n c · |Nk |. Every time a tree
modification to T was made on line-15 of AugDFS, at least one vertex in Sk would lose a tree edge
and at most 2 + log1+ǫ n vertices would gain a tree edge, and then the total loss of φ(T) would be
at least
(ck − ck−1 ) − (2 + log1+ǫ n) · (ck−1 − ck−2 ) ≥ (ck−1 − ck−2 )(c − 2 − log1+ǫ n)
2 + log1+ǫ n
1 5
1
) ≥ ck · (1 − ) · > 0.8 · ck
= ck · (1 − ) · (1 −
c
c
c 6
Note that dk has decreased by d′k − dk ≥
ǫ2
′
2 log n dk
≥
ǫ2
2 log n
· k|Nk′ |, and so there are at least
ǫ2
(d′k − dk )/(2k) ≥ 4 log n · |Nk′ | such modifications to T, this is because a tree modification can at
most move two vertices from Sk to Sk−1 thus decreasing dk by 2k. Therefore,
φ(T) ≤ φ′ (T) − (0.8 · ck ) · (
ǫ2
0.2ǫ2 ′
)φ (T)
· |Nk′ |) ≤ (1 −
4 log n
log2 n
The second inequality holds by maximality of ck · |Nk′ | which implies ck · |Nk′ | ≥ log1 n · φ′ (T)
In a nutshell, whenever AugPathDegRed returns true, φ(T) has decreases by a factor of at
0.2ǫ2
∆−1 , n · c∆ ), and
most 1 − log
2 . As long as ∆ has not changed, φ(T) belongs to the interval (c
n
3
consequently, φ(T) could suffer at most − log1− 0.2ǫ2 n · c = O( logǫ2 n ) rounds of AugPathDegRed
log2 n
that returns true. Therefore, the overall running time of SmallStepScheduling before ∆ decreases
by 1 would be upper bounded by O( ǫ13 m log5 n).
17
4.3.2
Approximation guarantee
We prove when SmallStepScheduling returns false, it must be ∆ ≤ (1 + 8ǫ)∆∗ + log n. Consider
the most recent execution of AugPathDegRed before returning. By the previous subsection, this
instance of AugPathDegRed has created a sequence of disjoint vertex subsets B0 , B1 , · · · , B1+log1+ǫ n
that satisfies the blocking property.
′
Let d′k−1 , Sk−1
and Sk′ be snapshots of dk−1 , Sk−1 and Sk right before this instance of Aug′
|. Then,
PathDegRed started. By maximality of ck · |Nk′ |, we have |Nk′ | ≥ 1c · |Nk−1
P∆
′
d′k
k|Nk′ |
1
1
1
i=k i|Ni |
≥
≥
=
>
>
P
′
′
′
∆
c(k−1)
′
dk−1
k|Nk | + (k − 1)|Nk−1 |
c+1
ǫ(k − 1)
1+ k
i=k−1 i|Ni |
n
The last inequality holds by k ≥ 19 log
and c = 6 log1+ǫ n + 12 ≤ 18 log1+ǫ n.
ǫ2
2
ǫ
′
As dk ≥ (1 − 2 log n )dk by the branching condition on line-6, using Lemma 15 and Lemma 14,
we can derive a lower bound
Ph
∗
i=0 |Bi |
, ∀0 ≤ h ≤ log1+ǫ n
∆ ≥ k(1 − 4ǫ) · Ph+1
i=0 |Bi |
By the pigeon-hole principle, there exists an h such that
∆∗ ≥ k(1 − 4ǫ) ·
or equivalently, ∆ ≤
4.4
1+ǫ
∗
1−4ǫ ∆
Ph
i=0 |Bi |
Ph+1
i=0 |Bi |
≥
1
1+ǫ ,
and hence
1 − 4ǫ
1
>
(∆ − log n)
1+ǫ
1+ǫ
1
+ log n < (1 + 8ǫ)∆∗ + log n for ǫ ∈ (0, 48
).
Correctness of the main algorithm
As a conclusion we prove Theorem 2. If the main algorithm terminates in the large-step phase,
then in previous subsections, we have proved ∆ ≤ (1 + 8ǫ) · ∆∗ ; if it returns in the small-step phase ,
n
n
. So, in sum ∆ ≤ (1 + 8ǫ)∆∗ + 10 log
. Reassigning
then either ∆ ≤ (1 + 8ǫ)∆∗ + log n or ∆ < 10 log
ǫ2
ǫ2
δ ← 8ǫ we can obtain the approximation guarantee promised in Theorem 2.
For the running time, the large-step phase can invoke at most O( 1ǫ log n) rounds of LargeStepScheduling because each invocation that returns true reduces ∆ by a factor of at most 1 − ǫ; thus
the total running time of the first while-loop never exceeds O( ǫ14 m log5 n). The small-step phase
2
can invoke at most O( logǫ3 n ) rounds of SmallStepScheduling because each invocation that returns
true reduces ∆ by at least 1; thus the total running time of the small-step phase is O( ǫ16 m log7 n)
which is dominant. So the running time of the main algorithm is O( ǫ16 m log7 n).
References
[1] Nikhil Bansal, Rohit Khandekar, and Viswanath Nagarajan. Additive guarantees for degreebounded directed network design. SIAM Journal on Computing, 39(4):1413–1431, 2009.
18
[2] Kamalika Chaudhuri, Satish Rao, Samantha Riesenfeld, and Kunal Talwar. What would
edmonds do? augmenting paths and witnesses for degree-bounded msts. Lecture notes in
computer science, 3624:26, 2005.
[3] Ted Fischer. Optimizing the degree of minimum weight spanning trees. Technical report,
Cornell University, 1993.
[4] Pierre Fraigniaud. Approximation algorithms for minimum-time broadcast under the vertexdisjoint paths mode. AlgorithmsESA 2001, pages 440–451, 2001.
[5] Martin Fürer and Balaji Raghavachari. An nc approximation algorithm for the minimum
degree spanning tree problem. In Proc. of the 28th Annual Allerton Conf. on Communication,
Control and Computing, pages 274–281, 1990.
[6] Martin Furer and Balaji Raghavachari. Approximating the minimum-degree steiner tree to
within one of optimal. Journal of Algorithms, 17(3):409–423, 1994.
[7] Harold N Gabow and Robert Endre Tarjan. A linear-time algorithm for a special case of disjoint
set union. In Proceedings of the fifteenth annual ACM symposium on Theory of computing,
pages 246–251. ACM, 1983.
[8] Michel X Goemans. Minimum bounded degree spanning trees. In Foundations of Computer
Science, 2006. FOCS’06. 47th Annual IEEE Symposium on, pages 273–282. IEEE, 2006.
[9] Philip N Klein, Radha Krishnan, Balaji Raghavachari, and R Ravi. Approximation algorithms
for finding low-degree subgraphs. Networks, 44(3):203–215, 2004.
[10] Jochen Könemann and R Ravi. A matter of degree: Improved approximation algorithms for
degree-bounded minimum spanning trees. In Proceedings of the thirty-second annual ACM
symposium on Theory of computing, pages 537–546. ACM, 2000.
[11] Jochen Könemann and R Ravi. Primal-dual meets local search: approximating mst’s with
nonuniform degree bounds. In Proceedings of the thirty-fifth annual ACM symposium on Theory
of computing, pages 389–395. ACM, 2003.
[12] Radha Krishnan and Balaji Raghavachari. The directed minimum-degree spanning tree problem. In FSTTCS, volume 2245, pages 232–243. Springer, 2001.
[13] R Ravi. Rapid rumor ramification: Approximating the minimum broadcast time. In Foundations of Computer Science, 1994 Proceedings., 35th Annual Symposium on, pages 202–213.
IEEE, 1994.
[14] R Ravi, Madhav V Marathe, SS Ravi, Daniel J Rosenkrantz, and Harry B Hunt III. Many birds
with one stone: Multi-objective approximation algorithms. In Proceedings of the twenty-fifth
annual ACM symposium on Theory of computing, pages 438–447. ACM, 1993.
[15] Mohit Singh and Lap Chi Lau. Approximating minimum bounded degree spanning trees to
within one of optimal. In Proceedings of the thirty-ninth annual ACM symposium on Theory
of computing, pages 661–670. ACM, 2007.
19
[16] Daniel D Sleator and Robert Endre Tarjan. A data structure for dynamic trees. In Proceedings
of the thirteenth annual ACM symposium on Theory of computing, pages 114–122. ACM, 1981.
[17] Guohui Yao, Daming Zhu, Hengwu Li, and Shaohan Ma. A polynomial algorithm to compute the minimum degree spanning trees of directed acyclic graphs with applications to the
broadcast problem. Discrete Mathematics, 308(17):3951–3959, 2008.
20
| 8cs.DS
|
arXiv:1803.07792v1 [math.AC] 21 Mar 2018
SUBFIELDS OF ALGEBRAICALLY MAXIMAL
KAPLANSKY FIELDS
F.-V. KUHLMANN
Abstract. Using the ramification theory of tame and Kaplansky
fields, we show that maximal Kaplansky fields contain maximal
immediate extensions of each of their subfields. Likewise, algebraically maximal Kaplansky fields contain maximal immediate
algebraic extensions of each of their subfields. This study is inspired by problems that appear in henselian valued fields of rank
higher than 1 when a Hensel root of a polynomial is approximated
by the elements generated by a (transfinite) Newton algorithm.
1. Introduction
For real functions, the Newton Algorithm is a nice tool to approximate their zeros. An analogue works in valued fields, such as the fields
of p-adic numbers. Take a complete discretely valued field K with valuation ring O and a polynomial f in one variable with coefficients in
O. If b ∈ O satisfies vf (b) > 2vf ′(b), then set x0 := b and
xi+1 := xi −
f (xi )
f ′ (xi )
for i ≥ 0. It can easily be shown, using the Taylor expansion of f ,
that this sequence is a Cauchy sequence, and if a limit exists, then it
is a root of f . This fact can be used to prove Hensel’s Lemma in the
complete fields of p-adic numbers, which states the existence of p-adic
roots of suitable polynomials.
Date: 21. 12. 2017.
2000 Mathematics Subject Classification. Primary 12J10; Secondary 12J25,
13A18, 12Y05.
Key words and phrases. valued field, Kaplansky field, tame field, Newton
algorithm.
The author would like to thank Anna Blaszczok for the careful proofreading
of an earlier version of this manuscript, and Pablo Cubides Kovacsics for helpful
remarks.
The initial work for this paper was done while the author was a visiting professor at
the Institute of Mathematics at the University of Szczecin; he would like to thank
his colleagues at this Institute for their great hospitality and support.
1
2
F.-V. KUHLMANN
Hensel’s Lemma also holds in other valued fields (such as power
series fields with arbitrary ordered abelian groups of exponents), which
can be much larger than the fields of p-adic numbers. Still, it can be
proved using the Newton algorithm, but if the value group of the field
is not archimedean ordered (i.e., if it is not an ordered subgroup of
R), then the algorithm may not deliver the root after the first ω many
iterations and transfinite induction is needed. For those who do not like
the technicalities of transfinite induction, S. Prieß-Crampe presented
in 1990 an elegant alternative (see [8]). The setting of the Newton
algorithm gives rise to a contracting function
Φ(x) := x −
f (x)
f ′ (x)
defined on a suitable subset of the field, whose unique fixed point, if
it exists, is a root of the polynomial. The existence is guaranteed if
f satisfies the assumptions of Hensel’s Lemma and the valued field
from which its coefficients are taken is spherically complete (which
is a property shared by all power series fields). While it is not the
original definition, it suffices here to say that a valued field is spherically
complete if and only if every pseudo Cauchy sequence in the sense of
[3] has a (pseudo) limit.
The members of the sequence obtained from the Newton algorithm
are approximations to a root of f . If the value group of the field in
which we are working is not archimedean ordered, then the problem
can appear that the sequence is not Cauchy, but only a pseudo Cauchy
sequence. The limits of pseudo Cauchy sequences which are not Cauchy
are not uniquely determined. In a recent paper [9], Prieß-Crampe considers the question whether this situation can be repaired, for a given
polynomial f , by passing to a suitable subfield of K in which the sequence becomes Cauchy and has a limit. She proves that this can
indeed be done when K is a spherically complete Kaplansky field. Let
us provide the necessary definitions and background.
We will work with (Krull) valuations v and write them in the classical
additive way, that is, the value group of v on a field K, denoted by vK,
is an additively written ordered abelian group, and the ultrametric
triangle law reads as v(a + b) ≥ min{va, vb}. We denote by Kv the
residue field of v on K, by va the value of an element a ∈ K, and by
av its residue.
A polynomial f over a field of characteristic p > 0 is called additive
if f (a + b) = f (a) + f (b) for all elements a, b in any extension of the
P
i
field. This holds if and only if f is of the form 0≤i≤n ci X p . Following
SUBFIELDS OF KAPLANSKY FIELDS
3
I. Kaplansky, we call a polynomial g a p-polynomial if g = f +c where
f is an additive polynomial and c is a constant.
A valued field (K, v) is called a Kaplansky field if char Kv = 0 or
if it satisfies Kaplansky’s hypothesis A: for char Kv = p > 0,
(A1) every p-polynomial with coefficients in Kv has a zero in Kv,
(A2) vK is p-divisible.
Prieß-Crampe proves her main result in [9] by a direct application
of this formulation of hypothesis A. While this proof is not without
interest, a more modern approach can lead to more insight, as we will
show in the present paper. A lot of important work has been done
over time by several authors in order to better understand Kaplansky’s
hypothesis A, e.g. by G. Whaples in [11], by F. Delon in her thesis,
and in the paper [7]. The final touch to this development was given
by Kaplansky himself, in cooperation with D. Leep (see [4, Section 9]).
We will use the following characterization. By Theorem 1 of [11] and
the other cited sources, hypothesis A is equivalent to the conjunction
of the following three conditions, where p denotes the characteristic of
the residue field:
(K1) if p > 0, then the value group is p-divisible,
(K2) the residue field is perfect,
(K3) the residue field admits no finite separable extension of degree
divisible by p.
A valued field is henselian if it satisfies Hensel’s Lemma, or equivalently, admits a unique extension of its valuation to its algebraic
closure. Note that every algebraic extension of a henselian field is
again henselian, with respect to the unique extension of the valuation.
A henselization of a valued field (K, v) is an algebraic extension which
is henselian and minimal in the sense that it can be embedded over K
in every other henselian extension field of (K, v). Henselizations exist
for every valued field (K, v), in fact, every henselian extension field of
(K, v) contains a henselization of (K, v). Henselizations are unique up
to valuation preserving isomorphism over K. Therefore, we will speak
of “the henselization of (K, v)” and denote it by (K h , v).
An extension (L|K, v) of valued fields is called immediate if the
canonical embeddings of vK in vL and of Kv in Lv are onto, in other
words, value group and residue field remain unchanged. A valued field
is called maximal if it does not admit any nontrivial immediate extensions; by [3, Theorem 4], this holds if and only if it is spherically
complete. A valued field is called algebraically maximal if it does not
4
F.-V. KUHLMANN
admit any nontrivial immediate algebraic extension. Since henselizations are immediate algebraic extensions, every algebraically maximal
field is henselian.
The notion of “purely wild extension” (an algebraic extension of a
henselian field that is linearly disjoint from all tame extensions) was
introduced in [7]. We will give the precise definitions for both types of
extensions in the Preliminaries Section.
The following is our main theorem:
Theorem 1.1. Take a valued field extension (L|K, v), with (L, v) an
algebraically maximal Kaplansky field. Then L contains a maximal
immediate algebraic extension of (K, v), as well as a maximal purely
wild extension of the henselization of (K, v) inside of (L, v).
If (L, v) is maximal, then it also contains a maximal immediate extension of (K, v).
Modulo the fact that a valued field is maximal if and only if it is
spherically complete, the last assertion of this theorem is proved in [9,
Theorem 3.5]. As already indicated, we present a different proof, which
will be based on the ramification theory of Kaplansky and tame fields.
A crucial tool in this proof is the following analogue of Lemma 3.7 of
[6], which deals with the case of tame fields. Interestingly, in the case of
algebraically maximal Kaplansky fields we do not need the assumption
of that lemma that the residue field extension Lv|Kv be algebraic.
Proposition 1.2. Take an algebraically maximal Kaplansky field (L, v)
and let K be a relatively algebraically closed subfield of L. Then also
(K, v) is an algebraically maximal Kaplansky field, with its residue field
relatively algebraically closed in that of L.
These results are a nice complement to the theory of tame fields
as developed in [6]. The proofs of the above results will be given in
Section 3, along with some more facts about algebraically maximal
Kaplansky fields.
2. Preliminaries
We start with the following well known fact:
Lemma 2.1. If (L, v) is henselian and K a relatively algebraically
closed subfield of L, then also (K, v) is henselian.
Proof. Assume that f is a polynomial with coefficients in the valuation
ring of (K, v) which satisfies the conditions of Hensel’s Lemma. Since
the valuation ring of (K, v) is contained in that of (L, v) and (L, v)
SUBFIELDS OF KAPLANSKY FIELDS
5
is henselian, f admits a root a in L which satisfies the assertions of
Hensel’s Lemma. Being a root of f , a is algebraic over K and since K
is relatively algebraically closed in L, we have that a ∈ K.
Take a finite extension (L|K, v) of valued fields. The Lemma of
Ostrowski says that whenever the extension of v from K to L is
unique, then
(1)
[L : K] = pν · (vL : vK) · [Lv : Kv]
with ν ≥ 0 ,
where p is the characteristic exponent of Lv, that is, p = char Lv if
this is positive, and p = 1 otherwise. For the proof, see [10, Theoreme
2, p. 236]) or [12, Corollary to Theorem 25, p. 78]). If pν = 1, then
we say that the extension (L|K, v) is defectless. Note that (L|K, v) is
always defectless if char Kv = 0. We call a henselian field a defectless
field if all of its finite extensions are defectless. Each valued field of
residue characteristic 0 is a defectless field.
An algebraic extension (L|K, v) of henselian fields is called tame if
every finite subextension E|K of L|K satisfies the following conditions:
(TE1) if char Kv = p > 0, then the ramification index (vE : vK) is
prime to p,
(TE2) the residue field extension Ev|Kv is separable,
(TE3) the extension (E|K, v) is defectless.
A tame valued field (in short, tame field) is a henselian field
for which all algebraic extensions are tame. From the definition of a
tame extension it follows that (K, v) is a tame field if and only if it
is a defectless field satisfying conditions (K1) and (K2). It is an easy
observation that every extension of a defectless field (K, v) satisfying
conditions (K1) and (K2) must be tame and so (K, v) must be a tame
field. The converse can be derived from the arguments in the second
part of the proof of Proposition 3.1 below.
By (K1) and (K2), the perfect hull of a tame field is an immediate
extension, and by (TF3), this extension must be trivial. This shows
that every tame field is perfect.
If char Kv = 0, then conditions (TE1) and (TE2) are void, and every
finite extension of (K, v) is defectless. Hence very algebraic extension
of a henselian field of residue characteristic 0 is a tame extension, and
every henselian field of residue characteristic 0 is a tame field.
Lemma 2.2. Every algebraically maximal valued field satisfying conditions (K1) and (K2), and in particular every algebraically maximal
Kaplansky field, is a tame field and hence a perfect and defectless field.
6
F.-V. KUHLMANN
Proof. This follows from [6, Theorem 3.2] and the facts about tame
fields we have mentioned above.
The converse of this lemma does not hold since for a tame field it
is admissible that its residue field has finite separable extensions with
degree divisible by p.
Take a valued field (K, v), fix an extension of v to K sep and call it
again v. The fixed field of the closed subgroup
Gr := {σ ∈ Gal (K sep |K) | v(σa − a) > va for all a ∈ OK sep \ {0}}
of Gal(K sep |K) (cf. [1, Corollary (20.6)]) is called the absolute ramification field of (K, v). For the following fact, see [1, (20.15 b)].
Lemma 2.3. Take a henselian field (K, v) and denote by Z the absolute
ramification field of (K, v). If (K ′ , v) is an extension of (K, v) inside
of Z, then Z is also the absolute ramification field of (K ′ , v).
The next result follows from [1, Theorem (22.7)] (see also [7, Proposition 4.1]).
Proposition 2.4. The absolute ramification field of (K, v) is the maximal tame extension of (K, v).
An algebraic extension of a henselian field is called purely wild if it
is linearly disjoint to the absolute ramification field and thus to every
tame extension. In [7] it was shown that maximal purely wild extensions are field complements to the absolute ramification field. In the
case of Kaplansky fields, they are at the same time maximal immediate
algebraic extensions, and they are unique up to isomorphism. See the
cited paper for details.
Finally, we will need the following results about the absolute ramification field.
Lemma 2.5. Take an algebraic extension (L|K, v) of henselian fields.
Denote by Z the absolute ramification field of (K, v). Then (Z ∩ L, v)
is a tame extension of (K, v) maximal with respect to being contained
in L, and (L, v) is a purely wild extension of (Z ∩ L, v).
Proof. By Proposition 2.4, (Z ∩ L, v) is a tame extension of (K, v).
Assume that (E, v) is a tame extension of (K, v) inside of (L, v) which
is an extension of (Z ∩ L, v). Then (E, v) is also a tame extension of
(Z ∩ L, v). By Lemma 2.3, Z is also the absolute ramification field of
(Z ∩ L, v). Thus by Proposition 2.4, Z is the maximal tame extension
of (Z ∩ L, v). Hence E ⊆ Z and therefore, E = Z ∩ L, which proves
SUBFIELDS OF KAPLANSKY FIELDS
7
that (Z ∩ L, v) is a tame extension of (K, v) maximal with respect to
being contained in L.
Now we wish to prove that (L|Z ∩ L, v) is a purely wild extension.
Since Z is the absolute ramification field of (Z ∩ L, v), it suffices to
show that L|Z ∩ L is linearly disjoint from Z|Z ∩ L. But this follows
from the fact that according to [2, Theorem 5.3.3 (2)], Z|K is a Galois
extension.
3. Results on algebraically maximal Kaplansky fields
We start with the following characterization of algebraically maximal
Kaplansky fields:
Proposition 3.1. A valued field (K, v) is an algebraically maximal
Kaplansky field if and only if it is henselian and does not admit any
finite extension of degree divisible by p.
Proof. If char Kv = 0, then the assertion is trivial since in this case,
(K, v) is an algebraically maximal Kaplansky field if and only if it is
henselian, and 0 does not divide the degree of any finite extension.
Now we consider the case of char Kv = p > 0. Assume first that
(K, v) is an algebraically maximal Kaplansky field. Then in particular,
it is henselian. By Lemma 2.2, it is a defectless field, that is, every
finite extension (K ′ |K, v) satisfies
[K ′ : K] = (vK ′ : vK) · [K ′ v : Kv] .
Since (K, v) satisfies (K1), (vK ′ : vK) is not divisible by p. Because of
(K2) and (K3), the same holds for [K ′ v : Kv]. Hence also [K ′ : K] is
not divisible by p.
Now assume that (K, v) is henselian and does not admit any finite
extension of degree divisible by p. By the Lemma of Ostrowski, every
finite extension (K ′ |K, v) satisfies
[K ′ : K] = pν · (vK ′ : vK) · [K ′ v : Kv]
with ν ≥ 0. Since [K ′ : K] is not divisible by p, it follows that ν = 0.
This implies that (K, v) is a defectless field and thus also an algebraically maximal field. Further, if there were an element α ∈ vK not
divisible by p, then adjoining to K the p-th root of any element of K
having value α would generate an extension of K of degree p. This
shows that vK must be p-divisible, i.e., (K, v) satisfies (K1). Finally, if
there were a finite extension of Kv with a degree divisible by p, then it
could be lifted to an extension of K of the same degree, which should
not exist. This shows that Kv does not admit any finite extension of
degree divisible by p, so (K, v) satisfies (K2) and (K3).
8
F.-V. KUHLMANN
Proof of Proposition 1.2. The first assertion is trivial if char Kv = 0
since a relatively algebraically closed subfield of a henselian field is
again henselian. Suppose that Kv is not relatively algebraically closed
in Lv. Then there exists some ā ∈ Lv\Kv separable-algebraic over Kv.
Lift its minimal polynomial f¯ over Kv up to a polynomial f over K of
the same degree. Since ā is a simple root of f¯, Hensel’s Lemma yields
the existence of a root a ∈ L of f with av = ā. Then a is algebraic over,
but not in K, which contradicts our assumption that K is relatively
algebraically closed in L. Hence Kv is relatively algebraically closed in
Lv.
Let us now assume that char Kv = p > 0. Then we can use the
assertion of Proposition 3.1. Since L is perfect by Lemma 2.2, the
same holds for its relatively algebraically closed subfield K. Since the
algebraically maximal field (L, v) is henselian, the same holds for (K, v)
by Lemma 2.1.
Take any finite extension K ′ |K. Since K is perfect and relatively algebraically closed in L, K ′ |K is linearly disjoint from L|K. Therefore,
[K ′ : K] = [L.K ′ : L], which is not divisible by p. In view of Proposition 3.1, this proves that (K, v) is an algebraically maximal Kaplansky
field. In particular, Kv is perfect. As in the case of char Kv = 0 it is
now shown that Kv is relatively algebraically closed in Lv.
Proposition 3.2. Take an algebraically maximal Kaplansky field (L, v)
and a henselian subfield K such that L|K is algebraic. Denote the
absolute ramification field of (K, v) by Z and set Z0 := Z ∩ L. Then
the following assertions hold:
a) (Z0 , v) is a tame extension of (K, v) maximal with respect to being
contained in L,
b) (L, v) is a maximal purely wild extension of (Z0 , v),
c) (K3) also holds for (Z0 , v).
Proof. The assertions are trivial if char Kv = 0 (in which case Z0 = L),
so let us assume again that char Kv = p > 0.
It follows from Lemma 2.5 that (Z0 , v) is a tame extension of (K, v)
maximal with respect to being contained in L, and that (L|Z0 , v) is
purely wild. In fact, (L, v) is a maximal purely wild extension of (Z0 , v)
since by Lemma 2.2 it is a tame field. Since (L|Z0 , v) is purely wild,
the extension of the respective residue fields is purely inseparable. As
L satisfies (K3), the same is consequently true for Z0 .
Proposition 3.3. Take an algebraically maximal Kaplansky field (L, v)
and a subfield K such that L|K is algebraic. Then (L, v) contains:
a) a maximal immediate algebraic extension of (K, v),
SUBFIELDS OF KAPLANSKY FIELDS
9
b) a henselization (K h , v) of (K, v) and a maximal purely wild extension of (K h , v).
Proof. As before, the assertions are trivial if char Kv = 0, so let us
assume again that char Kv = p > 0. Since the algebraically maximal
field (L, v) is henselian, it contains a henselization K h of (K, v), and
this is an immediate algebraic extension of (K, v). Thus we may from
now on assume that (K, v) itself is henselian.
For the proof of part a), we take (K ′ , v) to be an immediate extension
of (K, v), maximal with respect to being contained in L. Then (K ′ , v) is
henselian since it is an algebraic extension of the henselian field (K, v).
We set Z0 := Z ∩L, where Z is the absolute ramification field of (K ′ , v).
Suppose that (K ′ , v) is not algebraically maximal. Then there exists
a nontrivial immediate algebraic extension (K ′ (a)|K ′ , v). Since it is
purely wild, it is linearly disjoint from the tame extensions Z|K ′ and
Z0 |K ′ , and it follows that Z0 (a)|Z0 is also linearly disjoint from Z|Z0 .
By Lemma 2.3, Z is also the absolute ramification field of Z0 , so this
shows that (Z0 (a)|Z0 , v) is purely wild. Since by Proposition 3.2, Z0
satisfies (K3), all maximal purely wild extensions of (Z0 , v) are isomorphic over Z0 by Proposition 3.2 of [7]. As (L, v) is a maximal purely
wild extension of (Z0 , v) by Proposition 3.2, it follows that Z0 (a) admits an embedding in L over Z0 . Hence there is some a′ ∈ L such that
Z0 (a) and Z0 (a′ ) are isomorphic over Z0 . This isomorphism induces an
isomorphism K ′ (a) ≃ K ′ (a′ ) over K ′ . Since (K ′ , v) is henselian, this
isomorphism preserves the valuation, so (K ′ (a′ )|K ′ , v) is a nontrivial
immediate algebraic extension contained in L. As this contradicts the
maximality of (K ′ , v), we find that (K ′ , v) must be algebraically maximal.
In order to prove part b), take (K ′ , v) to be a purely wild extension
of K, maximal with respect to being contained in K. Again, (K ′ , v)
is henselian. By the same proof as before, just replacing “immediate”
by “purely wild”, one shows that (K ′ , v) is a maximal purely wild
extension of (K, v).
Proof of Theorem 1.1. In order to prove the first part of the theorem,
we use Proposition 1.2 to replace L by the relative algebraic closure of
K in L. Then the assertions follow from Proposition 3.3.
For the proof of the second part of the theorem, assume that (L, v) is
maximal, and take an immediate extension (K ′ , v) of (K, v), maximal
with respect to being contained in L. By the first part of our theorem,
the relative algebraic closure of K ′ in L contains a maximal immediate
algebraic extension of (K ′ , v); by the maximality of K ′ , it must be
10
F.-V. KUHLMANN
equal to K ′ . Hence (K ′ , v) does not admit any nontrivial immediate
algebraic extensions.
Suppose that (K ′ , v) is not maximal. Then by [3, Theorem 4], (K ′ , v)
admits a pseudo Cauchy sequence without a limit in K ′ . This must be
of transcendental type, because if it were of algebraic type, then by [3,
Theorem 3], there would exist a nontrivial immediate algebraic extension of (K ′ , v), contradicting its algebraic maximality. Since the pseudo
Cauchy sequence we took in K ′ is also a pseudo Cauchy sequence in
(L, v) and (L, v) is maximal, [3, Theorem 4] shows the existence of a
limit a in L. It follows from [3, Theorem 2] that (K ′ (a)|K ′ , v) is an
immediate transcendental extension. As it is contained in L, this contradicts the maximality of (K ′ , v), We have now proved that (K ′ , v), is
a maximal immediate extension of (K, v).
References
[1] Endler, O. : Valuation theory, Springer, Berlin (1972)
[2] Engler, A.J. – Prestel, A.: Valued fields, Springer Monographs in Mathematics.
Springer-Verlag, Berlin, 2005
[3] I. Kaplansky, Maximal fields with valuations I, Duke Math. Journ. 9 (1942),
303–321
[4] F.-V. Kuhlmann, Additive Polynomials and Their Role in the Model Theory of
Valued Fields, Proceedings of the Workshop and Conference on Logic, Algebra,
and Arithmetic, held October 18-22, 2003. Lecture Notes in Logic 26 (2006),
160–203
[5] F.-V. Kuhlmann, Approximation of elements in henselizations, manuscripta
math. 136 (2011), 461-474
[6] F.-V. Kuhlmann, The algebra and model theory of tame valued fields, J. reine
angew. Math. 719 (2016), 1–43
[7] F.-V. Kuhlmann, M. Pank, and P. Roquette, Immediate and purely wild extensions of valued fields, manuscripta math. 55 (1986), 39–67
[8] Prieß-Crampe, S. : Der Banachsche Fixpunktsatz für ultrametrische Räume,
Results in Mathematics 18 (1990), 178–186
[9] Prieß-Crampe, S. : Reaching fixed points as limits in subspaces, to appear in
J. Geometry
[10] Ribenboim, P.: Théorie des valuations. Les Presses de l’Université de Montréal,
Montréal, 2nd ed. (1968)
[11] G. Whaples, Galois cohomology of additive polynomial and n-th power mappings of fields, Duke Math. J. 24 (1957), 143–150
[12] Zariski, O. – Samuel, P.: Commutative Algebra, Vol. II, New York–Heidelberg–
Berlin (1960)
Institute of Mathematics, University of Szczecin, ul. Wielkopolska
15, 70-451 Szczecin, Poland
E-mail address: [email protected]
| 0math.AC
|
Visualization of Hyperspectral Images Using
Moving Least Squares
Danping Liao
Siyu Chen
Yuntao Qian
arXiv:1801.06635v1 [cs.CV] 20 Jan 2018
College of Computer Science
College of Computer Science
College of Computer Science
Zhejiang University, HangZhou, China Zhejiang University, HangZhou, China Zhejiang University, HangZhou, China
[email protected]
[email protected]
[email protected]
Abstract—Displaying the large number of bands in a hyperspectral image (HSI) on a trichromatic monitor has been an
active research topic. The visualized image shall convey as much
information as possible from the original data and facilitate image
interpretation. Most existing methods display HSIs in false colors,
which contradict with human’s experience and expectation. In
this paper, we propose a nonlinear approach to visualize an input
HSI with natural colors by taking advantage of a corresponding
RGB image. Our approach is based on Moving Least Squares
(MLS), an interpolation scheme for reconstructing a surface from
a set of control points, which in our case is a set of matching
pixels between the HSI and the corresponding RGB image. Based
on MLS, the proposed method solves for each spectral signature a
unique transformation so that the nonlinear structure of the HSI
can be preserved. The matching pixels between a pair of HSI and
RGB image can be reused to display other HSIs captured by the
same imaging sensor with natural colors. Experiments show that
the output images of the proposed method not only have natural
colors but also maintain the visual information necessary for
human analysis.
I. I NTRODUCTION
The high spectral resolution of a hyperspectral image (HSI)
enables accurate target detection and classification. Human
interaction with HSIs is crucial for image interpretation and
analysis. Displaying an (HSI) is a challenging task since
an HSI contains much more bands than the capability of a
trichromatic monitor. One common solution is to consider
HSI visualization as a dimension reduction problem where
the number of bands is reduced to 3 for a color visualization.
Although the requirements of HSI visualization are task dependent, there are some common goals such as information
preservation, consistent rendering and natural palette [1].
A straightforward way for color representation is to select
three of the original bands and map them to the RGB space [2],
[3]. However, band selection methods only take the selected
spectrum into account. As a result, the information in other
bands is ignored. To preserve the information across the
full wavelength range, dimension reduction methods such
as independent component analysis (ICA) [4] and principal
component analysis (PCA) [5] have been applied to HSI
visualization. These linear transformation models are based on
the global distribution of data, thus the intrinsic characteristics
of HSIs such as the nonlinear and local structures are not well
preserved.
To preserve the nonlinear structure of HSIs, manifold learning methods such as isometric feature mapping (ISOMAP) [6]
and locally linear embedding (LLE) [7] have been applied.
To preserve the pairwise spectral distance in HSIs, Mignotte
proposed a nonstationary Markov random field model [8] and
later extended this approach to address the tradeoff between
the preservation of spectral distances and the separability
of features [9]. Kotwal and Chaudhuria used a nonlinear
bilateral filter with edge preserving characteristic to compute the weights of bands at each pixel for band image
fusion [10]. These nonlinear approaches demonstrate excellent
performance in preserving the intrinsic information of HSIs.
While most existing methods try to convey as much information as possible from the original data to the visualized
image, most of them display HSIs in false colors. In general
false color visualizations can be hard to interpret when object colors are very different from their natural appearance.
Moreover, most data-adaptive methods suffer the problem of
“inconsistent rendering”, i.e., the same objects/materials being
displayed with very different colors in different images, which
also hinders the interpretation of HSIs. Therefore, “natural
palette” and “consistent rendering” gradually become two important criteria for evaluating the quality of HSI visualization.
To generate natural looking visualizations, Jacobson et
al. [1] proposed a linear spectral weighting function to fuse
all the bands in an HSI. The weighting function is a stretched
version of the CIE 1964 tristimulus color matching functions
in the visible range. As the weighting function is fixed, this
model can generate consistent colors. However, compared
with other data adaptive methods, the model preserves less
information for each specific images. Another approach for
displaying HSIs with natural colors was proposed by Connah
et al. [11], which takes advantage of a corresponding RGB
image. To obtain a natural-looking visualization, the gradient
of each pixel in the visualization is constrained to be the same
with that of its matching pixel in the corresponding RGB
image. This method requires pixel-wise matching between the
HSI and the corresponding RGB image, which is sometimes
difficult to achieve especially when the HSI and the RGB
image are acquired by sensors mounted on different platforms,
such as the airplane and satellite, with different geometrical
distortions. Liao et al. [12] proposed an approach to display
HSIs with natural colors based on manifold alignment between
an HSI and a high resolution RGB image. The HSI and
the RGB image are firstly projected to a common space
by manifold alignment, Then the HSI is projected from the
common space to the RGB space to generate a final image
that has similar colors to the corresponding RGB image. One
advantage of the approach is that it only requires a small
number of matching pixels. However, the estimated mapping
function from the spectral space to RGB space is linear, which
shows some limitation in representing the local structure of the
HSI.
In this paper, we propose a nonlinear method to display
HSIs with natural colors. Our goal is to produce an output
image which not only maintains the information of the input
HSI but also has similar colors to a corresponding RGB image.
Our method is based on Moving Least Squares (MLS), which
is a local and nonlinear interpolation scheme for reconstructing
a surface from a set of control points [13]. In the task of HSI
visualization, we use MLS to generate a color representation
of an input HSI through a set of matching pixels between the
HSI and the corresponding RGB image. Compared with the
methods that use a single transformation to map all the pixels
in the HSI to the RGB space, the proposed method estimates a
transformation for each pixel individually by solving a unique
weighted least squares problem for each spectral signature.
The strong locality of MLS allows our model to preserve the
local structure of HSIs very well.
One advantage of the proposed method is that the matching
pixels between a pair of HSI and RGB image can be reused
to visualize other similar HSIs acquired by the same imaging
sensor. Since the matching pixels set up the mapping relation
between the high dimensional spectral space and the RGB
space, the same/similar objects in different images are presented with consistent colors. Another advantage is that our
model does not require precise pixel-wise matching between
the HSI and RGB image. Instead, it only requires a small
number of matching pairs that represent the same/similar
objects, which greatly increases the flexibility of the approach.
The rest of the paper is organized as follows. Section II gives
a brief description of MLS and some related works. Section III
presents the MLS based HSI visualization approach. Some
application scenarios are also discussed in this section. The
experimental results are described in Section IV. Finally,
conclusions are drawn in Section V.
II. M OVING L EAST S QUARES
Moving Least Squares (MLS) is a scattered point interpolation technique for generating surfaces [13]. Within the
framework of MLS, a continuous function can be reconstructed from a set of point samples (control points). The
technique is attractive since the surface is reconstructed by
local computations and it generates a surface that is smooth
everywhere.
Based on the distances between the control points and an
input point, MLS solves a unique weighted least squares to
evaluate each input point individually. Let X be a set of control
points and Y the given function values of the control points.
Given an arbitrary point x to be evaluated, MLS aims to find
a transformation fx to map it to the target space. The optimal
function fx is acquired by minimizing the following weighted
least squares
X
wk | fx (Xk ) − Yk |2
(1)
k
where Xk is the kth control point and Yk is its given function
value. The summation is taken over all the control points. The
weights wk have the form
wk =
1
,
|x − Xk |2α
(2)
which are larger for the control points that are closer to x.
The name Moving Least Squares comes from the fact that the
weights wk change depending on the point x to be evaluated.
Therefore, MLS obtains a different transformation fx for
different x.
The MLS approach has been successfully applied in image
deformation [14], surface reconstruction [15], and image superresolution and denoising [16]. Recently, Hwang et al. [17]
used MLS to transfer the colors from a source RGB image to
a target RGB image. The matching pixels between the input
images serve as the control points for interpolation. For each
RGB value in the target image, the method computes an affine
function to map it to a desired color. The goal for estimating
the optimal affine function for each color is formalized as
minimizing the summation of a set of weighted least squares
over the control points.
In this paper, we apply MLS to display HSIs in natural
colors. Our work shares the same idea with [17] in that we
try to “transfer” the colors of a corresponding RGB image
to the visualization of an HSI. The main contribution of this
work is that we further formalize the objective function in
Eq.(1) in matrix form, and derive a solution in the form of
matrix multiplication, which can be easily parallelized and
be computed more efficiently compared to the solutions in
summation form.
III. H YPERSPETRAL I MAGE V ISUALIZATION BASED ON
MLS
We introduce an MLS based mechanism to display an HSI
with natural colors. The key idea is to generate an output
3D image through a set of matching pixels between the HSI
and a corresponding RGB image. In this section we first
introduce the MLS algorithm for HSI visualization, followed
by a coarse-to-fine searching process to find matching pixels.
Finally, two other visualization scenarios are discussed.
A. Moving Least Squares Framework
Let U ∈ Rp×n and V ∈ R3×n be a set of matching pixel
values from an HSI and an RGB image respectively, where
p is the number of bands in the HSI, and n is the number
of matching pixel pairs. These matching pixels will serve as
the control points to construct the output color image. Let
Uk and Vk denote the kth pair of matching pixels. Given a
high dimensional spectral signature x in the HSI, we aim to
find a transformation fx to transform it to the 3D RGB space.
Within the framework of MLS, the goal is to find the fx that
minimizes the following weighted least squares
G(fx ) =
n
X
wk | fx (Uk ) − Vk |2 .
(a) The 50th band of the HSI
(3)
k=1
As the spectral angle distance (SAD) is commonly used to
measure the distance between a pair of pixels in an HSI, the
weight wk for an input x is defined as
1
(4)
wk =
SAD(x, Uk )
where
SAD(x, Uk ) = arccos(
x · Uk
).
kxkkUk k
where Fx is a p×3 matrix representing a linear transformation
and bx is a 3 × 1 vector representing a translation. Taking the
partial derivative of Eq.(3) with respect to bx , we get
bx = v − FTx u
(7)
with u and v being the weighted centroids:
P
P
k wk Uk
k wk Vk
u= P
, v= P
.
w
k
k
k wk
(8)
Eq.(6) can then be rewritten as
fx = FTx (x − u) + v
(10)
k=1
where U and V are acquired by subtracting u and v from each
column of U and V respectively. Eq.(10) can be rewritten in
matrix form as
T
G(Fx ) = tr(FTx UWU Fx − 2FTx UWV + VWV )
(11)
where W is an n × n diagonal matrix with Wkk = wk , and
tr(·) denotes the trace of a matrix. This objective function is
convex and differentiable. Its derivative with respect to Fx is
∂G
T
= 2UWU Fx − 2UWV.
∂Fx
(b)
(d)
(e)
(c)
(f)
Fig. 2. Three HSIs and their corresponding RGB images. (a) The 50th band
of the Moffett Field. (b) The first band of D04. (c) The first band of G03. (d)
The RGB image of Moffett Field. (e) The RGB image of D03. (f) The RGB
image of G03.
The optimal Fx that minimizes the objective function can be
obtained by
T
T
Fx = (UWU )−1 UWV .
(13)
When Fx and bx are acquired, a spectral signature x is
transformed to an RGB color by Eq.(6). A natural-looking
representation of the HSI can be obtained after mapping all
the pixels in the HSI to the RGB space.
B. Finding the Matching Pixels Between an HSI and an RGB
Image
wk |FTx Uk − Vk |2
T
(a)
(9)
and Eq.(3) becomes
n
X
Fig. 1. An HSI and its corresponding RGB image of the Washington D.C.
mall.
(5)
Note that weights wk change depending on the spectral
signature x to be evaluated, thus the proposed method solves
a different weighted least problem for different x. Therefore,
the optimal transformation fx varies for each x, which allows
our method to preserve the local and nonlinear structure of
HSIs.
In the image deformation work [14], a rigid transformation
fx was chosen since it maintains the geometric properties with
less degrees of freedom. In the color transfer task [17], fx was
set to an affine transformation to model different elements of
the color deformation. In this work, we also choose affine
transformation to model the nonlinearities in HSIs. Thus fx is
represented by
(6)
fx (x) = FTx x + bx
G(Fx ) =
(b) The corresponding RGB image
(12)
Image registration is a common way to find the matching
pixels between two images. However, as an HSI and the
corresponding RGB image are often captured by different
cameras with different angles and image distortions, it is very
difficult to obtain a precise image registration. In this paper
we propose a two-step coarse-to-fine searching process. Firstly,
the HSI and the RGB image are registered via a homography
transform to get a coarse matching. Scale-invariant feature
transform (SIFT) is widely used to detect the matching key
points between images due to its robustness to changes in
scale, orientation, and illumination [18]. In this paper, we first
extract the SIFT key points from each band image of the HSI
and the RGB image, and then select a few of the most similar
(a) LP band selection
(a) LP band selection.
(b) Manifold alignment
(c) Stretched CMF
(d) Bilateral filtering
(e) Bicriteria optimization
(f) MLS
(b) Manifold alignment
(c) Stretched CMF
(d) Bilateral filtering
Fig. 4. Visual comparison of different visualization approaches on the Moffett
Field data set.
(e) Bicriteria optimization
(f) MLS
TABLE I
P ERFORMANCE COMPARISON ON ENTROPY.
Visualization methods
LP Band Selection
Manifold Alignment
Stretched CMF
Bilateral Filtering
Bicriteria Optimization
MLS
Washington
6.59
7.40
3.89
3.31
5.75
7.00
Moffett
6.36
6.89
5.28
5.87
6.00
6.30
G03
4.73
6.64
6.31
6.05
5.73
7.01
D04
5.58
7.02
7.13
7.27
6.06
7.39
Fig. 3. Visual comparison of different visualization approaches on the
Washington DC Mall data set.
pairs as the matching key points. With the matching key points
we estimate a homography transform using the Direct Linear
Transformation (DLT) algorithm [19].
To obtain a more accurate matching, for each pixel in the
HSI, we search for its final matching pixel in the neighborhood
window of the current matching one with SIFT distance. In
our experiments, the size of the window is set to 9×9. To find
a set of matching pairs, we randomly select 1% of pixels in
the HSI and then find their matching pixels in the RGB image
using the above process.
C. Other Visualization Scenarios
Reusing the Obtained Matching Pixels to Visualize Similar
HSIs: Based on the fact that the same types of objects shall
have similar spectral responses with the same sensor, the
matching pixels between a pair of HSI and RGB image can be
directly applied to visualize other semantically similar HSIs
captured by the same imaging sensor. This scenario is very
helpful for a quick overview of a batch of similar HSIs since
the matching process between the HSI and the RGB image
only needs to be done once.
Displaying an HSI Aided by an RGB Image Captured on
a Different Site: The MLS-based visualization method only
requires a small number of matching pixel pairs for color
constraint. Thus in the absence of a corresponding RGB image
captured on the same site as the input HSI, a semantically
similar RGB image can be used to help visualization. In this
case, interactive tools can be developed for users to manually
select a number of matching pixels between the two images.
TABLE II
P ERFORMANCE COMPARISON ON RMSE.
Visualization methods
LP Band Selection
Manifold Alignment
Stretched CMF
Bilateral Filtering
Bicriteria Optimization
MLS
Washington
68.35
58.41
92.46
102.43
98.47
46.85
Moffett
49.93
27.30
58.50
56.70
57.94
19.07
G03
95.19
32.88
44.09
78.54
69.60
16.45
D04
83.35
23.08
28.70
61.14
75.70
29.78
(a) LP band selection
(b) Manifold alignment
(a)
(b)
(c) Stretched CMF
(d) Bilateral filtering
(c)
(d)
Fig. 7. Visualizations of similar HSIs generated by reusing a set of matching
pixels. (a) The first band of F02. (b) The visualization of F02. (c) The first
band of G02. (d) The visualization of G02.
(e) Bicriteria optimization
(f) MLS
Fig. 5. Visual comparison of different visualization approaches on the G03
data set.
(a)
(a) LP band selection
(b) Manifold alignment
(b)
(c)
Fig. 8. Visualization of D04 generated with the aid of an RGB image captured
on a different site. (a) The first band of D04. (b) An RGB image of different
site. (c) The visualization of D04.
A. Hyperspectral Imaging Data
(c) Stretched CMF
(e) Bicriteria optimization
(d) Bilateral filtering
(f) MLS
Fig. 6. Visual comparison of different visualization approaches on the D04
data set.
IV. E XPERIMENTS AND D ISCUSSIONS
We compare our results against several recently proposed visualization approaches including band selection using LP [3],
stretched CMF [1], bilateral filtering [10], bicriteria optimization [9] and manifold alignment [12]. We also give the
experimental results on two other application scenarios.
Four HSI data sets were used to evaluate the performance
of different visualization methods. One remote sensing HSI
data set was taken over the Washington D.C. mall by the
Hyperspectral Digital Imagery Collection Experiment (HYDICE) sensor. The data consists of 191 bands after noisy
bands removed. Each band image has a size of 1208 × 307.
Fig. 1(a) shows its 50th band image. Fig. 1(b) shows the
registered corresponding RGB image obtained from Google
Earth. Another remote sensing HSI data was captured by
the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS)
over Moffett Field, California. The data consists of 224 bands,
and each band has a size of 501 × 651. Its 50th band image
is shown in Fig. 2(a). The corresponding RGB image shown
in Fig. 2(d) was also obtained from Google Earth.
The other two HSI data sets named “G03” and “D04”
were captured by a ground-based OKSI hyperspectral imaging
system mounted with a tunable LCTF filter. Each data set has
18 bands ranging from 460nm to 630nm at 10nm interval.
The size of each band is 960 × 1280. Fig. 2(b) and Fig. 2(c)
show the first band images of D04 and G03 respectively.
Their corresponding RGB images were taken by a Nikon D60
SLR digital camera and are shown in Fig. 2(e) and Fig. 2(f)
respectively.
B. Comparison of Visualization Methods
Fig. 3 shows the visual comparison of different visualization
approaches on the Washington D.C. mall data. It can be seen
that output image of the proposed method is very easy to
understand since it has very natural colors similar to that of
the corresponding RGB image in Fig. 1(b). Also, the details in
the HSI are clearly represented. The comparative results on the
Moffett field, G03 and D04 data sets are provided in Fig. 4-6
respectively. Likewise, in these experiments our approach not
only shows great performance on displaying the HSIs with
natural colors but also preserves the local structure of the HSIs.
Quantitative assessment of HSI visualization does not have
a universally accepted standard. In this paper, we adopt two
quantitative metrics: entropy of the output image [10] and root
mean square error (RMSE) between the output image and the
corresponding RGB image [4].
Entropy is a statistical measure of randomness that can be
used to characterize the texture of an image. An image with
higher entropy generally contains richer information. Table I
shows the comparative results in terms of entropy. It can be
found that the results of the proposed method have relatively
large entropy compared to other approaches, which suggests
that the proposed method is able to preserve the information
of the HSI for analysis.
The RMSE between the output image and the corresponding
RGB image is a straightforward way to evaluate the “naturalness” of the output image is [4]. The comparative RMSE
results are given in Table II. We can see that the output images
of the proposed method have smaller RMSE compared to
other methods in most cases, which shows clearly that the
proposed method is an excellent approach for displaying HSIs
with natural colors.
C. Generalizing the Matching Pixels to Visualize Other HSIs
Once we have obtained the matching pixels between an
HSI and a corresponding RGB image, the matching pixels
can be reused to visualize semantically similar HSIs captured
by the same hyperspectral imaging sensor. In this experiment,
the matching pixels between the G03 data set and its corresponding RGB image are applied to visualize three HSI
data sets named “F02” and “G02”, which were captured by
the same imaging sensor. Fig. 7 shows the first band images
and the visualizations of these HSIs. It can be seen that the
proposed method has nice performance in displaying these
HSIs with natural colors. Besides, the same classes of objects
are presented with consistent colors across the three images.
D. Displaying an HSI with an RGB Image of a Different Site
In this experiment, we visualize the D04 data set (see
Fig. 8(a)) aided by an RGB image captured on a different
site (see Fig. 8(b)). Fourteen matching pixel pairs between
the two images were manually selected to serve as control
points. Fig. 8(c) shows the color representation of D04. We
can see that this image has natural colors and maintains the
basic information in the HSI, which shows that our algorithm
performs well with a limited number of matching pixels
between images captured on different sites.
V. C ONCLUSIONS
We have presented a novel approach to visualize HSIs with
natural colors based on the MLS interpolation scheme. The
key idea is to construct a 3D color representation of an input
HSI through a set of matching pixels between the HSI and
a corresponding RGB image. Our method solves for each
spectral signature a unique transformation so that the nonlinear
structure of the HSI can be preserved. One advantage of the
proposed method is that the matching pixels between an HSI
and an RGB image can be directly applied to visualize similar
HSIs captured by the same imaging sensor. Our method also
works well when the HSI and the corresponding RGB image
are captured on different sites. In the future, we plan to take the
spatial information into account to better preserve the spatial
structure of HSIs.
R EFERENCES
[1] N. P. Jacobson and M. R. Gupta, “Design goals and solutions for display
of hyperspectral images,” IEEE Trans. Geosci. Remote Sens., vol. 43,
no. 11, pp. 2684–2692, 2005.
[2] S. Le Moan, A. Mansouri, Y. Voisin, and J. Y. Hardeberg, “A constrained
band selection method based on information measures for spectral image
color visualization,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 12,
pp. 5104–5115, 2011.
[3] H. Su, Q. Du, and P. Du, “Hyperspectral image visualization using band
selection,” IEEE J. Select. Topics Appl. Earth Observ. Remote Sens.,
vol. 7, no. 6, pp. 2647–2658, 2014.
[4] Y. Zhu, P. K. Varshney, and H. Chen, “Evaluation of ica based fusion of
hyperspectral images for color display,” in Proc. Int. Conf. Inf.Fusion,
2007, pp. 1–7.
[5] V. Tsagaris, V. Anastassopoulos, and G. A. Lampropoulos, “Fusion of
hyperspectral data using segmented pct for color representation and
classification,” IEEE Trans. Geosci. Remote Sens., vol. 43, no. 10, pp.
2365–2375, 2005.
[6] C. M. Bachmann, T. L. Ainsworth, and R. A. Fusina, “Exploiting
manifold geometry in hyperspectral imagery,” IEEE Trans. Geosci.
Remote Sens., vol. 43, no. 3, pp. 441–454, 2005.
[7] M. M. Crawford, L. Ma, and W. Kim, “Exploring nonlinear manifold
learning for classification of hyperspectral data,” in Opt. Remote Sens,
2011, vol. 3, pp. 207–234.
[8] M. Mignotte, “A multiresolution markovian fusion model for the color
visualization of hyperspectral images,” IEEE Trans. Geosci. Remote
Sens., vol. 48, no. 12, pp. 4236–4247, 2010.
[9] ——, “A bicriteria-optimization-approach-based dimensionalityreduction model for the color display of hyperspectral images,” IEEE
Trans. Geosci. Remote Sens., vol. 50, no. 2, pp. 501–513, 2012.
[10] K. Kotwal and S. Chaudhuri, “Visualization of hyperspectral images
using bilateral filtering,” IEEE Trans. Geosci. Remote Sens., vol. 48,
no. 5, pp. 2308–2316, 2010.
[11] D. Connah, M. S. Drew, and G. D. Finlayson, “Spectral edge image
fusion: Theory and applications,” in European Conference on Computer
Vision (ECCV). Springer, 2014, pp. 65–80.
[12] D. Liao, Y. Qian, J. Zhou, and Y. Y. Tang, “A manifold alignment
approach for hyperspectral image visualization with natural color,” IEEE
Trans. Geosci. Remote Sens., vol. 54, pp. 3151–3162, 2016.
[13] P. Lancaster and K. Salkauskas, “Surfaces generated by moving least
squares methods,” Mathematics of computation, vol. 37, no. 155, pp.
141–158, 1981.
[14] S. Schaefer, T. McPhail, and J. Warren, “Image deformation using
moving least squares,” in ACM transactions on graphics (TOG), vol. 25,
no. 3. ACM, 2006, pp. 533–540.
[15] S. Fleishman, D. Cohen-Or, and C. T. Silva, “Robust moving leastsquares fitting with sharp features,” in ACM Transactions on Graphics
(TOG), vol. 24, no. 3, 2005, pp. 544–552.
[16] N. K. Bose and N. A. Ahuja, “Superresolution and noise filtering using
moving least squares,” IEEE Trans. Image Process., vol. 15, no. 8, pp.
2239–2248, 2006.
[17] Y. Hwang, J.-Y. Lee, I. S. Kweon, and S. J. Kim, “Color transfer using
probabilistic moving least squares,” in IEEE Conf. Computer Vision and
Pattern Recognition (CVPR), 2014, pp. 3342–3349.
[18] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,”
Int. J. Comput. Vis., vol. 60, no. 2, pp. 91–110, 2004.
[19] R. Hartley and A. Zisserman, Multiple view geometry in computer vision.
Cambridge university press, 2003.
| 1cs.CV
|
Composition by Conversation
Donya Quick
Stevens Institute of Technology
[email protected]
arXiv:1709.02076v1 [cs.SD] 7 Sep 2017
ABSTRACT
Most musical programming languages are developed
purely for coding virtual instruments or algorithmic compositions. Although there has been some work in the domain of musical query languages for music information retrieval, there has been little attempt to unify the principles
of musical programming and query languages with cognitive and natural language processing models that would
facilitate the activity of composition by conversation. We
present a prototype framework, called MusECI, that merges
these domains, permitting score-level algorithmic composition in a text editor while also supporting connectivity to
existing natural language processing frameworks.
1. INTRODUCTION
We have reached an age where natural language interfaces
are increasingly available for our technology. These interfaces provide opportunities to interact more naturally with
our devices. For example, it is becoming common for people to speak directly to mobile devices to accomplish tasks
with no human on the other end of the interactions. However, to date there has been little work on the support of
conversation and interactions involving musical concepts
and scores.
Deep learning with neural networks is currently a popular strategy in the domain of natural language processing
(NLP), and has been applied to developing dialog systems
where a machine attempts to carry on a conversation with
a human [1]. However, neural nets remain difficult to analyze and extend once trained. In contrast, the classic approach of explicit knowledge representation symbolically
represents ideas and thought processes that can be directly
analyzed and extended, and provides a natural framework
for explaining decisions.
Falling under the general strategy of knowledge representation, elementary composable ideas, or ECIs, are simple,
atomic concepts that can be combined, or composed, to
form more complex concepts and relationships. For example, there is some evidence to suggest that “surface” is one
such concept that infants have and use to build other concepts, such as collections of surfaces forming objects and
internal spaces [2]. This style of representation has been
proposed to underly human natural language processing
tasks such as describing and referring to items and events
in visual environments [3, 4].
Clayton T. Morrison
University of Arizona
[email protected]
Programming languages already follow an ECI-like
model: there are primitives, such as integer and floatingpoint numbers, and a collection of basic operations and
data structures over those entities that are used to represent more complex concepts. Object-oriented programming seeks to model computational problems much in the
way we view our environment as composed of objects,
each of which consists of parts that interact and can be
manipulated in particular ways. However, not all ideas
fit neatly into this approach and there are many languages
that use very different strategies, such as functional programming languages. Additionally, appropriate language
features for a computer are not necessarily straightforward
derivations from verbal descriptions—there is a reason why
programming a computer is a complex skill that requires
specialized training.
We propose a basic collection of ECIs for music with an
accompanying Python implementation called MusECI 1 .
In this paper we describe ECI-style musical primitives and
structures for grouping these primitives. We also provide
basic operations to facilitate conversational manipulation
of music through a query-then-operation pattern of interaction. This lays the foundation for eventually building
larger platforms that would facilitate real-time conversations with machines about music. The goal of MusECI is
not to replace traditional score editing software, but rather
to enable the creation of new systems to augment those
environments and open new avenues for human-computer
interaction through music.
2. PROGRAMMING LANGUAGES
A central goal of the MusECI project is to create a language for score-level music representation that can be used
in an algorithmic or automated composition setting while
still adhering closely to models of music cognition and
expression in natural language. We adopt the constraints
that the data structures used to represent music both have
a straightforward interpretation for editing and playback
and that they can also be easily queried to identify features based on natural language descriptions and manipulated according to instructions.
Many musical programming languages focus on the
sound, or signal-level aspects of music generation. Examples of these include Csound [5], SuperCollider[6] 2 ,
and Max/MSP[8]. While these languages are extremely
1
Copyright: c 2017 Donya Quick et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License 4.0
Unported, which permits unrestricted use, distribution, and reproduction
in any medium, provided the original author and source are credited.
Our implementation is available at musica.ml4ai.org/.
SuperCollider also represents another language paradigm known as
live coding[7], which allows compilation and execution of fragments of
a larger program on demand. However, while gaining traction in music
performance, this kind of language paradigm has not been interfaced with
NLP systems.
2
useful in the context of music creation, they do not offer
a rich variety of representations for higher level features
in music. On the other extreme are languages that focus
on creating visually appealing musical scores. Languages
such as LilyPond [9] focus on type-setting aspects of a
score, with emphasis on details of visual presentation and
formatting and less on relationships between musical features that would be useful for modeling music cognition.
We seek a representation between the extremes of visual
detail and sound synthesis, focusing on common natural
language descriptors people use to talk about music: note,
melody, chord, phrase, and so on.
Other languages for representing score-level information
favor grouping of musical features, such as melodies and
harmonies, over the details of their visual presentation.
MusicXML[10] is one such language. It is a markup language that describes score-level features in music in an
application-independent way such that a composition can
be passed between different score rendering programs. Musical features are highly nested and groupings occur by visual features in the score. Unlike MIDI format, time is
measured in MusicXML by measures and beats, which are
relative to a particular tempo. MusECI’s implementation
supports MusicXML as an input format to specify musical
structures; the method for converting MusicXML’s standard into our musical ECI data structures is described in
section 3.3.
2.1 Algorithmic Score Analysis and Generation
None of the languages discussed so far support operations
for searching or manipulating the music. MuSQL [11]
is a structured query language for operating on musical
databases. It permits creation of, selection from, and alteration to musical structures. Its representations of musical
features are generalized to address multiple kinds of pitch
data, although time representations are limited to seconds—
there is no method for describing more abstract metrical
structures and containers such as measures and beats.
MuSQL uses standard SQL-style commands, which, although intuitive to programmers with the right training, are
not necessarily easily derived from natural language statements. One of MusECI’s goals is to bridge that gap.
Music21 [12] is a Python library that has become popular for hypothesis testing on large musical copora. It is
able to read a variety of formats, including MIDI and MusicXML, and uses an object-oriented style that is very close
to the structures used by MusicXML. Music21 has been
used in several computational musicology projects, including analysis of the large Bach chorales MusicXML corpus.
However, Music21 is primarily structured based on music
theoretic concepts, which, although potentially useful to
analyzing questions about music cognitive models, is not
always natural as a vehicle for mapping from natural language descriptions to operations on music.
Representations for score-level structures in programming
languages intended for algorithmic composition typically
revolve around the two classic computer science data structures of lists and trees. The Jython Music library [13],
a Python version of JMusic [14], favors list representations for constructing melodies and chords using notes and
rests. Melodies are lists of notes and rests with the assumption that one element’s onset is determined by the
previous element’s end time. Chords are also represented
as lists of notes having the same start time and duration.
Python’s support for nested lists allows for chords to be
created within a larger sequential, melodic structure. While
useful in pedagogical and algorithmic composition settings,
this list-based approach makes representation of features
such as arpeggiated chords tricky and its style of nesting
musical concepts like parts and phrases is very rigid.
In the Euterpea library [15] for representing musical structures in Haskell, musical ideas are represented as binary
trees, where leaves are notes and rests that store duration
information but not onset time. The onset for a particular leaf is determined by its placement in the larger tree,
whose intermediate nodes consist of two types of connectors: parallel and sequential. Sequentially composed subtrees are interpreted as following each other in time, and
those composed in parallel are assumed to have the same
start time (but not necessarily the same end time). Any
pattern of notes that can be represented in the standard
MIDI file format (which follows a list-based representation
of time-stamped events) can also be represented as one of
Euterpea’s trees. 3 However, it is frequently the case that
a given musical concept can have more than one tree representation, making comparison of equality between tree
structures and identification of shared features problematic. It may also not be cognitively natural to nest relationships in a strictly binary way—sometimes we may think of
three or more items as having the same level of hierarchy,
more like a list. Here, we use an n-ary tree representation
(each node essentially being a list) to mitigate this issue in
developing a framework to connect to NLP systems.
The structures and operations we use here are also very
similar to some representations used in Kulitta [16], which
is a Haskell-based framework for automated composition
in multiple styles. This lends our implementation to being
useful for automated composition tasks, while also allowing querying of data structures in a NLP-compatible way.
2.2 Modeling Natural Language Semantics
A key constraint for MusECI is that it must accommodate
the semantics of how music is described in natural language. To accomplish this, we take inspiration from computational languages designed for representing the natural
language semantics of visual and physical environments,
such as VoxML [4].
The VoxML modeling language encodes semantic knowledge of real-world objects represented as three-dimensional
physical models along with events and attributes that are
related and change through actions and interactions. The
framework encodes background knowledge about how interactions between objects affect their properties, allowing simulation to fill in information missing from language
input. With this representation, natural language expressions describing scenes, events, and actions on objects can
be specified in enough detail to be simulated and visualized. We similarly seek a language in which musical
concepts can be expressed in building-blocks that stand in
close correspondence to natural language expressions. The
3 This is true for notes as they would appear on a score. Performance
features, such as pitch bends and aftertouch, do not have a corresponding
representation in Euterpea.
framework must capture enough detail to support expressing background music knowledge typically assumed in order to interpret natural language descriptions of music as
well-defined musical structures that can be represented as
scores and played—a kind of “music simulation”.
3. MUSICAL PRIMITIVES
Our lowest level of representations for Musical ECIs are
called musical primitives: concepts that represent fundamental musical units. Because we are focused on music
at the level of a paper score, our primitives are based on
features that appear on scores: notes and rests.
3.1 Symbolic Primitives
N (P itch(0, ), , Onset(2, ), )
MusECI’s atomic musical concepts are related to
pitch and duration, although these, by themselves, cannot
be directly represented on a musical score. Currently we
only consider Western tonal systems based on the chromatic scale and therefore represent pitch information using
integers. Since some of the atomic concepts are best represented as numbers, we make use of two number types: Int
and F loat. We begin with the following two pitch-related
atomic primitives, both of which are simply musical interpretations of numbers:
P itchClass(Int)
Octave(Int)
(1)
These then yield the following derived primitives, which
are built on atomic primitives.
P itch(P itchClass, Octave)
ScaleIndex(Int, Contexts)
(2)
where Contexts refers to a collection of environmental
information and other labels used for resolving ambiguity.
For ScaleIndex, the Context would need to include information about the current Scale. The pitch number for a
pitch can be computed by P itchClass + 12(Octave + k),
where k has different standards in different areas of the literature. 4
The following two atomic primitives are building-blocks
for metrical information:
Beat(F loat)
M easure(Int)
We denote the set of all such symbols as P . For brevity,
in examples in this paper we will use N to refer to the
Python-style constructor for N ote and, in later sections,
both pitch and metrical values will be shown only as tuples.
An important difference between our approach and many
other musical representations it that we do not require values to be declared for all properties of these concepts. The
high degree of optional information in these constructs is
needed to reflect the various ways in which we speak about
music, which can involve many levels of abstraction and
also may be ambiguous. For example, using “ ” to indicate a blank field, we may capture the concept of “the C in
measure 3” as follows 5 :
(6)
This specifies a template that can be matched against a
musical representation with more concrete information to
identify the particular note being referred to. In MusECI,
we represent “ ” using Python’s None value. This variable degree of information is useful for specifying queries
over musical structure and also for communicating musical ideas at a variety of levels of abstraction that may not
have a uniform level of detail over the entire score.
The Contexts field of Note and Rest indicates additional
labels or other contextual information that may be useful
for querying. Our implementation also supports volume as
a note property, but it is omitted from the descriptions and
examples in this paper.
3.2 Connecting Primitives
MusECI uses Euterpea-inspired representations for connecting musical structures in sequence and in parallel. To
minimize the issue of having multiple tree structures associated with Euterpea’s binary trees, we use n-ary trees.
This permits notes in a melody or chord that have the same
level of conceptual hierarchy to appear at the same level
within the representation. We also define normal forms
for grouping structures from a score format such as MusicXML. The set of all possible sequential and parallel
structures, which we will call M , is defined recursively.
Seq({M |P }+, Contexts)
P ar({M |P }+, Contexts)
(7)
(3)
Derived primitives also exist for metrical information:
We will use the square bracket notation, [...], to denote
lists of items. Therefore, “the three note melody in part B”
could be represented as:
Onset(M easure, Beat)
Duration(Beat)
Seq([N ( ), N ( ), N ( )], “Part B”)
(4)
Pitch and metrical primitives are composed to produce
the concepts of Note and a Rest, which are similar to the
lowest levels of representation in Euterpea, JythonMusic
and Music21. We define the set of all notes and rests to
include the following primitives:
N ote(P itch, Duration, Onset, Contexts)
Rest(Duration, Onset, Contexts)
4
(5)
Octave zero on a piano keyboard is not standardized and varies between musical programming languages. Sometimes k = 0 such that
(C,0) is pitch number 0, but it is also common to have k = 1 and occasionally even k = −1 depending on the particular application.
(8)
We will use the term Music value to refer to any value that
is in M ∪ P . In other words, a Music value is some entity
that could be drawn on a score, whether a single note, a
rest, a melody, a chord, etc. Later we describe a collection
of operations that can be applied to any of these values.
3.3 MIDI and MusicXML Conversion
MusECI is able to parse both MIDI and MusicXML files
into a normal form with the representations discussed so
5 In MusECI, measure and beat numbers both index from zero. Therefore, “measure 3” produces a value of 2.
far. We use the following algorithm to convert MIDI and
MusicXML information into our system’s representations:
1. Greedily group Notes with the same onset and duration with P ar.
2. Greedily group temporally adjacent structures (potentially including chords from step 1) with Seq.
3. Group any remaining temporally separated, but still
sequential structures with Seq, using Rests to fill
temporal gaps.
4. Group any leftover items under a global P ar.
Importantly, this normal form for reading MusicXML is
not the only way to represent musical features. Other normal forms are possible, and we hope to improve our normal forms through learning in later work, such that nested
structures within melodies and phrases are identified in
genre-specific ways.
or according to scale degrees depending on the tonal
context given.
6. INTEGRATION WITH TRIPS PARSER
In our work we use the TRIPS natural language parser.
TRIPS is a best-first bottom-up chart parser that integrates
a grammar and lexicon to encode both syntactic and semantic features. These features are used to disambiguate
and produce a logical form representation that captures the
semantic roles of terms in the utterance [17]. Figure 1
shows the logical form graph that TRIPS produces after
parsing the phrase, “Move the B in measure 2 up an octave.” (Many details of the TRIPS logical form representation have been removed for presentation.) Labels on arcs
(surrounded by ‘< ... >’) represent semantic roles, and
bold-face labels represent lexical terms preceded by their
semantic type (e.g., B is a MusicNote).
SpeechAct REQUEST
4. SELECTION
<CONTENT>
An important operation for composition by conversation is
selection. Much like the SELECT statement from SQL,
which searches a database of tables for entries matching a
query, we wish to scan over a musical data structure and
return references to the correct portions of it.
We use pattern matching as an integral part of our selection process, which involves checking to see whether concrete definitions match features present in an incomplete or
abstract definition.
MusECI defines the select operation as a function that
takes two M usic values: one to use as a pattern to search
for (a query) and the other to pattern match against. The
“ ” values match any concrete value. For example, the
statement select(N ( , , ...), m) will select all notes from
the M usic value,, m and return a collection of references
to the notes. Returning references in an object-oriented
style allows for m, to be updated immediately after relevant parts of it are located:
query, operation = toQuery(parse)
operation(select(query, m))
where toQuery turns a parse of a natural language sentence into a query and an operation to perform on its result.
By leveraging Python’s functional language features,
MusECI also permits conditionals and classifiers as part of
query statements. For example, finding notes above octave
3, can be expressed as N (( , > 3), ...). Selection can also
be done for more complex queries using Seq and P ar.
5. OPERATIONS
Support for music creation and generation requires that
certain standard operators be defined to manipulate the musical data structures. The following are some examples of
operations in MusECI that we use in examples later on:
• invert(M usic) and invertAt(P itch, M usic): generalized musical inversion (flipping upside down
along the pitch axis) at the first pitch and a specified
pitch respectively.
• retrograde(M usic): reverse a musical structure.
• transpose(Int, M usic): transpose a Music value
by some number of half steps, either chromatically
CauseMove MOVE
<AGENT>
PERSON
*YOU*
<AFFECTED>
<RESULT>
<TENSE>
MusicNote B
DirectionUp UP
<LOCATION>
<GROUND>
InLoc IN
MusicOctave OCTAVE
<GROUND>
<SIZE>
MusicDomain MEASURE
Number
<NAME-OF>
<VALUE>
Measure 2
1
PRESENT
Figure 1. TRIPS parser-like representation of the phrase “Move the B in
measure 2 up an octave.”
The logical form produced by TRIPS allows us to map
from graphical patterns in the logical form to classes of
operations on musical representations. From the example
in Figure 1, a M OVE action in the direction U P affecting a
N OTE corresponds to a transposition operation applied to
our music representation. The source location of the Note
specifies the query for a note with a pitch class of B and
an onset within measure 2. Finally, moving up one octave
corresponds to transposing up 12 half steps. Together, this
information specifies the following operation:
transpose(12, select(N ((B, ), , (1, )), m))
(9)
6.1 Resolving Ambiguities
Natural language is extremely sensitive to conversational
history. Sentences such as “give that to him” are impossible to semantically resolve without context for mapping
pronouns to entities. In a musical setting, references such
as “the C next to it” also suffer from this issue.
Even with sufficient context to resolve references like
pronouns, there may be insufficient detail to locate a precise portion of a musical concept. Consider the pitch classes
in the opening refrain for “Twinkle Twinkle Little Star:”
C, C, G, G, A, A, G. Requesting to “move the G up a
half step” creates an ambiguity about which G should be
moved, since select(N ((G, ), , ...)) will find three
4
Figure 2. Starting melody.
4
Figure 3. The result of applying the change “Move the F up a whole step”
to the melody in Figure 2.
4
Figure 4. The result of applying the change “Move the C on the first beat
of measure two down a half step.” to the melody in Figure 3.
matches. However, “the G” implies we only want to operate on one of them.
A system for addressing referential ambiguity requires
two features: a working memory of recent references, and
an assumer algorithm or module that explores the working
memory using features from a parse tree to resolve ambiguities. That resolution step may be straightforward or
it may involve requesting additional information from the
user before proceeding. Disambiguation, therefore, has the
following workflow, where m is the target Music value.
query, operation = toQuery(parse)
x = assumer.resolve(select(query, m))
operation(x)
The creation of an appropriate assumer module is an area
of ongoing work in conjunction with the TRIPS parser integration. Because of this, the examples presented in the
next section are designed to avoid the two types of referential ambiguity described here. This means that all of the
examples can be parsed directly to code statements of the
form operation(select(...)) with no need for ambiguity
resolution before applying the operation derived from the
natural language sentence.
6.2 Examples
Consider the two-measure melody shown in Figure 2, and
suppose we wish to transform it into the melody shown
in Figure 4 through natural language commands. A precise way to describe this would be “move the F up a whole
step” followed by “move the C on the first beat of measure
two down a half step.” This will create the following interactions between the computer (C) and user (U).
Conversation 1
C: m = the melody in Figure 2.
U: “Move the F up a whole step.”
C: transpose(2, select(N ((F, ), , , ...), m))
U: “Move the C on the first beat of measure two down a
half step.”
C: transpose(−1, select(N ((C, ), , (1, 0), ...), m))
Applied sequentially, these commands produce the
melodies shown in Figures 3 and 4 respectively. The data
structure for the starting melody is shown in Figure 7, along
4
Figure 5. The result of applying the change “invert the notes in measure
two around G4” to the melody in Figure 3.
4
Figure 6. The result of applying the change “reverse the notes in measure
one” to the melody in Figure 5.
with the selection statements created by each of the commands and the resulting modification to the data structure.
Statements may also operate over collections of notes.
The following conversation illustrates this with inversion
and reversal of sections of the music (retrograde). Figures
5 and 6 show the result of each command in this conversation, and the data structure transformation is shown in
Figure 8.
Conversation 2
C: m = the melody in Figure 2.
U: “Invert the notes in measure two around G4.”
C: invertAt((G, 4), select(N ( , , (1, ), ...), m))
U: “Reverse the notes in measure one.”
C: retro(select(N ( , , (0, ), ...), m))
7. CONCLUSION AND FUTURE WORK
We have proposed the MusECI framework for modeling
music in a way that allows querying based on parsing of
natural language statements. Our Python implementation
of the musical data structures supports normal algorithmic
composition in a text editor in addition to the composition
by conversation use case in a way that is easily integrated
with parsing systems like TRIPS.
Although our data structures are robust for representing
complex musical structures, the overall system is currently
a prototype limited to a fairly narrow collection of musical
operations. This collection of operations needs to be expanded to feature more tree manipulation algorithms, such
as different strategies for removing and re-arranging notes
within a larger structure. Operations such as removal of
individual notes can have several potential interpretations,
such as replacement with a rest of the same duration and
removal followed by time-shifting other elements of the
data structure to close the gap. Both operations are valid,
but which is most appropriate depends on the larger context in which it is used. We are currently working on a
more diverse array of operations such as this to support a
wider range of common score manipulations.
Expansion of this framework for NLP-based music composition and manipulation will require adaptation of NLP
toolkits like the TRIPS parser. Standard parsers often do
not have a suitable dictionary of musically-relevant definitions to draw on when assigning semantics to nouns
and verbs. While terms like “transpose” and “reverse” can
be interpreted correctly due to their usage in non-musical
settings, correct interpretation of other terms is trickier.
Without incorporation of music-specific terminology, the
Seq
N (C,4) (0,0)
N (E,4) (0,1)
N (F,4) (0,2)
N (A,4) (0,3)
N (A,4) (1,0.5)
N (G,4) (1,1)
N (C,5) (1,2)
2
1
N((F, _ ), _ )
N (C,5) (1,0)
N (G,4) (0,2)
N (B,4) (1,0)
N((C, _ ), (1,0) )
Figure 7. Data structure representation of the melody in Figure 2 and the two alterations produced by the operations from conversation 1 that yield the
melodies in Figures 3 and 4 respectively. Duration information is not referenced directly by any of the selection operations and is therefore omitted from
notes in this diagram; N is shorthand for a Note object and constructor.
Seq
N (C,4) (0,0)
N (E,4) (0,1)
N (F,4) (0,2)
N (A,4) (0,3)
N (C,5) (1,0)
N (A,4) (1,0.5)
N (F,4) (0,1)
N (E,4) (0,2)
N (C,5) (1,2)
1
2
N (A,4) (0,0)
N (G,4) (1,1)
N (C,4) (0,3)
N (D,4) (1,0)
N (F,4) (1,0.5)
N (G,4) (1,1)
N (D,5) (1,2)
Figure 8. Data structure representation of the melody in Figure 2 and the two alterations produced by the operations from conversation 2 that yield the
melodies in Figures 3 and 4 respectively. Duration information is omitted for the same reasons as in Figure 7.
TRIPS parser identifies“C” and “F” as temperature scales
rather than as pitch classes. Just as humans learning about
music must have their vocabulary expanded, we are developing a music-specific ontology that can be incorporated in
the TRIPS parser to support a more diverse range of musical concepts and operations to achieve a correct parse tree
for sentences in a composition by conversation setting.
The addition of new representations for contexts is important for the creation of an assumer module. This requires
a way to infer the referents of ambiguous words like “it,”
“this,” and so on as well as keeping track of what spaces or
metrics are currently in use.
Our prototype framework is the beginning of an integrated
approach for handling natural language and musical concepts. Currently, MusECI and its integration into NLP systems is still a work in progress. However, we hope to eventually achieve real-time interactive programs capable of interacting with real musicians in a musical setting, similar to
how AI on computers and mobile devices is currently able
to respond to basic spoken commands in other domains.
We also aim for this communication to ultimately become
bi-directional and incorporate other aspects of musical artificial intelligence, perhaps even allowing the machine to
critique its human user’s musical ideas or offer its own suggestions to help complete a musical project.
8. REFERENCES
[1] I. V. Serban et al., “Building End-to-end Dialogue Systems Using Generative Hierarchical Neural Network
Models,” in Proceedings of the AAAI Conference on
Artificial Intelligence, 2016, pp. 3776–3783.
[2] J. Mandler, The foundations of mind: Origins of conceptual thought. Oxford University Press, 2004.
[3] R. St.Amant et al., “An Image Schema Language,” in
Proceedings of the International Conference on Cognitive Modeling, 2006.
[4] J. Pustejovsky and N. Krishnaswamy, “VoxML: A
Vistualization Modeling Language,” in Language Resources Evaluation Conference (LREC), 2016.
[5] R. Boulanger, The Csound Book: Perspectives in Software Synthesis, Sound Design, Signal Processing,and
Programming. MIT Press, 2000.
[6] J. McCartney, “Rethinking the Computer Music Language: SuperCollider,” Computer Music Journal,
vol. 26, no. 4, pp. 61–68, 2002.
[7] A. R. Brown and A. Sorensen, “Interacting with Generative Music through Live Coding,” Contemporary Music Review, vol. 28, no. 1, pp. 17–29, 2009.
[8] Cycling ’74. (2017) Tools for Sound, Graphics, and
Interactivity. [Online]. Available: https://cycling74.
com/
[9] H. Nienhuys and J. Nieuwenhuizen, “LilyPond, a system for automated music engraving,” in Proceedings of
the Colloquium on Musical Informatics, 2003.
[10] M. Good, “MusicXML for Notation and Analysis,” in
The Virtual Score: Representation, Retrieval, Restoration. Cambridge, MA: MIT Press, 2001, pp. 113–124.
[11] C. Wang et al., “MuSQL: A Music Structured Query
Language,” in Proceedings of the International Multimedia Modeling Conference, 2006, pp. 216–225.
[12] Music21. [Online]. Available:
music21/
http://web.mit.edu/
[13] B. Manaris and A. Brown, Making Music with Computers: Creative Programming in Python. New York,
NY, USA: Chapman and Hall, 2014.
[14] A. Brown, Making Music with Java.
lulu.com, 2009.
[15] P. Hudak et al. (2017) Euterpea. [Online]. Available:
http://hackage.haskell.org/package/Euterpea
[16] D. Quick, “Composing with Kulitta,” in Proceedings of
International Computer Music Conference, 2015, pp.
306–309.
[17] J. F. Allen, M. Swift, and W. de Beaumont, “Deep Semantic Analysis for Text Processing,” in Proceedings
of the Symposium on Semantics in Systems for Text Processing, 2008.
| 6cs.PL
|
arXiv:1801.05129v1 [math.AC] 16 Jan 2018
THE RELEVANCE OF FREIMAN’S THEOREM FOR
COMBINATORIAL COMMUTATIVE ALGEBRA
∗
JÜRGEN HERZOG, TAKAYUKI HIBI AND GUANGJUN ZHU
Abstract. Freiman’s theorem gives a lower bound for the cardinality of the
doubling of a finite set in Rn . In this paper we give an interpretation of his theorem
for monomial ideals and their fiber cones. We call a quasi-equigenerated monomial
ideal a Freiman ideal, if the set of its exponent vectors achieves Freiman’s lower
bound for its doubling. Algebraic characterizations of Freiman ideals are given,
and finite simple graphs are classified whose edge ideals or matroidal ideals of its
cycle matroids are Freiman ideals.
Introduction
Let X be a finite subset of Zn , and let A(X) be the affine hull of the set X, that
is, the smallest affine subspace of Qn containing X. The doubling of X is the set
2X = {a + b : a, b ∈ X}. The starting point of this paper is the following celebrated
theorem of Freiman [7]:
d+1
(1)
,
|2X| ≥ (d + 1)|X| −
2
where d is the dimension of the affine space A(X).
Now let K be a field and let I be a graded ideal in the polynomial ring S =
K[x1 , . . . , xn ]. We denote by µ(I) the minimal number of generators of I and by
ℓ(I) the analytic spread of I, that is, the Krull dimension of the fiber cone F (I) =
⊕k≥0 I k /mI k . Here m = (x1 , . . . , xn ) is the graded maximal ideal of S. It has been
noticed in [10, Theorem 1.9] that Freiman’s theorem has an interesting consequence
regarding the minimal number of generators of the square of a monomial ideal.
Namely, it was shown that if I ⊂ S is a monomial ideal with the property
that all
ℓ(I)
2
generators of I have the same degree, then µ(I ) ≥ ℓ(I)µ(I) − 2 .
For the application of Freiman’s theorem to a monomial ideal I ⊂ S and its fiber
cone F (I), it is sufficient to require that I is quasi-equigenerated, by which we mean
that the exponent vectors of all generators of I lie in a hyperplane of Zn . This
guarantees that the set of exponent vectors of the monomials u ∈ G(I 2 ) is obtained
from the set of exponent vectors of the monomials v ∈ G(I) by doubling. Here,
G(I) denotes the unique minimal set of monomial generators of a monomial ideal I.
2010 Mathematics Subject Classification. Primary 13C99; Secondary 13A15, 13E15, 13H05,
13H10.
Key words and phrases. Monomial ideal, Freiman ideal, Freiman graph, Freiman matroid, fiber
cone.
* Corresponding author.
1
We call a quasi-equigenerated
monomial ideal I ⊂ S a Freiman ideal, if µ(I 2 ) =
ℓ(I)µ(I) − ℓ(I)
. Of course, I is Freiman if and only if equality holds in (1) for the
2
set X of exponent vectors
of the monomials u ∈ G(I). The sets X ⊂ Zn for which
|2X| = (d+1)|X|− d+1
are characterized by Stanescu in [19]. Here, in Theorem 1.3,
2
we give several equivalent conditions for a quasi-equigenerated monomial ideal I ⊂ S
to be a Freiman ideal. For example, it is shown that I is Freiman if and only if the
fiber cone F (I) of I is Cohen–Macaulay and the defining ideal of F (I) has a 2-linear
resolution. This homological characterization of Freiman ideals can be deduced
from
ℓ(I)
2
the fact that for a Freiman ideal not only one has µ(I ) = ℓ(I)µ(I) − 2 , but also
ℓ+k−2
µ(I k ) = ℓ+k−2
µ(I)
−
(k
−
1)
for all k ≥ 1. Indeed, this fact is a consequence
k−1
k
of a result of Böröczky, Santos and Serra [4, Corollary 7]. Now this formula for the
minimal number of monomial generators of I k of a Freiman ideal I easily yields that
hi = 0 for all i ≥ 2, where (1, h1 , h2 , . . .) is the h-vector of the fiber cone F (I) of
I. Then we apply a result of Eisenbud and Goto [5, Corollary 4.5] and obtain that
F (I) is normal. Since F (I) is a toric ring, Hochster’s theorem [14] finally yields that
F (I) is Cohen–Macaulay.
Even more surprising than the inequality µ(I 2 ) ≥ ℓ(I)µ(I) − ℓ(I)
, and seemingly
2
not noticed before in the study of toric rings, are the following consequences of the
above mentioned results about the h-vector of the fiber cone F (I) and of Freiman’s
theorem, which may phrased as follows: let A be a standard graded toric K-algebra
with h-vector (1, h1 , h2 , . . .). Then h2 ≥ 0, and if h2 = 0, then hi = 0 for all i ≥ 2.
The inequality h2 ≥ 0 was first observed in [10, Corollary 2.6].
Being a Freiman ideal is a very restrictive condition. Therefore, one can expect
that for natural classes of monomial ideals arising in combinatorial context a nice
classification of Freiman ideals is possible. We illustrate this in two cases. In both
cases we consider a finite simple graph G. We call G a Freiman graph, if its edge
ideal I(G) is a Freiman ideal. This classification can be reduced to the case that G is
connected, see Corollary 2.2. In Theorem 2.3 it is shown that if G is connected and
H is the subgraph of G whose edges are the edges of all the 4-cycles of G, then G is
a Freiman graph if and only if there exist no primitive even walks in G, or otherwise
H is bipartite of type (2, s) for some integer s, and there exist no primitive even
walks in G of length > 4. The proof is inspired by the classification of edge ideals
with 2-linear resolution, due to Ohsugi and Hibi [16]. In the special case that G
is bipartite this classification is even more explicit. Indeed, in Corollary 2.4 it is
shown, that if G is a connected bipartite graph, then G is Freiman if and only if
G is a tree, or otherwise there exists a bipartite subgraph H of G of type (2, s) for
some integer s, a subset S of V (H) and for each w ∈ S an induced tree Tw of G
with V (H) ∩ V (Tw ) = {w}.
The other case studied here is that of the matroidal ideal IM of the cycle matroid
M of G. We call a matroid M whose matroidal ideal IM is a Freiman ideal, a
Freiman matroid. In Theorem 2.5 we show that the cycle matroid of G is Freiman
if and only if G contains at most one cycle. It is a challenging open problem to
classify all Freiman matroids. A few results in this direction, even for polymatroids,
can be found in [11].
2
As an interesting consequence of the classification of the cycle matroids which are
Freiman, it is shown in Theorem 2.13 that if r denotes the regularity of the base
ring of the cycle matroid of a graph G, then, unless the base ring is a polynomial
ring, one has 3 ≤ r ≤ e, where e is the number of the edges of G.
1. Freiman ideals and their fiber cones
Let K be a field and S = K[x1 , . . . , xn ] the polynomial ring in n indeterminates
over K, and let I ⊂ S be a monomial ideal. The unique minimal set of monomial
generators of I will be denoted by G(I). The number µ(I) = |G(I)| is the minimal
number of generators of I. We write xc for the monomial xc11 · · · xcnn where c =
(c1 , . . . , cn ).
Let a = (a1 , . . . , an ) ∈ Zn be an integer vector with all ai > 0. We say that
I is equigenerated (of degree d) with respect to a or simply say that I is quasiequigenerated, if there exists an integer d such that
d = ha, ci for all xc ∈ G(I).
P
Here ha, ci = ni=1 ai ci denotes the standard scalar product of a and c. The ideal I
is called equigenerated if a = (1, 1, . . . , 1).
L
The Krull dimension of the fiber cone F (I) = k≥0 I k /mI k is called the analytic
spread of I, and denoted ℓ(I). One always has
height I ≤ ℓ(I) ≤ min{µ(I), n}.
Let G(I) = {u1 , . . . , um }, and let T = K[y1 , . . . , ym] be the polynomial ring over
K in the indeterminates y1 , . . . , ym . Consider the K-algebra homomorphism ϕ : T →
F (I) with yj 7→ uj + mI for j = 1, . . . , m. Then F (I) ∼
= T /J, where J = Ker ϕ.
Since I is quasi-equigenerated, it follows that F (I) ∼
= K[u1 t, . . . um t] ⊂ T [t], where
t is a new indeterminate over T . Thus we see that J is a toric prime ideal generated
by binomials in T which are homogeneous with respect to the standard grading of
T.
The following theorem is a consequence of a famous theorem of Freiman [7] and
its generalizations by Böröczky et al [4].
Theorem 1.1. Let I ⊂ S be a quasi-equigenerated monomial ideal whose analytic
spread is ℓ(I). Then
ℓ(I) + k − 2
ℓ(I) + k − 2
k
µ(I) − (k − 1)
µ(I ) ≥
k
k−1
for all k ≥ 1.
Proof. Let S(I) ⊂ Zn be the set of exponent vectors of the elements of G(I). Since I
is a quasi-equigenerated monomial ideal, I k are quasi-equigenerated monomial ideals
for all k ≥ 1. It follows that µ(I k ) = |S(I k )| = |kS(I)| for all k ≥ 1. Let d be the
3
Freiman dimension of S(I), then from [4, Corollary 2],
d+k−1
d+k−1
|S(I)| − (k − 1)
|kS(I)| ≥
k
k−1
d+k−1
d+k−1
.
µ(I) − (k − 1)
=
k
k−1
On the other hand, by [10, Theorem 1.9] one has ℓ(I) = d + 1. Thus the assertion
follows.
The above theorem implies in particular that µ(I 2 ) ≥ l(I)µ(I)− l(I)
. In [11], the
2
first and third author of this paper called an equigenerated monomial ideal a Freiman
ideal, if equality holds. Here we extend this definition and call a quasi-equigenerated
ideal I a Freiman ideal, if
l(I)
2
.
µ(I ) = l(I)µ(I) −
2
In the next theorem we will give some characterizations of Freiman ideals in terms
of the Hilbert series of their fiber cones. To this end we recall some concepts from
commutative algebra.
Let R be a standard graded K-algebra. We denote its Krull dimension by
dim
P R and itsk embedding dimension by emb dim R. The Hilbert series HilbR (t) =
dimK Rk t of R is of the form
k≥0
HilbR (t) = h(R; t)/(1 − t)dim R ,
where h(R; t) = 1 + h1 t + h2 t2 + · · · is a polynomial with h1 = emb dim R − dim R.
The polynomial h(R; t) is called the h-polynomial of R, and the finite coefficient
vector of h(R; t) is called the h-vector of R.
P
The multiplicity of R is defined to be e(R) = i≥0 hi . By Abhyankar [1] it is
known that
emb dim R ≤ e(R) + dim R − 1,
if R is a domain. The same inequality holds if R is Cohen–Macaulay [17]. The
K-algebra R is said to have minimal multiplicity if emb dim(R) = e(R) + dim R − 1.
We first observe
Corollary 1.2. Let I ⊂ S be a quasi-equigenerated monomial ideal with analytic
spread ℓ, and let (1, h1 , h2 , . . .) be the h-vector of the fiber cone F (I) of I. Then
k
X
ℓ+k−i−1
hi ≥ 0 for all k ≥ 2.
k−i
i=2
In particular, h2 ≥ 0.
Proof. The Hilbert series of the fiber cone F (I) of I is
X
1 + h1 t + h2 t2 + h3 t3 + h4 t4 + · · ·
HilbF (I) (t) =
µ(I k )tk =
(1 − t)ℓ
k≥0
4
ℓ+2 3
ℓ+1 2
t + ···)
t +
= (1 + h1 t + h2 t + · · · )(1 + ℓt +
3
2
ℓ+1
+ h2 )t2 + · · ·
= 1 + (h1 + ℓ)t + (h1 ℓ +
2
X
k
ℓ+k−i−1
ℓ+k−1
hi ]tk + · · · .
+
+ [
k−i
k
i=1
2
It follows that
ℓ
ℓ+1
+ h2 .
+ h2 = ℓµ(I) −
µ(I) = h1 + ℓ, and µ(I ) = h1 ℓ +
2
2
2
Moreover, for all k ≥ 3 we have
ℓ
ℓ+k−3
ℓ+k−2
ℓ+k−1
k
hk−1 + hk
h2 + · · · +
h1 +
+
µ(I ) =
1
k−2
k−1
k
k
X
ℓ+k−i−1
ℓ+k−2
ℓ+k−1
hi
(µ(I) − ℓ) +
+
=
k−i
k−1
k
i=2
X
k
ℓ+k−2
ℓ+k−2
ℓ+k−i−1
+
µ(I) − (k − 1)
=
hi .
k
k−1
k
−
i
i=2
Therefore, Theorem 1.1 implies that h2 ≥ 0 and
k
P
i=2
ℓ+k−i−1
k−i
hi ≥ 0 for all k ≥ 2.
Let I be a graded ideal. An ideal J ⊆ I is called a reduction of I if I k+1 = JI k for
some nonnegative integer k. The reduction number of I with respect to J is defined
to be
rJ (I) = min{k | I k+1 = JI k }.
A reduction J of I is called a minimal reduction if it does not properly contain any
other reduction of I. If I is equigenerated and |K| = ∞, then a graded minimal
reduction of I exists. In this case the reduction number of I is defined to be the
number
r(I) = min{rJ (I) | J is a minimal reduction of I}.
Now we are ready to present the main result of this section.
Theorem 1.3. Let I ⊂ S be a quasi-equigenerated monomial ideal with analytic
spread ℓ, and let (1, h1 , h2 , . . .) be the h-vector of the fiber cone F (I) of I. Write
F (I) = T /J, where T is a polynomial ring over K, and J is contained in the square
of the graded maximal ideal of T . Then the following conditions are equivalent:
(a) I is a Freimanideal.
ℓ+k−2
(b) µ(I k ) = ℓ+k−2
µ(I)
−
(k
−
1)
for all k ≥ 1.
k−1
k
ℓ+k−2
ℓ+k−2
(c) µ(I k ) = k−1 µ(I) − (k − 1) k
for some k ≥ 2.
(d) h2 = 0.
5
(e) hi = 0 for all i ≥ 2.
(f) F (I) has minimal multiplicity.
(g) F (I) is Cohen–Macaulay and the defining ideal of F (I) has a 2-linear free
T -resolution.
Moreover, if |K| = ∞, then the above conditions are equivalent to
(h) F (I) is Cohen–Macaulay and r(I) = 1.
Proof. (a) ⇒ (b) follows from [4, Corollary 7], (b) ⇒ (c) is trivial, and (c) ⇒ (b)
also follows from [4, Corollary 7].
2
In the proof of Corollary 1.2, we have seen that µ(I) = h1 + ℓ, µ(I ) = ℓµ(I) −
ℓ
+ h2 and
2
X
k
ℓ+k−i−1
ℓ+k−2
ℓ+k−2
k
hi .
+
µ(I) − (k − 1)
µ(I ) =
k−i
k
k−1
i=2
By definition, I is a Freiman ideal, if and only if µ(I 2 ) = ℓµ(I) − 2ℓ . This is
equivalent to saying that h2 = 0. This proves (a) ⇐⇒ (d).
(b) ⇐⇒ (e) follows from [11, Proposition
1.8 (b)].
P
(e) ⇒ (f): Notice that e(R) = i≥0 hi . If hi = 0 for i ≥ 2, then
e(R) = h0 + h1 = 1 + h1 = 1 + µ(I) − ℓ.
This means that F (I) has minimal multiplicity.
(f) ⇒ (g): After a base field extension, we may assume that K is algebraically
closed. Since F (I) has minimal multiplicity, it follows that F (I) is normal by [5,
Corollary 4.5]. Since F (I) is a toric ring, Hochster’s theorem [14] implies that F (I)
is Cohen–Macaulay and hi = 0 for all i ≥ 2. Thus from [11, Corollary 1.5] we obtain
that the defining ideal of F (I) has a 2-linear free T -resolution.
(g) ⇒ (d): We may assume that K is infinite. Since F (I) is Cohen–Macaulay we
can choose a regular sequence t1 , . . . , tℓ of linear forms of T1 which is also regular
on F (I). We let T̄ = T /(t1 , . . . , tℓ ) and denote by J¯ the image of J in T̄ . Since
t1 , . . . , tℓ is a regular sequence on T /J and J has a 2-linear resolution, it follows
that dim T̄ /J¯ = 0 and J¯ has a 2-linear resolution. It follows that J¯ = m2T̄ , where
m2T̄ is the graded maximal ideal of T̄ , since the only mT̄ -primary ideals with linear
resolution are powers of the maximal ideal. Thus we see that F (I)i = 0 for i ≥ 2,
where F (I) = F (I)/(y1, . . . , yℓ ), and yi = ti + J for i = 1, . . . , ℓ. Since y1 , . . . , yℓ is
a regular sequence of linear form, the h-vector of F (I) and that of F (I) coincide.
Hence the conclusion follows.
(g) ⇒ (h): We assume that K is infinite, and let y1 , . . . , yℓ be as in the proof (g) ⇒
(d). We write yi = fi + mI with fi ∈ I for i = 1, . . . , ℓ, and let J = (f1 , . . . , fℓ ). Now
F (I)i = 0 for i ≥ 2 is equivalent to saying that I 2 = JI + mI 2 . Thus, Nakayama’s
lemma yields I 2 = JI, as desired.
(h) ⇒ (d): Since I is quasi-equigenerated, K is infinite and F (I) is Cohen–
Macaulay, a reduction ideal J of I with rJ (I) = 1 can be chosen of the form J =
(f1 , . . . , fℓ ) such that the sequence yi = fi + mI with i = 1, . . . , ℓ is a regular
sequence on F (I). Since I 2 = JI it follows that F (I)i = 0 for i ≥ 2, where
6
F (I) = F (I)/(y1, . . . , yℓ ). Since h-vector of F (I) and that of F (I) coincide, the
assertion follows.
2. Classes of Freiman ideals
2.1. Freiman graphs. Let K be a field and G be a finite simple graph on [n]. The
ideal I(G) ⊂ S = K[x1 , . . . , xn ] generated by the monomials xi xj with {i, j} ∈ E(G)
is called the edge ideal of G. We say that G is a Freiman graph, if I(G) is a
Freiman ideal. In this subsection we want to classify all Freiman graphs. Notice
that F (I(G)) is isomorphic to the edge ring K[G] = K[xi xj : {i, j} ∈ E(G)] of G.
Let T = K[ze : e ∈ E(G)] be the polynomial ring in the indeterminates ze , and let
IG be the kernel of the K-algebra homomorphism T → K[G] with ze 7→ xi xj for
e = {i, j}. Then K[G] ∼
= T /IG .
Certainly G is Freiman, if K[G] is the polynomial ring. The monomials xi xj with
{i, j} ∈ E(G) are algebraic independent if and only if K[G] is a polynomial ring,
which is the case if and only if IG = 0.
In order to describe the generators of IG one has to consider walks in G. A walk
W of length r of G is a sequence of vertices i0 , i1 , . . . , ir such that {ij , ij+1 } is an
edge of G for j = 0, . . . , r − 1. If i0 = ir , then W is called a closed walk. The closed
walk is called even (odd), if r is even (odd). It is called a cycle, if the vertices ij
with 0 ≤ j ≤ r − 1 are pairwise distinct.
Now we have
Proposition 2.1. The edge ring K[G] of G is a polynomial ring, if and only if each
connected component of G contains at most one cycle, and this cycle is odd.
N
Proof. Let G1 , . . . , Gs be the connected components of G. Then K[G] = si=1 K[Gi ].
Thus K[G] is a polynomial ring if and only if each K[Gi ] is a polynomial ring. Hence
we may assume that G is connected and have to show that K[G] is a polynomial
ring if and only if G contains at most one cycle, and this cycle is odd.
Let W : i0 , i1 , . . . , i2s be an even closed walk, and set ej = {ij , ij+1 } for j =
0, . . . , 2s − 1. Then
s−1
s−1
Y
Y
fW =
ze2k −
ze2k+1 ∈ IG ,
k=0
k=0
and all binomial generators of IG are of this form, see [16, Lemma 1.1]. Thus K[G]
is a polynomial ring if and only if fW = 0 for each even closed walk in G.
Suppose that K[G] is a polynomial ring and G contains an even cycle C. Then
fC 6= 0, a contradiction. Therefore, G contains no even cycles. Suppose G contains
at least two odd cycles, say C1 and C2 . Let i be a vertex of C1 and j be a vertex of
C2 . Then, since G is connected, there exists a walk W connecting i with j. Now we
consider the following even closed walk W ′ : we first walk around C1 starting with
i. Then we walk along W from i to j. Then we continue our walk around C2 to
come back to j. From there we walk back to i along W , but backwards. Obviously,
fW ′ 6= 0. Thus there cannot be two odd cycles. This proves one direction of the
proposition.
7
Conversely, suppose that G contains only one odd cycle, it is enough to prove that
IG = 0. By [16, Lemma 3.1 and Lemma 3.2], the toric ideal IG is generated by the
binomials fW , where W is an even closed walk of G of one of the following types:
(1) W is an even cycle of G; (2) W = (C1 , C2 ), where C1 and C2 are odd cycles of
G having exactly one common vertex; (3) W = (C1 , W1 , C2, W2 ), where C1 and C2
are odd cycles of G having no common vertex and where W1 and W2 are walks of
G both of which combine a vertex v1 of C1 and a vertex v2 of C2 . Walks of these
three types are called primitive.
By the hypothesis that G contains only one odd cycle, there cannot exist primitive
even closed walks of G. Therefore, IG = 0.
Corollary 2.2. Let G be a finite simple graph with r connected components. Then
following conditions are equivalent:
(a) G is a Freiman graph.
(b) All connected components of G are Freiman graphs and at least r − 1 of its
connected components contain at most one cycle, and this cycle is odd.
Proof. Let G1 , . . . , Gr be the connected components of G, and let N
HilbK[Gi] (t) =
r
di
h(K[Gi ]; t)/(1 − t) with di = dim K[Gi ]. Then, since K[G] =
i=1 K[Gi ], it
follows that
r
Y
HilbK[G](t) =
HilbK[Gi] (t) = h(K[G]; t)/(1 − t)d ,
i=1
Qr
P
where h(K[G]; t) = i=1 h(K[Gi ]; t) and d = ri=1 di = dim K[G].
(a) ⇒ (b): Since G is Freiman, Theorem 1.3 implies that deg h(K[G]; t) ≤ 1.
Therefore, deg h(K[Gi ]; t) ≤ 1 with equality for at most one i. Thus (b) follows
from Proposition 2.1 together with Theorem 1.3.
(b) ⇒ (a): The assumptions in (b) together with Theorem 1.3 and Proposition 2.1
imply that deg h(K[Gi ]; t) ≤ 1 and deg h(K[Gi ]; t) = 0 for all h(K[Gi ]; t) but possibly one of them. Thus deg h(K[G]; t) ≤ 1, which by Theorem 1.3 implies that G is
Freiman.
Due to Corollary 2.2 it is enough to consider only connected graphs for the classification of Freiman graphs.
Theorem 2.3. Let G be a finite simple connected graph, and let H be the subgraph
of G whose edges are the edges of all the 4-cycles of G. Then G is Freiman, if and
only if there exist no primitive even walks in G, or otherwise H is bipartite of type
(2, s) for some integer s, and there exist no primitive even walks in G of length > 4.
Proof. If G has no primitive even walks, then K[G] is the polynomial ring, and hence
G is Freiman. Suppose now that H is bipartite of type (2, s) for some s, and there
exist no primitive even walks in G of length > 4. Since G admits no primitive even
walks of length > 4, the ideal IG is generated by the binomials fC with C a 4-cycle
of G. In particular, IG = IH , and K[G] is a polynomial extension of K[H]. Since H
is bipartite of type (2, s), the ideal IH may be viewed as the ideal of 2-minors of a
8
2 × s-matrix of indeterminates. Hence the Eagon-Northcott complex [6] provides a
2-linear resolution of IH (= IG ). By Theorem 1.3, this implies that G is Freiman.
Conversely, suppose that G is Freiman. Then Theorem 1.3 implies that IG has no
generators of degree > 2, and hence G admits no primitive even walks of length > 4.
We may assume that H is bipartite of type (r, s) with r ≤ s. Since the edge set of
H is the union of the edges of the 4-cycles of G, we must have r ≥ 2. Also it follows
from this fact that H is a complete bipartite graph if r = 2. Indeed, suppose that
one edge is missing, say the edge {2, r}. Then {1, r} ∈ E(H) and {1, r} belongs to
a 4-cycle. Any 4-cycle containing the edge {1, r} is of the form r, 1, i, 2, r for some
i 6= r. Hence {2, r} ∈ E(H), a contradiction.
It remains to be show that r = 2. Suppose r > 2, and let {v1 , . . . , vr } ∪
{w1 , . . . , ws } be the bipartition of the graph H. Since H is the union of the
edges of the 4-cycles of G, we may assume that, after a relabeling of the vertices,
C1 : v1 , w1 , v2 , w2, v1 is a 4-cycle of H. Since r > 2, there must exist another 4-cycle
C2 of H which contains the vertex v3 . Then C1 and C2 have only one vertex or one
edge in common, or the cycles C1 and C2 are disjoint. But in the last case there is a
walk in G connecting a vertex of C1 with a vertex of C2 since G is connected. In any
case we have in this situation a primitive even walk of length > 4, a contradiction.
Corollary 2.4. Let G be a finite connected bipartite graph. Then G is Freiman if
and only if G is a tree, or there exists a bipartite subgraph H of G of type (2, s), a
subset S of V (H) and for each w ∈ S an induced tree Tw of G with V (H) ∩ V (Tw ) =
{w}.
Proof. If G is a tree or if there exists a bipartite subgraph H of G satisfying the
conditions of the corollary, then the subgraph H is the union of all 4-cycles of G,
and moreover there exist no primitive even walks in G of length > 4. Therefore, G
is Freiman, by Theorem 2.3.
Conversely, assume that G is Freiman and let H be as defined in Theorem 2.3.
Then Theorem 2.3 implies that H = ∅ or H is bipartite of type (2, s) for some s. If
H = ∅, then Theorem 2.3 implies that G has no even cycles of any length. Since G
is bipartite and connected, this implies that G is a tree.
Suppose now that H is bipartite of type (2, s). If G = H, then there is nothing to
prove. Otherwise there exists v ∈ V (G) \ V (H). Since G is connected there exists a
walk P1 outside of H connecting v with w1 ∈ V (H). Assume there is another walk
P2 outside H connecting v with w2 ∈ V (H) with w1 6= w2 . Then there exists w
belonging to P1 and P2 such that the subwalk of P1 connecting w with w1 and the
subwalk of P2 connecting w with w2 have only the vertex w in common. Then we
get a cycle walking from w1 to w to w2 , and from w2 back to w1 (inside H). Since
G is bipartite, this cycle must be even, a contradiction because by Theorem 2.3, G
has even cycles of length at most 4. But all even cycles of length 4 belong to H. It
follows that each vertex v of G not belonging to H is connected with at most one
vertex w ∈ H.
Now let w ∈ H, and let V be the set of vertices v ∈ V (G) \ V (H) which are
connected to w via a walk. We denote by Tw the restriction of G to V . As shown
9
above, V (Tw1 ) ∩ V (Tw2 ) = ∅. Since G has no even closed walks of length > 4, it
follows that there is for each v ∈ V (Tw ) a unique walk, whose edges are all distinct
from one another, and which connects v with w. In fact, if there exist two such
walks connecting v with w, say P1 , P2 , then their lengths must be both odd or both
even, since G is bipartite. By Theorem 1.3, the lengths of P1 and P2 are both 2.
In fact, since G is Freiman, there exist no primitive even walks in G of length > 4,
thus the lengths of P1 and P2 are both 1 or 2. If their lengths are both 1, then
G have a multiple edge connecting v with w. But the graph G is simple without
multiple edge, thus the union of P1 and P2 is a 4-cycle of G. This implies v ∈ V (H),
a contradiction. Now the property that for each v ∈ V (Tw ) there exists a unique
walk connecting v with w, implies that Tw is a tree.
A typical example of a bipartite Freiman graph is shown in the next figure.
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
Figure 1
2.2. Cycle matroids. Given a finite simple graph G on [n]. The cycle matroid of
G is the matroid whose ground set is E(G), and whose bases are the sets E(F ) with
F a spanning forest of G. A spanning forest of G is a maximal acyclic subgraph of
G.
A matroid M is called graphic, if for some simple graph G, M is isomorphic to
the cycle graph of G.
Let K be a field. Let M be a matroid on the ground set [m], and B the set of its
bases. To M we attach theQideal IM ⊂ S = K[x1 , . . . , xm ], whose generators are the
monomials uB with uB = i∈B xi and B ∈ B. The ideal IM is called the matroidal
ideal of M. The fiber cone of IM is the base ring of M. We say that M is a Freiman
matroid, if IM is a Freiman ideal.
Let E(G) = {e1 , . . . , em }, and M the cycle matroid
Q of G. Then IM ⊂ K[x1 , . . . , xm ]
is the ideal generated by the monomials uF = ei ∈F xi with F a spanning forest
of G.
As the main result of this subsection we have
10
Theorem 2.5. Let M be the cycle matroid of G. Then the following conditions are
equivalent:
(a) M is a Freiman matroid.
(b) G contains at most one cycle.
(c) F (IM ) is a polynomial ring.
We will need the following lemmas before proving Theorem 2.5. First, we have
Lemma 2.6. Let G be a finite simple graph whose cycle matroid is M. Let E be a
subset of the edges of G, Γ the graph with E(Γ) = E and N be the cycle matroid of
Γ. Then N is a Freiman matroid.
Proof. We observe that F (IN ) is a combinatorial pure subring of F (IM ) in the
sense of [15]. We can write F (IM ) = B/J and F (IN ) = A/I, where A and B
are polynomial rings and J ⊂ B and I ⊂ A are the binomial relation ideals of
F (IM ) and F (IN ), respectively. By [15, Corollary 2.5] we have βijA (I) ≤ βijB (J) for
all i, j. Since M is a Freiman matroid, it follows from Theorem 1.3 that J has a
2-linear B-resolution. Therefore the above inequality of Betti numbers implies that
I has a 2-linear A-resolution. Furthermore, since F (IN ) is normal, F (IN ) is Cohen–
Macaulay, by Hochster’s theorem, see [14]. We now apply again Theorem 1.3 and
conclude that N is a Freiman matroid.
Lemma 2.7. Let G be a finite simple graph consisting of two cycles C1 and C2
which have at most one vertex in common. Then the cycle matroid of G is not a
Freiman matroid.
Proof. Let Mi be the cycle matroid of Ci , and let ri = |E(Ci )|. Then IMi can be identified with the squarefree Veronese ideal Iri ,ri −1 of degree ri − 1 in ri indeterminates.
Hence the fiber cone of IMi is isomorphic to K[Iri ,ri −1 ] and this ring is isomorphic to
a polynomial ring Si over K in ri indeterminates. Since C1 and C2 have at most one
vertex in common, each spanning forest of G is obtained from G by removing one
edge from C1 and another edge from C2 . This shows that IM = IM1 IM2 . Hence the
fiber cone of IM is isomorphic to the Segre product T = S1 #S2 of the polynomial
rings S1 and S2 .
By [8, Theorem 4.2.3] we know that dim T = r1 + r2 − 1. Thus HilbT (t) =
h(T ; t)/(1 − t)r1 +r2 −1 , and so the a-invariant a(T ) of T is equal to deg h(T ; t) − (r1 +
r2 − 1) (see [3, Definition 4.4.4]). Hence we get
(2)
deg h(T ; t)(t) = a(T ) + (r1 + r2 − 1).
On the other hand, if ωT is the canonical module of T , then
(3)
a(T ) = − min{i : (ωT )i 6= 0},
see [8, Definition 3.1.4]. We now use another result of Goto and Watanabe [8,
Theorem 4.2.3] which says that ωT ∼
= ωS1 #ωS2 . Since ωSi = Si (−ri ), we have
ωT ∼
= S1 (−r1 )#S2 (−r2 ). The ith graded component (ωT )i of ωT is isomorphic to
(S1 (−r1 )#S2 (−r2 ))i = (S1 )i−r1 ⊗ (S2 )i−r2 .
11
Therefore, min{i : (ωT )i 6= 0} = max{r1 , r2 }. Combining (2) with (3), we get
deg h(T ; t) = r1 + r2 − 1 − max{r1 , r2 }.
Since each cycle has at least 3 edges, we see that ri ≥ 3. It follows that deg h(T ; t) ≥
2. Thus Theorem 1.3 implies that M is not a Freiman matroid.
In the case that C1 and C2 are connected by a walk P , we have IM = uIM1 IM2 ,
where u is the monomial whose factors correspond to the edge of P . Then K[IM ] =
K[uIM1 IM2 ] ∼
= K[IM1 IM2 ], and the rest of the proof is as before.
Lemma 2.8. Let G be a finite simple graph consisting of a cycle C of length r and
a walk W of length s which connects two non-consecutive vertices of C. Then the
cycle matroid M of G is not a Freiman matroid.
Proof. By hypothesis, the walk W divides cycle C into two walks W1 and W2 , where
W1 , W2 and W have common starting point and endpoint. Thus C = W1 ∪ W2 . For
simplicity we set W3 = W . Let Mi be the cycle matroid of Wi , ri = |E(Wi )|, and
let M4 be the cycle matroid of C and r4 = |E(C)|. Then r3 = s, r4 = r1 + r2 = r.
By the proof of Lemma 2.7, the fiber cone F (IMi ) of IMi is isomorphic to K[Iri ,ri −1 ]
and this ring is isomorphic to a polynomial ring Si over K in ri indeterminates.
Therefore, we get IM = Ir3 ,r3 −1 Ir1 +r2 ,r1 +r2 −1 + JIr3 ,r3 , where J = Ir1 ,r1 −1 Ir2 ,r2 −1 . We
have Ir3 ,r3 ⊂ Ir3 ,r3 −1 and Ir1 +r2 ,r1 +r2 −1 ⊂ J, Let S = S3 ⊗ S4 and HilbF (IM ) (t) =
h(S; t)/(1 − t)ℓ , where ℓ = dim(F (IM )). By [13, Theorem 1.6], we have
HilbS/IM (t) = HilbS/(Ir3 ,r3 −1 Ir1 +r2 ,r1 +r2 −1 +JIr3 ,r3 ) (t)
= HilbS3 /Ir3 ,r3 −1 (t) HilbIr1 +r2 ,r1 +r2 −1 (t) + HilbS4 /J (t) HilbIr3 ,r3 (t)
+ HilbS3 /Ir3 ,r3 (t) HilbS4 /Ir1 +r2 ,r1 +r2 −1 (t)
= HilbIr1 +r2 ,r1 +r2 −1 (t) + HilbS4 /J (t) HilbIr3 ,r3 (t) + HilbS3 /Ir3 ,r3 (t).
Hence
1
− HilbS/IM (t)
(1 − t)r3 +r4
∞
∞
X
1
m + r4 − 1 m
h(S4 ; t) X m
1
t +(
−[
−
)
t
=
m
(1 − t)r3 +r4
(1 − t)r4
(1 − t)r4 −1 m=0
m=0
HilbF (IM ) (t) =
+
=
=
=
=
1
]
(1 − t)r3 −1
1
1
h(S4 ; t)
1
1
1
−[
+[
−
]
+
]
r
+r
r
r
r
−1
3
4
4
4
4
(1 − t)
(1 − t)
(1 − t)
(1 − t)
1 − t (1 − t)r3 −1
1
(1 − t)r3 + (1 − t)r3 −1 [1 − (1 − t)h(S4 ; t)] + (1 − t)r4 +1
−
(1 − t)r3 +r4
(1 − t)r3 +r4
1 − (1 − t)r3 − (1 − t)r3 −1 [1 − (1 − t)h(S4 ; t)] − (1 − t)r4 +1
(1 − t)r3 +r4
1 − (1 − t)s − (1 − t)s−1 [1 − (1 − t)h(S4 ; t)] − (1 − t)r+1
,
(1 − t)r+s
12
where deg h(S4 ; t) = min{r1 , r2 } − 1, as seen in the proof of Lemma 2.7.
We consider the following two cases:
r
(1) If s = 1, then HilbF (IM ) (t) = h(S4 ;t)−1−(1−t)
and h(S; t) = h(S4 ; t)−1−(1−t)r ,
(1−t)r
Since deg h(S4 ; t) = min{r1 , r2 } − 1 and r = r1 + r2 , deg h(S; t) ≥ max{r1 , r2 } + 1.
Notice that ri ≥ 2, we have that deg h(S; t) ≥ 3;
(2) If s ≥ 2, then h(S; t) = 1 − (1 − t)s − (1 − t)s−1 [1 − (1 − t)h(S4 ; t)] − (1 − t)r+1
such that h(S; 1) = 1 and deg h(S; t) ≥ 5.
Thus Theorem 1.3 implies that the cycle matroid M of G is not a Freiman matroid.
Now we are ready to prove Theorem 2.5.
Proof. (a) ⇒ (b): Suppose there is a cycle C for which there exists a walk W which
connects two non-consecutive vertices of C. Then let Γ be the graph with edges
E(W ) ∪ E(C). By Lemma 2.8 the cycle matroid of Γ is not a Freiman matroid, and
so M is not a Freiman matroid by Lemma 2.6, a contradiction. Thus there is no
cycle C of G such that there exists a walk W which connects two non-consecutive
vertices of C.
Let us now assume that G contains at least two cycles C1 and C2 . We discuss two
cases: (i) V (C1 ) ∩ V (C2 ) = ∅, and (ii) V (C1 ) ∩ V (C2 ) 6= ∅.
Let Γ be the graph with E(Γ) = E(C1 ) ∪ E(C2 ). In case (i), the cycle matroid
of Γ is not a Freiman matroid, see Lemma 2.7. Therefore, according to Lemma 2.6,
also M is not a Freiman matroid.
In case (ii), by using Lemma 2.6, we may assume that Γ = G. Let S = V (C1 ) ∩
V (C2 ), and Σ the graph which is obtained from G by restriction to S. Then Σ is
a disjoint union of walks and some isolated vertices. If Σ is just a single vertex, M
is not a Freiman matroid by Lemma 2.7, and if Σ is just a single walk, then the
edges E(C1 ) ∪ E(C2 ) \ E(Σ) form a cycle C and Σ is a walk which connects two
non-consecutive vertices of C. Thus Lemma 2.8 implies that M is not a Freiman
matroid. Finally assume that Σ has more than one connected component. Let
V (C1 ) = {v1 , . . . , vr } and V (C2 ) = {w1 , . . . , ws } with a counter-clockwise labeling
of the vertices. We may assume without loss of generality that v1 = w1 and v2 6= w2 ,
since Σ is not connected. There exist integers a, b with the property that va = wb
and vi 6= wj for all i, j with 1 < i < a and 1 < j < b. It then follows that the walk
W ′ with V (W ′ ) = {w1 , w2 , . . . , wb} is a walk connecting the vertices v1 and va of
C1 . Thus M is not a Freiman matroid.
(b) ⇒ (c): Let G1 , . . . , Gs be the connected components of G and let Mi be the
cycle matroid of Gi . Since G contains at most one cycle, we may assume that the
components Gi are trees for i ≥ 2. Then, as in the proof of Lemma 2.8, we see
that IM1 = uIr,r−1 if G1 contains a cycle C with |E(C)| = r, or otherwise IM1 is a
principal monomial ideal. Moreover, all IMi for i ≥ 2 are principal monomial ideals.
Since IM = IM1 · · · IMs , we have K[IM ] ∼
= K[IM1 ], and hence K[IM ] is a polynomial
ring.
(c) ⇒ (a) is obvious.
13
We close this paper with discussing the regularity of the base ring of the cycle
matroid of a simple graph. Let, as before, G be a finite simple graph on [n] and
E(G) = {e1 , . . . , em } its edge set. Let e1 , . . . , em denote the canonical
P coordinate
m
vectors of Rm . Given a subset F ⊂ E(G), one defines ρ(F ) =
ei ∈F ei ∈ R .
For example, ρ({e2 , e3 , e5 }) = e2 + e3 + e5 = (0, 1, 1, 0, 1, 0, . . . , 0) ∈ Rm . The base
polytope P(M) of the cycle matroid M of G is the convex polytope which is the
convex hull of the finite set {ρ(F ) : ρ(F ) ∈ B} in Rm . One has dim F (IM ) =
dim P(M) + 1.
Recall that, when G is connected, a vertex i ∈ [n] of G is said to be a cut vertex if
the induced subgraph G|[n]\{i} of G on [n] \ {i} is disconnected. A finite connected
simple graph G is called 2-connected if no vertex of G is a cut vertex of G. It then
follows from [20, Theorem 5] that
Lemma 2.9. Let G be a finite 2-connected simple graph on [n] with e edges. Let M
be the cycle matroid of G and P(M) the base polytope of M. Then
ℓ(IM ) = dim P(M) + 1 = e.
Corollary 2.10. Let G be a finite connected simple graph on [n] with e edges and
write c for the number of cut vertices of G. Let M be the cycle matroid of G and
P(M) the base polytope of M. Then
ℓ(IM ) = dim P(M) + 1 = e − c.
Proof. Let, say, n ∈ [n] be a cut vertex of G. Let G′ and G′′ be the connected
components of the induced subgraph G|[n−1] . Write U for the vertex set of G′ and
V that of G′′ . Let M ′ be the cycle matroid of G|U ∪{n} and M ′′ that of G|V ∪{n} .
Let e′ be the number of edges of G|U ∪{n} and e′′ that of G|V ∪{n} . Since the base
ring F (IM ) is the Segre product of F (IM ′ ) and F (IM ′′ ), it follows that dim F (IM ) =
dim F (IM ′ ) + dim F (IM ′′ ) − 1. Let c′ be the number of cut vertices of G′ and c′′ that
of G′′ . Then using induction yields dim F (IM ′ ) = e′ − c′ and dim F (IM ′′ ) = e′′ − c′′ .
Hence dim F (IM ) = (e′ − c′ ) + (e′′ − c′′ ) − 1. Since e = e′ + e′′ and c = c′ + c′′ + 1,
the desired formula dim F (IM ) = dim P(M) + 1 = e − c follows.
Corollary 2.11. Let G be a finite disconnected simple graph on [n] with e edges and
G1 , . . . , Gs its connected component. Let ci denote the number of cut vertices of Gi
for 1 ≤ i ≤ s. Let M be the cycle matroid of G and P(M) the base polytope of M.
Then
ℓ(IM ) = dim P(M) + 1 = e − c − s + 1.
Proof. Let Mi be the cycle matroid of Gi for 1 ≤ i ≤ s. Since the base ring F (IM ) is
the Segre product of the base rings F (IM1 ), . . . , F (IMs ), the required formula follows
immediately from Corollary 2.10.
Let G be a finite 2-connected simple graph on [n] with e edges and P(M) the
base polytope of the cycle matroid M of G. Thus dim P(M) = e − 1. We introduce
the sequence of integers (δ0 , δ1 , δ2 , . . .) by the formula
∞
∞
X
X
e
i
(1 − λ)
δi t =
|iP(M) ∩ Zm | ti ,
i=0
i=0
14
where iP(M) is the ith dilated polytope {iα : α ∈ P(M)} of P(M). It then follows
that δi = 0 for i > dim P(M) = e − 1. We say that δ(P(M)) = (δ0 , δ1 , . . . , δe−1)
Pe−1 i
is the δ-vector of P(M) and δ(P(M); t) = i=0
δi t is the δ-polynomial of P(M).
Furthermore, introducing the integer r0 ≥ 1 by
r0 = min{r ∈ Z : r ≥ 1 and r(P(M) \ ∂P(M)) ∩ Zm 6= ∅},
where P(M) \ ∂P(M) is the relative interior of P(M), one has the formula
(4)
deg δ(P(M); t) = e − r0 .
We refer the reader to [12] for the detailed information about δ-vectors.
Since the base polytope P(M) possesses the integer decomposition property ([9,
p. 250]), it follows that
Lemma 2.12. The δ-polynomial δ(P(M); t) of P(M) coincides with the h-polynomial
of F (IM ).
Since the base ring F (IM ) is Cohen–Macaulay, it is well-known that the regularity
reg(F (IM )) of F (IM ) is one more than the degree of the h-polynomial h(F (IM ); t) of
F (IM ). In other words, the regularity reg(F (IM )) is one more than the degree of the
δ-polynomial δ(P(M); t) of P(M). Since the hyperplanes xi = 0 and xi = 1 of Rm
are supporting hyperplanes of P(M) for 1 ≤ i ≤ n, it follows that no lattice point of
Rm belongs to the relative interior of P(M). Hence, by using the formula (4), one has
deg δ(P(M); t) = deg h(F (IM )); t) ≤ e − 2. Thus, in particular reg(F (IM )) ≤ e − 1.
Furthermore, Theorem 2.5 guarantees that, unless F (IM ) is the polynomial ring,
one has deg h(F (IM ), t) ≥ 2 and reg(IM ) ≥ 3. As a result,
Theorem 2.13. Let G be a finite 2-connected simple graph on [n] with e edges. Let
M be the cycle matroid of G. Unless F (IM ) is the polynomial ring, one has
3 ≤ reg(F (IM )) ≤ e − 1.
It is natural to ask, given integers 3 ≤ r < e, if there exists a 2-connected simple
graph G with e edges for which reg(F (IM )) = r, where M is the cycle matroid of G
and F (IM ) is the base ring of M.
Corollary 2.14. Let G be a finite disconnected simple graph on [n] with e edges and
G1 , . . . , Gs its connected components. Let ci denote the number of cut vertices of Gi
for 1 ≤ i ≤ s. Let M be the cycle matroid of G. Unless F (IM ) is the polynomial
ring, one has
3 ≤ reg(F (IM )) ≤ e − c − s.
Finally, since the toric ideal of the base ring of a cycle matroid of a finite simple
graph is generated by quadratic binomials [2], it follows that no base ring of a cycle
matroid has a linear resolution.
Acknowledgement. This paper is supported by the National Natural Science
Foundation of China (11271275) and by the Foundation of the Priority Academic
Program Development of Jiangsu Higher Education Institutions.
15
References
[1] S. S. Abhyankar, Local rings of high embedding dimension, Amer. J. Math. 89 (1967), 1073–
1077.
[2] J. Blasiak, The toric ideal of a graphic matroid is generated by quadrics, Combinatorica 28
(2008), 283–297.
[3] W. Bruns and J. Herzog, Cohen-Macaulay rings, Cambridge University Press, Cambridge,
UK, 1998.
[4] K. J. Böröczky, F. Santos and O. Serra, On sumsets and convex hull, Discrete Comput. Geom.
52 (2014), 705–729.
[5] D. Eisenbud and S. Goto, Linear free resolutions and minimal multiplicity, J. Algebra 88
(1984), 89–133.
[6] J. A. Eagon and D. G. Northcott, Ideals defined by matrices and a certain complex associated
with them, Proc. R. Soc. Lond. Ser. A 269 (1962), 188–204.
[7] G. A. Freiman, Foundations of a structural theory of set addition, Translations of mathematical monographs 37, Amer. Math. Soc. Providence, Phode Island, 1973.
[8] S. Goto, K. Watanabe, On graded rings, I, J. Math. Soc. Japan 30 (1978), 179–213.
[9] J. Herzog and T. Hibi, Monomial ideals, Graduate Texts in Mathematics 260, Springer, London, 2010.
[10] J. Herzog, M. Mohammadi Saem and N. Zamani, On the number of generators of powers of
an ideal, arXiv: 1707.07302v1.
[11] J. Herzog and G. J. Zhu, Freiman ideals, arXiv: 1709.02827v1.
[12] T. Hibi, “Algebraic Combinatorics on Convex Polytopes,” Carslaw Publications, Glebe, N. S.
W., Australia, 1992.
[13] L. T. Hoa and N. D. Tam, On some invariants of a mixed product of ideals, Arch. Math. 94
(2010), 327–337.
[14] M. Hochster, Rings of invariants of tori, Cohen–Macaulay rings generated by monomials, and
polytopes, Ann. of Math. 96 (1972), 228–235.
[15] H. Ohsugi, J. Herzog and T. Hibi, Combinatorial pure subrings, Osaka J. Math. 37 (2000),
745–757.
[16] H. Ohsugi and T. Hibi, Toric Ideals generated by Quadratic Binomials, J. Algebra 218 (1999),
509–527.
[17] J. D. Sally, On the associated graded rings of a local Cohen–Macaulay ring, J. Math. Kyoto
Univ. 17 (1977), 19–21.
[18] K. Shah, On the Cohen-Macaulayness of the fiber cone of an ideal, J. Algebra 143 (1991),
156–172.
[19] Y. V. Stanchescu, On the simplest inverse problem for sums of sets in several dimensions,
Combinatorica 18 (1998), 139–149.
[20] N. L. White, The basis monomial ring of a matroid, Advances in Math. 24 (1977), 292–297.
Jürgen Herzog, Fachbereich Mathematik, Universität Duisburg-Essen, Campus
Essen, 45117 Essen, Germany
E-mail address: [email protected]
Takayuki Hibi, Department of Pure and Applied Mathematics, Graduate School
of Information Science and Technology, Osaka University, Toyonaka, Osaka 5600043, Japan
E-mail address: [email protected]
Guangjun Zhu, School of Mathematical Sciences, Soochow University, Suzhou
215006, P. R. China
E-mail address: [email protected]
16
| 0math.AC
|
arXiv:1803.07678v1 [math.GR] 20 Mar 2018
Lagrange Theorem For Hom-Groups
Mohammad Hassanzadeh
University of Windsor
Windsor, Ontario, Canada
[email protected]
Abstract
Hom-groups are nonassociative groups where the unitality and associativity are
twisted by a map. We show that a Hom-group (G, α) is a quasigroup. We use Cayley
table of Hom-groups to introduce some examples. We prove Lagrange theorem for
Hom-groups which says that the order of any Hom-subgroup H divides the order of
G. As an application we show that the dimension of a Hom-sub Hopf algebra of
the finite dimensional Hom-group Hopf algebra KG divides the the order of G. The
new tools introduced in this paper could potentially have applications in quasigroups,
nonassociative Hopf algebras, Hom-type objects, combinatorics, and cryptography.
1
Introduction
Nonassociative objects such as quasigroups, loops, non-associative algebras and Hopf algebras have many applications in several context. Among all, Hom-type objects have been
under intensive research and progress in the last decade. Hom-Lie algebras were appeared
in quantum deformations of Witt and Virasoro algebras [AS], [CKL], [CZ]. Hom-Lie algebras are a generalization of Lie algebras were Jacobi identity is twisted by a linear map
and first introduced in [HLS]. The Witt algebra is complexification of the Lie algebra of
∂
and the Lie bracket which
polynomial vector fields on a circle with a basis Ln = −z n+1 ∂z
is given by [Lm , Ln ] = (m − n)Lm+n . This Lie algebra is also can be viewed as the Lie
algebra of derivations D of the ring C[z, z −1 ], where D(ab) = D(a)b + aD(b). The Lie
bracket of two derivations D and D ′ is given by [D, D ′ ] = D ◦D ′ − D ′ ◦D. This algebra has
a central extension called the Virasoro algebra which is appearing in two-dimensional conformal field theory and string theory. One can define a quantum deformation of D given by
(z)
Dq (f )(z) = f (qz)−f
. This class of linea operators is different from usual derivations and
qz−z
they satisfy σ-derivation property Dq (f g) = gDq (f ) + σ(f )Dq (g) where σ(f )(z) = f (qz).
An example of a σ-derivation is the Jackson derivative on polynomials in one variable.
The set of σ-derivations with the classical bracket is a new type of algebra so called
σ-deformations of the Witt algebra. This algebra does not satisfy the Jacobi identity.
Instead it satisfies Hom-Jacobi identity and it is called Hom-Lie algebra. The corresponding associative algebras, called Hom-associative algebras, were introduced in [MS1]. Any
1
Hom-associative algebra with the bracket [a, b] = ab−ba is a Hom-Lie algebra. Later other
nonassociative objects such as Hom-coalgebras [MS2], Hom-bialgebras [MS2], [MS3], [Ya2],
[GMMP], and Hom-Hopf algebras [MS2], [Ya3], [Ya4] were introduced and studied.
Hom-groups are nonassociative objects which recently were appeared in the study of grouplike elements of Hom-Hopf algebras [LMT]. Having this notion will help us to find out
more about Hom-Hopf algebras. One knows that groups and Lie algebras have important rules to develop many concepts related to Hopf algebras. During last years HomLie algebras had played important rules to understand structures of Hom-Hopf algebras.
However the lack of having notion of Hom-groups could affect to miss some concepts
which potentially can be extended from Hom-groups to Hom-group Hopf algebras and
therefeore possibly to all Hom-Hopf algebras. A Hom-group (G, α) is a set G with a map
α : G −→ G which is endowed with a multiplication which satisfies Hom-associativity
property α(a)(bc) = (ab)α(c). Furthermore G has a Hom-unit element 1 which satisfies
a1 = 1a = α(a). Every element of g ∈ G has an inverse g−1 satisfying gg−1 = g−1 g = 1.
Although the twisting map α of a Hom-group (G, α) does not need to be inventible in the
original works [LMT], [H1], however more interesting results, including main results of this
paper, are obtained if α is invertible. As a result through the paper we assume that the
twisting map α is bijective. Since any Hom-group gives rise to a Hom-Hopf algebra, called
Hom-group Hopf algebra [H2], it is interesting to know what properties will be enforced
by the invertibility of α on Hom-Hopf algebras. It is shown in [CG] that the category
of modules over a Hom-Hopf algebra with invertible twisting is monoidal. For this reason they called them monoidal Hom-Hopf algebras. Also many interesting properties of
Hom-Hopf algebras, such as integrals, modules, comodules, and Hopf representations are
obtained when α is bijective, see [CWZ], [H2], [PSS] [ZZ], [MP]. The author in [H1] introduced some basics of Hom-groups, their representations and Hom-group (co)homology.
They showed that the Hom-group (co)homology is related to the Hochschild (co)homology
[HSS] of Hom-group algebras.
In Section 2, we introduce some fundamental notions of Hom-groups such as Hom-subgroups,
cosets, center, and centralizer of an element. In Theorem 2.12 we show that a Hom-group
is a quasigroups. This means division is always possible to solve the equations ax = b
and ya = b. However if a Hom-group is a loop then 1x = α(x) = x which means α = Id
and therefore G should be a group. We use properties of Cayley table of Hom-groups
to present some examples. Studying the Cayley table of quasigroups have applications
in combinatorics and cryptography, see [CPS], [DK], [SC], [BBW]. In Section 3, we use
cosets to partition a Hom-group G. Then we prove the Lagrange theorem for Hom-groups
showing that for any finite Hom-group, the order of any Hom-subgroup divides the order
the Hom-group. In Section 4, we apply Lagrange theorem to show that if A is a Hom-sub
Hopf algebra of KG, then dim(A) divides |G|. We finish the paper by presenting some
conjectures about Hom-groups.
2
Contents
1
Introduction
1
2
Hom-groups and Hom-subgroups
3
3 Hom-cosets and Lagrange theorem
7
4 Hom-group Hopf algebras
9
2
Hom-groups and Hom-subgroups
In this section, we introduce basic notions of Hom-groups such as Hom-subgroups, cosets,
and the center of a Hom-group (G, α). Our definition of a Hom-group is a special case of the
one discussed in [LMT] and [H1]. Through the paper, we assume the map α is invertible.
Therefore our axioms will be different from the ones in the original definition. We show
that some of the axioms can be obtained by Hom-associativity when α is invertible.
Definition 2.1. A Hom group consists of a set G together with a distinguished member
1 ∈ G, a bijective set map: α : G −→ G, a binary operation µ : G × G −→ G, where these
pieces of structure are subject to the following axioms:
i) The product map µ : G × G −→ G is satisfying the Hom-associativity property
µ(α(g), µ(h, k)) = µ(µ(g, h), α(k)).
For simplicity when there is no confusion we omit the multiplication sign µ.
ii) The map α is multiplicative, i.e, α(gk) = α(g)α(k).
iii) The element 1 is called unit and it satisfies the Hom-unitality conditions
g1 = 1g = α(g),
α(1) = 1.
v) For every element g ∈ G, there exists an element g −1 ∈ G which
gg −1 = g−1 g = 1.
One notes that in the definition of the Hom-groups in [LMT], [H1] and [H2] the condition
(v) was replaced by the one that for any g ∈ G there exists a natural number n satisfying
the Hom-invertibility condition αn (gg−1 ) = αn (g−1 g) = 1, where the smallest such n is
called the invertibility index of g. Clearly if α is invertible this condition will be simplified
to the condition (v).
Proposition 2.2. Let (G, α) be a Hom-group. Then
i) The inverse of any element is unique.
ii) (ab)−1 = b−1 a−1 for all a, b ∈ G.
3
Proof. i) First we show that the right inverse is unique. Let us assume an element g ∈ G
has right inverses a, b ∈ G. So ga = 1 and gb = 1. Since there exists g−1 ∈ G where
g−1 g = 1, then
α(g−1 )(ga) = α(g−1 )1.
By Hom-associativity we have (g−1 g)α(a) = α2 (g−1 ). By Hom-unitality we obtain α2 (a) =
α2 (g−1 ). Since α is invertible then a = g−1 . Similarly b = g −1 and therefore a = b.
Likewise we can prove that left inverse is unique. Now we show that left and right inverse
are the same. Let g ∈ G, ag = 1 and gb = 1. Therefore (ag)α(b) = 1α(b). By Homassociativity we have α(a)(gb) = α2 (b). Then α(a)1 = α2 (b). So α2 (a) = α2 (b). By
invertibility of α we get a = b. Therefore the inverse element is unique.
ii) The following compuatation shows that ab is the inverse of b−1 a−1 .
(ab)(b−1 a−1 )
= α(α−1 (ab))[b−1 a−1 ]
= [α−1 (ab)b−1 ]α(a−1 )
= [α−1 (a)α−1 (b)]b−1 α(a−1 )
= [α−1 (a)α−1 (b)]α(α−1 (b−1 )) α(a−1 )
= a α−1 (b)α−1 (b−1 ) α(a−1 )
= (a1)α(a−1 )
= α(a)α(a−1 ) = 1.
We used the Hom-associativity in the third equality, multiplicity of α in the fourth equality,
and Hom-associativity in the fifth equality.
Proposition 2.3. [H1] Let (G, µ) be a group and α : G −→ G a group automorphism.
Then (G, α ◦ µ, α) is a Hom-group.
Definition 2.4. [H1] A subset H of a Hom-group (G, α) is called a Hom-subgroup of G
if (H, α) is itself a Hom-group. We denote a Hom-subgroup H of G by H G.
Remark 2.5. If (G, α) is a Hom-group and H G, then for any h ∈ H we have α(h) =
1h ∈ H. Therefore α(H) ⊆ H. When α is a bijection we have α(H) = H.
Definition 2.6. The set gH = {gh, h ∈ H} is called the left coset of the Hom-subgroup
H in G with respect to the element g. Similarly the set Hg = {hg, h ∈ H} is called the
right coset of H in G.
We denote the number of elements of a Hom-group G by | G | .
Definition 2.7. The Center Z(G) of a Hom-group (G, α) is the set of all x ∈ G where
xy = yx for all y ∈ G.
Proposition 2.8. Let (G, α) be a Hom-group. Then Z(G) G.
4
Proof. Let x, y ∈ Z(G). Then for all a ∈ G we have
(xy)a = (xy)(α(α−1 (a))) = α(x)(yα−1 (a)) = α(x)[α−1 (a)y]
= [xα−1 (a)]α(y) = [α−1 (a)x]α(y) = a(xy).
Then Z(G) is closed under multiplication of G. Also if x ∈ Z(G) and a ∈ G then
xa−1 = a−1 x. Then (xa−1 )−1 = (a−1 x)−1 . Therefore ax−1 = x−1 a. So x−1 ∈ Z(G).
Therefore Z(G) is a Hom-subgroup of G.
Definition 2.9. The centralizer of an element x ∈ G is the set of all elements g ∈ G
where gx = xg and it is denoted by CG (x).
Proposition 2.10. Let (G, α) be a Hom-group. Then CG (x) G for all x ∈ G.
Proof. The proof is similar to previous Proposition.
Let (G, α) and (H, β) be two Hom-groups. The morphism f : G −→ H is called a
morphism of Hom-groups [H1], if β(f (g)) = f (α(g)) and f (gk) = f (g)f (k) for all g, k ∈ G.
Two Hom-groups G and H are called isomorphic if there exist a bijective morphism of
Hom-groups f : G −→ H.
Example 2.11. Let (G, α) and (G′ , α′ ) be two Hom-groups. Then (G × G′ , α × α′ ) is a
Hom-group by the multiplication given by (g, h)(g′ , h′ ) = (gg ′ , hh′ ).
Here we recall the definition of a quasigroup. A quasigroup (Q, ∗) is a set Q with a
multiplication ∗ : Q × Q −→ Q, where for all a, b ∈ Q, there exist unique elements
x, y ∈ Q such that both
a ∗ x = b,
y ∗ a = b.
Theorem 2.12. Every Hom-group (G, α) is a quasigroup.
Proof. First we show that the x = α−1 (a−1 )α−2 (b) satisfies the equation ax = b.
ax
= a α−1 (a−1 )α−2 (b)
= α(α−1 (a)) α−1 (a−1 )α−2 (b)
= α−1 (a)α−1 (a−1 ) α−1 (b)
= α−1 (aa−1 )α−1 (b)
= α−1 (1)α−1 (b)
= 1α−1 (b) = b.
We used the Hom-associativity in the third equality, invertibility of a in the fifth and the
Hom-unitality in the last equality. Similarly y = α−2 (b)α−1 (a−1 ) satisfies the equation
ya = b. Therefore G is a quasigroup.
5
Remark 2.13. (Cayley table of finite Hom-groups)
Since every Hom-group is a quasi group, then the Cayley table of a Hom-group has all
the properties of the one for quasigroup. However having invertibility and Hom-unitality
conditions, one obtains more properties. We put all the different elements of G in the first
row and column such that the Hom-unit 1 is in the first place. For simplicity we use the
matrix notation [cij ] for the Cayley matrix.
Some of the properties of Hom-groups are as follows:
i) Every row and column of the Cayley table of a Hom-group (G, α) is a permutation of
the set G. This is because Hom-groups have the cancelation property. In fact the Cayley
table of a Hom-group is an example of a Latin square.
ii) Rows and columns can not be the identity permutation of G except α = Id, which
means G is a group.
iii) If the element in the ith row and j th column is 1 then the element in the j th row and
ith column also should be 1. This is because of the invertibility condition.
iv) The Cayley table is symmetric if and only if (G, α) is abelian.
v) The first row and first column are the same. Because 1a = a1 = α(a).
Example 2.14. (Classification of Hom-groups of order 3).
In this example we show that there is only one Hom-group of order 3. We use the Cayley
table to classify all the Hom-groups of order 3. Let G = {1, a, b}. To begin filling out
the Cayley table, we start from the first row. Since by property (ii), the first row can not
be the identity permutation of G, therefore the only possibility is c12 = b and c13 = a.
Since the first column is the same as the first row therefore there will be only 4 spots
c22 , c23 , c32 , c33 to find. Now we argue on the place of the unit 1 in the second row. We
note that c22 6= 1 because otherwise c23 = a which is a contradiction because the third
column will have two copies of a which will not be an permutation. Therefore c22 = a and
c23 = 1. Now since the second and the third columns should be a permutation of G then
there is only one case left which is the following Cayley table:
G
1
a
b
1
1
b
a
a
b
a
1
b
a
1
b
This gives α(a) = b and α(b) = a. One can check that (G, α) with above Cayley table
satisfies the Hom-associativity and therefore it is an abelian Hom-group. This Hom-group
is isomorphic to Z3α where the multiplication is defined to be twisting the multiplication of
the additive cyclic group Z3 by the group automorphism α : Z3 −→ Z3 given by α(1) = 2,
as explained in Proposition 2.3.
We finish this section by introducing a non abelian Hom-group of order 6.
6
Example 2.15. (Hom-Dihedral group D3 )
We recall that the Dihedral group D3 is the smallest non-abelian group which is given by
{r, s, | r 3 = s2 = 1,
srs = r −1 }. In fact the elements of D3 are {1, r, r 2 , s, sr, rs}. We
consider the conjugation automorphism ϕs (x) = sxs−1 . Now we twist the multiplication
of D3 by ϕs , as explained in Proposition 2.3, to obtain a Hom-group D3α where α = ϕs .
The Cayley table is given by
D3α 1 r
1
1 r2
r
r2 r
2
r
r 1
s
s rs
rs sr s
sr rs sr
3
r 2 s rs
r s sr
1 sr rs
r 2 rs s
sr 1 r
rs r 2 1
s r r2
sr
rs
s
sr
r2
r
1
Hom-cosets and Lagrange theorem
In this section we focus on finite Hom-groups and we prove Lagrange theorem for them.
The following lemma shows that the number of elements in a Hom-subgroup and an its
cosets are the same.
Lemma 3.1. Let (G, α) be a finite Hom-group. If H G then |gH| = |H| for all g ∈ G.
Proof. It is enough to show that for hi 6= hj the elements ghi and ghj are different in gH.
Suppose ghi = ghj = b. By Theorem 2.12 we have
hi = α−1 (g −1 )α−2 (b) = hj .
Lemma 3.2. Let (G, α) be a finite Hom-group and H G. Then gH = H if and only if
g ∈ H.
Proof. If g ∈ H then gH ⊆ H. However by Lemma 3.1 we have |gH| = |H|. Therefore
gH = H. Conversely of gH = H then g1 ∈ H. So α(g) ∈ H. Since α is invertible and
α(H) = H then g ∈ H.
Lemma 3.3. Let (G, α) be a finite Hom-group and H G. For all x, y ∈ G if xH ∩ yH 6=
∅ then xH = yH.
Proof. Since xH ∩yH 6= ∅, there exists h1 , h2 ∈ H such that xh1 = yh2 . Then by Theorem
2.12 we have
x = α−2 (yh2 )α−1 (h−1
1 ).
7
By invertibility of α we obtain
x = [α−2 (y)α−2 (h2 )]α(α−2 (h−1
1 )).
By Hom-associativity we have
−1
−1
−2
x = α−1 (y)[α−2 (h2 )α−2 (h−1
1 )] = α (y)[α (h2 h1 )].
Now we show that xH ⊆ yH. Let xh ∈ xH. Then
xh = α−1 (y)[α−2 (h2 h−1
1 )] h
−1
= α−1 (y)[α−2 (h2 h−1
1 )] α(α (h))
−1
= y[α−2 (h2 h−1
1 )α (h)].
We used invertibility of α in the second equality, and Hom-associativity in the third
equality. Since α(H) = H then
−1
x = y[α−2 (h2 h−1
1 )α (h)] ∈ yH.
Therefore xH ⊆ yH. By Lemma 3.1 we have |xH| = |H| = |yH|. Therefore xH = yH.
As a consequence of the previous discussions we obtain the following result.
Proposition 3.4. Let (G, α) be a finite Hom-group and H G. Then the set of all cosets
of H in G gives a partition of the set G.
Theorem 3.5. (Lagrange theorem for Hom-groups)
Let (G, α) be a finite Hom-group and H G. Then |H| divides |G|.
Proof. By the previous Proposition the cosets of H in G gives a partition of G. By Lemma
3.1 the size of all cosets are the same as the size of H. Since G = ∪x∈G xH, then |G| is a
multiplication of |H|.
Example 3.6. Let G be a group, α : G −→ G a group automorphism and H G. We
twist the multiplication of G by α to obtain the Hom-group Gα as explained in Proposition
2.3. Then α(H) Gα where α(x)−1 = α(x−1 ). Specially if x ∈ Gα be an element
of of order n and < x > the cyclic group generated by x, then α(< x >) Gα and n
divides |Gα |. One notes that due to the nonassociativity of the product the Hom-subgroup
α(< x >) does not need to be cyclic in the usual sense.
Example 3.7. All cyclic groups of order 6 are isomorphic to Z6 . We define the group
automorphism a Z6 −→ Z6 given by
α(1) = 5, α(2) = 4, α(3) = 3, α(4) = 2, α(5) = 1, α(0) = 0
8
Now we twist the multiplication of Z6 by α as explained in Proposition 2.3 to obtain a
Hom-group Z6α given by the following table of multiplication
Z6α
0
1
2
3
4
5
0
0
5
4
3
2
1
1
5
4
3
2
1
0
2
4
3
2
1
0
5
3
3
2
1
0
5
4
4
2
1
0
5
4
3
5
1
0
5
4
3
2
It can be verified that Z2α = {0, 2} and Z3α = {0, 2, 4} are the only non-trivial Homsubgroups of the Hom-group Z6α . Their sizes are 2 and 3 which both divides 6. However
the Hom-subgroup Z3α is not cyclic in the usual sense. In fact 2 + 2 in Z3α is 2 and 4 + 4
is 4. Therefore 2 and 4 can not be the generators of Z3α in the usual sense. This example
shows that the proper notion of power of an element in the Hom-groups is not clear.
4
Hom-group Hopf algebras
In this section we apply Lagrange theorem for finite Hom-groups to find out about dimensions of Hom-sub-Hopf algebras of Hom-group Hopf algebras. First we recall the
definitions of Hom-algebras, Hom-coalgebras, Hom-bialgebras and Hom-Hopf algebras.
By [MS1], a Hom-associative algebra A over a field K is a K-vector space with a bilinear
map µ : A ⊗ A −→ A, called multiplication, and a linear homomorphism α : A −→ A
satisfying the Hom-associativity condition
µ(α(a), µ(b, c)) = µ(µ(a, b), α(c)),
for all elements a, b, c ∈ A. A Hom-associative algebra A is called unital with unit 1 if
α(1) = 1, and a1 = 1a = α(a). By [MS2], [MS3], a Hom-coalgebra is a triple (C, ∆, ε, β),
where C is a K-vector space, ∆ : C −→ C ⊗ C a linear map, called comultiplication, with
a Sweedler notation ∆(c) = c(1) ⊗ c(2) , counit ε : K −→ C, and β : C −→ C a linear map
satisfying the Hom-coassociativity condition,
β(c(1) ) ⊗ c(2)(1) ⊗ c(2)(2) = c(1)(1) ⊗ c(1)(2) ⊗ β(c(2) ),
and
c(1) ε(c(2) ) = ε(c(1) )c(2) = β(c),
ε(β(c)) = ε(c).
A (α, β)-Hom-bialgebra is a tuple (B, m, 1, α, ∆, ε, β) where (B, m, 1, α) is a unital Homalgebra and (B, ε, ∆, β) is a counital Hom-coalgebra where ∆ and ε are morphisms of
Hom-algebras, that is
9
i) ∆(hk) = ∆(h)∆(k).
ii) ∆(1) = 1 ⊗ 1.
iii) ε(xy) = ε(x)ε(y).
iv) ε(1) = 1.
v) ε(α(x)) = ε(x).
Here we recall the definition of Hom-Hopf algebras from [MS2] and [MS3]. A Hombialgebra (B, m, η, α, ∆, ε, β) is called a (α, β)-Hom-Hopf algebra if it is endowed with a
morphism S : B −→ B, called antipode, satisfying
a) S ◦ η = η and ε ◦ S = ε.
b) S is an inverse of the identity map Id : B −→ B for the convolution product, i.e, for
any x ∈ B,
S(x(1) )x(2) = x(1) S(x(2) ) = ε(x)1B .
(4.1)
This definition of Hom-Hopf algebra and specially antipodes is different from the one in
[LMT]. For more details see [H2]. However if α is invertible, both definitions will be
equivalent.
Example 4.1. For any Hom-group (G, α), the Hom-group algebra KG is a (α, Id)-HomHopf algebra. It is a free algebra on G where the coproduct is given by ∆(g) = g ⊗ g,
counit by ε(g) = 1, the antipode by S(g) = g−1 , and β = Id with α which is linearly
extended from G to KG. One notes that elements g ∈ KG are group-like elements. Also
KG is a cocommutative Hom-Hopf algebra. If G is an abelian Hom-group then KG is a
commutative Hom-Hopf algebra.
Lemma 4.2. Let (G, α) be a Hom-group. The vector space A is a Hom-sub Hopf algebra
of KG if and only if there exists H G where A = KH.
Proof. Let A be a Hom-sub-Hopf algebra of KG. Then the set of group-like elements of
A forms a Hom-group H, see [LMT], and clearly A = KH. Conversely if H G then by
the structure of product, coproduct and the antipode explained in the previous example,
KH is a Hom-sub Hopf algebra of KG.
Theorem 4.3. Let (G, α) be a Hom-group and KG be the Hom-group Hopf algebra. If A
is a Hom-sub-Hopf algebra of KG, then dim(A) divides |G|.
Proof. By previous Lemma there exists H G where A = KH. Since dim(A) = |H|,
then by Lagrange theorem dim(A) divides |G|.
Example 4.4. Let us consider the cyclic group Z5 . We define a group automorphism
α : Z5 −→ Z5 given by
α(1) = 2, α(2) = 4, α(3) = 1, α(4) = 3, α(0) = 0.
10
One twists the multiplication of Z5 by α as explained in Proposition 2.3 to obtain a
Hom-group Z5α given by the following table of multiplication
Z5α
0
1
2
3
4
0
0
2
4
1
3
1
2
4
1
3
0
2
4
1
3
0
2
3
1
3
0
2
4
4
3
0
2
4
1
Since the order of KZ5α is prime, by previous theorem it does not have any non-trivial
Hom-sub Hopf algebra.
Example 4.5. Consider the Hom-group Hopf algebra KG. Since Z(G) G then the
center of the Hom-algebra KG is the same as KZ(G) and its dimension divides |G|.
Remark 4.6. Conjectures
A challenge in studying a Hom-group (G, α) is defining a proper notion of power of an
element. Since G is not associative we can define two different types of powers called left
and right powers. Following the context of nonassociative objects such as quasigroups, an
approach to define a right power of an element x in a Hom-group (G, α) is as follows. We
set x1 = x, x2 = xx. Now x3 = (x2 )x and inductively we can define other right powers.
In fact one can define the right multiplication function Ra (x) = xa. So x2 = Rx (x),
x3 = Rx (x2 ) and generally xn = Rx (xn−1 ). Similarly if La (x) = x then left powers of x
can inductively be defined by xn = Lx (xn−1 ). However this method has some problems
such as defining cyclic Hom-subgroups. The notion of power and order of an element of G
, and therefore cyclic Hom-subgroups are not clear for Hom-groups. Consequently some
fundamental theorems of group theory such as Cauchy’s theorem will be left as a conjecture
for Hom-groups; if (G, α) is a finite Hom-group and p is a prime number dividing the order
of G then G contains an element, and therefore a Hom-subgroup of order p.
References
[AS] N. Aizawa and H. Sato, q-deformation of the Virasoro algebra with central extension,
Phys. Lett. B, 256, (1991), p. 185-190.
[BBW] D. Bryanta, M. Buchanana, and I. M. Wanlessb The spectrum for quasigroups
with cyclic automorphisms and additional symmetries, Discrete Mathematics, Volume
309, Issue 4, (2009), Pages 821-833.
[CG] S. Caenepeel, and I. Goyvaerts, Monoidal Hom-Hopf algebras, Comm. Algebra 39(6),
(2011), p. 2216-2240.
11
[CKL] M. Chaichian, P. Kulish, and J. Lukierski , q-deformed Jacobi identity, q-oscillators
and q-deformed infinite-dimensional algebras, Phys. Lett. B, 237 (1990), p. 401-406.
[CPS] O. Chein, H. Pflugfelder, and J. D. H. Smith, Quasigroups and loops: theory and
applications, Heldermann Verlag, Berlin, (1990).
[CZ] T. L. Curtright and C. K. Zachos , Deforming maps for quantum algebras, Phys.
Lett. B, 243 (1990), p. 237-244.
[CWZ] Y. Chen, Z. Wang, and L. Zhang, Integrals for monoidal Hom-Hopf algebras and
their applications, Journal of Mathematical Physics 54, 073515, (2013).
[DK] J. Denes and A. Keedwell, Latin squares and their applications, Academic Press,
Inc., Budapest, (1974).
[GMMP] G. Graziani, A. Makhlouf, C. Menini, and F. Panaite BiHom-associative algebras, BiHom-Lie algebras and BiHom-bialgebras, SIGMA 11, (2015), 086, 34 pages.
.
[H1] M. Hassanzadeh, Hom-groups, Representations
https://arxiv.org/abs/1801.07398, (2018).
[H2] M.
Hassanzadeh,
On
antipodes
https://arxiv.org/abs/1803.01441, (2018).
of
and
homological
Hom-Hopf
algebra,
algebras,
[HLS] J. T. Hartwig, D. Larsson, and S. D. Silvestrov, Deformations of Lie algebras using
σ-derivations, J. Algebra 295 (2006), no. 2, p. 314.361.
[HSS] M. Hassanzadeh, I. Shapiro and S. Sütlü, Cyclic homology for Hom- algebras, Journal of Geometry and Physics, Volume 98, December (2015) , P. 40-56.
[LMT] C. Laurent-Gengouxa, A. Makhlouf, and J. Teles, Universal algebra of a Hom-Lie
algebra and group-like elements , Journal of Pure and Applied Algebra, Volume 222,
Issue 5, (2018), P. 1139-1163.
[MS1] A. Makhlouf and S. D. Silvestrov, Hom-algebra structures, J. Gen. Lie Theory Appl.
2 (2008), no. 2, p. 51-64.
[MS2] A. Makhlouf and S. Silvestrov, Hom-Lie admissible Hom-coalgebras and Hom-Hopf
algebras, Generalized Lie theory in mathematics, physics and beyond, Springer, Berlin,
(2009), p. 189-206.
[MS3] A. Makhlouf and S. Silvestrov , Hom-algebras and Hom-coalgebras, J. Algebra Appl.
9 (2010), no. 4, p. 553-589.
[MP] A. Makhlouf, and F. Panaite Integrals for monoidal Hom-Hopf algebras and their
applications, Journal of Algebra Volume 441, 1 November (2015), Pages 314-343
12
[PSS] F. Panaite, P. Schrader, and M. D. Staic, Hom-Tensor Categories and the HomYang-Baxter Equation, https://arxiv.org/abs/1702.08475, (2017).
[SC] B. Schneier, Applied cryptography, John Wiley and Sons, Inc., New York, NY, (1996).
[Ya2] D. Yau, Hom-bialgebras and comodule Hom-algebras, Int. Electron. J. Algebra 8
(2010), p. 45-64.
[Ya3] D. Yau, Hom-quantum groups: I. Quasi-triangular Hom-bialgebras, J. Phys. A 45
(2012), no. 6.
[Ya4] D. Yau, Enveloping algebras of Hom-Lie algebras, Journal of Generalized Lie Theory
and Applications, 2, (2008), p. 95-108.
[ZZ] X. Zhao, and X. Zhang, Lazy 2-cocycles over monoidal Hom-Hopf algebras, Colloq.
Math. 142(1), (2016), p. 61-81.
13
| 4math.GR
|
arXiv:1802.04735v1 [cs.CV] 13 Feb 2018
Semantic Scene Completion Combining Colour and Depth:
preliminary experiments
André B. S. Guedes
Teófilo E. de Campos
Universidade de Brası́lia
Brası́lia-DF, 70910-900, Brazil
Adrian Hilton
CVSSP, University of Surrey
Guildford, GU2 7XH, UK
[email protected]
[email protected]
http://www.cic.unb.br/˜teodecampos/
Abstract
opsis and top-down is the use of prior knowledge from related scenes and objects. If both types of cues are combined,
it is possible to estimate a complete scene geometry by using a single visual and depth map of a scene. This is well
illustrated in Figure 2 of [13]. A visual sensor captures a
single view of a scene which provides measurements (e.g.
RGB and Depth) of the visible objects but it is not possible
to measure the geometry of occluded regions. However, if
the class of the objects is identified, it is possible to infer the
complete scene geometry, enabling a full 3D representation
to be proposed.
Semantic scene completion is the task of producing a
complete 3D voxel representation of volumetric occupancy
with semantic labels for a scene from a single-view observation. We built upon the recent work of Song et al. [13], who
proposed SSCnet, a method that performs scene completion
and semantic labelling in a single end-to-end 3D convolutional network. SSCnet uses only depth maps as input, even
though depth maps are usually obtained from devices that
also capture colour information, such as RGBD sensors and
stereo cameras. In this work, we investigate the potential of
the RGB colour channels to improve SSCnet.
Solid computational demonstrations of this have started
to be published recently. Notably, Song et al. [13] introduced the problem of Semantic Scene Completion (SSC),
i.e., given an depth map, the goal is to generate a 3D image where each voxel is associated to one out of N + 1 labels, where there are N known object labels plus an ‘empty
space’ label.
1. Introduction
The task of reasoning about scenes in 3D is one of the
seminal goals of Computer Vision [8]. If the 3D geometry of a scene is known, robots are able to plan trajectories, avoid collisions or clean surfaces. If the semantic
labels of each surface or voxel is also known a robot can
also figure interact with the environment and perform more
complex tasks, such as moving objects from one location to
another; opening/closing doors, drawers, windows; operating kitchen appliances etc. Three-dimensional maps with
labelled voxels have several other applications, including
surveillance, assistive computing, augmented reality and so
on. One issue is that capturing the full geometry of a scene
can be time consuming (if it is done using a scanning technique [9]) or expensive (if it is done using a rig of calibrated
sensors).
It is well known that vision is a combination of so called
bottom-up and top-down processes [8]. Bottom-up information can be obtained by matching local features for stere-
In [13], this problem is approached using a Deep 3D
Convolutional Neural Network coined SSCNet. That paper demonstrates impressive results on completing and labelling a full 3D scene generated from a single depth map.
Using a combination of bottom-up dues (from the depth
sensor) and top-down cues (learnt from the training set),
their method is able to infer the geometry and labels of the
whole scene, including heavily occluded regions, such as
the regions under tables and behind sofas, as illustrated in
Figure 1 of [13].
However, one of the main limitations of SSCNet is that
it was not designed to use any colour information, only
depth maps are used. This clearly impairs the method as
indoor scenes generally include various sources of error in
depth and geometry estimation. Highly reflective scenes
with glass, mirrors or shiny surfaces usually induce false
depth. If depth is captured using stereo cameras, textureless and non-Lambertian surfaces often result in errors in
feature detection and matching. Colour information also
The work described in this extended abstract and in the attached poster
was presented at the ICCV 2017 Workshop on 3D Reconstruction meets
Semantics (3DRMS).
1
data - SuncgData
disambiguates between different objects that have similar
shape or that are co-planar, like posters on the wall. Furthermore, it is clear that colour offers crucial information
for semantic labelling that strongly complement depth information, as seen in papers that focus on semantic segmentation from RGBD images, e.g. [11, 12, 7, 4, 14, 1, 2].
In this paper, we propose to use colour in addition to
depth for Semantic Scene Completion. For that, we propose
modifications of the SSCnet architecture in order to fuse
RGB and depth. A new input layer was proposed to encode
colour in the visible frustum and we combined a feature extraction training technique for multiple view learning.
data
seg_label
conv1_1 - Convolution (7, 2, 3, 1)
16
color_data
relu1_1 - ReLU
conv2_1 - Convolution (3, 1, 1, 1)
reduction2_1 - Convolution (1, 1, 0, 1)
32
32
conv2_1
reduction2_1
conv2_2 - Convolution (3, 1, 1, 1)
relu2_1 - ReLU
32
conv2_2
res2_2 - Eltwise
res2_2
pool2 - MAX Pooling (2, 2, 0)
2. Colour SSCNet
Depth maps are acquired using an RGB-D sensor and using the sensor’s intrinsic calibration parameters, a 3D point
cloud is generated. The observed geometry is then encoded
using flipped Truncated Signed Distance Function (fTSDF),
proposed in [13]. This method associates a value to each
point in the 3D space to a function of its distance to the
nearest surface point. The sign of this value indicates if it is
visible or occluded. Apart from the occlusion coding, this
method is viewpoint independent.
The fTSDF encoding of voxels describe the geometry of
the space, but it does not carry any information about the
colour or grey level of the visible objects. We propose to
encode the RGB values of the visible surfaces in another
voxel representation of the scene. The three channels are
normalised to range from 0 to 1. Empty spaces and occluded regions are coded with the -1 value for the three
colour channels.
We apply these two encoding techniques to RGB and
Depth signals and run them through a 3D CNN that learns
to map from RGBD to a labelled 3D volume. Labelled
volumes were obtained as described in [13], i.e., the binvox voxelisation technique [10] was applied to 3D models,
which accounts for both surface and interior voxels using a
space carving approach.
We built upon the 3D CNN architecture of SSCNet [13].
To combine RGB and Depth, we propose the two fusion
schemes described below.
Early fusion. The first layer of SSCNet was adapted so
that it takes as input a concatenation of fTSDF and the three
colour channels encoded as described above. The remaining
of the network is the same.
Mid-level fusion. This architecture is depicted in Figure 1, drawn using Caffe [6] (better viewed on a screen).
The numbers in brackets are: kernel size, stride, pad and
dilation factor, respectively. The branch on the left is essentially a copy of SSCNet. A colour 3D CNN was built following a similar architecture to SSCNet up to the concatenation layer. This layer originally aggregated the output of
five scales gathered from previous layers and it is followed
seg_weight
conv1_1
add2_2 - ReLU
pool2
reduction3_1 - Convolution (1, 1, 0, 1)
conv3_1 - Convolution (3, 1, 1, 1)
64
64
conv3_1
reduction3_1
relu3_1 - ReLU
conv3_2 - Convolution (3, 1, 1, 1)
64
conv3_2
res3_2 - Eltwise
color_conv1_1 - Convolution (7, 2, 3, 1)
16
res3_2
add3_2 - ReLU
color_conv1_1
relu3_2 - ReLU
conv3_3 - Convolution (3, 1, 1, 1)
color_pool2 - MAX Pooling (2, 2, 0)
color_relu1_1 - ReLU
64
conv3_4 - Convolution (3, 1, 1, 1)
conv3_3
color_pool2
relu3_3 - ReLU
color_conv3_3 - Convolution (3, 1, 1, 1)
64
64
conv3_4
color_conv3_3
res3_4 - Eltwise
color_conv3_4 - Convolution (3, 1, 1, 1)
color_relu3_3 - ReLU
64
res3_4
add3_4 - ReLU
color_conv3_4
relu3_4 - ReLU
conv3_5 - Convolution (3, 1, 2, 2)
color_conv3_5 - Convolution (3, 1, 2, 2)
64
64
conv3_5
relu3_5 - ReLU
color_conv3_5
conv3_6 - Convolution (3, 1, 2, 2)
color_relu3_5 - ReLU
color_conv3_6 - Convolution (3, 1, 2, 2)
64
64
conv3_6
color_conv3_6
res3_6 - Eltwise
conv3_7 - Convolution (3, 1, 2, 2)
add3_6 - ReLU
color_res3_6 - Eltwise
res3_6
color_res3_6
relu3_6 - ReLU
color_add3_6 - ReLU
color_relu3_6 - ReLU
64
color_conv3_7 - Convolution (3, 1, 2, 2)
64
conv3_7
conv3_8 - Convolution (3, 1, 2, 2)
color_conv3_7
relu3_7 - ReLU
color_relu3_7 - ReLU
color_conv3_8 - Convolution (3, 1, 2, 2)
64
64
conv3_8
color_conv3_8
res3_8 - Eltwise
color_res3_8 - Eltwise
res3_8
add3_8 - ReLU
color_res3_8
Concat - Concat
color_add3_8 - ReLU
sum
sum_relu - ReLU
color_conv4_1_new - Convolution (1, 1, 0, 1)
128
conv4_1
relu4_1 - ReLU
conv4_2 - Convolution (1, 1, 0, 1)
128
conv4_2
relu4_2 - ReLU
fc12 - Convolution (1, 1, 0, 1)
12
fc12
loss - SoftmaxWithLoss
loss
Figure 1. Network architecture for mid-level fusion.
by three 3D convolutional layers. Although this network
does not have any fully connected (FC) layer, the last three
layers perform the same class as FC layers in classification
CNNs, as they are closer to the output, which produces labelled data. A typical mid-level fusion using CNNs is done
at the input of the first FC layer. Therefore, we believe the
most appropriated layer to fuse SSCNet-like sub-networks
is at the concatenation layer.
However, the original SSCNet is very memory-intensive
and it requires about 7GB of GPU memory to process a
single depth image1 . If we were to duplicate all layers of
the network up to the concatenation layer, much more than
12GB would be required, whereas most GPU models available nowadays have up to 12GB of RAM. To add to the
challenge, our RGB coding uses three channels per image,
rather than one as in the fTSDF model, though this only affects the first convolutional layer. Therefore, some of the
convolutional layers were removed from the colour branch
of SSCNet, but we have preserved all the dilated convolution layers, as this is a significant feature of SSCNet which
widely expands the receptive field of the network [15].
In addition, we also evaluated a colour-only SSCNet,
which follows the same architecture as the colour branch
of the mid-level RGB-D fusion network, but it is followed
by the top three convolutional layers, without aggregating
activations from the fTSDF branch.
3. Experiments and Training Strategies
Our evaluations focused on the NYU depth v2
dataset [12], using the standard split of 795 training samples and 654 test samples2 . However, instead of the standard semantic segmentation labels, we used the labels devised for scene completion, where objects are grouped into
7 categories plus window, wall, floor, ceiling and another
category that identifies free space. This set of labels originated from [5]. As explained in [13], ground truth volumes
were obtained from 3D mesh annotations of [3]. Our implementation was developed using the Caffe framework [6].
We evaluated the two architectures proposed in Section 2: early and mid-level fusion and compared it against
the original SSCNet and colour-only. For all methods that
we proposed, training was done following these strategies:
• Random initialisation: all parameters were randomly
initialised and the whole network was trained from
scratch.
1 The original implementation from the authors, obtained from https:
//github.com/shurans/sscnet, actually requires almost 12GB.
We removed some redundancy from their code, freeing about 5GB of
memory.
2 We used the train+validation split for training and the test
split for testing, following the sample indices available from
https://github.com/shelhamer/fcn.berkeleyvision.
org/tree/master/data/nyud.
• Feature learning: we kept the original SSCNet parameters trained by Song et al. [13] for all the original
layers and optimised only the colour layers, i.e., the
original SSCNet parameters were frozen.
• Fine tuning: this is similar to the strategy above, except that instead of freezing the original layers, we also
enabled their parameters to be optimised, but with the
learning rate ratio of 0.2 times the ratio of the new layers.
• Surgery: was applied only for the early fusion approach. It is similar to fine tuning, except that the
weights of the input layer which related to depth were
set to the original parameters of the first layer of SSCNet and the other weights (linked to the colour channels) of the same convolutional kernel were initialised
randomly.
Voxel labelling is done by applying soft-max to the
scores of the last convolutional layer of the networks and
optimisation is done using cross-entropy as a loss function,
averaged out over all classes.
The results were evaluated using the Intersection over
the Union (IoU) between predicted class labels and ground
truth, averaging out over all voxels in the test set and all
classes. We followed [13] and evaluated our results both in
terms of completion (i.e., the ability to detect if an occluded
voxel is occupied or free space) and in terms of semantic
labelling of voxels of all classes.
4. Results and Discussion
Our results so far show that none of the proposed architectures and training strategies actually lead to results that
are better than the original SSCNet based only on depth observations, i.e., through the training iterations, our results
peaked at scene completion IoU of 56.6 and average semantic scene completion of 30.5, which are both results obtained by the original SSCNet on the test set of the NYU
depth v2 dataset. In other words, our experiments in the
NYU depth v2 dataset (with the 12 category labels [5])
show that the proposed method for coding colour information is not as discriminative as fTSDF, neither it complements depth information.
However, the performance of our colour-only network,
initialised with randrom weights, followed a monotonic increase as the number of training iterations increased, though
it did not converge with the same number of iterations as
the architectures that use depth. Therefore, there is certainly relevant information in RGB, but it should probably
be combined with fTSDF in a different way, perhaps using late fusion. Even if early or mid-level fusion are not
the ideal strategies in this problem, further investigation is
also needed to understand why RGB has not complemented
Depth at all. It might be an artefact of the dataset and set
annotated classes, as it is possible that geometry alone is
already very discriminative. A suggestion is to verify this
using more complex scenes with more occlusions or with
finer object class labels.
Our results have also shown that unconstrained Fine
Tuning leads to a higher decrease in the loss function than
the constrained optimisation methods (Feature Learning
and Surgery). However, after 1000 iterations, the test set
performance (measured by IoU) starts to decrease due to
over-fitting. Although the loss is lower for Fine Tuning, we
did not observe a significant difference between the methods in terms of test set performance.
5. Conclusion
In this paper we reported ongoing work that considers
the problem of Semantic Scene Completion in 3D from a
single RGBD image. Starting from the 3D CNN architecture of SSCNet [13], which used only depth maps as input,
we proposed to combine RGB and Depth information using
early and mid-level fusion schemes.
Our preliminary results were not better than the original depth-only method. Therefore, further investigation is
needed in order to verify if the dataset (NYU depth v2 with
12 labels obtained from [5]) prilidedges structural information such that depth is already very discriminative. A finer
set of labels or a more complex dataset should be evaluated.
Other directions of future work are to evaluate late fusion
scheme and investigate other ways to encode RGB information.
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
6. Acknowledgements
We are grateful for the valuable comments and suggestions provided by anonymous reviewers of the first
version of this manuscript. TEdC’s attendance to this
workshop is sponsored by Fundação de Apoio a Pesquisa
do Distrito Federal (FAP-DF), edital 01/2017, protocolo
no 18708.76.44500.14072017
References
[1] K. Chen, Y.-K. Lai, and S.-M. Hu. 3D indoor scene modeling
from RGB-D data: a survey. Computational Visual Media,
1(4):267–278, 2015.
[2] D. Eigen and R. Fergus. Predicting depth, surface normals
and semantic labels with a common multi-scale convolutional architecture. In Proc 15th Int Conf on Computer Vision, Santiago, Chile, 2015.
[3] R. Guo, C. Zou, and D. Hoiem. Predicting complete 3D
models of indoor scenes. arXiv preprint arXiv:1504.02437,
2015.
[4] S. Gupta, P. Arbelez, R. Girshick, and J. Malik. Indoor scene
understanding with RGB-D images: Bottom-up segmenta-
[13]
[14]
[15]
tion, object detection and semantic segmentation. Int Journal
of Computer Vision, pages 1–17, 2014.
A. Handa, V. Patraucean, V. Badrinarayanan, S. Stent, and
R. Cipolla. Understanding real world indoor scenes with
synthetic data. In Proc IEEE Conf on Computer Vision and
Pattern Recognition, Las Vegas, NV, June 26 - July 1, pages
4077–4085, 2016.
Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint
arXiv:1408.5093, 2014.
O. Kahler and I. Reid. Efficient 3D scene labeling using fields of trees. In Proc 14th Int Conf on Computer Vision, Sydney, Australia, pages 3064–3071, 2013.
DOI:10.1109/ICCV.2013.380.
D. C. Marr. Vision: a computational investigation into
the human representation and processing of visual information. MIT Press, 1982. ISBN-13: 9780262514620,
DOI:10.7551/mitpress/9780262514620.001.0001.
R. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim,
A. Davison, P. Kohli, J. Shotton, S. Hodges, and A. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and
tracking. In Proc ISMAR, 2011.
F. S. Nooruddin and G. Turk. Simplification and repair of
polygonal models using volumetric techniques. IEEE Transactions on Visualization and Computer Graphics, 9(2):191–
205, 2003. Used in the implementation of binvox, available
from Patrik Min’s website: http://www.patrickmin.
com/binvox/.
N. Silberman and R. Fergus. Indoor scene segmentation using a structured light sensor. In Proc ICCV Workshops, 2011.
N. Silberman, D. Hoiem, P. Kohli, and R. Fergus. Indoor segmentation and support inference from RGBD
images.
In Proc European Conf on Computer Vision, pages 746–760, 2012.
NYU depth v2 dataset
available from http://cs.nyu.edu/˜silberman/
datasets/nyu_depth_v2.html.
S. Song, F. Yu, A. Zeng, A. X. Chang, M. Savva, and
T. Funkhouser. Semantic scene completion from a single depth image. In Proc 30th IEEE Conf on Computer Vision and Pattern Recognition, Honolulu, Hawaii,
July 21-26, 2017. Preprint available as technical report
arXiv:1611.08974.
A. Wang, J. Lu, J. Cai, G. Wang, and T.-J. Cham. Unsupervised joint feature learning and encoding for RGBD scene labeling. IEEE Trans Image Processing, 2015.
DOI:10.1109/TIP.2015.2465133.
F. Yu and V. Koltun. Multi-scale context aggregation by dilated convolutions. In Proc Int Conf on Learning Representations ICLR, Toulon, France, April 24-26 2016. Preprint
available as arXiv technical report arXiv:1511.07122.
chair
floor
table
chair
wall
b) shape completion c) completion+labeling
Acknowledgements:
T. deCampos thanks FAPDF, process 18708.76.44500.14072017, for
partially funding his attendance to ICCV and the 3DMRS workshop.
A. Hilton thanks the EPSRC Programme Grant EP/L000539/1 (S3A).
We thank the authors above for sharing code and pre-trained
SSCnet parameters.
Main reference:
S. Song, F. Yu, A. Zeng, A. X. Chang, M. Savva, and T. Funkhouser.
Semantic scene completion from a single depth image. CVPR 2017.
This work
by Song et al. CVPR 2017
SSC from depth only
Previous method
a) surface labeling
floor
table
wall
observed surface
observed free
occluded
outside view
outside room
conv(32,1,1,1)
add
add
conv(32,1,1,1)
- For the early fusion network, we tried these strategies:
●
Feature learning: where only the first layer was trained to learn colour parameters and
the remaining parameters where locked.
●
Fine tuning: where a learning rate of 0.2 was used for the previously learnt layers and
1.0 for the new layer.
- For the mid-level fusion network, we also tried learning all parameters from scratch with
random initialisation.
Learning strategies
- The input point cloud is converted to a 3D volume aligned with gravity and following
Manhattan assumptions.
- Depth is encoded using the flipped Truncated Signed Distance Function (fTSDF).
- Colour is encoded for each voxel as an RGB triplet.
●
To compute the colour of a voxel, the inverse of the voxelation process is applied and
colour is computed by averaging out the original RGB region.
●
Colour values are normalised to be between 0 and 1.
●
Occluded or empty voxels have their colour set to -1.
Depth and Colour encoding
Encoding and learning
Mid-level fusion
add
add
conv(32,1,1,1)
add
dilated (64,3,1,2)
dilated (64,3,1,2)
dilated (64,3,1,2)
dilated (64,3,1,2)
dilated (64,3,1,2)
dilated (64,3,1,2)
conv(32,1,1,1)
conv (32,3,1,1)
conv (32,3,1,1)
dilated (64,3,1,2)
add
add
Early fusion: concatenation at the first layer
conv (16,7,2,1)
conv (16,7,2,1)
conv (16,7,2,1)
add
dilated (64,3,1,2)
dilated (64,3,1,2)
dilated (64,3,1,2)
dilated (64,3,1,2)
- Introduced by Song et al. CVPR 2017
(figure extracted from their paper)
conv (32,3,1,1)
conv (32,3,1,1)
add
add
dilated (64,3,1,2)
add
add
Semantic Scene Completion
pool
i
ng
pool
i
ng
pool
i
ng
conv (64,3,1,1)
conv (64,3,1,1)
conv (64,3,1,1)
conv (64,3,1,1)
conv (64,3,1,1)
conv (64,3,1,1)
conv (64,3,1,1)
conv (64,3,1,1)
conv (64,3,1,1)
conv (64,3,1,1)
c
onc
at
c
onc
at
Fusion schemes
conv (128,1,1,1)
conv (128,1,1,1)
Problem statement
conv (128,1,1,1)
45.60
Learning
from
scratch
59.3
62.46
Feature
learning
Song et al
CVPR’17
61.50
58.9
Feature
learning
Fine
tuning
59.4
Fine
tuning
Precision
(%)
92.9
100.0
82.26
84.11
93.1
92.4
56.6
46.60
54.31
54.20
56.1
56.3
IoU
(%)
Fine
tuning
Song et al
CVPR’17
–
21
–
Learning
from
scratch
Depth
only
36.92
–
Precision
(%)
Feature
learning
Fine
tuning
Learning
Strategy
Colour
only
Mid-level
fusion
–
46
18.30
45.03
34.23
Recall
(%)
- inferring labels in 3D
30.50
16
05.92
27.45
17.19
IoU
(%)
Semantic Scene Completion
Phantom cones are being
predicted.
It is likely that the data
balancing strategy is not
inducing a high penalty when
an empty voxel is predicted
as containing an object.
Colour information will be
more crucial on datasets with
finer object labels (e.g.,
poster on the wall).
Observations
Top view of prediction
The resuts bellow illustrates one of the potential causes for the degeneration of our results
Sample result
Depth
only
Mid-level
fusion
Early
Fusion
Learning
Strategy
- inferring geometry
Recall
(%)
Results on the NYU dataset
Scene completion
University of Brasilia, Brazil [email protected]
CVSSP, University of Surrey, UK [email protected]
http://www.cic.unb.br/~teodecampos
André B. S. Guedes, Teo E. de Campos1 and Adrian Hilton2
Semantic Scene Completion Combining Colour and Depth: preliminary experiments
conv (12,1,1,1)
conv (12,1,1,1)
conv (128,1,1,1)
| 1cs.CV
|
What Can This Robot Do?
Learning Models from Appearance and Experiments
1
1
arXiv:1712.05497v1 [cs.AI] 15 Dec 2017
Ashwin Khadke and Manuela Veloso
Abstract— When presented with an unknown robot (subject)
how can an autonomous agent (learner) figure out what this new
robot can do? The subject’s appearance can provide cues to its
physical as well as cognitive capabilities. Seeing a humanoid
can make one wonder if it can kick balls, climb stairs or
recognize faces. What if the learner can request the subject
to perform these tasks? We present an approach to make the
learner build a model of the subject at a task based on the
latter’s appearance and refine it by experimentation. Apart
from the subject’s inherent capabilities, certain extrinsic factors
may affect its performance at a task. Based on the subject’s
appearance and prior knowledge about the task a learner can
identify a set of potential factors, a subset of which we assume
are controllable. Our approach picks values of controllable
factors to generate the most informative experiments to test
the subject at. Additionally, we present a metric to determine
if a factor should be incorporated in the model. We present
results of our approach on modeling a humanoid robot at the
task of kicking a ball. Firstly, we show that actively picking
values for controllable factors, even in noisy experiments, leads
to faster learning of the subject’s model for the task. Secondly,
starting from a minimal set of factors our metric identifies
the set of relevant factors to incorporate in the model. Lastly,
we show that the refined model better represents the subject’s
performance at the task.
I. I NTRODUCTION
Consider the case where a robot is presented with another
robot that the former knows little about. How would it
go about interpreting what this new robot can do? The
appearance could provide some cues. If the new robot has
an arm, one might wonder if it can lift or throw objects.
Appearances can convey information about not just the physical tasks that a robot can potentially perform but also it’s
cognitive capabilities. Seeing a robot equipped with a camera
and microphone could make one wonder if can recognize
faces, detect hand gestures or interpret voice commands.
How would one determine if the new robot can perform any
or all of the appearance-deduced tasks? Moreover, although
physical appearances can provide rich cues in interpreting
what tasks a new robot can perform they are not sufficient
to identify the caveats and strengths at performing these
tasks. The robots Roomba and Braava appear similar and
are both used for cleaning floors. But the Roomba can clean
carpets and not wet floors, which is exactly the opposite of
what Braava can do. Building a model of the new robot’s
capabilities at a task by passively observing it operate, can
be a solution. But what if it were possible to communicate
with it and request it to perform a task in certain scenarios?
1
School of Computer Science, Carnegie Mellon University, 5000 Forbes
Avenue, Pittsburgh, Pennsylvania, USA
In this paper we tackle the problem of an autonomous
agent, learner, making a model of another autonomous agent,
subject, in performing a particular task, when it is possible
for the former to experiment with the latter at that task. The
outcomes of these experiments can be non-deterministic. We
assume that the learner knows the appearance of the subject
and can infer what tasks the subject can potentially perform
based on that. We thus name the model as Appearance
Based Model (ABM). Apart from the subject’s inherent
capabilities, certain extrinsic factors may affect the subject’s
performance at the task. We assume that a subset of these
factors are controllable and the learner can choose values
for such factors in the experiments it conducts. We provide
an approach for the learner to pick values for controllable
factors in an experiment that generate the most informative
outcomes. We call the assignment of controllable factors
in an experiment as a query and this approach of actively
picking the assignment as active-querying. Based on prior
knowledge about the task and the appearance of the robot, we
assume that the learner knows the set of relevant factors that
may affect the subject’s performance. First we present our
approach to learn a model when the learner knows exactly
the set of factors that need to be incorporated in the model.
Further, starting with a minimal set of factors, we present a
method for the learner to identify relevant factors that need
to be added to refine the model.
We demonstrate the usefulness of our active querying approach at modeling a soccer playing humanoid robot perform
a BallKick task. We show that active-querying leads to faster
learning of the subject’s model than passively observing it
perform. To show the effectiveness of our model refinement
approach, we introduce a shortcoming in the subject that
is triggered by a particular factor. We initialize the model
excluding this factor and, show that our algorithm identifies
the missing factor and including it leads to a better model.
II. R ELATED W ORK
The idea of learning by experimentation isn’t new. [1]
describes LEX, a program that learns heuristics to solve indefinite integration problems by generating practice problems
to solve. ABMs can deal with tasks and experiments with
non-deterministic outcomes which LEX can’t. Additionally,
an ABM can capture the notion of how good a subject is at
performing a certain task.
[2] presents EXPO, a framework for refining the internal
model of a planning domain by experimentation. Imperfections in the model may create discrepancies between what a
planner expects and what happens when a plan is executed.
EXPO hypothesizes corrections to the model and generates
experiments to test these hypotheses. Contrary to ABMs,
EXPO does not handle non-deterministic experiments or
planning environments.
DIDO [3] is framework for learning models that capture
the effects of applying certain Motor Operations on objects
in the environment. While learning these effects, rather than
considering the amount of information gained by experimenting with a particular object, DIDO randomly picks objects
to test on. Our approach to learn ABMs explicitly reasons
about information gain while selecting values for controllable
factors in an experiment.
[4] is one of the earliest works to pose the problem of
actively learning a world model in a reinforcement learning
context. Reasoning about information gain in deciding what
actions to perform, in building a world model, is a common
approach in SLAM. [5], [6] and [7] are some works that
address this problem.
III. A PPEARANCE BASED M ODELS
Suppose the learner is presented with a legged robot
(subject) equipped with cameras. Of the possible tasks the
learner may interpret as the subject being capable of, such
as kicking a ball, walking over obstacles or climbing stairs,
say the learner chooses to model the subject at the kicking
task. Based on prior knowledge about the task, the learner
might incorporate factors such as size of the ball, turf on
which the subject is executing the task, position of the ball
with respect to the player before it kicks. Turf is a factor
most likely beyond the learner’s control. Position and ball
size seem likely to be controllable. In every experiment the
learner commands the subject to kick a ball in a particular
direction from a particular position on the given turf and
observes the direction the ball is kicked.
Here we introduce the notation for Appearance Based
Model (ABM) for a task T .
A. Definition
Model MT = (SIT,OUT,FF,DK)
• OUT - Set of random variable(s) denoting the out1
come(s) of task T . Outcome is an instantiation of OUT.
• SIT= CNTX ∪ COMM.
– COMM - Variables denoting the commands given
to the subject for task T .
– CNTX - Set of extrinsic factors that affect OUT.
– Situation, Command and Context are instantiations
of SIT, COMM and CNTX respectively.
• Frequency Function (F F {X, Y }) returns the set of joint
probability distributions of observing variables in X
given variables in Y . X ⊂ OUT, Y ⊂ SIT ∪ OUT.
• Domain Knowledge (DK)
– Sets SIT and OUT for the task T .
– ATT o ∀o OUT are the sets of additional variables
that may affect the outcome variable o, but unac2
counted in the model .
– Dependencies amongst variables in SIT and OUT
MBallKick = (SIT, OUT, FF, DK)
SIT = {KDc, Position, Turf , BallSize}, OUT = {KDo}
CNTX = {Position, Turf , BallSize}, COMM = {KDc},
ATT KDo = {BallColor}, FF{OUT, S} = P(KDo∣S)∀S ⊂ SIT
DK KDo ⫪ KDc, Position, BallSize, Turf
The ABM for the BallKick task is as shown above. Here,
Position represents the location of the ball with respect
to the subject. KDc and KDo denote the commanded and
observed kick direction respectively. KDo is None if the
subject attempts but fails to kick the ball or doesn’t attempt
to kick. The former may be because of it’s inability to kick
in certain positions. The latter may mean it doesn’t detect
the ball or doesn’t know how to perform a kick.
Position {Left, Middle, Right}
Turf {Grass, Synthetic, Sand}, BallSize {Small, Large}
KDc {Left, Mid, Right}, KDo {Left, Mid, Right, None}
Here the set OUT is singleton. In general this may not
be the case. Moreover, a variable in OUT may depend on
another variable in the set. The learner requesting the subject
to perform the task in certain Context is equivalent to making
interventional queries to the model. We propose using a
Causal Bayesian Network (CBN) to represent FF. CBNs
capture the causal dependence of Outcomes on Commands
and Context and, simplify learning when queries are interventional.
IV. L EARNING ABM S
Here we describe our approach to learn an ABM for a particular task. The sets SIT and OUT are assumed part of DK.
Our approach learns FF by actively selecting Situations and
observing the subject perform the task in these Situations.
A. Using a CBN to represent the Frequency Function
•
•
•
•
2
1
An instantiation of a set of variables is a dictionary mapping variables
in the set to values in their domain. Context ← ∪xCNTX {x ∶ xval } where
xval Dom(x)
FF ≡ (G, θ) , where G ≡ (χ, E) is the underlying
directed acyclic graph and parameters θ quantify the
conditional probability tables associated with G.
χ = SIT ∪ OUT. Variables in OUT are dependent on
some or all variables in SIT and may also depend
on other variables in OUT. Specifically, ∀(xi , xj ) E,
xi χ and xj OUT. The set E is specified in DK.
Q ⊂ SIT is the set of variables the learner can control.
A Query is an instantiation of Q.
An Experiment is an instance of the subject performing
the task under a Query. An Observation is an instantiation of χ at the end of an experiment
3
FF captures distributions of o marginalized over the variables in ATT o .
In experiments with the subject these variables are assigned values randomly.
3
Overloaded notation. FF{X,Y} denotes P(X∣Y) but just FF denotes the
Causal Bayesian Network associated with the Frequency Function.
•
COMM⊂ Q. Q∩CNTX need not be empty. These are
the set of controllable factors.
We assume that all dependencies among variables, included in the model, are known and are part of DK associated
with the task. Learning FF is then equivalent to learning the
parameters θ for the underlying graph G.
with respect to p can be quantified as (1)
Errorp (θ ) = ∫θ KLχ (θ∣∣θ )p(θ)dθ
′
KLχ (θ∣∣θ ) = ∑
(1)
KL(Pθ (Xi ∣u)∣∣Pθ′ (Xi ∣u)
∑
′
(2)
Xi χ uD(Ui )
θ̃ = arg min
Errorp (θ )
′
′
(3)
Risk(p) = Errorp (θ̃)
(4)
θ
B. Active Learning for CBN
We assume that dependencies known in DK are sufficient
for the learner to construct a network structure G. Additionally, the learner has a prior p(θ) over the parameters of G.
We use the active learning algorithm presented in [8] to build
′
our posterior p . We describe it here in short.
We assume all random variables in FF are Multinomial
and θ parametrize the conditional probability tables. Specifically, θXi ∣ui = (θxi 1∣ui , ⋯, θxi ri ∣ui ) denotes the vector of
probabilities for different values of Xi when the parents of
Xi χ in G have value ui . We use Bayesian estimation to
learn parameters θ i.e. we learn a distribution p(θ) over possible parameter values. Dirichlet distribution, parameterized
+
by hyperparameters αj R and α⋆ = ∑j αj , is the conjugate
prior for Multinomial. We make the assumption of parameter
independence as in [9], i.e. parameters associated with each
node are independent as well as parameters of a node for
different states of parent variables are also independent.
With parameter independence and all variables in FF being
Multinomial, we have a Dirichlet distribution for every θXi ∣ui
where αxi j∣ui is the hyperparameter for θxi j∣ui .
In each Experiment the learner choses a Query and
requests the subject to perform the task. Based on the
Observation the learner updates the prior p(θ) to obtain the
′
posterior p (θ) and proceeds to the next Experiment by setting
′
p to p. The learning algorithm must address two issues, how
to update p(θ) based on Query and Observation and, what
Query to pick next based on current estimate of p(θ).
Since the nodes in Q are not sampled but assigned
(Query is interventional), it doesn’t make sense to update
parameters corresponding to these nodes. As the underlying
representation is a CBN, updating parameters of variables
that are ancestors of variables in Q also doesn’t make sense.
We call variables that are neither in Q nor are an ancestor to
a variable in Q as Updateable. Given a prior p(θ), Query and
the obtained Observation we do standard Bayesian update on
the parameter distribution of only the Updateable variables.
To generate a Query from the current estimate of p(θ)
we need a metric to evaluate how good the current estimate
is. We can then quantify the improvement in the estimate
brought about by different queries and pick the one which
⋆
leads to the most improvement. Let θ be the true param′
eters of the model and θ be the point estimate evaluated
⋆
′
using p(θ). KLχ (θ ∣∣θ ) ≡ KL(Pθ⋆ (χ)∣∣Pθ′ (χ)), where Pθ′ (χ)
represents the joint distribution of variables in χ when
′
parameterized with θ , can be used as measure of error for
⋆
our point estimate. We don’t know θ , but we do have p(θ),
which is our belief of what the optimal θ is given the prior
′
and Observations. Thus the Error in our point estimate θ
′
θ̃, the point estimate that minimizes the Error with respect
to our current estimate p, turns out to be ∫ p(θ)dθ i.e. Ep (θ).
We use Risk of a distribution, defined in equation (4), as
the measure of quality for the estimate p(θ). Lower the Risk,
better the estimate. For a CBN with Multinomial variables,
using the simplification in (2), the Risk of a distribution p
can be expressed as in (5).
Risk(p) = ∑
∑
i uD(Ui )
Pθ̃ (u)δ(αxi1 ∣u , ⋯, αxiri ∣u )
(5)
ri
δ(αxi1 ∣u , ⋯, αxiri ∣u ) = ∑ (HαXi ∣u +
j=1
αxi j∣u
αxi ⋆∣u
(Ψ(αxi j∣u + 1) − Ψ(αxi ⋆∣u + 1)))
(6)
ri
αxi j∣u
αxi j∣u
HαXi ∣u = − ∑ α
ln( α
)
xi ⋆∣u
j
(7)
xi ⋆∣u
In equation (6), Ψ(α) = Γ (α)/Γ(α), where Γ denotes the
′
Gamma function and Γ its derivative.
We would want to see Observations that reduce the Risk
associated with our estimate p. For a particular Query q we
take expectation over possible Observations to evaluate the
Expected Posterior Risk (EPR) for that Query. Lower the
EPR better the Query.
′
EPR(p, q) = EΘ∼p(θ) (EX∼PΘ (χ∣Q=q) (Risk(p )))
′
(8)
′
In (8) p represents the posterior obtained after updating prior
p with sample drawn from PΘ (χ∣Q = q). For a CBN with
only one child node X taking values xi ∀i {1, ⋯, r} and all
it’s parent nodes U being in Q, EPRX is as shown in (9)
EPRX (p, q) = ∑ Pθ̃ (xj ∣q) ∑ Pθ̃′ (u)δ(αx1 ∣u , ⋯, αxr ∣u ) (9)
′
′
u
j
EPR(p, q) = ∑
∑
Pθ̃ (u∣q)EPRXi (p, u)
(10)
i uD(Ui )
In (9), θ̃ and αxj ∣u denote the parameters for posterior p (θ)
corresponding to θ̃ and αxj ∣u in prior p(θ). For a general
CBN the EPR decomposes as in (10). Moreover, it is easier
to compute ”reduction in Risk” ∆(χ∣q) = Risk(p) - EPR(p,q)
than evaluating the EPR for a Query. Subtracting (10) from
(5) and approximating Pθ̃′ (u) in (9) by Pθ̃ (u) we get
′
′
′
∆(χ∣q) ≈ ∑
∑
Pθ̃ (u∣q)∆(Xi ∣u)
(11)
i uD(Ui )
∆(Xi ∣u) = Pθ̃ (u)(HαXi ∣u − ∑j Pθ̃ (xj ∣u)Hα′X ∣u ) (12)
i
Algorithm (1) picks the Query that gives the highest ∆(χ∣q).
Algorithm 1 LearnFF(DK, maxIter)
From DK
Construct G ≡ (χ, E)
Initialize Q, α
p(θ) ← Dirichlet(α), iter ← 0
while iter < maxIter do
Query ← BestQuery(p(θ), Q)
Observation, Context ← Experiment(Query)
p(θ) ← Update(p(θ), Query, Context, Observation)
end while
F F ← (G, p(θ))
Algorithm 2 BestQuery(p(θ), Q)
maxDelQ ← −∞
θ̃ = ∫ θp(θ)dθ
for q Dom(Q) do
∆(χ∣q) ← ∑ ∑ Pθ̃ (u∣q)∆(Xi ∣u)
Imax = min(En(P̂o∣CObs ), En(P̂Aoj ))
(15)
In equation (14), D(Aoj ) denotes domain of Aoj and En(P)
returns the entropy of distribution P. P̂I∣j (i) denotes the
fraction of times variable I had value i when J⊂Parents(I)
took values as specified in j, in the past Experiments. Since
variables in ∪i ATT i are sampled randomly and, independently of the Context, P̂Aoj (a) ≈ P̂Aoj ∣CObs (a). Thus I in
equation (14) is approximately the Mutual Information of ’o’
and Aoj given CObs. This makes R the Coefficient of Mutual
Information. If Roj,CObs is greater than a threshold, the learner
incorporates Aoj in CNTX, updates G by adding Aoj as a
parent to variable o and creates conditional probability tables
of o for the updated graph G. We identify Attributes solely
based on Observations in the Experiments.
Algorithm 3 IdenDep(P̂ , AT To , CNTObs, RTh, MinSamp)
AttIncluded ← [ ]
for Aoj AT To do
for c CNTObs do
V ← Valid(∪aD(Aoj ) {P̂o∣a,c }, MinSamp)
if Size(V ) > 2 then
valid
P̂Aoj ∣c ← NormalizeValid(P̂Aoj ∣c , V )
i uD(Ui )
if ∆(χ∣q) > maxDelQ then
maxDelQ ← ∆(χ∣q)
qbest ← q
end if
end for
return qbest
valid
Po∣c ← ∑ P̂o∣a,c P̂Aoj ∣c (a)
aV
valid
I ← En(Po∣c ) − ∑ En(P̂o∣a,c )P̂Aoj ∣c (a)
aV
C. Model Refinement
Every subject may have a slightly different set of Factors
that affect their performance at a task. Consider the ball kick
task. If the subject only detects balls of a certain color, a
variable BallColor should be added to CNTX. The learner
may not always know what exactly to include in the contexts
(CNTX). We present an approach to identify variables to
include in the set CNTX.
We assume ∀o OUT the learner knows a set of Attributes
(ATT o ). These are variables that may influence the distribution of o, but are not included in the model. Moreover,
ATT i ∩ ATT j = ∅ for i ≠ j. Including all of them would
result in a model that is unnecessarily large and difficult
to learn. Starting with a minimal set of variables in CNTX
and observing Outcomes in the Experiments, we present an
approach to identify variables that need to be added to the
model. In every Experiment, values for variables in ∪i AT Ti
are randomly chosen, independent of the choice of Context.
Let CNTObs be the set of Contexts observed till the current
Experiment. At the end of every Experiment the learner
estimates P (o∣Aoj ,CObs) ∀o OU T and Aoj AT To , where
CObs CNTObs. The learner includes Aoj in CNTX if the
distributions of o for different values of Aoj are drastically
different in at least one observed Context, CObs. We define
a metric Roj,CObs to evaluate the variation in distributions
of o for different values of Aoj in a particular CObs.
Roj,CObs =
I = En(P̂o∣CObs ) −
∑
aD(Aoj )
I
Imax
En(P̂o∣a,CObs )P̂Aoj (a)
Imax ← min(En(Po∣c ), En(P̂Aoj ∣c ))
R← II
max
if R > RTh then
AttIncluded.append(Aoj )
break
end if
end if
end for
end for
return AttIncluded
valid
Algorithm (3) presents the above idea. Input set of distributions P̂ includes ∪Aoj ATT o (∪aD(Aoj ),cCNTObs {P̂o∣a,c }) and
∪Aoj ATT o ,cCNTObs {P̂Aoj ∣c }. We call a distribution valid if the
number of samples used to estimate it is at least MinSamp.
The function ’Valid’ returns a set V , such that ∀aV P̂o∣a,c is
valid. The function ’NormalizeValid’ normalizes the distribution of Aoj given Context c, over subset of values identified
by V . Algorithm (4) summarizes the model refinement approach. ’Modify’ adds nodes corresponding to the identified
Attributes to the graph G, makes connections such that the
added Attributes are parents of appropriate outcome nodes,
removes conditional probability tables corresponding to the
modified outcome nodes and creates conditional probability
tables corresponding to the new connections.
(13)
V. E XPERIMENTS
(14)
We used the humanoid robot, NAO, as subject for our
experiments. We show results for modeling the robot at the
Algorithm 4 FFMCorrect(DK, maxIter, RTh, MinSamp)
From DK
Construct G ≡ (χ, E)
Initialize Q, α, AT To ∀ o OUT
Initialize P̂ , p(θ) ← Dirichlet(α)
iter ← 0, CNTObs ← ∅
while iter < maxIter do
AttAdded ← {}
Query ← BestQuery(p(θ), Q)
Attr ← SampleUniform(∪oOU T AT To )
Observation, Context ← Experiment(Query, Attr)
CNTObs ← CNTObs ∪ Context
p(θ) ← Update(p(θ), Query, Context, Observation)
P̂ ← Update(P̂ , Query, Context, Observation)
for o OU T do
AttAdded[o] ← IdenDep(P̂ , ATT o , CNTObs, RTh,
MinSamp)
end for
p(θ), G, Q ← Modify(p(θ), G, Q, AttAdded)
end while
F F ← (G, p(θ))
BallKick task discussed before. Tryouts are a common way of
selecting players for a soccer team. Tryouts for a robot soccer
team are an apt scenario to apply our active experimentation
approach and hence the choice of this particular task. Our
approach builds a model, but without knowing the true model
of the subject we cannot say how good the learned model
is. So we programmed the robot to behave according to
a predefined model. Specifically, we defined a Frequency
Function for the predefined model to reflect the behavior
of a right-footed player at the BallKick task. To determine
how close a learned model is to the predefined model, we
used the KL Divergence of the joint distribution of variables
in OUT, KLOUT (θpredefined ∣∣θlearned ) as defined in equation (2),
as an error measure. We show that active experimentation
leads to faster convergence to the true model than passively
observing the subject perform the task. Moreover, we show
that our model refinement approach is able to identify the
correct set of missing factors and that the refined model better
represents the true model.
KDc
P
KDo
Fig. 1.
CBN of predefined subject model at BallKick task. P stands
for Position, KDc and KDo for kick direction commanded and observed
respectively.
P
KDc
BS
KDo
Fig. 2. CBN of the predefined subject model at the BallKick task. BS
stands for BallSize and the rest are as defined in Fig (1).
perception and actuation, sometimes the robot ended up
kicking in directions it wasn’t commanded to. Fig. (3) shows
that despite the noisy experiments our approach converged
faster to the predefined model.
Fig. 3. Trend in KLOUT (θpredefined ∣∣θactive ) and KLOUT (θpredefined ∣∣θrandom ).
A. Passive vs Active
Passively observing the subject perform the task, where
you don’t control the scenarios in which you witness the
subject, is equivalent to randomly picking Contexts to build a
model. We learned a model using our approach and another
model by randomly picking Contexts. Fig. (1) depicts the
underlying CBN for this model. We experimented with a
single ball on a synthetic turf. Thus the variables Turf and
BallSize are dropped from the model. A subtle point to
note, the Experiments are noisy. The robot was programmed
to behave according to the predefined model i.e. when the
learner chose a particular Context, the value of KDo sampled
from the Frequency Function of this predefined model was
given as command to the robot. However, owing to noisy
Fig. 4. Trend in KLOUT (θpredefined ∣∣θlearned ). Red point depicts the Experiment in which the learner added attribute BallSize to the ABM. Learning
curve for a model initialized with the correct set of factors is in green.
P
KDc
BS
BC
KDo
Fig. 6. Initial CBN of the learner’s model of the subject at the BallKick
task. BC stands for BallColor and the rest are as defined in Fig. (2). The
dotted nodes belong to ATT KDo and are initially not incorporated in the
model. The learner randomly picks values for variables in ATT KDo
Fig. 5. Trend in KLOUT (θpredefined ∣∣θlearned ). Red points depict Experiments
at which the learner incorporated an attribute in the ABM. Green curve
shows learning trend for a model initialized with the correct set of factors.
This plot shows results from a simulated subject.
B. Model Refinement
We programmed the robot to detect balls of any color but
only a specific size. If presented with a ball of a different
size, the subject wouldn’t detect and thus won’t kick i.e.
KDo would be None. Fig. (2) depicts the underlying CBN
for the predefined model. The learner starts with a model as
depicted in Fig. (6). Initially, ball size and color are part
of the Attribute set i.e. the learner choses ball size and
color randomly. Using Observations from the experiments,
the learner correctly identified the missing variable to be
BallSize and learned the Frequency Function for the refined
model. The blue curve in Fig. (4) shows the trend in KL
Divergence while learning this model. To show that incorporating relevant Attributes yields a model, better representative
of the subject, we learned another model including BallSize
from the start. The trend in KL Divergence for this model
is depicted by the green curve in Fig. (4). As is seen, the
model that includes the variable BallSize gets closer to the
predefined model. Moreover, post refinement the learning
trends for both models are almost the same. We tested
in simulation, how our approach fares as the number of
Attributes increase. The robot could only detect balls of a
certain size and KDo was sampled randomly when playing
on Sand. Experiments had noise (10% of the time KDo was
sampled randomly). As seen in Fig (5), including all the
relevant Attributes gives a better model. More the number of
missing factors, more the number of Experiments needed to
identify them all. Moreover, it becomes progressively harder.
As an Attribute gets included, possible Contexts increase.
The active learning algorithm avoids repeating Contexts and
thus estimating valid P̂o∣a,c requires more Experiments.
VI. C ONCLUSION AND F UTURE W ORK
As robots become increasingly ubiquitous the problem of
building models of unknown robots becomes all the more
relevant. An autonomous repairman may need to test a
subject to quickly identify defects. A moderator or a coach
robot may need to identify strengths of a particular team
member in deciding what tasks to assign it to. To the best of
our knowledge, the problem of building models of unknown
robots, based on appearance and active experimentation is
novel and yet unexplored. We proposed an approach to solve
this problem and presented results of experiments with a
NAO robot. Even when experiments are noisy, the learner
is able to identify relevant factors that dictate the subject’s
behavior and build a model of the same.
How to utilize these learned models in planning and
decision making is a very relevant future direction that we
intend to explore. Moreover, we assumed that experimenting
with the subject in any Context has the same cost. Depending
upon the task this may be quite far from reality. Learning
models in such scenarios is another interesting problem.
R EFERENCES
[1] T. M. Mitchell, P. E. Utgoff, and R. Banerji, “Readings in
knowledge acquisition and learning,” B. G. Buchanan and D. C.
Wilkins, Eds. San Francisco, CA, USA: Morgan Kaufmann
Publishers Inc., 1993, ch. Learning by Experimentation: Acquiring and
Refining Problem-solving Heuristics, pp. 504–517. [Online]. Available:
http://dl.acm.org/citation.cfm?id=170641.170690
[2] Y. Gil, “Acquiring domain knowledge for planning by experimentation,”
CARNEGIE-MELLON UNIV PITTSBURGH PA DEPT OF COMPUTER SCIENCE, Tech. Rep., 1992.
[3] P. D. Scott and S. Markovitch, “Experience selection and problem
choice in an exploratory learning system,” Machine Learning,
vol. 12, no. 1, pp. 49–67, Aug 1993. [Online]. Available:
https://doi.org/10.1007/BF00993060
[4] J. Schmidhuber, “Curious model-building control systems,” in Neural
Networks, 1991. 1991 IEEE International Joint Conference on. IEEE,
1991, pp. 1458–1463.
[5] H. J. S. Feder, J. J. Leonard, and C. M. Smith, “Adaptive mobile
robot navigation and mapping,” The International Journal of Robotics
Research, vol. 18, no. 7, pp. 650–668, 1999. [Online]. Available:
http://dx.doi.org/10.1177/02783649922066484
[6] F. Bourgault, A. A. Makarenko, S. B. Williams, B. Grocholsky, and
H. F. Durrant-Whyte, “Information based adaptive robotic exploration,”
in in Proceedings IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS, 2002, pp. 540–545.
[7] C. Stachniss and W. Burgard, “Exploring unknown environments
with mobile robots using coverage maps,” in Proceedings of
the 18th International Joint Conference on Artificial Intelligence,
ser. IJCAI’03. San Francisco, CA, USA: Morgan Kaufmann
Publishers Inc., 2003, pp. 1127–1132. [Online]. Available: http:
//dl.acm.org/citation.cfm?id=1630659.1630821
[8] S. Tong and D. Koller, “Active learning for parameter estimation
in bayesian networks,” in Proceedings of the 13th International
Conference on Neural Information Processing Systems, ser. NIPS’00.
Cambridge, MA, USA: MIT Press, 2000, pp. 626–632. [Online].
Available: http://dl.acm.org/citation.cfm?id=3008751.3008842
[9] D. Heckerman, D. Geiger, and D. M. Chickering, “Learning bayesian
networks: The combination of knowledge and statistical data,” Machine
Learning, vol. 20, no. 3, pp. 197–243, Sep 1995. [Online]. Available:
https://doi.org/10.1023/A:1022623210503
| 2cs.AI
|
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
arXiv:1602.04842v3 [math.GR] 10 Nov 2016
DANIEL FRANZ
Abstract. Normal residual finiteness growth measures how well a finitely generated residually finite group is approximated by its finite quotients. We show that any finitely generated
linear group Γ ≤ GLd (K) has normal residual finiteness growth asymptotically bounded
2
above by (n log n)d −1 ; notably this bound depends only on the degree of linearity of Γ. If
char K = 0 or K is a purely transcendental extension of a finite field, then this bound can
2
be improved to nd −1 . We also give lower bounds on the normal residual finiteness growth of
Γ in the case that Γ is a finitely generated subgroup of a Chevalley group G of rank at least
2. These lower bounds agree with the computed upper bounds, providing exact asymptotics
on the normal residual finiteness growth. In particular, finite index subgroups of G(Z) and
G(Fp [t]) have normal residual finiteness growth ndim(G) . We also compute the non-normal
residual finiteness growth in the above cases; for the lower bounds the exponent dim(G) is
replaced by the minimal codimension of a maximal parabolic subgroup of G.
1. Introduction
Let Γ be a finitely generated residually finite group with finite generating set X. If γ ∈ Γ,
let ||γ||X be the word length of γ with respect to X. If γ ∈ Γ is nontrivial, we say a finite
E
(n) to be the
quotient Q of Γ detects γ if the image of γ in Q is nontrivial. Define FΓ,X
smallest natural number N such that for all γ ∈ Γ with ||γ||X ≤ n, γ is detected by a
quotient of size at most N.
E
We call the function FΓ,X
the normal residual finiteness growth function of Γ. This function
was first studied by Bou-Rabee in [3], and its asymptotics have been studied for virtually
nilpotent linear groups [7], arithmetic groups [6], linear groups [8], and free groups [4] [13],
≤
with the best current estimate for free groups given in [22]. A related function is FΓ,X
, the
non-normal residual finiteness growth function of Γ, defined as the smallest natural number
N such that for all γ ∈ Γ with ||γ||X ≤ n, there exists H ≤ G with γ 6∈ H and [G : H] ≤ N.
This function has also been studied for certain classes of groups, in particular for virtually
special groups in [5] and for free groups in [7] [9] [15]. Our goal in this paper is to obtain
≤
E
and FΓ,X
when Γ is a linear group.
better estimates of the functions FΓ,X
While these functions depend on the choice X of generating set, their asymptotic growths,
which we call the normal residual finiteness growth of Γ and non-normal residual finiteness
growth of Γ, respectively, are independent of the choice of generating set ([3], Lemma 1.1).
We thus drop the reference to X for the remainder of the introduction. We compare the
asymptotic growth of functions by writing f g if for some C, f (n) ≤ Cg(Cn) for all n.
It was shown in [8] that if Γ is a finitely generated linear group over an infinite field, then
E
FΓ (n) nk for some k depending on the field and the degree of linearity. A natural question
is whether the dependence on the field of coefficients is necessary. Our first result is that in
2010 Mathematics Subject Classification. 20F69 (primary), 20E26, 20G15 (secondary).
Key words and phrases. Residual finiteness growth, residual finiteness, linear groups.
1
2
DANIEL FRANZ
fact there is a uniform bound on the residual finiteness growth of finitely generated linear
groups with a fixed degree of linearity.
Theorem 1.1. Let Γ ≤ GLd (K) be a finitely generated linear group with d ≥ 2.
2
(i) FΓE (n) (n log n)d −1 and FΓ≤ (n) (n log n)d−1 .
(ii) If char K = 0 or K is a purely transcendental extension of a finite field, then
2
FΓE (n) nd −1 and FΓ≤ (n) nd−1 .
One potential application of normal residual finiteness growth is in showing a group is
nonlinear. For a finitely generated group Γ, one can show FΓ (n) is super-polynomial to
conclude Γ is nonlinear. If Γ is infinitely generated, the uniform bound of Theorem 1.1
provides another method for establishing nonlinearity. In particular, this result has potential
applications in the study of profinite groups.
Corollary 1.2. Let G be a group such that for each k ∈ N, G has a finitely generated
subgroup H with FHE (n) nk . Then G is nonlinear.
The proof of Theorem 1.1 easily generalizes to certain algebraic groups, yielding the following more specific result. By a Chevalley group we mean a split simple group scheme defined
over Z, not necessarily simply connected, with irreducible root system Φ. For such a group
G, let dim(G) be its dimension and a(G) be the minimal codimension of a proper parabolic
subgroup; these values are given in Table 1 and justified in Lemma 2.4.
Φ
dim(G) a(G)
Al , l ≥ 2 l2 + 2l
l
Bl , l ≥ 2 2l2 + l 2l − 1
Cl , l ≥ 3 2l2 + l 2l − 1
Dl , l ≥ 4 2l2 − l 2l − 2
G2
14
5
F4
52
15
E6
78
16
E7
133
27
E8
248
57
Table 1.
Theorem 1.3. Let G be an affine algebraic group scheme defined over Z, let K be a field,
and let Γ ≤ G(K) be finitely generated.
(i) FΓE (n) (n log n)dim(G) and, if G is a Chevalley group, FΓ≤ (n) (n log n)a(G) .
(ii) If char K = 0 or K is a purely transcendental extension of a finite field, then
FΓE (n) ndim(G) and, if G is a Chevalley group, FΓ≤ (n) na(G) .
When G is a Chevalley group of rank at least 2, we can also determine lower bounds for
residual finiteness growth, which, coupled with Theorem 1.3, yield precise asymptotics for
residual finiteness growth. In [6], Bou-Rabee and Kaletha determined the lower bound for
normal residual finiteness growth of arithmetic groups of G. We generalize this statement to
non-normal residual finiteness growth and the characteristic p setting.
Theorem 1.4. If O = Z or Fp [t] and Γ ≤ G(O) has finite index, where G is a Chevalley
group of rank at least 2, then FΓE (n) ndim(G) and FΓ≤ (n) na(G)
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
3
Normal and non-normal residual finiteness growth can only decrease when passing to a
subgroup, so Theorem 1.4 also gives lower bounds for all finitely generated subgroups of
G(K), G a Chevalley group of rank at least 2 and K a field. Combining this lower bound
with the upper bound from Theorem 1.3 then gives exact asymptotics for normal and nonnormal residual finiteness growth.
Corollary 1.5. Let G be a Chevalley group of rank at least 2, let K be a field of characteristic
0 or a purely transcendental extension of a finite field, and let Γ ≤ G(K) be finitely generated.
Put O = Z if char K = 0 and O = Fp [t] if char K = p > 0.
If Γ ∩ G(O) ≤ G(O) has finite index, then FΓE (n) ≈ ndim(G) and FΓ≤ (n) ≈ na(G) .
The main tool used to provide the uniform upper bound in characteristic 0 is a higher
dimensional version of the Chebotarev density theorem formulated by Serre [20]. In positive
characteristic the needed analogue is not available, so we use an effective form of the usual
Chebotarev density theorem [23]. Specifically, we need a statement about natural density,
not Dirichlet density. Over certain fields, this causes the bounds to be powers of n log n
instead of n. For the lower bounds we use properties of Chevalley groups and associated
graded Lie algebras, as well as the congruence subgroup property.
The paper is organized as follows. After collecting some lemmas in Section 2, we prove
Theorem 1.3 in Section 3. This is done in stages, beginning with the case of a purely
transcendental extension of a finite field. We then collect results on the Chebotarev Density
Theorem which are used to prove Theorem 1.1 and the remaining parts of Theorem 1.3
together.
In Section 4, we consider graded Lie algebras arising from Chevalley groups and relate
them to the problem of finding lower bounds for normal and non-normal residual finiteness
growth. The characteristic 0 part of Theorem 1.4 is then proved in Section 5, and in Section
6 the proof is completed in the positive characteristic setting.
Acknowledgments. I would like to thank my advisor Mikhail Ershov for his great advice
and support while working on this topic. I would also like to thank Martin Kassabov for
discussions which greatly simplified some of the proofs of lower bounds. My thanks also
to Andrei Rapinchuk for suggesting a reference that was key in proving the upper bound
statement and to Khalid Bou-Rabee for providing helpful comments on an early draft of this
paper.
2. Preliminaries
Let Γ be a finitely generated group, generated by a finite symmetric set X. If γ ∈ Γ is
nontrivial, define
DΓE (γ) = min{[Γ : N] : N E Γ, γ 6∈ N},
DΓ≤ (γ) = min{[Γ : H] : H ≤ Γ, γ 6∈ H}.
Then the normal and non-normal residual finiteness growth of Γ are determined by the
functions
E
FΓ,X
(n) = max{DΓE (γ) : ||γ||X ≤ n, γ 6= 1},
≤
FΓ,X
(n) = max{DΓ≤ (γ) : ||γ||X ≤ n, γ 6= 1}.
4
DANIEL FRANZ
We will measure asymptotic growth by writing f g if there exists C such that f (n) ≤
Cg(Cn) for all n. If f g and g f we will write f ≈ g.
≤
E
The asymptotic growths of FΓ,X
(n) and FΓ,X
(n) are independent of the generating set
(Lemma 1.1, [3]), so the reference to X will be dropped. Another consequence of Lemma 1.1
in [3] that will be used tacitly for the remainder of the paper is that FHE (n) FΓE (n) and
FH≤ (n) FΓ≤ (n) if H ≤ Γ.
We will need the following result when proving lower bounds; it is contained in Lemma 2.4
in [6]. In particular it will allow us to pass from a Chevalley group to its simply connected
cover. We include the proof for completeness.
Lemma 2.1. Assume Γ and ∆ are finitely generated, residually finite groups. If f : Γ → ∆
is a homomorphism with finite kernel, then FΓE (n) F∆E (n) and FΓ≤ (n) F∆≤ (n).
Proof. Since f (Γ) ≤ ∆, FfE(Γ) (n) F∆E (n). Hence it suffices to show FΓE (n) FfE(Γ) (n).
Assume Γ = hXi, |X| ≤ ∞. Then f (Γ) is generated by f (X) = {f (x) : x ∈ X}.
Since the kernel of f is finite, if n is sufficiently large then f (γ) 6= 1 for all γ ∈ Γ with
||γ||X = n. Let n be large enough to ensure this and let γ ∈ Γ with ||γ||X = n. We have
||f (γ)||f (X) ≤ n and f (γ) 6= 1, so there exists a normal subgroup N Ef (Γ) such that f (γ) 6∈ N
and [f (Γ) : N] ≤ FfE(Γ),f (X) (n). Hence N ′ = N ker(f ) E Γ satisfies
γ 6∈ N ′ and [Γ : N ′ ] ≤ FfE(Γ),f (X) (n),
E
so FΓ,X
(n) ≤ FfE(Γ),f (X) (n) and thus FΓE (n) FfE(Γ) (n).
The same argument with N replaced by an arbitrary subgroup H shows that FΓ≤ (n)
F∆≤ (n).
If f ∈ Fp [t][x1 , · · · , xs ], we treat f as a polynomial with coefficients in Fp [t] and consider
the degree of f to be the total degree of the x′i s. Define the height of f to be ht(f ) =
max{deg g(t) : g(t) is a coefficient of f }.
Lemma 2.2. Let f ∈ Fp [t][x1 , · · · , xs ] be nonzero with deg f ≤ 2m . Then there exist
g1 (t), · · · , gs (t) ∈ Fp [t] with deg gi (t) ≤ m for each i such that f (g1 (t), · · · , gs (t)) =
6 0.
Proof. We induct on s. Suppose s = 1. Since Fp [t] is an integral domain and deg f ≤ 2m ,
f (x) has at most 2m roots. There are at least 2m+1 elements of Fp [t] with degree at most m,
so f (g(t)) 6= 0 for some g(t) with deg g(t) ≤ m.
Now assume the lemma is true for s = n − 1 and suppose s = n. When considered as a
polynomial over xs with coefficients in Fp [t][x1 , · · · , xs−1 ], f has at most 2m roots, so there
is some gs (t) ∈ Fp [t] with deg g(t) ≤ m such that
f (x1 , · · · , xs−1 , g(t)) 6= 0.
Applying the inductive hypothesis finishes the proof.
We will also need the following size estimates.
Lemma 2.3. Let G be an affine algebraic group scheme defined over Z, q a prime power.
There exists a constant C independent of q such that |G(Fq )| ≤ Cq dim(G) .
Proof. Let A be the Hopf algebra representing G, finitely generated over a field k, so that
G(Fq ) = Homk (A, Fq ). By Noether normalization, A is a finitely generated module over a
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
5
polynomial ring k[x1 , · · · , xd ], where d = dim(G). If A is generated as a module by y1 , · · · , ym,
then each yi is integral over k[x1 , · · · , xd ], so for each 1 ≤ i ≤ m we can find a polynomial
fi (x1 , · · · , xd , Y ) ∈ k[x1 , · · · , xd ][Y ]
such that fi (x1 , · · · , xd , yi ) = 0. Let c = max deg fi . An element ϕ ∈ Homk (A, Fq ) is
1≤i≤m
determined by the images of the xi and yj . Given choices of ϕ(xi ), which can be made
arbitrarily, for each 1 ≤ j ≤ m there are at most c choices of ϕ(yj ) that will satisfy
fj (ϕ(x1 ), · · · , ϕ(xd ), ϕ(yj )) = 0. Thus |Homk (A, Fq )| ≤ cm q d , so C = cm is the required
constant.
Lemma 2.4. Let G be a Chevalley group with an embedding into SLd , q be a prime power,
1 dim(G)
and H ≤ G(Fq ) be a proper subgroup of minimal index. Then |G(Fq )/Z(G(Fq ))| ≥ 2d
q
1 a(G)
a(G)
≤ [G(Fq ) : H] ≤ 2q
.
and 2 q
Proof. The size bound of |G(Fq )/Z(Fq )| follows from Theorem 25, §9, in [21]. The index of
the largest maximal subgroup of G(Fq ) can be found in [14] (Theorem 5.2.2) for the classical
groups, and in [24], [25] for the exceptional groups.
3. Upper Bounds
3.1. Purely transcendental extensions of finite fields. We begin this section by proving
Theorems 1.1 and 1.3 in the case that K is a purely transcendental extension of a finite field.
Lemma 3.1. Let f (t) ∈ Fq [t] be nonzero with degree at most n. Then there exists a finite
field F with 2n < |F| ≤ 2nq and a homomorphism φ : Fq [t] → F such that φ(f (t)) 6= 0.
Proof. It is a well known generalization of a result of Gauss [19] that the number of irreducible
polynomials in Fq [t] of degree k is
Iq (k) =
1X
µ(d)q k/d,
k
d|k
where µ is the Mobius inversion function. It is easy to check that kIq (k) ≥ 21 q k for k ≥ 2.
Given f (t) ∈ Fq [t], we wish to find an irreducible polynomial of appropriate degree that
does not divide f (t). To that end, note that if f (t) is divisible by all irreducible polynomials
of degree k, then
1
deg f (t) ≥ kIq (k) ≥ q k .
2
So now let f (t) ∈ Fq [t] have degree at most n. Find M ∈ N with 21 q M −1 ≤ n < 12 q M . Then
by the above observation and since deg f ≤ n, there is some irreducible polynomial h(t) with
degree M such that h(t) does not divide f (t). From the choice of M we have
2n < q M ≤ 2nq,
so f (t) is not zero in the field F = Fq [t]/(h(t)), which satisfies 2n < |F| ≤ 2nq.
Proposition 3.2. Let G be an affine algebraic group scheme defined over Z and let K be
a purely transcendental extension of Fq (t) for some prime power q. If Γ ≤ G(K) is finitely
generated, then FΓE (n) ndim(G) and, if G is a Chevalley group, FΓ≤ (n) na(G) . If G = GLd ,
2
then FΓE (n) nd −1 and FΓ≤ (n) nd−1 .
6
DANIEL FRANZ
Proof. Fix an embedding G ֒→ GLd , allowing us to treat elements of Γ as invertible matrices
with entries in K. Because Γ is finitely generated, we may assume the transcendence basis
of K is finite, so write K = Fq (t)(x1 , · · · , xs ) for some indeterminates xi . For notational
convenience write R = Fq [t][x1 , · · · , xs ]. Again using the fact that Γ is finitely generated,
Γ ≤ G(S) for some S = R[g −1 ], g ∈ R.
Let X be a symmetric finite generating set of Γ. Let m > 0 such that g m γ ∈ Matd (R) for
all γ ∈ X. Now let A ∈ Γ with ||A||X = n and put B = g mn A ∈ Matd (R).
Since A is a word of length n in the elements of X, we may view B as a word of length n
in the elements of g m X = {g mγ : γ ∈ X}. Let N be larger than the degree or height of any
entry of an element of g m X.
If A is not a scalar matrix, then B has a nonzero off-diagonal entry or two diagonal entries
with nonzero difference; in this case put f equal to one of these nonzero values. We can
ignore the finitely many instances where A is a scalar matrix of determinant 1. If A = aId is
scalar with determinant not equal to 1, put f = g mnd (ad − 1). Our general strategy is to map
R[x1 , · · · , xs ] to an appropriately sized finite field F so that f g is not mapped to 0. This map
will then extend to a homomorphism ϕ : S → F with ϕ(f ) 6= 0, so that under the induced
homomorphism
ϕ∗ : G(S) → G(F),
the image of A is not a scalar matrix or has determinant not equal to 1.
To use our lemmas, we must first bound the degrees of the entries of B. Recall that B
can be represented as a word of length n in g mX, and each entry of an element of g mX has
degree less than or equal to N. Thus each entry of B has degree bounded above by nN; in
particular, deg f ≤ ndN, so if we set h = f g, then deg h ≤ 2ndN for sufficiently large n.
Similar reasoning shows ht(f ) ≤ 2ndN.
Since h is nonzero, it has some nonzero coefficient h0 (t) ∈ Fq [t] with deg h0 (t) ≤ 2ndN.
By Lemma 3.1, there exists a field F and homomorphism τ : Fq [t] → F such that
2n(2dN) ≤ |F| ≤ 2qn(2dN)
and τ (h0 ) 6= 0. Extending τ in the natural way to
τ : Fq [t][x1 , · · · , xs ] → F[x1 , · · · , xs ],
note that τ (h) 6= 0 and deg τ (h) ≤ 2ndN < |F|. Hence there exist α1 , · · · , αs ∈ F so that
τ (f )(α1 , · · · , αs ) ∈ F× , as is easily shown by induction on s.
Composing this evaluation map with τ yields a homomorphism θ : R → F such that
θ(h) 6= 0. Since the image of θ is a field and h = f g, g is mapped to a unit by θ, so θ
extends to a ring homomorphism ϕ : S → F satisfying ϕ(f ) 6= 0. Finally, ϕ induces a group
homomorphism
ϕ∗ : G(S) → G(F)
with ϕ∗ (A) nontrivial; by the choice of f , if det (ϕ∗ (A)) = 1 then ϕ∗ (A) is not a scalar
matrix. By Lemma 2.3, |G(F)| ≤ C|F|dim(G) for some constant C depending only on G, so
|F| ≤ 4qdNn and FΓE (n) ndim(G) .
If G is a Chevalley group, let P be a maximal subgroup of minimal index in G(F), so
[G(F) : P ] ≤ 2|F|a(G) by Lemma 2.4. The intersection of all conjugates of P is normal, so
since G(F)/Z(G(F)) is simple, this intersection is contained in Z(G(F)). But det A = 1, so
ϕ∗ (A) is not a scalar matrix, and hence ϕ∗ (A) 6∈ Z(G(F)). Thus ϕ∗ (A) is not in a subgroup
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
7
of G(F) of index at most 2|F|a(G) . Hence
A 6∈ H ≤ Γ and [Γ : H] ≤ 2|F|a(G) ,
so FΓ≤ (n) na(G) .
Now assume G = GLd . If det(ϕ∗ (A)) 6= 1, then the image of ϕ∗ (A) in F∗ under the
determinant map is nontrivial. If det(ϕ∗ (A)) = 1 then ϕ∗ (A) ∈ GLd (F) is not a scalar
matrix, so the image of ϕ∗ (A) is nontrivial in GLd (F)/Z(GLd (F)), the size of which is of
2
2
order |F|d −1 . Hence FΓE (n) nd −1 .
In addition, the image of ϕ∗ (A) in GLd (F)/Z(GLd (F)) is in the image of SLd (F), which is
isomorphic to PSLd (F). Applying the Chevalley group argument from above to PSLd and
using the fact that [PGLd (F) : PSLd (F)] ≤ d, we find that FΓ≤ (n) nd−1 .
3.2. Chebotarev density theorem. To work with coefficients in arbitrary fields, we need
to use the Chebotarev density theorem, which also plays an important role in [3], [6], [8].
In characteristic 0 we use the following higher dimensional generalization of the Chebotarev
density theorem, a consequence of Lemma 9.3 and Theorem 9.11 in [20].
Lemma 3.3. Let F = Q(x1 , · · · , xs ) and let K/F be Galois with Galois group G. Let π(x)
be the number of degree 1 maximal ideals in Z[x1 , · · · , xs ] with residue field of size at most x,
and let π1 (x) be the number of such ideals which are unramified over K such that |OK /m′ | =
xs+1
1
xs+1
|OF /m| for all m′ lying over m. Then π(x) ∼
and
π
(x)
∼
.
1
log(xs+1 )
|G| log(xs+1 )
In the characteristic p case, we use an effective version of the Chebotarev density theorem
for function fields. Let L be a Galois extension of Fp (t) with Galois group G and define
π(x, L/Fp (t)) to be the number of primes in Fp [t] of degree x which are unramified and
split completely
in L. Let P be the set of primes in Fp [t] which ramify over L, and set
Q
D = deg( q∈P q).
Theorem 3.4 ([23], Theorem 1). Let Fpm be the algebraic closure of Fp in L. If m divides
x, then
px/2 (2 + D)
1
m
Ip (x)| ≤
+D 1+
|π(x, L/Fp (t)) −
|G|
|G|x
x
Lemma 3.5. Fix c1 , c2 > 0. Let f (y) ∈ Fp [t][y] be separable with degree k. If n is sufficiently
large, h(t) ∈ Fp [t] has degree at most c1 n log n, and f (y) has discriminant ∆(f ) ∈ Fp [t] of
degree less than c2 log n, then there exists c ≤ 2c1 (k!)pk! , dependent on n, so that there exists
an irreducible polynomial g(t) ∈ Fp [t] of degree at most logp (cn log n) not dividing h(t) such
that f (y) factors into distinct linear factors mod g(t).
Proof. Let h(t) and f (y) satisfy the assumptions of the lemma, and let L be the splitting
field of f (y). Let Qx be the set of primes in Fp [t] of degree x which are unramified and split
completely in L, so π(x, L/Fp (t)) = |Qx |. If x > deg(∆(f )), then no element of Qx divides
∆(f ), and hence f (y) factors into distinct linear factors mod q(t) for all q(t) ∈ Qx .
We want to find x of an appropriate size so that some g(t) ∈ Qx does not divide h(t). To
that end, observe that
Y
q(t) = xπ(x, L/Fp (t)).
deg
q(t)∈Qx
8
DANIEL FRANZ
Let Fpm be the algebraic closure of Fp in L. Since [L : Fp (t)] ≤ k!, |G| ≤ k! and m|k!, so if
m|x then Theorem 3.4 yields
m
px/2 (2 + D)
1
π(x, L/Fp (t)) ≥ Ip (x) −
−D 1+
k!
x|G|
x
≥
1
px/2 (2 + D)
Ip (x) −
− 2D.
k!
x
Since deg(h(t)) ≤ c1 n log n, if xπ(x, L/Fp (t)) > c1 n log n, then some g(t) ∈ Qx will not
px
and D ≤ deg(∆(f )) ≤ c2 log n, we have
divide h(t). Using the estimates Ip (x) ≥
2x
xIp (x)
− px/2 (2 + D) − 2Dx
k!
px
≥
− px/2 (2 + c2 log n) − 2c2 x log n.
2(k!)
xπ(x, L/Fp (t)) ≥
Then
(1)
xπ(x, L/Fp (t)) − c1 n log n ≥
px
− px/2 (2 + c2 log n) − 2c2 x log n − c1 n log n.
2(k!)
If we set x = logp (c′ n log n) for some c′ > 0, then the right hand side of (1) becomes
c′ n log n p ′
− c n log n(2 + c2 log n) − 2c2 logp (n log n) log n − c1 n log n.
2(k!)
c′ n log n
and c1 n log n. Hence if c′ > 2c1 (k!) and n is
2(k!)
sufficiently large, then the above expression is positive.
However, we also need x to be an integer divisible by m, while logp (c′ n log n) may not even
be an integer. Since m divides k!, it is enough to have k! divide x. For any n, the interval
(logp (2c1 (k!)n log n), logp (2c1 (k!)pk! n log n)] has length k!, so it contains an integer multiple
of k!. Thus there exists c > 0 satisfying
The highest order terms in n are
2c1 (k!) < c ≤ 2c1 (k!)pk!
such that x = logp (cn log n) ∈ k!Z. Note that while the choice of c depends on n, its absolute
value is bounded independent of n.
For n sufficiently large, the above choice of c yields xπ(x, L/Fp (t)) > deg(h(t)), so we
conclude that there is some irreducible g(t) ∈ Fp [t] such that g(t) does not divide h(t) and
f (y) mod g(t) factors into distinct linear factors.
3.3. Proof of Theorems 1.1 and 1.3. The primitive element theorem plays a key role in
the proof of Theorem 1.1; this always applies to number fields, but finite extensions of Fp (t)
can be inseparable. The following lemma allows us to always work in the separable situation.
Lemma 3.6. Let p be a prime and put E0 = Fp (x1 , · · · , xs ) for some x1 , · · · , xs algebraically
independent over Fp . If L0 /E0 is a finite extension, then there is some positive integer m
1/pm
such that if x̃j = xj
for 1 ≤ j ≤ s, then L = L0 (x̃1 , · · · , x̃s ) is a separable extension of
E = Fq (x̃1 , · · · , x̃s ).
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
9
Proof. First note that since L0 /E0 is finite, there are some α1 , · · · , αk ∈ L0 such that L0 =
E0 (α1 , · · · , αk ). Each αi is the root of an irreducible polynomial fi (y) ∈ E0 [y]. In turn, each
m
fi (y) = gi (y p i ) for some irreducible, separable gi (y) ∈ E0 [y] and some positive integer mi .
1/pm
Set m = max{mi }, put x̃j = xj
for 1 ≤ j ≤ s, and let E = Fq (x̃1 , · · · , x̃s ). For each i,
m−mi
1/pmi
form g̃i (y) ∈ E[y] by replacing each xj in gi (y) by x̃pj
= xj
.
Since we are in characteristic p, we then have
m
m
fi (y) = gi (y p i ) = g̃i (y)p i ,
so g̃i (αi ) = 0. Each g̃i (y) is separable, so L = L0 (x̃1 , · · · , x̃s ) is separable over E.
We now prove the remaining parts of Theorems 1.1 and 1.3 together.
Theorem 3.7. Let G be an affine algebraic group scheme defined over Z, K be a field, and
Γ ≤ G(K) be a finitely generated subgroup. Put g(n) = n if char K = 0 and g(n) = n log n
if char K > 0.
Then FΓE (n) g(n)dim(G) and, if G is a Chevalley group, FΓ≤ (n) g(n)a(G) . If G = GLd ,
2
then FΓE (n) g(n)d −1 and FΓ≤ (n) g(n)d−1 .
Proof. Fix an embedding G ֒→ GLd (G ֒→ SLd if G is a Chevalley group) and let Γ = hXi ≤
G(K), where X is finite and symmetric. We may assume K is the field generated by the
entries of the elements of X. We first consider the case char K = 0, though we will see later
that most of the arguments leading up to the use of the Chebotarev density theorem apply
when char K > 0.
Since K is a finitely generated field, K is a finite extension of F = Q(x1 , · · · , xs ) for some
algebraically independent elements x1 , · · · , xs . Replacing K by its Galois closure if necessary,
we may assume K/F is Galois. By the primitive element theorem, K = Q(x1 , · · · , xs )[α] for
some α ∈ K, which we can choose to be integral over Z[x1 , · · · , xs ] = OF . Let f (y) ∈ OF [y]
be the minimal polynomial for α over Q(x1 , · · · , xs ) and set k = deg f (y).
The entries of the elements of X generate a ring contained in OF [g(x1 , · · · , xs )−1 ][α] = R[α]
for some g ∈ OF . Let J be the ideal of R[y] generated by f (y). If the ring homomorphism
ǫα : F [y] → F [α] = K
is evaluation of y to α, then clearly ker ǫα is the ideal generated by f (y). We claim the kernel
of ǫα |R[y] is J, so that R[α] ∼
= R[y]/J. This follows from the fact that f (y) is monic; if some
element of R[y] is a multiple of f (y) in K[y], then it must in fact be a multiple in R[y], as is
seen by an easy computation of coefficients.
We now present an outline of the proof. Let A ∈ Γ with ||A||X = n. Using the above
argument, we consider Γ as being embedded in G(R[y]/J). By using appropriate coset representatives and multiplication to eliminate inverses, we examine the entries of A as elements
of OF [y], i.e. as polynomials with integer coefficients. We then produce a homomorphism
OF [y] → Fp [y] under which an entry A remains nontrivial and the image of f (y) splits into
distinct linear factors. The end result is a homomorphism R[y] → Fp which factors through
J, inducing a homomorphism G(R[y]/J) → G(Fp ) which detects A.
Accomplishing this with no regard for the size of G(Fp ) is fairly straightforward, but to
achieve the desired bound, we must keep track of certain details. This is the reason we prefer
to work in OF [y]; these are simply polynomials with integer coefficients, with easily tracked
“size” properties.
10
DANIEL FRANZ
So let A ∈ Γ be nontrivial with ||A||X = n. If h ∈ R[y], set e
h to be the element of R[y]
e
e
e
with h ≡ h mod J and degy h < k. If b = h + J ∈ R[α], set b = e
h. For each γ ∈ X, let
e
γ be the element of Matd (R[y]) with e
e
γij = γf
γ |γ ∈ X}. Let m > 0 such that
ij . Put X = {e
m
g γe ∈ Matd (OF [y]) for all γ ∈ X. Let N be the maximum degree of the entries of all the
g mγe as polynomials in x1 , · · · , xs .
e = γe1 · · · γen . Then (g m )n A
e = B is a product of n elements
If A = γ1 · · · γn , γi ∈ X, let A
e so B ∈ Matd (OF [y]). For simplicity, suppose B has a nonzero off-diagonal
chosen from g m X,
entry h(x1 , · · · , xs , y) = h which is not divisible by f (y); the other cases can be treated as in
the proof of Proposition 3.2. Then for some constant α0 depending on g m , X, and s, we have
degy h ≤ (k − 1)n, degx1 ,··· ,xs h ≤ Nn, and ht(h) ≤ α0n ,
where ht(h) is the height of h, the largest absolute value of a coefficient of h.
We want to ensure that h continues to not be divisible by f (y) when we evaluate the
xi ; the easiest way to accomplish this is by degree considerations, so we now replace h by
e
h. Since our goal is to map this element to something nonzero, it will then suffice to clear
denominators and map the resulting polynomial to something nonzero. We need to do this
carefully to keep track of how the x degrees and coefficient sizes change.
Write
k−1
X
k
f (y) = y +
aj (x1 , · · · , xs )y j ,
j=0
where aj ∈ OF , and put M to be the maximum degree of the aj . Then for r > k, the
coefficients of yer will be sums of products of the aj . For example,
y k+1 = y · y k ≡ y −
k−1
X
aj y j
j=0
k
= −ak−1 y −
≡ ak−1
k−1
X
j=0
k−2
X
mod J
aj y j+1
j=0
aj y j −
= a0 ak−1 +
k−1
X
j=1
k+1 .
= yg
!
k−2
X
aj y j+1 mod J
j=0
(aj ak−1 − aj−1 )y j
As the above example helps illustrate, each yer ∈ OF [y] and the coefficients of yer will include
(k−1)n includes products of at most
products of at most r − (k − 1) coefficients of f (y), so y^
(k − 1)(n − 1) terms. Hence if a(x1 , · · · , xs ) is a coefficient of yer with r ≤ (k − 1)n, then
deg a ≤ M(n − 1)(k − 1) and ht(a) ≤ β (n−1)(k−1)
for some β independent of n.
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
11
We obtain e
h ∈ OF [y] by replacing each y r by yer . Using the size and degree estimates on
yer , we have
h ≤ N0 n, and ht(e
h) ≤ α0n ,
degy e
h < k, degx1 ,··· ,xs e
where N0 and α0 are independent of n.
Viewing e
h as a polynomial with coefficients in OF = Z[x1 , · · · , xs ], some coefficient
b(x1 , · · · , xs ) = b of e
h is nonzero. Let ∆(f (y)) be the discriminant of f (y), an element
of OF . Consider the polynomial
(2)
b′ (x1 , · · · , xs ) = g(x1 , · · · , xs )∆(f (y))b(x1 , · · · , xs ) ∈ OF .
Since the only term in this product that depends on n is b(x1 , · · · , xs ), b′ retains the properties
from e
h that its degree is linear in n, bounded by cn, and its height is exponential in n, bounded
n
by α .
Let OK be the integral closure of OF in K, and recall the definitions of π(x) and π1 (x)
from Lemma 3.3. We note that for each ideal m counted in π1 (x), f (y) mod m factors into
linear factors.
We wish to find m E OF with |OF /m| ≤ Cn for some constant C independent of n such
that b′ (x1 , · · · , xs ) mod m 6= 0 and f (y) mod m factors into distinct linear factors. There are
π1 (Cn) ideals that satisfy the latter condition. We count the number that fail the first.
Each degree one maximal ideal of OF = Z[x1 , · · · , xs ] is of the form (p, x1 −a1 , · · · , xs −as )
for some 0 ≤ ai ≤ p −1. For such an ideal m, b′ (x1 , · · · , xs ) ∈ m if and only if b′ (a1 , · · · , as ) ≡
0 mod p. Set
Xp (b′ ) = {a ∈ Fsp : b′ (a) ≡ 0 mod p}.
Then |Xp (b′ )| = ps if p|b′ , and a straightforward induction on s shows |Xp (b′ )| ≤ s deg(b′ )ps−1
if p does not divide b′ . So if M > 0,
X
X
X
|Xp (b′ )| ≤ s deg(b′ )
ps−1 +
ps .
p≤M
p≤M
We can split the second sum into two as
X
X
ps =
(3)
√
p≤ n,p|b′
p≤M,p|b′
p≤M,p|b′
ps +
√
X
ps .
n<p≤M,p|b′
√
n, so if there are k terms
Each prime in the second sum of (3) is greater
√than
√ in the sum,
the product of the involved primes is at least ( n)k . Since b′ has height αn , ( n)k ≤ αn , so
k ≤ 2n log α/ log n. Hence
X
2n log α s
M .
ps ≤
log
n
√
′
n<p≤M,p|b
Letting n be sufficiently large, Lemma 3.3 gives
√
X
( n)s+1
s
√
p ≤2
,
log(( n)s+1)
√
p≤ n
so if we put M = Cn for some C > 1 to be determined, we conclude that
X
4n log α s
M .
ps ≤
log n
′
p≤M,p|b
12
DANIEL FRANZ
If we let m0 = |Gal(K/F )|, then Lemma 3.3 also gives
X
1
Ms
M s+1
s−1
π1 (M) ≥
and
p
≤
2
,
2m0 log(M s+1 )
log(M s )
p≤M
so recalling that deg(b′ ) ≤ cn, we have
X
4n log α s
2 log α
c
Ms
s
′
.
+
M = 2nM
+
|Xp (b )| ≤ 2scn
s)
log(M
log
n
log
M
log
n
p≤M
Then we want
1
c
2 log α log M
2 log α
M s+1
s
⇔ M > 4m0 (s + 1)n c +
.
> 2nM
+
2m0 log(M s+1 )
log M
log n
log n
We can choose
log C
C > 4m0 (s + 1) c + log α 1 +
log n
independent of n when n isP
large since log C/ log n becomes arbitrarily small. Hence with
M = Cn, we have π1 (M) > p≤M |Xp (b)|. The number on the right side of this inequality is
the number of degree one maximal ideals containing b′ (x1 , · · · , xs ), so we can in fact choose
a maximal ideal m of OF with b′ (x1 , · · · , xs ) 6= 0 mod m, f (y) mod m factoring into distinct
linear factors, and OF /m ∼
= Fp , where p ≤ M = Cn.
Now consider the homomorphism
ψ : OF [y] → (OF /m)[y] ∼
= Fp [y].
Since ψ(b′ ) 6= 0 and g|b′ , we have ψ(g) 6= 0, so ψ extends to
π : R[y] = OF [g −1 ][y] → Fp [y]
with π(b′ ) 6= 0 and π(f (y)) a product of distinct linear polynomials.
Recalling the definition of b′ = b′ (x1 , · · · , xs ), π(b′ ) 6= 0 implies π(e
h) 6= 0. By our choice
e
e
e
of h, deg π(h) < deg π(f ), so π(f ) does not divide π(h). In particular, π(f ) has some
linear factor y − λ ∈ Fp [y] that does not divide π(e
h). Hence under the evaluation map
ǫλ : Fp [y] → Fp that sends y to λ, π(f ) is sent to 0 and π(e
h) remains nontrivial. Thus we
have a homomorphism
ǫλ ◦ π : R[y] → Fp
which maps f (y) to 0 and maps e
h(y) to a nonzero element of Fp . This map thus factors
through J = (f (y)) E R[y], yielding the commutative diagram below.
R[y]
ǫa
π //
// Fp
Fp [y]
<<
❍❍
①
①
❍❍
①①
❍❍
①
❍❍
①① ϕ
❍##
①①
R[y]/J
Recall from the beginning of the proof that h = Bij ≡ g mn Aij mod J for some i 6= j,
and that e
h ≡ h mod J. Then by the above diagram, the homomorphism ϕ : R[y]/J → Fp
satisfies
0 6= ϕ(e
h + J) = ϕ(h + J) = ϕ(g mn Aij + J).
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
13
Since g ∈ R× , ϕ(g mn + J) 6= 0, so we conclude that ϕ(Aij + J) 6= 0. Thus the ring
homomorphism ϕ induces a group homomorphism
ϕ∗ : G(R[y]/J) → G(Fp )
with ϕ∗ (A) a nontrivial, non-diagonal matrix. Restricting ϕ∗ to Γ yields the desired homomorphism.
By the choice of m we have |Fp | ≤ Cn, so we conclude FΓE (n) ndim(G) . The remaining
bounds are proved as at the end of the proof of Proposition 3.2.
Now consider the case char K = p. Then K is a finite extension of F = Fp (t, x1 , · · · , xs )
for some algebraically independent elements t, x1 , · · · , xs . By Lemma 3.6 we can assume K
is a separable extension of F . As in the characteristic 0 case we may then replace K by its
Galois closure and assume K/F is Galois. By the primitive element theorem, K = F [α] for
some α ∈ K. We again can assume α is integral, and we let f (y) ∈ Fp [t][x1 , · · · , xs ][y] be the
minimal monic polynomial for α over Fp (t)(x1 , · · · , xs ), with deg f (y) = k. In this context
OF = Fp [t][x1 , · · · , xs ] and R = OF [g −1 ], g ∈ OF .
One can now perform the same steps as in the characteristic 0 case, replacing Z by Fp [t]
and replacing the exponential size bounds on the coefficients by linear degree bounds. Indeed,
the first place where the characteristic p argument diverges is just after (2). So we pick up
the argument at that point, using the same notation as before.
We have a polynomial b′ (x1 , · · · , xs ) defined similarly as in (2),
b′ (x1 , · · · , xs ) = g(x1 , · · · , xs )∆(f (y))b(x1 , · · · , xs ) ∈ OF ,
with degree m ≤ c0 n for some c0 independent of n. Then by Lemma 2.2, there exist
g1 (t), · · · , gs (t) each of degree at most log m such that b′ (g1 (t), · · · , gs (t)) 6= 0. If ǫ : OF →
Fp [t] is the evaluation homomorphism with ǫ(xi ) = gi (t), we have ǫ(g) 6= 0, so ǫ extends to
ǫ : R → Fp [t] and thus induces a homomorphism
ψ : R[y] → Fp [t][y]
satisfying ψ(e
h) 6= 0, ϕ(f ) 6= 0, and ψ(∆(f )) 6= 0.
We are now in a position to use Lemma 3.5. We observe that
deg ǫ(b′ ) ≤ c′ n + m log m ≤ c1 n log n
for some constants c′ , c1 . Also, the discriminant of ψ(f (y)) has degree at most c2 log n for
some constant c2 since f (y) is independent of n and each gi (t) has degree at most log(c0 n).
Thus by Lemma 3.5, we can find c > 0 independent of n and an irreducible polynomial
F (t) of degree less than cn log n such that F (t) does not divide ψ(b′ ) and ψ(f (y)) factors
completely mod F (t).
Put F = Fp [t]/(F (t)), let πF be the homomorphism πF : Fp [t][y] → F[y] induced by
Fp [t] → F, and define π = πF ◦ ψ : R[y] → F[y]. Following the same arguments as in the
characteristic 0 case, one can then show the desired residual finiteness growth bounds.
4. Lower Bound Preliminaries
For the remainder of the paper we will assume G is a (simple) Chevalley group of rank at
least 2, and for this section we will in addition assume G is simply connected.
Let Φ be the irreducible root system of rank l ≥ 2 associated to G, and let g(C) be the
corresponding Lie algebra, with Chevalley basis {eα : α ∈ Φ} ∪ {h1 , · · · , hl }. For a field F,
14
DANIEL FRANZ
let g(F) be the Lie algebra over F with the given Chevalley basis. For background on root
systems and Chevalley bases, see [12].
Fix an embedding of G in SLd . Then there is a Lie algebra embedding g(C) into sld (C)
such that the action of G on g(C) by conjugation, via matrix multiplication, is the same as
the adjoint action.
Let K be the field Q or Fp (t) with ring of integers O = Z or Fp [t], respectively. , We have
G(O) = G(K) ∩ SLd (O). Fix a maximal ideal m and k ∈ N. Set R = O/mk and let p be the
characteristic of the field F = O/m.
Let Gi be the kernel of the projection G(R) → G(O/mi ) for 1 ≤ i ≤ k (note that
Gk = {1}). We now use the Gi to construct a graded Lie algebra (see [2] or [17], Chapter
7 for more details; these kernels were also used in [3] to compute normal residual finiteness
growth). Each Gi /Gi+1 is an elementary abelian p-group for 1 ≤ i ≤ k − 1, so we can define
an Fp vector space
L(G1 ) =
k−1
M
Gi /Gi+1 .
i=1
Defining the bracket on homogeneous elements to be
[xGi+1 , yGj+1] = (x, y)Gi+j+1,
where (x, y) is the group commutator, gives L(G1 ) the structure of a Lie algebra over Fp . We
have a Lie algebra isomorphism
L(G1 ) ∼
= g(F) ⊗ xF[x]/(xk ) =
Now let H ≤ G(R). Continuing to follow [2], define
L(H) =
k−1
M
i=1
k−1
M
xi g(F).
i=1
(H ∩ Gi )Gi+1 /Gi+1 .
Then L(H) is a graded Lie subalgebra of L(G1 ) and the codimension of L(H) in L(G1 ),
viewed as vector spaces over Fp , is logp [G1 : H ∩ G1 ].
Using the realization of L(G1 ) as g(F) ⊗ xF[x]/(xk ), write
L(H) =
k−1
M
xi hi ,
i=1
where
hi = (H ∩ Gi )Gi+1 /Gi+1 ∼
= (H ∩ Gi )/(H ∩ Gi+1 )
is viewed as an Fp -subspace of g(F). We note that G1 H/G1 ∼
= H/(H ∩ G1 ) acts on each hi
by conjugation and [hi , hj ] ⊆ hi+j if i + j < k.
In addition, we have the following result when H is a normal subgroup.
Lemma 4.1. Assume G(R) is perfect and H E G(R). If H 6= G(R), then G1 H 6= G(R).
Proof. Recall that R = O/mk . For any 1 ≤ j ≤ k − 1, there is a natural surjective homomorphism
Gj /Gj+1 → Gj H/Gj+1 H.
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
15
If Gj H = G(R), then G(R)/Gj+1H is the image of the abelian group Gj /Gj+1. Since G(R)
is perfect, G(R)/Gj+1H must be trivial, so Gj H = Gj+1H.
In particular, if G1 H = G(R), then the above argument implies Gk H = G(R). Since
Gk = 1, we conclude that H = G(R) if G1 H = G(R).
Set
E(F) =
M
Feα .
α∈Φ
We will write E for E(F) when context makes clear what F is.
Lemma 4.2. Fix α ∈ Φ and assume that α is a short root if Φ is of type Cl , l ≥ 2. Suppose
U, V ≤ g(F) satisfy Feα 6⊆ [U, V ]. Then
codimE (U ∩ E) + codimE (V ∩ E) ≥ 2[F : Fp ].
Proof. Write [U, V ] ∩ Feα = Aeα , A ≤ F. Let Tr: F → Fp be the usual trace form. By
assumption, A is proper, so there exists some nonzero a0 ∈ F such that Tr(a0 a) = 0 for all
a ∈ A. Replacing U by a0 U, we may assume Tr(a) = 0 for all a ∈ A.
Let B = {b1 , · · · , bn } be an Fp -basis of F, and let B ′ = {b′1 , · · · , b′n } be the dual basis of B
with respect to the trace, so that
(
1 if i = j
′
Tr(bi bj ) =
0 if i 6= j.
We now construct subspaces of E which intersect U and V trivially. Let β ∈ Φ such that
α − β ∈ Φ and [eα−β , eβ ] = eα (such a β always exists because we exclude α from being a
long root if Φ is of type Cl ).
Set X = {bi eα−β , b′i eβ : 1 ≤ i ≤ n} and define an involution on X by bi eα−β = b′i eβ . Then
if w1 6= w2 ∈ X, [w1 , w2 ] = aeα with Tr(a) = 0.
Let XU ⊆ X be maximal with respect to the property hXU i ∩ U = 0, and set XV = {w :
w ∈ X \ XU }. We show hXV i ∩ V = 0. P
If not, there is some nonzero v =
w∈XV sw w ∈ V , where each sw ∈ Fp . We now
construct u ∈ U such that the coefficient of [u, v] ∈ Feα has nonzero trace, contradicting
[U, V ] ∩ Feα ⊆ Aeα .
Some coefficient sw0 is nonzero, and we can assume sw0 = 1. Then w0 6∈ XU , so u =
w0 + z ∈ U for some z ∈ hXU i. By the choices of XU and XV , [z, v], [w0 , v − w0 ] ∈ Fxα each
have coefficients with trace 0, so for some a ∈ F with Tr(a) = 0,
[u, v] = [w0 , w0] + axα = (±bi b′i + a)xα ∈ [U, V ]
for some 1 ≤ i ≤ n. But Tr(bi b′i + a) = ±1 6= 0, yielding a contradiction, so hXV i ∩ V = 0.
Since hXU i, hXV i ⊆ E, we conclude that
codimE (U ∩ E) + codimE (V ∩ E) ≥ |XU | + |XV | = 2[F : Fp ].
For a commutative ring R, we write {xα (r) : r ∈ R} for the root subgroup of G(R)
corresponding to α ∈ Φ. The following result, proved in [21] for the case R is a Euclidean
domain and in [1] when R is semi-local, allows us to use a nice generating set of G(R) for
the rings R we are interested in.
Lemma 4.3. If R is a Euclidean domain or a semi-local ring and G is a simply connected
Chevalley group, then G(R) = hxα (r) : α ∈ Φ, r ∈ Ri.
16
DANIEL FRANZ
Lemma 4.4. Let G be a Chevalley group of type Φ. Then G(Fp [t]) is perfect unless p = 2
and Φ is of type B2 or G2 .
Proof. The statement is proved in chapter 11 of [10] for Chevalley groups over fields, but the
same arguments apply to the polynomial ring Fp [t].
If g ∈ O, we write G(O, g) = ker(G(O) → G(O/g)), and we denote the gcd of π and g as
(π, g). We call G(O, g) a principal congruence subgroup; any subgroup of G(O) containing
a principal congruence subgroup is called a congruence subgroup.
If the rank of G is at least 2, then G(O) has the congruence subgroup property: every
finite index subgroup of G(O) is a congruence subgroup (see Chapter 9 of [18] for details).
We note that it is necessary that G be simply connected for this to be true.
Using the congruence subgroup property we will be able to reduce to the case of considering
principal congruence subgroups. The next two statements will help us work with their images
in G(R).
Lemma 4.5. Let R = O/π k for some irreducible π ∈ O and set ∆ = G(O, g) for some
g ∈ O. Let ∆ be the image of ∆ in G(R).
(i) If (π, g) = 1, then ∆ = G(R).
(ii) If (π k , g) = π s with s < k, then E ⊆ (∆ ∩ Gi )Gi+1 /Gi+1 for s ≤ i ≤ k − 1.
Proof. First assume (π, g) = 1. Then for any f ∈ O, there exist h1 , h2 ∈ O such that
h1 π k + h2 g = f . Thus if α ∈ Φ,
xα (h2 g) = xα (f )xα (−h1 π k ) ∈ ∆,
so xα (f mod π k ) ∈ ∆. Lemma 4.3 then implies ∆ = G(R).
Now assume (π k , g) = π s with s < k. Then using similar reasoning as above, for any α ∈ Φ
and any f ∈ O, xα (π s f mod π k ) ∈ ∆. Hence for s ≤ i ≤ k − 1,
proving the lemma.
{xα (π i f mod π k ) : α ∈ Φ, f ∈ O} ⊆ ∆ ∩ Gi ,
We note that the statement in (ii) of the above lemma is not optimal. In fact (∆ ∩
Gi )Gi+1 /Gi+1 = G(R) in this case, but we only require the weaker statement to prove the
following corollary.
Corollary 4.6. With the same setup as in Lemma 4.5, let H ≤ ∆ and fix α ∈ Φ, a short
root if Φ is type Cl . Assume (π k , g) = π s and Feα 6⊆ hj for some 1 ≤ j ≤ k − 1 such that
s < j/2. If s = 0, then
codimL(∆) L(H) ≥ [F : Fp ](j − 1).
If s ≥ 1, then
codimL(∆) L(H) ≥ [F : Fp ](j − 2s + 1).
Proof. Since Feα 6⊆ hj and [hi , hj−i ] ⊆ hj for 1 ≤ i ≤ j − 1, we have Feα 6⊆ [hi , hj−i ]. Put
di = (∆ ∩ Gi )Gi+1 /Gi+1 .
If s = 0, then ∆ = G(R) by Lemma 4.5, so E ⊆ di for 1 ≤ i ≤ k − 1. Then Lemma 4.2
implies
codimdi (hi ) + codimdj (hj−i ) ≥ 2[F : Fp ]
for 1 ≤ i ≤ j − 1. Hence
codimL(∆) L(H) ≥ [F : Fp ](j − 1).
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
17
If s ≥ 1, Lemma 4.5 gives that E ⊆ di for s ≤ i ≤ k − 1. There are j − 2s + 1 integers in
the interval [s, j − s], so the previous reasoning yields the desired inequality.
Lemma 4.7. Let F be a finite field of characteristic p and G a simply connected Chevalley
group. For all but finitely many p, the adjoint action of G(F) on g(F) is irreducible. The
exceptions are given in Table 2, along with the largest possible dimension of a proper ideal
I ⊆ g(F) invariant under the action of G(F) in those cases. If G is type B2 and p = 2, then
any invariant ideal I is either the center or contains Feα for all short roots α.
Proof. See Theorem 2.1 in [11].
Φ
p
max dim(I) min codim(I)
Al , l ≥ 2 p|(l + 1)
1
l2 + 2l − 1
Bl , l ≥ 3
2
2l + 2
2l2 − l − 2
2
Cl , l ≥ 2
2
2l − l
2l
2
Dl , l ≥ 4
2
2
2l − l − 2
G2
3
7
7
F4
2
26
26
E6
3
1
77
E7
2
1
132
Table 2.
The final lemma of this section enables us to apply Lemma 4.7 to the situation where we
consider g(F) as a vector space over Fp . Because the proof is technical, we postpone it to the
end of the paper.
Lemma 4.8. Let F be a finite field of characteristic p such that |F| ≥ 4, and let G be a simply
connected Chevalley group with root system Φ. Let V be a proper Fp -subspace of g(F). If V
is G(F)-invariant, then FV , the F subspace spanned by V , is a proper ideal of g(F) which is
invariant under the action of G(F).
5. Lower bounds in characteristic 0
We continue with the notation of the previous section, with G remaining a simply connected
Chevalley group with a fixed embedding into SLd . Fix α ∈ Φ, a short root if G is type Cl .
We first provide lower bounds for the normal and non-normal residual finiteness growth of
G(Z). The values of dim(G) and a(G) can be found in Table 1.
Lemma 5.1. Let R = Z/pk for a prime p, k ≥ 1. Let ∆ = G(Z, N) and ∆ be the image of
∆ in G(R). Assume (pk , N) = ps . Let r ≥ N be sufficiently large and set
Lr = (lcm(1, 2, · · · , r))3(dim(G)+s) ,
Mr = xα (Lr mod pk ).
1
1
If Mr 6∈ H ≤ ∆, then [∆ : H] ≥ r a(G) . If in addition H E ∆, then [∆ : H] ≥ r dim(G) .
2
2d
Proof. Let Mr , H be as in the statement and suppose pm−1 ||Lr , by which we mean pm−1 is
the largest power of p dividing Lr . We have m ≤ k since Mr 6= 1 and Mr ∈ ∆ since N|Lr .
18
DANIEL FRANZ
The argument splits into a few cases. We will consider H as an arbitrary subgroup and as a
normal subgroup in each case.
Case 1: k = 1. Since k = 1, we have R ∼
= Fp and p > r ≥ N, so (p, N) = 1. By Lemma 4.5,
∆ = G(Fp ), so H is a proper subgroup of G(Fp ). Then by Lemma 2.4, [G(Fp ) : H] ≥ 21 pa(G) .
Since Mr is nontrivial, p does not divide Lr , so by construction of Lr , p > r. Hence [G(Fp ) :
1
H] ≥ r a(G) , as desired.
2
If in addition H is normal, then H ⊆ Z(G(Fp )) since G(Fp )/Z(G(Fp)) is simple. Thus by
Lemma 2.4,
1
1
[G(Fp ) : H] ≥ |G(Fp )/Z(G(Fp))| ≥ pdim(G) > r dim(G) .
2d
2d
Case 2: k ≥ 2, m = 1. Let G1 be the kernel of the projection G(R) → G(Fp ), and recall
the graded Lie algebras L(G1 ) and L(H) defined in the previous section. Since m = 1, p
does not divide Lr , so again p > r and ∆ = G(Fp ). We also have Mr 6∈ G1 . If in addition
G1 H 6= G(R), then the image of H in G(R)/G1 ∼
= G(Fp ) is proper, so
1
1
[G(R) : H] ≥ pa(G) > r a(G) .
2
2
1
If H is normal, then by the same reasoning as before we see that [G(R) : H] > r dim(G) .
2d
If G1 H = G(R), then since p > r is large, G1 H/G1 ∼
= G(Fp ) acts irreducibly on g(Fp )
by Lemma 4.7, so each hj is trivial or all of g(Fp ). If all are g(Fp ), this forces H = G(R),
contradicting Mr 6∈ H. Thus hj is trivial for some j and
codimL(G1 ) L(H) ≥ codim hj = dim(G),
so [G(R) : H] ≥ pdim(G) > r dim(G) .
Case 3: k ≥ 2, m ≥ 2. Since Mr 6∈ H and pm−1 ||Lr , we have Mr ∈ Gm−1 \ Gm , so
Fp eα 6⊆ hm−1 . If pl || lcm(1, · · · , r), then
m − 1 = 3(dim(G) + s)l and p(l+1) dim(G) > r dim(G) .
In particular, s < j/2, so by Corollary 4.6, if s ≥ 1 then
Since
we conclude that
codimL(∆) (L(H)) ≥ m − 2s.
m − 2s = 3(dim(G) + s)l − 2s + 1 ≥ dim(G)(l + 1),
[∆ : H] ≥ [∆ ∩ G1 : H ∩ G1 ] ≥ pdim(G)(l+1) > r dim(G) .
A similar argument works when s = 0, using the corresponding inequality from Corollary
4.6.
Theorem 5.2. Let G be a Chevalley group of rank at least 2, not necessarily simply connected,
and let ∆ be a finite index subgroup of G(Z). Then F∆E (n) ndim(G) and F∆≤ (n) na(G) .
Proof. Let Gsc be the simply connected cover of G. Then the natural map Gsc (Z) → G(Z)
has finite kernel, so by Lemma 2.4 in [6], the residual finiteness growth of G(Z) is bounded
below by that of Gsc (Z). Thus we may assume G is simply connected, with irreducible root
system Φ.
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
19
Then G(Z) satisfies the congruence subgroup property, so ∆ contains some principal congruence subgroup. Since residual finiteness growth can only decrease by passing to a subgroup, we may assume ∆ = G(Z, N) for some N ∈ Z. Let s be the largest power of a prime
dividing N.
Fix r ≥ N sufficiently large and put Lr = (lcm(1, 2, · · · , r))3(dim(G)+s) . Fix some α ∈ Φ,
a short root if G is type Cl . We show that Mr = xα (Lr ) is in every subgroup of G(R) of
sufficiently small index. First we need to determine the word length of Mr in ∆.
By Theorem A in [16], there exists a generating set X of G(R) so that
||Mr ||X ≤ C1 log |Lr |
for some C1 > 0. By the Prime Number Theorem, lcm(1, · · · , r) ∼ er , so log |Lr | ≤ C2 (d+s)r
for some absolute constant C2 . Since G(Z) is quasi-isometric to ∆, we conclude that
||Mr ||Y ≤ Cr
for some generating set Y of ∆ and some constant C independent of r.
Now suppose Mr 6∈ H ≤ ∆. By the congruence subgroup property of G(Z), H ⊇ G(Z, N ′ )
Q
for some N ′ ∈ Z. Let R = Z/N ′ and let N ′ = ki=1 pki i be the prime factorization of N ′ .
Write G(i) = G(Z/pki i ) for each i, so that
G(R) ∼
=
k
Y
G(i) .
i=1
Let πN ′ be the natural projection G(Z) → G(R). Then πN ′ (Mr ) 6∈ πN ′ (H), so in some G(i) ,
Mr = xα (LR mod pki i ) 6∈ H, where these are the images in G(i) . So by Lemma 5.1,
1
[∆ : H] ≥ [∆, H] ≥ r a(G) ,
2
1 dim(G)
r
if H E ∆. Recalling that Mr has word length n ≤ Cr finishes the
and [∆ : H] ≥
2d
argument.
6. Lower bounds in characteristic p
We continue using the same setup as in the previous section but now deal with the groups
G(Fp [t]) ⊆ SLd (Fp [t]). We first prove the following lemma allowing us to handle the case when
G is type B2 and p = 2. Let the root system of type B2 have roots {±ǫ1 , ±ǫ2 , ±(ǫ1 ± ǫ2 )}, as
usual.
Lemma 6.1. Fix k ≥ 1 and set R = F2 [t]/f (t)k for some irreducible f (t). Let G be a simply
connected Chevalley group of type B2 and let H be a proper normal subgroup of G(R). If
G(R)′ 6⊆ H, then G1 H 6= G(R).
Proof. We first set up some notation to make the computations more clear. For 1 ≤ j ≤ k,
put G(j) = G(F2 [t]/π j ), and for 1 ≤ i ≤ j, set G(j)i = ker(G(j) → G(i)). We will continue
writing Gi for G(k)i . Note that G(k) = G(R).
We first show Gk−1 H 6= G(k). Assume otherwise. Then
G(k)/H = Gk−1 H/H ∼
= Gk−1 /(H ∩ Gk−1 )
is a nontrivial abelian quotient of G(k), so [G(k), G(k)] ⊆ H, a contradiction.
Recall that we can view hj = (H ∩ Gj )Gj+1 /Gj+1 as a subspace of g(F).
20
DANIEL FRANZ
We now assume for the sake of contradiction that G1 H = G(k). Since Gk−1 H 6= G(k),
there exists some 2 ≤ j ≤ k − 1 such that Gj−1H = G(k) and Gj H 6= G(k). Then
g(F)/hj−1 ∼
= Gj−1 /(H ∩ Gj−1)Gj ∼
= Gj−1H/Gj H
is nontrivial, so hj−1 6= g(F). We now show that G1 H = G(k) also implies that hj−1 = g(F).
Put H(j) = Gj H/Gj and observe H(j) is properly contained in G(j). Since Gj−1 H = G(k),
G(j)j−1H(j) = G(j), so G(j)′ ⊆ H(j). Hence
xǫ1 (π j−1)xǫ2 +ǫ1 (π j−1) = (xǫ2 (1), xǫ1 −ǫ2 (π j−1 )) ∈ H(j) ∩ G(j)j−1 ,
so eǫ1 + eǫ2 +ǫ1 ∈ hj−1 .
The subspace hj−1 is invariant under the action of G1 H/G1 = G(F), so by Lemma 4.8,
Fhj−1 is a proper ideal of g(F). Then by Lemma 4.7, Fhj−1 is the center of g(F) or contains
Feα for each short root α. Clearly Fhj−1 is not the center, so it contains eǫ1 and thus also
contains eǫ2 +ǫ1 . This then forces Fhj−1 to be all of g(F), a contradiction.
Let α ∈ Φ be a short root if G is type Cl .
Lemma 6.2. Let R = Fp [t]/f (t)k for an irreducible polynomial f (t), k ≥ 1. Let ∆ =
ker(G(Fp [t]), g(t)) and let ∆ be the image of ∆ in G(R). Assume (f (t)k , g(t)) = f (t)s . Fix
r ≥ deg(g(t)), and set
Lr (t) = (lcm{h(t) ∈ Fp [t] : deg(h(t)) ≤ r})3(dim(G)+s) .
If p = 2 and G is type C2 , let
Mr = xǫ1 (Lr (t) mod f (t)k )xǫ1 +ǫ2 (Lr (t) mod f (t)k ),
and otherwise set
Mr = xα (Lr (t) mod f (t)k ).
1
1
If Mr 6∈ H ≤ ∆, then [∆ : H] ≥ pra(G) . If in addition H E ∆, then [∆ : H] ≥ pr dim(G) .
2
2d
Proof. Let Mr , H be as in the statement, put q = pdeg(f (t)) , and suppose f (t)m−1 ||Lr (t),
where m ≤ k since Mr 6= 1. Observe that Mr ∈ ∆ since g(t)|Lr (t). The argument splits into
a few cases. As in the proof of Lemma 5.1, we will treat H as an arbitrary subgroup and
then as a normal subgroup in each case. The arguments are very similar to the characteristic
0 case, so details will sometimes be skipped.
Case 1: k = 1. Since k = 1, we have R ∼
= Fq and deg(f (t)) > r ≥ deg(g(t)), so f (t) and g(t)
are relatively prime. Then ∆ = G(Fq ) by Lemma 4.5, so H is a proper subgroup of G(Fq ).
1
By Lemma 2.4, [G(Fq ) : H] ≥ q a(G) . Hence
2
1
[∆ : H] = [G(Fq ) : H] ≥ pra(G) .
2
If H is normal, then H ⊆ Z(G(Fq )) since G(Fq )/Z(G(Fq )) is simple, so Lemma 2.4 gives
1
1
∆ : H] ≥ q dim(G) > pr dim(G) .
2d
2d
Case 2: k ≥ 2, m = 1. Since m = 1, we again have deg(f (t)) > r and ∆ = G(R). Let G1
be the kernel of the projection G(R) → G(Fq ), and recall the graded Lie algebras L(G1 ) and
L(H).
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
21
We first consider the case H EG(R). If G(R) is perfect, then by Lemma 4.1, G1 H 6= G(R).
Hence the image of H in G(Fq ) is proper and
1
[∆ : H] ≥ pr dim(G)
2d
as before. Otherwise, by Lemma 4.4, p = 2 and G is type B2 or G2 . In the former case,
Mr = [xǫ1 (1), xǫ1 −ǫ2 (Lr (t) mod f (t)k )] ∈ G(R)′ ,
so G(R)′ 6⊆ H and thus G1 H 6= G(R) by Lemma 6.1, yielding the desired bound as shown
above. If G is type G2 and G1 H = G(R), then G1 H/G1 ∼
= G(Fq ) acts irreducibly on each hi
by Lemmas 4.7 and 4.8. Hence some hi is trivial, so
codimL(G1 ) L(H) ≥ dim(G) deg(f (t))
and
[∆ : H] ≥ pdim(G) deg(f (t)) ≥ pr dim(G) .
If H is an arbitrary subgroup of G(R), then the case G1 H 6= G(R) again reduces to a
previous argument. So assume G1 H = G(R). Then G1 H/G1 ∼
= G(Fq ) acts on each hi , so
for each i, either hi = g(F) or Fhi is a proper ideal, using Lemma 4.8. Since hi ⊆ Fhi , by
examining Table 2 and Table 1 we see that each hi is all of g(Fq ) or has codimension at least
a(G) deg(f (t)). Since H is proper, not all the hi can be g(Fq ), so
codimL(G1 ) L(H) ≥ a(G) deg(f (t)) > ra(G).
Thus [G(R) : H] ≥ pra(G) .
Case 3: k ≥ 2, m ≥ 2. We handle H being normal and arbitrary simultaneously. Since
Mr 6∈ H and f (t)m−1 ||Lr (t), we have Mr ∈ Gm−1 \Gm , so Fq eα 6⊆ hj for some m−1 ≤ j ≤ k−1
(Fq (eǫ1 + eǫ1 +ǫ2 ) 6⊆ hj if G is type B2 , p = 2).
By the construction of Lr (t), f (t)m−1 ||Lr (t) implies m − 1 = 3(dim(G) + s)l for some
integer l ≥ 1 satisfying deg(f (t))(l + 1) > r. In particular, s < j/2, so by Corollary 4.6, if
s ≥ 1 then
codimL(∆) (L(H)) ≥ deg(f (t))(j − 2s + 1)
We have
so
and hence
≥ deg(f (t))(m − 2s).
m − 2s = 3(dim(G) + s)l − 2s + 1 ≥ dim(G)(l + 1),
deg(f (t))(m − 2s) ≥ dim(G) deg(f (t))(l + 1) > r dim(G),
[∆ : H] ≥ [∆ ∩ G1 : H ∩ G1 ] ≥ pr dim(G) .
A similar argument works when s = 0, using the corresponding inequality from Corollary
4.6.
We note that while Corollary 4.6 does not directly apply in the case G is type B2 , p = 2,
the same arguments in Lemma 4.2 work when using eǫ1 + eǫ1 +ǫ2 in place of eα because
eǫ1 + eǫ1 +ǫ2 = [eǫ1 + eǫ2 , eǫ1 −ǫ2 + eǫ2 −ǫ1 ].
Theorem 6.3. Let G be a Chevalley group, not necessarily simply connected, of rank at least
2, let p be a prime, and let ∆ be a finite index subgroup of G(Fp [t]). Then F∆E (n) ndim(G)
and F∆≤ (n) na(G) .
22
DANIEL FRANZ
Proof. As in the proof of Theorem 5.2, we may assume G is simply connected and ∆ =
G(Fp [t], g(t)) for some g(t) ∈ Fp [t]. Let s be the largest power of an irreducible polynomial
dividing g(t).
Fix r ≥ deg(g(t)) and set
Lr (t) = (lcm{h(t) : deg(h(t)) ≤ r})3(dim(G)+s) .
Let Φ be the root system of G, and let α ∈ Φ, with the extra condition that α is a short root
if Φ is of type Cl , l ≥ 2. Set
(
xǫ1 (Lr (t))xǫ1 +ǫ2 (Lr (t)) if Φ = C2 , p = 2
Mr =
.
xα (Lr (t))
otherwise
By Theorem A in [16], there exists a generating set X of G(Fp [t]) so that
||Mr ||X ≤ C1 deg(Lr (t))
for some constant C1 . The degree of lcm{h(t) ∈ Fp [t] : deg(h(t)) ≤ r} is at most 2pr , so
deg(Lr (t)) ≤ 6(dim(G) + s)pr . Hence ||Mr ||X ≤ C2 pr for some constant C2 . Since G(Fp [t])
is quasi-isometric to ∆, we conclude that Mr has word length n ≤ Cpr for some constant C
with respect to some generating set of ∆.
The remaining argument is the same as in the proof of Theorem 5.2. Substituting Lemma
1
6.2 for Lemma 5.1, one shows that if Mr 6∈ H ≤ ∆, then [∆ : H] ≥ pra(G) , and if H is
2
1 r dim(G)
.
normal then [∆ : H] ≥ p
2d
Appendix A. Proof of Lemma 4.8
Here we give the postponed proof of Lemma 4.8. We will use the descriptions of the
irreducible root systems given in section 12 of [12] except for G2 ; in this case we fix a base
{αS , αL }, where αS and αL are short and long roots, respectively, so that the short roots are
{±αS , ±( αS + αL ), ±(2αS + αL )} and the long roots are {±αL , ±(3αS + αL ), ±(3αS + 2αL )}.
Lemma A.1. Let Φ be an irreducible root system.
(1) If α, γ ∈ Φ and α is a long root, then γ − 2α ∈ Φ if and only if γ = α.
(2) If Φ is not of type Cl , l ≥ 2, then there exist long roots α, β ∈ Φ such that α + β ∈ Φ and
α − β 6∈ Φ.
Proof. We first prove (1). Let α, γ ∈ Φ with α long. If γ = α then γ − 2α = −α ∈ Φ. So
assume that γ − 2α ∈ Φ. If θ is the angle between α and γ, then (α, γ) = |α||γ| cos θ, so
(α, γ) ≤ |α||γ|, with equality if and only if γ = α. Then
|γ − 2α|2 = (γ − 2α, γ − 2α)
= |γ|2 − 4(α, γ) + 4|α|2
≥ |γ|2 − 4|α||γ| + 4|α|2
= (|γ| − 2|α|)2,
with equality if and only if γ = α. But (|γ| − 2|α|)2 ≥ |α|2 and α is a long root, so
|γ − 2α|2 > (|γ| − 2|α|)2 is not possible since γ − 2α ∈ Φ. Hence γ = α.
We now prove (2) case by case. If Φ is a simply laced root system, then there are no root
strings of length greater than 2, so any choice of α, β ∈ Φ with α + β ∈ Φ will suffice.
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
23
If Φ is of type Bl , l ≥ 3, or F4 , set α = ǫ1 − ǫ2 and β = ǫ2 − ǫ3 . Then α + β = ǫ1 − ǫ3 ∈ Φ
and α − β 6∈ Φ.
If Φ is of type G2 , then put α = αL and β = 3αS + αL . Then α + β = 3αS + 2αL ∈ Φ and
α − β = −3αS 6∈ Φ.
Let Φ be an irreducible root system, let G be a Chevalley group of type Φ and let g be the
Lie algebra of type Φ with Chevalley basis B = {eα : α ∈ Φ} ∪ {h1 , · · · , hl }. The following
equations and more information on Chevalley groups can be found in [10].
Let g(Z) be the Z-span of B; this is a Lie algebra over Z. If α ∈ Φ and v ∈ g(Z), then
by the properties of Chevalley bases, 21 [eα , [eα , v]] and 61 [eα , [eα , [eα , v]]] are both in g(Z). If p
is a prime, then using the natural map g(Z) → g(Fp ), we can interpret these expressions as
elements in g(Fp ), regardless of the choice of p. In particular, these expressions make sense
even if p = 2 or p = 3.
Using this interpretation, if F is a field, α ∈ Φ, and t ∈ F, then
(4)
1
1
xα (t) · v = v + t[eα , v] + t2 [eα , [eα , v]] + t3 [eα , [eα , [eα , v]]],
2
6
where the final term is always 0 if Φ is not of type G2 .
We will need the following specific instances of (4).
xα (t) · eα = eα ,
xα (t) · e−α = e−α + thα − t2 eα ,
xα (t) · hα = hα − 2teα .
If α, β ∈ Φ are linearly independent, i.e. β 6= ±α, then
xα (t) · hβ = hβ − hα, βieα,
q
X
Mα,β,i ti eiα+β ,
xα (t) · eβ = eβ +
i=1
where Mα,β,i ∈ {±1, ±2, ±3}.
We are now ready to prove Lemma 4.8.
Lemma 4.8. Let F be a finite field of characteristic p such that |F| ≥ 4, and let G be a simply
connected Chevalley group with root system Φ. Let V be a proper Fp -subspace of g(F). If V
is G(F)-invariant, then FV , the F-subspace spanned by V , is a proper ideal of g(F) which is
invariant under the action of G(F).
Proof. Let Φ have rank l and fix a Chevalley basis B = {eα : α ∈ Φ} ∪ {h1 , · · · , hl } of g(F).
The F-subspace FV is an ideal of g(F) if [g(F), FV ] ⊆ FV , but it is sufficient to check that
[eα , FV ] ⊆ FV for all α ∈ Φ, as we now show.
Let Π = {α1 , · · · , αl } be the base for Φ associated to the Chevalley basis B, so that
hi = [eαi , e−αi ] for 1 ≤ i ≤ l. Then using the Jacobi identity, for v ∈ FV and 1 ≤ i ≤ l we
have
[hi , v] = [[eαi , e−αi ], v] = [eαi , [e−αi , v]] − [e−αi , [eαi , v]].
Thus if [eα , FV ] ⊆ FV for all α ∈ Φ, then [hi , FV ] ⊆ FV as well, so [g(F), FV ] ⊆ FV and FV
is an ideal. We now proceed to the proof of the lemma.
24
DANIEL FRANZ
First assume that V is actually an F-subspace of g(F), so V = FV . If α ∈ Φ, λ ∈ F, and
v ∈ V , then we can write (4) as
1
1
(5)
xα (λ) · v − v = λ[eα , v] + λ2 [eα , [eα , v]] + λ3 [eα , [eα , [eα , v]]] ∈ V.
2
6
Since |F| ≥ 4, there exist three distinct nonzero elements s, t, u ∈ F. Fix v ∈ V and α ∈ Φ
and write the right hand side of (5) as λz1 + λ2 z2 + λ3 z3 ∈ V . Since V is an F-subspace, this
implies z1 + λz2 + λ2 z3 ∈ V . Using s, t, u in place of λ, we have
v1 = z1 + sz2 + s2 z3 ∈ V,
v2 = z1 + tz2 + t2 z3 ∈ V,
v3 = z1 + uz2 + u2 z3 ∈ V.
The matrix of this linear system is Vandermonde and hence invertible. Then z1 = [eα , v] is
a linear combination of v1 , v2 , and v3 , and thus [eα , v] ∈ V .
Therefore V is an ideal of g(F). Since V is assumed to be proper, V = FV is a proper
ideal of g(F), so the lemma is proved in this case.
We now assume for the remainder of the proof that V is an Fp -subspace, but not necessarily
an F-subspace, of g(F). Then FV is a G(F)-invariant F-subspace of g(F) and thus an ideal
by the above argument. It remains to show that FV 6= g(F). We split the proof into three
cases which cover different restrictions on p and Φ.
Case 1: p 6= 2, Φ not of type G2 . If s ∈ F, α ∈ Φ, and v ∈ V , then by (4) we have
xα (s) · v − xα (−s) · v = 2s[eα , v] ∈ V,
so s[eα , v] ∈ V . Thus s[hi , v] ∈ V for all s ∈ F, 1 ≤ i ≤ l, v ∈ V by the argument at the
beginning of this proof.
Therefore the F-span of {[x, v] : x ∈ g(F), v ∈ V } is contained in V and hence is not equal
to g(F). But this set is just [g(F), FV ]. Since char F 6= 2, [g(F), g(F)] = g(F), so we must
have FV 6= g(F).
Case 2: p any prime, Φ not of type Cl , l ≥ 2. Assume FV = g(F); we will show that this
implies V = g(F), a contradiction.
Let EL and ES be the F-subspaces of g(F) spanned by {eα : α long} and {eα : α short},
respectively, so g(F) = H ⊕ ES ⊕ EL , with the convention that ES = 0 if Φ is simply laced.
Since V is G(F)-invariant and xα (t) · e−α = e−α + thα − t2 eα for α ∈ Φ, t ∈ F, to show that
V = g(F) it suffices to show ES ⊕ EL ⊆ V .
Fix v ∈ V , which we write as
X
(6)
v =h+
sβ eβ ∈ V,
β∈Φ
where h ∈ H and sβ ∈ F. If γ ∈ Φ is a long root and δ ∈ Φ, then 2γ + δ ∈ Φ if and only if
δ = −γ by Lemma A.1(1), so for any t ∈ F,
(7)
xγ (t) · v − v = t[eγ , v] − t2 s−γ eγ ∈ V.
By Lemma A.1(2), we can find long roots α and β such that α + β ∈ Φ and α − β 6∈ Φ. If
γ ∈ Φ, then by Lemma A.1(1), γ − 2α ∈ Φ if and only if γ = α.
By assumption, FV = g(F), so there exists v ∈ V with sα 6= 0. Hence to show that
Fe−α ⊆ V , it is sufficient to prove the following claim.
QUANTIFYING RESIDUAL FINITENESS OF LINEAR GROUPS
25
Claim: Let γ1 = β, γ2 = −α, and γ3 = −(α + β). For any t ∈ F and any v ∈ V written as
in (6),
t[eγ3 , [eγ2 , [eγ1 , v] − s−γ1 eγ1 ]] = ±tsα e−α ∈ V.
Proof. Fix t ∈ F, v ∈ V . Set v1 = [eγ1 , v] − s−γ1 eγ1 , which is in V by (7). Since
−(γ1 + γ2 ) = α − β 6∈ Φ,
the coefficient of e−γ2 in v1 is 0, and thus v2 = [eγ2 , v1 ] = [eγ2 , [eγ1 , v]] ∈ V by (7). In addition,
v2 ∈ ES ⊕ EL . Similarly, the coefficient of e−γ3 in v2 is 0 since −(γ1 + γ2 + γ3 ) = 2α 6∈ Φ, so
v3 = t[eγ3 , v2 ] ∈ V by (7) and v3 ∈ ES ⊕ EL .
For any γ ∈ Φ,
γ + γ1 + γ2 + γ3 = γ − 2α,
and γ − 2α ∈ Φ if and only if γ = α. We also have γ2 + γ3 = β − 2α 6∈ Φ and γ1 + γ2 + γ3 =
−2α 6∈ Φ, so in fact v3 = ±tsα e−α as claimed.
The Weyl group W ≤ G(F) of Φ acts transitively on {eγ : γ long}, so since Fe−α ⊆ V and
α is a long root, we conclude that EL ⊆ V . If Φ is simply laced, this immediately implies
V = g(F), the desired contradiction.
If Φ is of type Bl , l ≥ 3 or F4 , then for t ∈ F,
xǫ1 (t) · eǫ2 −ǫ1 − eǫ2 −ǫ1 = ±teǫ2 ± t2 eǫ1 +ǫ2 ∈ V.
Since ǫ1 + ǫ2 is a long root and EL ⊆ V , we have Feǫ2 ⊆ V . By the transitive action of W on
{eγ : γ short}, ES ⊆ V and hence V = g(F).
If Φ is of type G2 , then for t ∈ F,
x−αS −αL (t) · eαL − eαL = ±te−αS ,
so Fe−αS ⊆ V . Hence ES ⊆ V and V = g(F).
Case 3: p = 2, Φ of type Cl , l ≥ 2. We again assume FV = g(F) for the sake of contradiction.
Let γ1 = 2ǫ2 and γ2 = −2ǫ1 . These are long roots with γ1 + γ2 6∈ Φ and γ + γ1 + γ2 ∈ Φ if
and only if γ = ǫ1 − ǫ2 . Then by the same reasoning as in the argument for Case 2, if t ∈ F
and v ∈ V is written as in (6), we have
t[eγ2 , [eγ1 , v] − s−γ1 eγ1 ] = ±tsǫ1 −ǫ2 eǫ2 −ǫ1 ∈ V.
Since FV = g(F), sǫ1 −ǫ2 6= 0 for some v ∈ V , so Feǫ2 −ǫ1 ⊆ V and hence ES ⊆ V .
To show
PEL ⊆ V , let v ∈ V with s2ǫ2 6= 0. Since ES ⊆ V , we can assume v is of the form
v = h + α long sα eα . The only long roots α satisfying ǫ1 − ǫ2 + α ∈ Φ are α = −2ǫ1 and
α = 2ǫ2 , and
xǫ1 −ǫ2 (1) · e−2ǫ1 − e−2ǫ1 = ±e−ǫ1 −ǫ2 ± e2ǫ1 ,
Therefore
xǫ1 −ǫ2 (1) · e2ǫ2 − e2ǫ2 = ±eǫ1 +ǫ2 ± e−2ǫ2 .
xǫ1 −ǫ2 (1) · v − v = seǫ1 −ǫ2 ± s−2ǫ1 e−ǫ1 −ǫ2 ± s−2ǫ1 e2ǫ1 ± s2ǫ2 eǫ1 +ǫ2 ± s2ǫ2 e−2ǫ2 ∈ V,
where [eǫ1 −ǫ2 , h] = seǫ1 −ǫ2 for some s ∈ F. Again using the fact that ES ⊆ V , we conclude
that
v1 = ±s2ǫ2 e2ǫ1 ± s−2ǫ1 e−2ǫ2 ∈ V.
Then if t ∈ F,
xe−ǫ1 −ǫ2 (t) · v1 − v1 = ±ts2ǫ2 eǫ1 −ǫ2 ± t2 s2ǫ2 e−2ǫ2 ∈ V,
26
DANIEL FRANZ
so t2 s2ǫ2 e−2ǫ2 ∈ V. But F is a finite field with characteristic 2, so F2 = F, and thus we conclude
that EL ⊆ V and hence V = g(F), a contradiction.
References
[1] E. Abe and K. Suzuki, On normal subgroups of Chevalley groups over commutative rings, Tohoku Math.
J. 28 (1976) no. 1, 185-198.
[2] Y. Barnea and R. Guralnick, Subgroup growth in some pro-p groups, Proceedings of the AMS. 130
(2001), 653-659.
[3] K. Bou-Rabee, Quantifying residual finiteness, J. of Algebra 323 (2010), 729-737.
[4] K. Bou-Rabee, Approximating a group by its solvable quotients, N.Y.J. of Math 17 (2011), 699-712.
[5] Bou-Rabee, Hagen, Patel, Residual finiteness growths of virtually special groups, Math. Z. 279 (2015)
no. 1-2, 297-310.
[6] K. Bou-Rabee and T. Kaletha, Quantifying residual finiteness of arithmetic groups, Compos. Math. 148
(2012), 907-920.
[7] K. Bou-Rabee and D.B. McReynolds, Asymptotic growth and least common multiples in groups, Bull.
Lond. Math. Soc. 43 (2011), 1059-1068.
[8] K. Bou-Rabee and D.B. McReynolds, Extremal behavior of divisibility functions, Geometriae Dedicata,
175 (2015), 407-415.
[9] N. Buskin, Economical separability in free groups, Sib. Math. J. 50 (2009) no. 4, 603-608.
[10] R. Carter, Simple Groups of Lie Type, Pure Appl. Math., vol 28, Wiley, London (1972).
[11] G. M. D. Hogeweij, Almost classical Lie algebras: I, II, Indag. Math. 44 (1982) no. 4, 441-460.
[12] J. Humphreys, Introduction to Lie Algebras and Representation Theory, Springer, New York (1972).
[13] M. Kassabov and F. Matucci, Bounding the residual finiteness of free groups, Proc. Am. Math. Soc. 139
(2011), 2281-2286.
[14] P. Kleidman and M. Liebeck, The subgroup structure of the finite classical groups, Cambridge University
Press, 1990.
[15] G. Kozma and A. Thom, Divisibility and laws in finite simple groups, Math. Ann. 364 (2016) no. 1-2,
79-95.
[16] Lubotzky, Mozes, Raghunathan, The word and Riemannian metrics of semisimple groups, Publ. Math.
Inst. Hautes Etudes Sci. 91 (2000), 5-53.
[17] A. Lubotzky and D. Segal, Subgroup Growth, Progress in Mathematics, 212. Birkhauser Verlag, Basel,
2003.
[18] V. Platonov and A. Rapinchuk, Algebraic groups and number theory. Translated from the 1991 Russian
original by Rachel Rowen. Pure and Applied Mathematics, 139. Academic Press, Inc., Boston, MA,
1994.
[19] S. Roman, Field Theory, Springer-Verlag, 1995.
[20] J.-P. Serre, Lectures on Nx (p), Research Notes in Mathematics 11, CRC Press, 2012.
[21] R. Steinberg, Lectures on Chevalley groups, Yale University, 1968.
[22] A. Thom, About the length of laws for finite groups, http://arxiv.org/abs/1508.07730.
[23] V. Murty and J. Scherk, Effective versions of the Chebotarev density theorem for function fields, C. R.
Acad. Sci. Paris Sér. I Math. 319 (1994), 523-528.
[24] A. V. Vasilyev, Minimal permutation representations of finite exceptional groups of types G2 and F4 ,
Algebra and Logic. 35 (1996) no. 6, 371-383.
[25] A. V. Vasilyev, Minimal permutation representations of finite exceptional groups of types E6 , E7 , and
E8 , Algebra and Logic. 36 (1997) no. 5, 302-310.
| 4math.GR
|
SUMS OF TWISTED CIRCULANTS
arXiv:1607.05716v1 [math.PR] 19 Jul 2016
AARON ABRAMS, HENRY LANDAU, ZEPH LANDAU, AND JAMIE POMMERSHEIM
Abstract. The rate of convergence of simple random walk on the Heisenberg group over Z/nZ
with a standard generating set was determined by Bump et al [1, 2] to be O(n2 ). We extend this
result to random walks on the same groups with an arbitrary minimal symmetric generating set.
We also determine the rate of convergence of simple random walk on higher-dimensional versions
of the Heisenberg group with a standard generating set. We obtain our results via Fourier analysis,
using an eigenvalue bound for sums of twisted circulant matrices. The key tool is a generalization
of a version of the Heisenberg Uncertainty Principle due to Donoho-Stark [4].
1. Introduction
In this paper we establish upper bounds on the norms of certain sums of Hermitian matrices.
Our interest in this result is twofold. First, using standard Fourier analytic techniques developed by
Diaconis [3], our bounds enable an analysis of rates of convergence for a family of random walks on
various finite Heisenberg groups, building on work of Bump et al. [1]. Second, we obtain our bound
using a type of “uncertainty principle” that may be useful in analyzing other related problems.
1.1. Heisenberg groups. The Heisenberg group
form
1 x
0 1
0 0
over a ring R is the set of 3 × 3 matrices of the
z
y ,
1
with x, y, z ∈ R, under the usual operation of matrix multiplication. In this paper we take R = Z/nZ
for an odd prime integer n. Letting (x, y, z) denote the above matrix, the resulting Heisenberg group
H(n) is generated by X = (1, 0, 0) and Y = (0, 1, 0), since XY X −1 Y −1 = (0, 0, 1). This group has
order n3 and center {(0, 0, z)} of size n.
The stipulations that n be odd and prime are mainly for convenience.
The authors of [1] establish the rate of convergence of random walk on H(n) with steps taken uniformly at random from {X ±1 , Y ±1 }. Their approach is to use the standard representation-theoretic
techniques of [3], though there are difficulties. The group H(n) has n2 one-dimensional representations and n − 1 irreducible representations of dimension n. To apply the standard machinery what is
needed are good bounds on the eigenvalues of the average of the images of the generators under the
n-dimensional representations of H(n). This leads to the study (in [1]) of the matrices M (r) defined
as follows. Let S be the n × n “shift” matrix which acts on the standard basis by Sei = e(i−1) mod n
Date: July 17, 2016.
1
2
ABRAMS, LANDAU, LANDAU, AND POMMERSHEIM
and for r = 1, . . . , n − 1, let D(r) be the diagonal matrix whose jth entry is 2 cos(2πrj/n). Then
for each r the matrix
M (r) =
1
(S + S −1 + D(r))
4
is the average of the images of X ±1 and Y ±1 in one of the n-dimensional irreducible representations of
H(n). The papers [1] and [2] present several different proofs that the norm of M (r) is bounded above
by 1 − O( n1 ). Once this is done, a straightforward analysis of the one-dimensional representations
of H(n) reveals that those eigenvalues can be as large as 1 − O( n12 ), so the bound on M (r) shows
that the behavior of random walk is governed by the one-dimensional representations and a mixing
time of O(n2 ) is established.
In this paper we generalize the first approach utilized in [2] in two different ways. Our main
contribution is the analysis of simple random walks on H(n) with different minimal, symmetric
generating sets. For s1 , s2 , r1 , r2 ∈ Z/nZ, consider random walk on H(n) with steps taken uniformly
from
G = {(s1 , r1 , 0), (−s1 , −r1 , 0), (s2 , r2 , 0), (−s2 , −r2 , 0)}.
Note that G generates H(n) if and only if r1 s2 6≡ r2 s1 mod n. In the case that G generates, the
matrices analogous to M (r) that arise in the representation theory are averages of matrices that
we call twisted circulants. We will define these matrices shortly. The main aim of this paper is to
establish the same bound 1 − O( n1 ) on the eigenvalues of these averages of twisted circulants. The
same reasoning then yields mixing times O(n2 ) for random walk on H(n) with generating set G.
Our approach, following [2], is to generalize a version of the Heisenberg Uncertainty Principle due
to Donoho and Stark [4]. We discuss this further in Section 1.3.
Another generalization of the main result of [2] applies to simple random walk on higher-dimensional
Heisenberg groups. The d-dimensional Heisenberg group consists of upper triangular (d + 2) × (d + 2)
matrices with ones on the diagonal and zeroes everywhere else except the top row and the rightmost
column. With entries from Z/nZ, this is a nilpotent group of order n2d+1 and class d + 1.
Using a tensor product decomposition of the representations of these groups, we show that the
rate of convergence of simple random walk on these groups with a standard set of generators can be
determined easily from the lower-dimensional case. The result is Corollary 4.1: with the standard
generators, dn2 steps are necessary and sufficient for convergence of simple random walk on the
d-dimensional Heisenberg group over Z/nZ. This works easily because the tensor product decomposition is naturally compatible with the standard basis. Probably the same rate occurs with other
generating sets, though we do not carry out this analysis.
1.2. Main theorem. We begin with the definition of the twisted circulants. We will index the rows
and colums of all matrices starting at 0. Let S denote the n × n matrix with ij entry 1 if i + 1 = j
mod n and 0 otherwise. Thus S enacts the shift operator on the standard basis of Cn . A circulant
is any non-negative power of S. If D is a diagonal matrix with entries di and C = S s is a circulant
then the matrix DC has the entries di on the sth (cyclic) diagonal above the main:
SUMS OF TWISTED CIRCULANTS
3
d0
dn−s
d1
..
.
..
.
dn−s−1
..
.
dn−1
Recall that n is a fixed odd prime. Let ω = e2πi/n and let R be the diagonal matrix with
Rj,j = ω j . Note that
SR = ωRS.
A twisted circulant is a unitary matrix of the form
A(r, s) := Rr S s ,
and these have Hermitian counterparts
M (r, s) :=
1 r s
1
(R S + (Rr S s )∗ ) = (A(r, s) + ω rs A(−r, −s))
2
2
with r, s ∈ Zn .
The matrices A(r, s) are scalar multiples of the matrices in the images of the n-dimensional
representations of H(n). Note that the matrix M (r) that arose in [1] is equal to 21 (M (r, 0)+M (0, 1)).
Our main result (proved in Section 2.3) is this:
Theorem 1.1. If r1 s2 6≡ r2 s1 mod n then the norm of
O( n1 ).
1
2
(M (r1 , s1 ) + M (r2 , s2 )) is at most 1 −
If r1 s2 ≡ r2 s1 mod n then the operators M (r1 , s1 ) and M (r2 , s2 ) commute, and the analysis is
different; we discuss this case in Section 3.
Corollary 1.2. If r1 s2 6≡ r2 s1 mod n then n2 steps are necessary and sufficient for random walk
on H(n) with generating set {(s1 , r1 , 0), (−s1 , −r1 , 0), (s2 , r2 , 0), (−s2 , −r2 , 0)} to become uniform.
The next observation can be used in the same way to determine mixing times for random walks
with larger sets of generators.
P
Corollary 1.3. Let ri , si ∈ Zn for i = 1, . . . , d and consider the matrix M = d1
M (ri , si ). If there
are 2k disjoint pairs of integers in {1, 2, . . . , d} such that ri sj 6≡ rj si mod n for each pair i, j, then
2k
the norm of M is at most 1 − dn
. In particular if a constant fraction of the indices can be paired in
this way then the norm of M is at most 1 − O( n1 ).
Proof. The matrix M is the average of k matrices of norm at most 2 −
norm 1.
2
n
and d − 2k matrices of
4
ABRAMS, LANDAU, LANDAU, AND POMMERSHEIM
1.3. Remarks on uncertainty and Gauss sums. Theorem 1.1 can be framed in the context of
the following general question: given two Hermitian matrices A and B what can one say about the
norm of the sum A + B? The triangle inequality yields an upper bound of ||A|| + ||B|| on the largest
eigenvalue of A+B, and of course this bound is sometimes achieved e.g. when the leading eigenvectors
of A and B coincide. Horn’s inequalities generalize this observation but due to their generality do
not address the “typical” situation. Specifically, suppose the diagonalizing bases are “well-mixed”
with respect to each other, meaning roughly that the eigenspaces for the leading eigenvalues of A
and B are very far from aligning. This is meant to suggest that the two eigenbases should satisfy a
kind of uncertainty principle: any vector well-concentrated in the eigenbasis of A should be spread
out in the eigenbasis of B, and vice versa. Applying this to the leading eigenvector of A + B gives
a bound on its eigenvalue as a weighted average of the largest parts of the spectra of A and B. See
Section 2.1 for precise statements; in the context of twisted circulants, this yields Theorem 1.1.
This is very similar to the uncertainty principle of Donoho-Stark [4]. In their case the eigenspaces
for A and B are related by the Fourier Transform, but this is not essential. Their beautiful and
simple proof of the uncertainty principle relies solely on the fact that the individual matrix entries
of the Fourier Transform are of size O( √1n ). In fact this is exactly what we mean by “well-mixed.”
Thus our uncertainty principle, Lemma 2.1, is just a slight generalization of the result from [4].
For the particular twisted circulants we are interested in, when we computed the change of basis
matrix (see Section 2.2) we were delightfully surprised by the appearance of Gauss sums in our
calculation. These enabled us to establish that the entries have norm exactly √1n , and hence the
uncertainty principle applies.
2. Proof of main theorem
2.1. Uncertainty Principle. For a set S ⊆ {0, . . . , n−1} and x ∈ Cn define xS to be the projection
of x onto the coordinates of S: (xS )j = xj for j ∈ S and (xS )j = 0 for j otherwise. Let S̄ denote
the complement {0, . . . , n − 1} \ S.
Lemma 2.1 (Uncertainty principle). Suppose U is a unitary matrix and c a positive real number
such that |Ui,j | ≤ √cn for all 1 ≤ i, j ≤ n. Then for any sets S, T ⊆ {0, . . . , n − 1} and any v ∈ Cn
of unit norm,
q
2
2
|S||T | 1 − k(U v)T̄ k − kvS̄ k
≥
.
n
ckvS k
Proof. The bound |Ui,j | ≤ √cn implies that the maximum absolute value for a coefficient of U vS is
|S| √cn √1 kvS k which implies that
|S|
(1)
ckvS k p
k(U vS )T k ≤ √
|S||T |.
n
p
2
2
k
√S
Now 1 − k(U v)T̄ k = k(U v)T k ≤ (k(U vS )T k + k(U vS̄ )T k)2 ≤ ( ckv
|S||T | + kvS̄ k)2 where we’ve
n
used (1) on the first term and the fact that U is unitary on the second term. Rearranging terms
gives the result.
SUMS OF TWISTED CIRCULANTS
5
Corollary 2.2. Let U be a unitary matrix with |Ui,j | ≤ √cn for all 1 ≤ i, j ≤ n. For any v ∈ Cn of
unit norm and sets S, T ⊆ {0, . . . , n − 1} with |S||T | ≤ 2cn2 , we have max(kvS̄ k, k(U v)T̄ k) ≥ 15 .
Proof. Since kvS k ≤ 1, the assumption on the size of |S||T | in Lemma 2.1 yields the inequality:
q
1
2
( 1 − k(U v)T̄ k − kvS̄ k)2 ≤ .
2
1
It can easily be verified that max(kvS̄ k, k(U v)T̄ k) ≥ 5 is a necessary condition for the inequality to
hold.
2.2. The eigenstructure of DC. Given an arbitrary matrix M , we’ll say a unitary U diagonalizes
M if U ∗ M U is a diagonal matrix. Note that U ei (ei being the ith coordinate basis vector) is the
eigenvector for M corresponding to the eigenvalue located on the ith diagonal element of U ∗ M U .
(Not all matrices have such a unitary but self-adjoint ones and unitaries do.)
Let F be the n × n Fourier transform matrix defined by Fk,l = √1n ω kl where ω = e2πi/n . Also,
define the permutation matrix Πs by (Πs )si,i = 1 for 0 ≤ i ≤ n − 1 with the remaining entries 0;
thus Πs (x0 x1 · · · xn−1 )∗ = (x0 xs · · · xs(n−1) )∗ (with indices taken mod n).
Lemma 2.3. Let C = S s be a circulant and D be diagonal with entries ai of unit norm. Let
α = a0 a1 · · · an−1 and let B be the diagonal matrix with entries
λk
Bk,k = Qk−10
l=0
asl
,
where λ0 is any fixed nth root of α. Then the unitary matrix Πs BF diagonalizes DC. The (j, j)
entry of the resulting diagonal matrix (Πs BF )∗ DC(Πs BF ) is ω j λ0 .
Proof. Define λ0 to be an nth root of α so that λ0 , ωλ0 , ω 2 λ0 , . . . ω n−1 λ0 are all the nth roots of α.
It is straightforward to verify that ω j λ0 is an eigenvalue for DC with eigenvector xj = Πs vj where
the kth coordinate of vj is given by
(2)
ω jk
λk
(vj )k = √ Qk−10
.
n l=0 asl
Define X (respectively V ) to be the matrix whose jth column is xj , (respectively vj ), so that X is
a unitary that diagonalizes DC and X = Πs V . From (2), since the second factor is independent of
j, we see V = BF with B defined in the statement of the lemma.
Given r, s, define X(r, s) = Πs BF where B is a diagonal matrix with Bk,k = ω −rs
k(k−1)
2
.
Corollary 2.4. The matrix X(r, s) is unitary, diagonalizes Rr S s , and has all entries of norm
√1 .
n
Lemma 2.5. Given nonzero elements r1 , r2 , s1 , s2 of Zn such that r1 s2 6= r2 s1 , each entry of the
matrix (X(r1 , s1 ))∗ X(r2 , s2 ) has norm √1n .
Proof. (X(r1 , s1 ))∗ X(r2 , s2 ) = F ∗ B1∗ Π∗s1 Πs2 B2∗ F = F ∗ B1∗ Πs−1 s2 B2∗ F . We compute that the (c, d)
1
entry of this product is
X
−1
−1
−1
[(X(r1 , s1 ))∗ X(r2 , s2 )]c,d =
ω −cj ω r1 s1 j(j−1)/2 ω −r2 s2 s1 s2 j(s1 s2 j−1)/2 ω s1 s2 jd .
j
6
ABRAMS, LANDAU, LANDAU, AND POMMERSHEIM
Let Ω be the nth root of unity satisfying ω = Ω2 . (We use here that n is odd.) Then the above sum
becomes
X
2
Ωαj +βj
j
−1
2
where α = r1 s1 − r2 s−1
2 s1 and β = −2c − r1 s1 + r2 s1 + 2s1 s2 d. As long as α 6= 0, i.e. r1 s2 6= r2 s1 ,
√
this is a Gauss sum whose norm is n.
In the case that r1 s2 = r2 s1 , we get α = 0 and the sum in the proof of the lemma is not a Gauss
sum. We discuss this situation in Section 3.
2.3. Putting it together. We now prove our main results. Lemma 2.5 does not apply in the case
s1 = 0, but here the analysis is easier and in fact doesn’t require that the second operator have
entries coming from R. A special case is M = 12 (M (r1 , 0) + M (r1 , s1 )).
Theorem 2.6. Set M = 12 M (r1 , 0) + 21 (DC + (DC)∗ ) with C = S s and D any diagonal matrix
with entries of unit norm. Then we have kM k ≤ 1 − O( n1 ).
Proof. Let v be a maximal eigenvalue for the self-adjoint operator M so that kM k = hM v, vi ≤
|hM1 v, vi| + |hDCv, vi| + 1 = |hM1 v, vi| + |hD0 U v, U vi| + 1, where M1 is shorthand for the matrix
M (r1 , 0) and where the matrix U diagonalizes the unitary matrix DC with resulting diagonal matrix
D0 , i.e. D0 = U ∗ (DC)U . By Lemma 2.3, U = Πs BF and therefore both U and U ∗ have all entries of
norm √1n and therefore satisfy the conditions for the uncertainty principle (Corollary 2.2). Choosing
√
√
S = [1, n/2] we have max(kvS̄ k, k(U ∗ v)T̄ k) ≥ 15 for any set T with |T | ≤ n. If this maximum
is achieved by the first term, the fact that all the eigenvalues of D on S̄ are bounded by 2 − O( n1 )
gives the result. We are left with showing the result under the assumption that k(U ∗ v)T̄ k ≥ 15 .
We note that by the proof of Lemma 2.3 the eigenvalues of U ∗ are spread evenly around the unit
circle. If we let α be the unit vector in the direction of hD0 U ∗ v, U ∗ vi and set fi = h(D0 )i , αi we
P
√
have |hD0 U ∗ v, U ∗ vi| = i fi |vi |2 . Choosing T to be the locations i for the top n values of fi (in
absolute value) and noting that the remaining values of fi must have absolute value below 1 − O( n1 )
yields the result.
Theorem 1.1. Let M = 12 (M (r1 , s1 ) + M (r2 , s2 )) with r1 , s1 , r2 , s2 nonzero elements of Zn . If
r1 s2 6≡ r2 s1 mod n then kM k ≤ 1 − O( n1 ).
Proof. The argument has lots of similarities to the previous result. For shorthand write Ai =
A(ri , si ) = Rri S si (for i = 1, 2) so that M (ri , si ) = Ai + A∗i . Let v be a maximal eigenvalue for
the self-adjoint operator M so that kM k = hM v, vi ≤ |hA1 v, vi| + |hA2 v, vi| + 2 = |hB1 U1 v, U1 vi| +
|hB2 U2 v, U2 vi| + 2 where the unitary matrix Ui = X(ri , si ) diagonalize Ai with resulting diagonal
matrices Bi , i.e., Bi = Ui Ai Ui∗ for i = 1, 2. We write w = U1 v and we consider the resulting
quantities |hB1 w, wi| and |hB2 (U2 U1∗ w), U2 U1∗ wi|. Lemma 2.5 establishes that U2 U1∗ has all entries
bounded by √1n and therefore satisfies the uncertainty principle. The remainder of the proof mirrors
that of the previous theorem.
SUMS OF TWISTED CIRCULANTS
7
3. Equal slopes
The hypothesis that r1 s2 6= r2 s1 in Zn is required for the proof of Lemma 2.5. If r1 s2 = r2 s1 then
the sum in the proof of that lemma is not a Gauss sum, because α = 0. For each c there is a single
value of d that gives a sum of 1, and the rest give zeroes. The resulting matrix is a permutation
matrix.
k(k−1)
From (1.2) it follows that (A(r, s))k = ω 2 rs A(kr, ks). Thus A(r, s) and ωI generate an abelian
group of matrices including A(kr, ks) for all k. So r1 s2 = r2 s1 implies that M (r1 , s1 ) and M (r2 , s2 )
commute and generate a group (usually) isomorphic to Z2n .
In this case, the matrix M = 21 (M (r1 , s1 ) + M (r2 , s2 )) will have norm close to 1 for some choices
of the parameters, and will not have norm close to 1 for other choices of parameters. Specifically,
we note that the eigenvalues of M are given as follows. Let k = r1−1 r2 = s−1
1 s2 , with the inverses
taken modulo n. Then the eigenvalues of M are given by:
!
2π( −k(k−1)
r1 s1 + kd)
2πd
1
2
cos
+ cos
, d = 0, . . . , p − 1.
λd =
2
n
n
We see that for a given n, these eigenvalues only depend on k and the product r2 s2 . For n = 401,
Figure 1 shows the values of these parameters for which the matrix M has norm greater than
1 − cos( 2π
n ). In the graph, the horizontal axis gives the value of r2 s2 while the vertical axis gives the
value of kr2 s2 .
On the other hand, one finds computationally, again for n = 401, that the norm of M is less than
1 − n1 for approximately half of the choices of the parameters.
4. Higher dimensional Heisenberg groups
An analysis very similar to what we have already done can be carried out for the higher dimensional Heisenberg groups defined in Section 1.2. We denote this group by H(p, d) where p = n is
still an odd prime. As p and d will not change we also refer to this simply as H. It is a p-group of
order p2d+1 and nilpotency class d + 1.
Here we analyze random walk on H. We briefly describe the representation theory of H, as
communicated to us by Persi Diaconis. There are p2d one-dimensional irreducible representations
and p − 1 irreducible representations of dimension pd . These latter representations are described as
follows. We view elements of H as triples (x, y, z), where x, y ∈ Zdp , and z ∈ Zp . Let q = exp 2πi
p
and let V be the vector space of all complex-valued functions on (Zp )d . Then for each 0 6= c ∈ Zp ,
there is an irreducible representation ρc of H on V given by
[ρc (x, y, z)f ](w) = q c(y·w+z) f (w + x).
The key to understanding these representations, and the random walk, is a tensor product decomposition. Let W denote the vector space of complex-valued functions on Zp , and note that the
V is naturally isomorphic to W ⊗d . Define operators S and R on W by
[Sg](u) = g(u + 1),
[Rg](u) = q u g(u),
>
8
ABRAMS, LANDAU, LANDAU, AND POMMERSHEIM
400
300
200
100
100
200
300
Figure 1. A mark indicates that M has eigenvalues greater than 1 −
given choices of r1 s1 and k. Here n = 401.
400
1
n
for the
where S and R are the matrices defined in Section 1.2. Then one easily verifies that the operator
ρc (x, y, z) decomposes as
ρc (x, y, z) = q cz Rcy1 S x1 ⊗ · · · ⊗ Rcyd S xd .
We are interested in bounding the top eigenvalue of the average of ρc on the size 4d generating
set consisting of the ei = (wi , 0, 0) and the fi = (0, wi , 0) and their negatives, where wi denotes the
SUMS OF TWISTED CIRCULANTS
9
ith vector in the standard basis of Zdp . We see that
ρc (ei ) = I ⊗ · · · I ⊗ S ⊗ I · · · ⊗ I,
with the S appearing in the ith tensor factor, and
ρc (fi ) = I ⊗ · · · I ⊗ Rc ⊗ I · · · ⊗ I.
So for fixed i the average 41 (ρc (ei ) + ρc (−ei ) + ρc (fi ) + ρc (−fi )) is precisely the sum 21 M (0, 1) +
1
1
2 M (c, 0), tensored with a bunch of identity matrices. Recall from Section 1.1 that 2 M (0, 1) +
1
2 M (c, 0) is the same as the matrix M (c) studied in [1]. Thus, by the results of [1] or [2], or by
Theorem 2.6, the top eigenvalue of such an operator is at most 1 − O(1/p). Averaging these over i
again produces an operator with top eigenvalue at most 1 − O(1/p).
By contrast, the one-dimensional representations of H = H(p, d) map each of the 4d generators
{±ei , ±fi } to a pth root of unity, so the largest value for the average of these 4d numbers arises
when all but one have value 1 and the other is exp(2πi/p). This comes out to something larger than
1 − O( dp12 ), which we have shown is much larger than the contribution from the high-dimensional
representations. Thus again in this case the rate of convergence is governed by the one-dimensional
representations.
Corollary 4.1. For simple random walk on H(p, d) with steps {±wi , 0, 0), (0, ±wi , 0)} each chosen
with the same probability 1/4d, the mixing time is O(dp2 ).
References
[1] Bump, Diaconis, Hicks, Miclo, Widom. An exercise (?) in Fourier analysis on the Heisenberg group. Ann Fac Sci
Toulouse Math, to appear. Available at http://arxiv.org/abs/1502.04160 .
[2] Bump, Diaconis, Hicks, Miclo, Widom. Useful bounds on the extreme eigenvalues and vectors of matrices for
Harper’s operators. Journal of Operator Theory, to appear.
[3] Persi Diaconis. Group Representations in Probability and Statistics, volume 11 of Institute of Mathematical Statistics Lecture Notes — Monograph Series, Hayward, CA, 1988.
[4] David L. Donoho and Philip B. Stark. Uncertainty principles and signal recovery. SIAM J. Appl. Math.,
49(3):906?931, 1989.
(Aaron Abrams) Washington and Lee University
E-mail address: [email protected]
(Henry Landau) AT&T Research
E-mail address: [email protected]
(Zeph Landau) University of California, Berkeley
E-mail address: [email protected]
(Jamie Pommersheim) Reed College
E-mail address: [email protected]
| 4math.GR
|
Parameter Estimation for a partially observed Ornstein-Uhlenbeck process
with long-memory noise
arXiv:1501.04972v2 [math.PR] 12 Oct 2016
Brahim El Onsy1 Khalifa Es-Sebaiy2 Frederi G. Viens
Cadi Ayyad University and Purdue University
3
Abstract: We consider the parameter estimation problem for the Ornstein-Uhlenbeck process X driven by a fractional Ornstein-Uhlenbeck process V , i.e. the pair of processes defined
by the non-Markovian continuous-time long-memory dynamics dXt = −θXt dt + dVt ; t > 0,
with dVt = −ρVt dt + dBtH ; t > 0, where θ > 0 and ρ > 0 are unknown parameters, and
B H is a fractional Brownian motion of Hurst index H ∈ ( 21 , 1). We study the strong
con
sistency as well as the asymptotic normality of the joint least squares estimator θbT , ρbT
of the pair (θ, ρ), based either on continuous or discrete observations of {Xs ; s ∈ [0, T ]} as
the horizon T increases to +∞. Both cases qualify formally as partial-observation questions
since V is unobserved. In the latter case, several discretization options are considered. Our
proofs of asymptotic normality based on discrete data, rely on increasingly strict restrictions
on the sampling frequency as one reduces the extent of sources of observation. The strategy
for proving the asymptotic properties is to study the case of continuous-time observations
using the Malliavin calculus, and then to exploit the fact that each discrete-data estimator
can be considered as a perturbation of the continuous one in a mathematically precise way,
despite the fact that the implementation of the discrete-time estimators is distant from the
continuous estimator. In this sense, we contend that the continuous-time estimator cannot
be implemented in practice in any naïve way, and serves only as a mathematical tool in the
study of the discrete-time estimators’ asymptotics.
Key words: Least squares estimator; fractional Ornstein Uhlenbeck process; Multiple integral; Malliavin calculus; Central limit theorem.
2010 Mathematics Subject Classification: 60F05; 60G15; 60H05; 60H07.
1
1.1
Introduction
Context and background
The question of drift parameter estimation for solutions of stochastic differential equations
driven by fractional Brownian noise goes back at least as far as the seminar work of Kleptsyna
and Le Breton in [17] (also see prior references therein), where a maximum likelihood estimator
(MLE) was proposed. This work was a first genuine attempt to show how to compute the
MLE in practice in the regular case (self-similarity ‘Hurst’ parameter H > 1/2), by relying on
continuous-time observation of a single path over a finite time interval, with strong consistency
and asymptotic normality as the horizon increases to infinity. This work was itself motivated
by – and is an extension of – now classical ideas of how to use the Girsanov theorem to compute
1
Faculty of Sciences and Techniques - Marrakech, Cadi Ayyad University, Morocco.
E-mail:
[email protected]
2
National School of Applied Sciences - Marrakesh, Cadi Ayyad University, Morocco.
E-mail:
[email protected]
3
Dept. Statistics and Dept. Mathematics, Purdue University, 150 N. University St., West Lafayette, IN
47907-2067, USA. E-mail: [email protected]
1
the MLE in the case of white noise, as presented in the 1978 treatise [19]. The continuoustime data was also invoked in [17] to justify that any diffusion-type parameter (any constant
multiplicative term in front of the equation’s driving noise) would then be directly observable.
That observation remains generally true in many contexts for continuous-time data, including
when the noise is fractional Brownian. It explains why so many authors since [17] have
continued to study the estimation of drift parameters for fractional-noise-driven problems.
We have listed some of these works below. Our paper inscribes itself in this line of work, and
can be viewed as a study of partial observation questions, as we will explain shortly.
Generally speaking, this effort to understand drift estimation for ergodic diffusions, even
Gaussian ones, is of fundamental importance for any quantitative study where mean-reverting
quantities are believed to be asymptotically stationary, and are either observed with noise or
are intrinsically stochastic. The best known class of examples, which also encapsulates the
question of whether or not the stochastic process of interest is observed directly or indirectly,
is that of stochastic volatility in quantitative finance. Some of the original ideas on how to
estimate this volatility’s drift parameters is given in the 2000 research monograph [12]. Similar
models with fractional noise were introduced in the late 1990’s, as in [10] for continuous time,
but to our knowledge, their statistical estimation was left unexplored for more than ten years.
The broadest question, applicable in finance and other fields, is to estimate jointly the drift,
diffusion, and memory parameters for fractional-noise-driven equations; it is typically nontrivial, and with the exception of one study in [4], largely unresolved by bona fide statistical
means for non-self-similar discretely observed continuous-time processes. We will not address
this issue here. In the case of partially observed data, we refer the interested reader to a study
in the context of high-frequency financial data, where a sequential Bayesian methodology is
combined with classical estimation techniques and a calibration method to find H: see [9].
We also mention a method described in [27] for a similar model with partial observation,
where it is shown that the minimax-sense optimal estimator has a very slow convergence rate,
and relies on high-frequency data; it can be argued that when applied to bona fide financial
data, this estimator cannot be implemented without leaving the realm where continuous-time
semi-martingale models with long-memory volatility are appropriate.
The present paper looks specifically at a difficulty which arises when one single path is
used to estimate more than one real drift-type parameter. Questions of identifiability can arise
in this context (see for example the treatment of the Generalized Method of Moments (GMM)
as described in [21]). Once such a question has been resolved, the main practical obstacle to
implementation is typically that of discretizing the data given by the path, i.e. using only
observations in discrete time. For the sake of conciseness, we look at a specific situation
where identifiability is resolved explicitly and in a natural way, within the task of deriving the
continuous-time and discrete-time estimators, without having to rely on abstract conditions
which would provide a priori identifiability of the vector of parameters. Our objective is to
describe conditions under which strongly consistent and asymptotically normal estimation
can be established quantitatively, based on discretely-observed data alone. Since access to
and analysis of continuous-time data is not typically a realistic assumption, we will view any
estimators based on continuous-time data as tools in the task of deriving strong consistency
and asymptotic normality for the discrete-time estimators. Our strategy is to attempt to
discretize the former, which includes a need to approximate Riemann and stochastic integrals.
In the process, we discover some fundamental differences between the discrete and continuous
time estimators. In particular, we find that while least squares (LS) estimation appears to
be the best tool based on continuous-time observation of a path, when converted to discrete
2
time data, the estimator’s interpretation as a LS optimizer is lost, and a GMM interpretation
seems more appropriate. This introduction, including our summary of results in Section 1.2,
contains specific details supporting these ideas.
We are largely motivated by the paper [6], which studies the drift-estimation problem
for the Ornstein-Uhlenbeck process driven itself by another, unobserved Ornstein-Uhlenbeck
process (OU-OU). Their work only deals with an underlying white noise. Specifically, let W
be a standard Brownian motion and let θ and ρ be positive real parameters. The OU-OU
process is the solution of the following system
X0 = 0; dXt = −θXt dt + dVt , t > 0;
(1)
V0 = 0; dVt = −ρVt dt + dWt , t > 0.
Here one may consider that X is observed, while V is not, or conversely, or that both processes
are observed; which interpretation is used makes a crucial difference, as we will see.
Since the quadratic variation of V is t, the classical Girsanov theorem implies that a natural
candidate to estimate θ is the MLE (recall that this idea was already contained in [19]), which
can be easily computed for this model: one gets
RT
− 0 Xt δXt
b
(2)
θT =
RT 2
X
dt
t
0
=
−XT2 + T
,
RT
2 0 Xt2 dt
(3)
In (2), the integral with respect to X must be understood in the Itô sense. Consequently,
line (3) follows from Itô’s formula, and the fact that X too has quadratic variation equal to t.
One also notes that this estimator’s construction is in fact non-dependent on the form of the
bounded-variation part in the definition of V , which can be interpreted as a form of robustness
of θbT with respect to model misspecification, although we are about to see that the behavior
of θbT depends heavily on V ’s drift specification.
b
On the other hand, it is worth noticing
R Tthat θT coincides formally with a least squares
estimator (LSE). Indeed, by interpreting 0 Xt Ẋt dt as the Itô integral in (2), θbT formally
minimizes
Z
T
θ −→
2
Ẋt + θXt dt.
0
By reversing the roles of V and X, one can obtain an estimator for ρ similar to (3). However,
if V is unobserved, one may recast that estimator by using the estimated value of Vt based on
θbT and the observed path of X, based on (1). In other words, one defines
ρbT =
−VbT2 + T
RT
2 0 Vbt2 dt
Rt
where Vbt = Xt + θbT 0 Xt dt for every t 6 T . As it turns out, this joint estimator θbT , ρbT ,
based on continuous observations of X alone, does not converge to (θ, ρ), but rather to an
explicit rational function of the pair of unknown parameters (θ, ρ). This was proposed and
proved in [6], wherein a semimartingale approach was used to study the asymptotic behavior.
Specifically they showed
3
• strong consistency: as T −→ +∞, almost surely,
θbT −→ θ + ρ
θρ(θ + ρ)
ρbT −→
(θ + ρ)2 + θρ
• asymtpotic normality: as T −→ +∞,
√
T θbT − (θ + ρ), ρbT −
θρ(θ + ρ)
(θ + ρ)2 + θρ
(4)
(5)
law
−→ N (0, Γ)
where Γ is a covariance matrix which has a explicit form as a function of θ and ρ. While
intuition gathered from the full-observation case is in fact erroneous when
is unobserved
V
(the naïve candidates for θbT and ρbT lead to modified limits rather than θbT , ρbT −→ (θ, ρ) ),
nevertheless we have the full picture for the asymptotic behavior of the MLEs/LSEs associated
with (1), at the minor cost of having to solve a non-linear system of two equations to obtain
consistent estimates of (θ, ρ).
In the present paper, our goal is to investigate what happens when, in (1), the standard
Brownian motion W is replaced by a fractional Brownian motion B H . Thus we assume from
now on that X is an Ornstein-Uhlenbeck process driven by a fractional Ornstein-Uhlenbeck
process V : this means the pair (X, V ) is the unique solution of the system of linear stochastic
differential equations
X0 = 0; dXt = −θXt dt + dVt , t > 0
(6)
V0 = 0; dVt = −ρVt dt + dBtH , t > 0,
where B H = BtH , t > 0 is a fractional Brownian motion (fBm) with Hurst index H ∈ ( 12 , 1),
and where θ > 0 and ρ > 0 are unknown parameters. Though X is defined for all H in (0, 1),
to keep technical difficulties to a reasonable level, we restrict ourselves to the case H ∈ (1/2, 1).
It turns out that we need the condition θ 6= ρ for identifiability; remarkably, this is not needed
when H = 1/2, as we saw in the system (4), (5). Details of our results are summarized in
Section 1.2.
We now provide some references to estimation with fBm noise, which are further motivations for our work. We mentioned that the single-drift parameter estimation problem for
fractional diffusion processes based on continuous-time observations was originally studied in
[17] via maximum likelihood; more recent work on this question includes, e.g., [28, 26]. Recently, the LSE for the fractional Ornstein-Uhlenbeck (fOU) process, i.e. the process V in
(6) was proposed in [14]: assuming V is fully observed in continuous time, the LSE for ρ is
defined by
RT
Vt δVt
,
ρT := − R0T
2
0 Vt dt
RT
where the integral 0 Vt δVt is interpreted in the sense of Skorokhod. This integral is the
extension to fBm of Itô’s integral for Brownian motion. In the case ρ > 0, [14] proved that ρT
is strongly consistent and asymptotically normal as T → ∞. Unlike in the case of Brownian
motion (H = 1/2) discussed above, ρT does not coincide with the MLE given in [17], because
the Girsanov theorem for fBm takes a different form than in the case H = 1/2. Given the
4
notorious fact that Skorohod integrals are difficult to interpret in practical terms for fBm, the
authors of [14] proposed in addition the following alternate estimator, which is arguably a
method of moments :
1
− 2H
Z T
1
;
(7)
V 2 dt
HΓ(2H)T 0 t
they proved it is strongly consistent and asymptotically normal. In the case ρ < 0, [5]
established that ρT of ρ is strongly consistent and asymptotically Cauchy-distributed.
The alternate choice of estimator (7) does not, however, avoid the use of continuous observations over discrete ones; this is a problem with many works on fBm-driven models, and is an
additional motivation for us to study the asymptotics of estimation for fBm-driven processes
based on discrete observations. There exists a rich literature on this practical problem for
ordinary diffusions driven by Brownian motions; we refer for instance to [26]. A handful of
authors are beginning to look at these questions with various fBm-driven models, starting with
[28] in 2007, and more recently [3, 7, 11, 20]. In particular, for the fOU process V , motivated
by the estimator given in (7), [15] studied its natural Riemann-sum-type time discretization
n
X
1
Xi2
nHΓ(2H)
i=1
!−
1
2H
(8)
providing strongly consistency and Berry-Esséen-type theorems for it. While we have no
doubt that this estimator is indeed strongly consistent and asymptotically normal, the proofs
in [15] rely on a possibly flawed technique, since the passage from line -7 to -6 on page 434
therein requires the condition H > 3/4, while one expects normal asymptotics only for the
case H 6 3/4.
In our paper, we focus our discussion on estimators which are derived from a basic LSE,
since that technique is known, at least in the Brownian case described in [6], to allow for a
straightforward bivariate extension, as mentioned previously. The LSE has also given rise to a
number of successful studies in the univariate case with fractional processes: we have already
cited [3, 7, 11, 20], while [2] is the continuous-time version of [3].
Herein, specifically, we begin our study of LSE for (θ, ρ) in (6) by using the formal least2
RT
squares interpretation mentioned above, i.e. looking for the minimizer of θ −→ 0 Ẋt + θXt dt;
this leads formally to the following estimator for θ
RT
Xt δXt
(9)
θbT = − R0 T
2 dt
X
t
0
and to the similar estimator
RT
ρbT = − R0T
0
Vbt δVbt
Vb 2 dt
(10)
t
Rt
for ρ, where, because of the fact that V is unobserved, one uses Vbt = Xt + θbT 0 Xt dt for every
t 6 T , instead of relying on the unobserved Vt in the construction of ρbT . These estimators
θbT , ρbT are no longer the MLEs, since, as we mentioned, the Girsanov theorem for fBm does not
have the same form as for Brownian motion, but they are still formally LSEs. Nevertheless,
there is a major difference with respect to the Brownian motion case. Indeed, since the process
5
X is no longer a semimartingale, in (9) and (10) one cannot interpret the numerators using the
Itô integral; the Skorohod integral turns out to be the correct notion to use here. We mentioned
above that Skorohod integrals are difficult to use in practice, but since our Hurst parameter H
exceeds 1/2, it is possible to reinterpret the Skorohod integrals as so-called Young integrals, a
pathwise notion, modulo a correction term which we will be able to compute explicitly thanks
to the Malliavin calculus.
Having succeeded in correctly interpreting the stochastic integrals in (9) and (10), the issue
of how to discretize them becomes paramount to practical implementation, and herein we will
propose several different options, some of which allow for strong consistency and asymptotic
normality under broader conditions than others.
Our discrete-observation study also applies to the case H = 1/2 as a limiting case. The
article [6] treats this case solely with continuous observations; our work thus covers an extension of their work to discrete observations. Checking the validity of this statement rigorously
is straightforward; for the sake of conciseness, we leave it to the interested reader.
1.2
Summary of results and heuristics
We now summarize our results, the structure of our article, and our main proof elements,
including useful heuristics when available.
• In Section 2 and in the Appendix we introduce the needed mathematical background
material for our study, including elements of the Malliavin calculus, a convenient criterion
for establishing normal convergence on Wiener chaos, and the relation between Skorohod
integrals and Young integrals with respect to fBm when H > 1/2.
• In Section 3, we concentrate on proving strong consistency and asymptotic normality
for the estimators θbT and ρbT with continuous observations.
– We first prove the following almost surely convergences:
θbT
ρbT
→ θ ∗ := ρ + θ,
→ ρ∗ :=
(11)
ρθ (θ + ρ)
ρ2−2H −θ 2−2H
θ −2H −ρ−2H
+ (θ + ρ)2
.
(12)
The proof relies on studying the numerator and the denominator of the expressions for θbT and ρbT separately. For the denominators, we rely on Birkhoff’s ergodic
theorem, and elementary covariance estimations for exponential convolutions with
fBm. For the numerators, we express the Skorohod integrals as Young integrals
plus their correction terms involving Malliavin derivatives which are explicit deterministic functions since our processes are Gaussian.
– The expression θ ∗ = ρ + θ in (11) is easy to explain: as noted below in line (16), X
satisfies a stochastic integro-differential equation in which the zero-mean-reversion
term − (ρ + θ) X (t) dt appears, and thus a natural candidate for a consistent estib
mator of ρ + θ is the LSE
one adds merely one mean-zero noise term
R θ, whether
t
H
dB or another term 0 Xs ds which is asymptotically small. This is why the
limiting behavoir of θb remains the same for us as in (4), which is the Brownian
case (H = 1/2) studied in [6]. The details of this heuristic are omitted, since the
6
full proof we present herein is needed to be convincing. On the other hand, the
expression in (12) for our ρ∗ is more opaque; there does not seem to be a direct
heuristic to explain it, beyond our computations. When one compares our ρ∗ in
(12) with the ρ∗ in (5) found in [6], one sees that the term ρθ in (5) is replaced by
the expression ρ2−2H − θ 2−2H / θ −2H − ρ−2H , which can help identify how the
case of fBm deviates from the case H =1/2.4 The expression for ρ∗ is analyzed
b ρb , which helps explain to some extent why
further in the context of discretizing θ,
this complicated expression arises, as the reader will find out in the first paragraph
of Section 4.2.
– We prove asymptotic normality of θbT , ρbT by expressing the Skorohod integrals as
iterated Wiener integrals, identifying dominant portions of these integrals, relying
on a criterion for normal convergence in law in Wiener chaos, combined with a
number of almost sure convergences. Our
main asymptotics
normality result is a
√
∗
∗
b
central limit theorem that holds for T θT − θ , ρbT − ρ as T → ∞, as soon as
H ∈ [1/2, 3/4). The asymptotic covariance is given explicitly. See Theorem 8. The
upper limit of validity of this theorem is a typical threshold in normal convergence
theorems in the second Wiener chaos. See for example a classical instance of this
situation in the Breuer-Major central limit theorem, as presented in [22, Chapter
7]. For H > 3/4, we conjecture that the estimators are asymptotically Rosenblattdistributed (again see [22, Chapter 7] for a classical example of such a phenomenon),
and that the convergence occurs almost surely; this point is not discussed further,
for the sake of conciseness.
• The topic of Section 4 is to construct estimators based solely on discrete observations.
The asymptotic results we prove still require increasing horizon. We also assume that X
is observed at evenly spaced intervals, with a time step ∆n , and we set the time horizon
to be Tn = n∆n → ∞ as n → ∞. Let tk = k∆n be the kth observation time. For
instance, the case ∆n = 1 corresponds to a fixed observation frequency; other conditions
on ∆n will include requiring ∆n to tend to 0 as fast as a certain negative power of
n, i.e. the observation frequency increases as the horizon increases. For some strong
consistency results, it will even be possible for us to relax conditions on ∆n where it
is allowed to tend to infinity like a power of n, i.e. with decreasing frequency as the
horizon increases.
Arguably, to be consistent with the assumption that only X is observed, the only estimators which are of practical use are those whichrely solely
on the values {Xtk : k = 1, . . . , n}.
b
Designing such an estimator by discretizing θT , ρbT turns out to be a difficult task, in
which the final expression solves a non-linear system in the spirit of that which would
follow from (11) and (12), but is rather distinct from this system because of the difficulty
in how to interpret the discretizations of the Skorohod or Young integrals.
The method we have chosen moves through several intermediate steps, where we gradually increase the number of terms in the estimators which are replaced by discretized
versions. This method has the advantage of clearly showing where the restrictions on the
4
As a way to compare these terms, which do coincide when H = 1/2, we can see that if ρ tends to θ,
the aforementioned expression in (5) tends to θ2 (1 − H) /H, which thus deviates significantly from the case
H = 1/2 quantitatively, particularly for H close to 1.
7
observation frequency ∆n come into play. Each intermediate estimator can be considered as a perturbation of the previous one, starting with the continuous-time estimator
of Section 3. Thus arguably all these estimators can be considered as tools used for
the final objective, attained in Section 4.3, of constructing a strongly consistent and
asymptotically normal estimator of (θ, ρ) based only on the data {Xtk : k = 1, . . . , n}.
Nonetheless, some of the other estimators are relevant in their own right, as they might
correspond to realistic partial or full observation cases.
– The main technical estimates which allow our discretization are given at the beginning of Section 4. These are Lemmas 10 and 11, based on applications of the
Borel-Cantelli lemma. For Z any stochastic process, we let
Qn (Z) := n−1
n
X
(Ztk )2 .
k=1
Rt
Rt
Let St := 0 Xs2 ds and Σt := 0 Xs ds. We show √
that the discrepancy between
STn /Tn and its discrete version Qn (X) is = o 1/ Tn almost surely. We then
compute threeR different discrepancies related to Σ: first we show that the
√ difference
T
between Tn−1 0 n Σ2t dt and its discrete version Qn (Σ) is also = o 1/ Tn almost
surely. Then we show that with Σ̂ the version of Σ which
depends only on X
Pk
2
b
observations, i.e. Σtk := ∆n i=1 Xti−1 , we get that Qn Σ̂ − Qn (Σ) tend to 0
almost surely. This is helpful to prove strong consistency
of discrete estimators. To
√
prove asymptotic normality, we need that Qn Σ̂ − Qn (Σ) = o 1/ Tn , which
we prove holds almost surely. Increasingly restrictive conditions on ∆n are needed
for these successive results.
– We first concentrate on discretizing the denominators of θbTn , ρbTn .
∗ We replace the denominator of θbTn by Qn (X), yielding an estimator θ̃n , and
we then replace the denominator of ρbT by Qn (X) + (θ̃n )2 Qn (Σ), yielding an
RT
estimator ρ̃n , because, as it turns out, 0 Vbt2 dt is asymptotically equivalent,
almost surely, to Qn (X) + (θ̃n )2 Qn (Σ). Thanks to this, to the almost sure
equivalence of Qn (X) with STn /Tn , and similarly for Qn (Σ), coming from
Lem
mas 10 and 11, the strong consistency and asymptotic normality of θ̃n , ρ̃n
follows from that of θbTn , ρbTn proved in Section 3. Here it is sufficient to
assume H ∈ (1/2, 1) and ∆n 6 nα for some α ∈ (−∞, 1/H) for the strong
consistency; note that ∆n is allowed to remain constant or even increase like a
moderate power in this case. For the asymptotic normality, it is sufficient that
→ 0.
H ∈ (1/2, 3/4) and n∆H+1
n
∗ A second result is obtained in which we forego havingP
access to the process Σ
b t = ∆n k Xt2 ; in this case,
itself, relying instead on its discrete version Σ
k
i=1
i−1
almost-sure converge of ρ̃n requires nα+1 ∆H+1
→
0
for
some
α > 0, and the
n
3
2H+3
central-limit result for ρ̃n requires n ∆n
→ 0.
– We are then able to define and study a bonafide estimator based on discrete data
alone.
8
∗ We begin with assuming that we have access to both Xtk and Σtk for all
k = 1, . . . , n. The stochastic integrals in θ̃n and ρ̃n were analyzed in Section 3,
and were found, under scaling by Tn−1 , to be asymptotically constant, where
the explicit constants depend on the parameters. By using these limits and a
discretization of the Riemann integrals in the denominators of θ̃n and ρ̃n , this
allows us, at the beginning
of Section 4.2, to motivate the definition of a pair
of estimators θ̌n , ρ̌n as solution of the non-linear system
F θ̌n , ρ̌n = (Qn (X), Qn (Σ))
where F is a positive function of the variables (x, y) in (0, +∞)2 defined by:
for every (x, y) ∈ (0, +∞)2
1
2−2H − x2−2H , x−2H − y −2H
y
if x 6= y
2
2
F (x, y) = HΓ(2H)× y −x
(1 − H)x−2H , Hx−2H−2
if y = x,
and the data statistics used in the system are Qn (X) and Qn (Σ). Strong
consistency and asymptotic normality follow for the uniquely defined θ̌n , ρ̌n .
The delicate computation of the asymptotic
covariance is given. The parameter
restrictions remain the same as for θ̃n , ρ̃n , namely strong consistency if H ∈
(1/2, 1) and ∆n 6 nα for some α ∈ (−∞, 1/H), and asymptotic normality if
→ 0.
H ∈ (1/2, 3/4) and n∆H+1
n
∗ By redefining the estimators θ̌n , ρ̌n using Σ̂ instead of Σ, one ensures that
only the data {Xtk : k = 1, . . . , n} is used. Here, the results from the previous case can be applied directly with the auxiliary results on how Σ̂ perturbs
Σ (Lemma 11), to obtain the same almost-sure convergence and central-limit
result, but these are now restricted respectively to the aforementioned ob→ 0 for some α > 0, and
servation frequency parameter ranges nα+1 ∆H+1
n
3
2H+3
→ 0.
n ∆n
– Finally, to illustrate how the complexity of the nonlinearities in the definition of
to the partial-observation problem, we define a pair
θ̌n , ρ̌n may be
attributable
of estimators θn , ρn under the assumption that both {Xtk : k = 1, . . . , n} and
{Vtk : k = 1, . . . , n} are available. The ρn is explicit given {Vtk : k = 1, . . . , n}, and
is identical to the one given in [15], i.e. (8), as it should be. The θ n satisfies the
following straightforward non-linear equation given ρn and {Xtk : k = 1, . . . , n}:
(θn )
2−2H
−
Qn (X)
HΓ(2H)
2−2H Q (X) 2
n
−
ρn .
(θn ) = ρn
HΓ(2H)
2
The parameter restrictions are
as when X and Σ are discretely observed:
the same
strong consistency holds for θn , ρn if H ∈ (1/2, 1) and ∆n 6 nα for some α ∈
(−∞, 1/H), and asymptotic normality holds if H ∈ (1/2, 3/4) and n∆H+1
→ 0.
n
Before we proceed with the details of our study, we provide needed mathematical tools in
Section 2.
9
2
Preliminaries
In this section we describe some basic facts on the stochastic calculus with respect to a
fractional Brownian motion. For a more complete presentation on the subject, see [23] and
[1].
The fractional Brownian motion (BtH , t > 0) with Hurst parameter H ∈ (0, 1), is defined as a
centered Gaussian process starting from zero with covariance
RH (t, s) = E(BtH BsH ) =
1 2H
t + s2H − |t − s|2H ; s, t > 0,
2
We assume that B H is defined on a complete probability space (Ω, F, P ) such that F is the
2
sigma-field generated by B H . By Kolmogorov’s continuity criterion and the fact E BtH − BsH =
|s − t|2H , we deduce that B H admits a version which has Hölder continuous paths of any order
γ < H.
Fix a time interval [0, T ]. We denote by H the canonical Hilbert space associated to the
fractional Brownian motion B H ; the book [23], among many other references, can be consulted
for the construction and properties of H. We use the following convenient notation for Wiener
integrals with respect to B H :
Z T
ϕ (s) dB H .
B H (ϕ) :=
0
Of interest to us is the fact that, with H > 1/2, for a pair of (non-random) functional elements
ϕ, ψ of H, its inner product satisfies
Z TZ T
H
H
hϕ, ψiH = E B (ϕ) B (ψ) = H(2H − 1)
ϕ(u)ψ(v)|u − v|2H−2 dudv.
0
0
It follows from [25] that the set |H| of functional elements in H is Banach and actually contains
1
L H ([0, T ]).
The Malliavin derivative D w.r.t. B H , which is an H-values operator, is defined first by
setting that
DB H (ϕ) = ϕ
for any ϕ ∈ H, and then by requiring that it satisfy a multi-parameter chain rule: for any f ∈
n
n
C∞
b ( R , R) (infinitely differentiable functions from R to R with bounded partial derivatives)
and any ϕ1 , ..., ϕn ∈ H, D operates on the cylinder r.v. F := f (B H (ϕ1 ), ..., B H (ϕn )) as
DF =
n
X
∂f
(B H (ϕ1 ), ..., B H (ϕn ))ϕi .
∂xi
i=1
The domain D 1,2 of D is then the the closure of the set of cylinder r.v.’s F with respect to
the norm
kF k21,2 := E(F 2 ) + E(kDF k2H ).
The divergence operator δ is the adjoint of the derivative operator D : an H-valued r.v.
u ∈ L2 (Ω; H) belongs to its domain Domδ if
E |hDF, uiH | 6 cu kF kL2 (Ω)
10
for some constant cu and every cylinder r.v. F . In this case δ(u) is uniquely defined by the
duality
E(F δ(u)) = E hDF, uiH
for any F ∈ D 1,2 . We will make use of the notation
δ(u) =
Z
T
0
us δBsH ,
u ∈ Domδ.
RT
In particular, δ extends the Wiener integral: for h ∈ |H|, B H (h) = δ(h) = 0 hs δBsH .
For every n > 1, let Hn be the nth Wiener chaos of B H , that is, the closed linear subspace
of L2 (Ω) generated by the random variables {Hn (B H (h)), h ∈ H, khkH = 1} where Hn is the
nth Hermite polynomial. The mapping In (h⊗n ) = n!Hn (B H (h)) provides a linear isometry
between the symmetric tensor product H⊙n (equipped with the modified norm k.kH⊙n =
√1 k.kH⊗n ) and Hn . It also turns out that In (h⊗n ) is the multiple Wiener integral of h⊗n
n!
w.r.t. B H . For every f, g ∈ H⊙n the following product formula holds
E (In (f )In (g)) = n!hf, giH⊗n .
For h ∈ H⊗n , the multiple Wiener integrals Iq (f ), which exhaust the set Hq , satisfy a hypercontractivity property (equivalence in Hq of all Lp norms for all p > 2), which implies that
for any F ∈ ⊕ql=1 Hl , we have
1/p
1/2
6 cp,q E |F |2
E |F |p
for any p > 2.
(13)
It is well-known that L2 (Ω) can be decomposed into the infinite orthogonal sum of the spaces
Hn . That is, any square integrable random variable F ∈ L2 (Ω) admits the following “Wiener
chaos” expansion
∞
X
In (fn ),
F = E(F ) +
n=1
H⊙n
where the fn ∈
are uniquely determined by F .
Finally, we will use the following central limit theorem for multiple stochastic integrals (see
[24]).
Theorem 1 Let {Fn , n > 1} be a sequence of random variables in the q-th Wiener chaos Hq ,
q > 2, such that limn→∞ E(Fn2 ) = σ 2 . Then the following conditions are equivalent:
(i) Fn converges in law to N (0, σ 2 ) as n tends to infinity.
(ii) kDFn k2H converges in L2 to a constant as n tends to infinity.
3
Asymptotic behavior of LSEs
Throughout the paper we assume that H ∈ ( 12 , 1), θ > 0 and ρ > 0 such that θ 6= ρ.
It is readily checked that we have the following explicit expression for Xt :
Xt =
θ
ρ
Xtρ +
Xθ
ρ−θ
θ−ρ t
11
(14)
where for m > 0
Xtm
=
Z
t
0
e−m(t−s) dBsH .
(15)
On the other hand, we can also write that the system (6) implies that X solves the following
stochastic integro-differential equation
Z t
Xs ds dt + dBtH .
(16)
dXt = − (θ + ρ) Xt dt − ρθ
0
For convenience, and because it will play an important role in the forthcoming computations,
we introduce the following processes related to Xt :
Z T
Z T
Z T
Z T
2
2
Xt Vt dt;
Vt dt; PT =
Xt dt; LT =
Xt dt; ΣT =
ST =
0
0
0
and
bT =
L
where for 0 6 t 6 T
Z
T
0
0
Vbt2 dt
Vbt = Xt + θbT Σt ,
(17)
and θbT is our continuous LSE for θ as given in (9). We will need the following lemmas.
Lemma 2 Assume H ∈
1
2, 1
. Then, as T → ∞
Z
1 T 2
Xt dt −→ η X ,
T 0
Z
1 T 2
Σt dt −→ η Σ ,
T 0
Z
1 T
Σt Xt dt −→ 0
T 0
(18)
(19)
(20)
almost surely, where
ηX =
and
ηΣ =
HΓ(2H) 2−2H
[ρ
− θ 2−2H ],
ρ2 − θ 2
HΓ(2H) −2H
[θ
− ρ−2H ].
ρ2 − θ 2
Proof. From (6) we can write
H
Xt
Xt
Bt
d
=A
dt + d
0
Σt
Σt
θ + ρ −θρ
Xt
where A =
is geometrically ergodic because the largest
. The process
1
0
Σt
eigenvalue of A is negative. Then to prove Lemma 2, using Birkhoff’s ergodic theorem (for
instance see [13]), it is sufficient to study the convergence of E[Xt2 ], E[Σ2t ] and E[Σt Xt ] as
12
t −→ ∞.
For the convergence of E[Xt2 ], (14) leads to
E[Xt2 ]
=
η
X
ρ
ρ−θ
Since
=
2
E[(Xtρ )2 ]
ρ
ρ−θ
2
+
λ(ρ, ρ) +
θ
θ−ρ
then by using 1) of Lemma 21 we obtain
2
θ
θ−ρ
2θρ
E[(X θ )t (Xtρ )].
(θ − ρ)2
E[(Xtθ )2 ] −
2
λ(θ, θ) −
2θρ
λ(θ, ρ)
(θ − ρ)2
η X − E[Xt2 ] 6 c(H, θ, ρ)e−t/2 .
Thus we deduce the convergence (18).
Using the same argument and the fact that
Σt =
Vt − Xt
X θ − Xtρ
= t
θ
ρ−θ
(21)
we deduce the convergence (19).
RT
Finally, the convergence (20) is satisfied by using 0 Σt Xt dt =
ΣT
2
and point 5) of Lemma 21.
Lemma 3 We have
1
T
Z
T
0
Xt δXt −→ −(ρ + θ)η X
(22)
almost surely as T −→ ∞.
Proof. From (6) and (A-2) we can write
Z
0
T
Xt δXt = −θ
= −θ
Z
Z
T
0
T
0
−αH
Xt2 dt
Xt2 dt
Z
T
0
Z
−ρ
Z
0
Z
T
0
Xt dBtH
=
=
Thus
Z
0
Z
T
0
T
0
0
T
0
Xt dBtH
Ds Xt (t − s)2H−2 dsdt
where αH = 2H(2H − 1). Moreover,
Z
Z
Xt δBtH
Z
Z T
2
Xt dt +
Xt Σt dt − ρ
Xt Vt dt +
0
− θρ
t
T
Z
T
T
Xt2 dt
Z
T
Xt Σt dt
+ θρ
Xt dXt + (θ + ρ)
0
0
0
Z T
Z T
XT2
2
Xt Σt dt.
Xt dt + θρ
+ (θ + ρ)
2
0
0
T
Xt δXt =
XT2
− αH
2
Z
T
0
13
Z
t
0
Ds Xt (t − s)2H−2 dsdt.
Since
Ds Xtm = e−m(t−s) 1[0,t] (s)
we deduce that
Z
1 T
Xt δXt =
T 0
Z Z
XT2
αH T t 1
−
(ρe−ρ(t−s) − θe−θ(t−s) )(t − s)2H−2 dsdt
2T
T 0 0 ρ−θ
Z Z
αH T t 1 −ρr
XT2
(23)
ρe
− θe−θr r 2H−2 drdt.
−
2T
T 0 0 ρ−θ
=
Thanks to l’Hôpital’s rule, as T −→ ∞
Z Z
αH T t 1 ρr
θr
r 2H−2 drdt −→
ρe − θe
T 0 0 ρ−θ
=
HΓ(2H) 2−2H
[ρ
− θ 2−2H ]
ρ−θ
(ρ + θ)η X
.
αH
Finally, combining this last convergence and point 5) of Lemma 21, the proof of Lemma 3 is
done.
We now have all the elements to obtain our strong consistency result for θbT .
Theorem 4 We have
θbT −→ θ ∗
almost surely as T −→ ∞, where θ ∗ = θ + ρ.
Proof. The proof follows directly from the convergence (18) and Lemma 3.
The next lemmas are additional elements needed to prove the strong consistency of ρbT .
Lemma 5 We have
bT
L
−→ η X + (ρ + θ)2 η Σ
T
almost surely as T −→ ∞.
Proof. The equation (17) ensures
Z
Z T
2
b
b
Xt dt + 2θT
LT =
0
T
0
Xt Σt dt + θbT2
Z
T
0
Σ2t dt,
and the desired conclusion follows by using Lemma 2 and Theorem 4.
Lemma 6 We have
1
T
almost surely as T −→ ∞.
Z
T
0
Vbt δVbt −→ −ρθ(ρ + θ)η Σ
Proof. From (6) and (17) we can write
Z
Z T
Z T
Z T
X 2 dt − θbT (θ + ρ)
Xt δXt + θbT
Vbt δVbt =
0
0
0
+
Z
0
T
Σt θbT δBtH + θbT2
t
Z
T
Σt Xt dt.
0
14
T
0
Σt Xt dt − ρθ θbT
Z
0
T
Σ2t dt
On the other hand
Z T
θbT Σt dBtH
0
Z T
Z T
b
b
Σt Vt dt
=
θT Σt dVt + ρθT
0
0
Z T
Z T
Z T
Z T
−1
−1 b
b
b
b
= −θ
θT Xt dVt + θ θT
Vt dVt + ρθT
Σt Xt dt + ρθ θT
Σ2t dt
0
0
0
0
Z
Z T
Z T
Z T
Z T
−1
2
−1 b
b
b
b
b
Σt Xt dt + ρθ θT
Vt dVt + ρθT
= −θ
Xt dt + θ θT
θT Xt dXt − θT
0
=
−1 2 b
X − θT
2θ T
Z
0
0
T
0
0
Z T
Z T
θbT 2
2
b
b
Σ2t dt.
Xt dt +
Σt Xt dt + ρθ θT
VT + ρθT
2θ
0
0
T
0
Now, applying (A-2), we obtain
Z T
Z T
Z T
Z T
1 2 θbT 2 b
2
b
b
b
Xt δXt − XT +
Vt δ Vt =
Σt Xt dt + θT
V − θT θ
Σt Xt dt
2θ
2θ T
0
0
0
0
Z TZ t
−αH
Ds (θbT Σt )(t − s)2H−2 dsdt.
0
On the other hand
It follows from (23) that
θbT
=
=
RT
0
Σ2t dt
(24)
0
Ds (θbT Σt ) = Σt Ds θbT + θbT Ds Σt .
Xt δXt
ST
1 2
2 XT
− αH
RT Rt
1
0 ρ−θ
0
ρe−ρr − θe−θr r 2H−2 drdt
ST
.
Hence, for s < T
Thus
αH
Z
T
0
=
XT
αH
ST
+αH θbT
Z
t
0
Z T
XT Ds XT − θbT Ds ST
.
Ds θbT =
ST
Ds (θbT Σt )(t − s)2H−2 dsdt
0
Z
T
0
Z
t
Σt Ds XT (t − s)
0
Z
t
0
2H−2
Z Z
θbT T t
dsdt − αH
Σt Ds ST (t − s)2H−2 dsdt
ST 0 0
Ds Σt (t − s)2H−2 dsdt
:= J1,T − J2,T + J3,T
We shall prove that for every ε > 0
|J1,T |
−→ 0,
Tε
|J2,T |
−→ 0,
Tε
θ+ρ
J3,T
−→
HΓ(2H)[(−ρ)1−2H − (−θ)1−2H ]
T
ρ−θ
15
(25)
(26)
(27)
almost surely as T → ∞.
We first estimate J1,T . Clearly, (14) implies
Z Z
XT T t
J1,T = αH
Σt Ds XT (t − s)2H−2 dsdt
ST 0 0
Z t
Z
αH XT T
=
ρe−ρ(T −s) − θe−θ(T −s) (t − s)2H−2 dsdt
Σt
ρ − θ ST 0
0
Z t
Z T
αH XT
ρe−ρ(T −t+x) − θe−θ(T −t+x) x2H−2 dxdt.
Σt
=
ρ − θ ST 0
0
The last equality comes from making the change of variable x = t − s.
Hence
|J1,T |
|XT |/T ε supt∈[0,T ] |Σt |
6
c(H,
θ,
ρ)
.
Tε
ST /T
Tε
Using (18), (21) and the point 5) of Lemma 21, the convergence (25) is obtained.
Next we estimate J2,T . By (14) we have
Z Z
θbT T t
J2,T = αH
Σt Ds ST (t − s)2H−2 dsdt
ST 0 0
Z T
Z Z
θbT T t
= 2αH
Xu Ds Xu (t − s)2H−2 dudsdt
Σt
ST 0 0
s
Z TZ t Z T
θbT
Xu ρe−ρ(u−s) − θe−θ(u−s) (t − s)2H−2 dudsdt.
Σt
= 2αH
(ρ − θ)ST 0 0
s
Then
|J2,T |
Tε
6
6
θbT supt∈[0,T ] |Xt | supt∈[0,T ] |Σt | Z
c(H, θ, ρ)
c(H, θ, ρ)
Tε
ST
θbT
0
T
Z
t
0
e− min(θ,ρ)(T −s) (t − s)2H−2 dsdt
supt∈[0,T ] |Xt | supt∈[0,T ] |Σt |
ST /T
Tε
−→ 0
almost surely as T −→ ∞. The last convergence comes from (18), (21), Theorem 4 and the
point 5) of Lemma 21. Thus, the convergence (26) is satisfied.
Finally, we estimate J3,T . Using (21) and (14)
Z TZ t
J3,T = αH θbT
Ds Σt (t − s)2H−2 dsdt
0
=
0
Z Z
αH θbT T t −θ(t−s)
e
− e−ρ(t−s) (t − s)2H−2 dsdt.
ρ−θ 0 0
By l’Hôpital rule we obtain
J3,T
θ+ρ
−→
HΓ(2H)[θ 1−2H − ρ1−2H ]
T
ρ−θ
almost surely as T −→ ∞.
Using the above estimations (25), (26), (27) together with (24), Lemma 2, Theorem 4, the
point 5) of Lemma 21 and Lemma 3 the desired result is then obtained.
16
Theorem 7 We have the almost sure convergence
ρbT −→ ρ∗
as T → ∞, where
ρ∗ =
θρ(θ + ρ)η Σ
.
η X + (θ + ρ)2 η Σ
Proof. The proof is a straightforward consequence of Lemma 5 and Lemma 6.
Our approach to prove the asymptotic normality for both estimators θbT and ρbT looks first
at the normal convergence of the T -indexed second-chaos sequence based on the kernel which
appears in the representation (15) of X. Thereafter, thanks to elementary stochastic calculus
in the second chaos, these double stochastic integrals will be identified in an expression for
the leading terms in θbT − θ ∗ in the proof of Theorem 8. A similar technique, plus the use of
the chain rule of Young integrals and their relation to Skorohod integrals, is used to find again
that the leading terms in ρbT − ρ∗ are also linear combinations of the same double integrals;
the analysis of the lower-order terms are less evident than for θbT − θ ∗ ; the proof of Theorem
8 records all the details.
We are ready to prove the asymptotic normality of θbT , ρbT .
Theorem 8 Assume that H ∈ ( 12 , 34 ). Then
√
law
T θbT − θ ∗ , ρbT − ρ∗ −→ N (0,t P Γ P )
where the matrices Γ and P are defined respectively in (A-5) and (35).
Proof. We express θbT − θ ∗ and ρbT − ρ∗ as linear combinations of the double stochastic
integrals identified in the previous theorem, plus lower-order terms.
The case of θbT − θ ∗ is rather straightforward. It follows from (16) that
RT
RT
ρθ 0 Xt Σt dt − 0 Xt δBt
∗
b
.
θT − θ =
ST
Since
Z
and
1
√
T
Z
0
T
Xt δBtH
=
=
we can write
where
T
0
1
Xt Σt dt = Σ2T
2
Z TZ t
1
√
(ρe−ρ(t−s) − θe−θ(t−s) )δBsH δBtH
(ρ − θ) T 0 0
1
√
ρI2 (fTρ ) − θI2 (fTθ )
(ρ − θ) T
√
T θbT − θ ∗ =
RTθ
1√
(ρ−θ) T
θI2 (fTθ ) − ρI2 (fTρ )
ST /T
√
ρθ Σ2T / T
:=
−→ 0
2 ST /T
17
+ RTθ .
(28)
(29)
almost surely as T −→ ∞.
For ρbt − ρ∗ , the situation is significantly more complex. We have for every 0 6 t 6 T
which leads to
Vbt = Xt + θbT Σt = Vt + (θbT − θ)Σt = Vt + (θbT − θ ∗ )Σt + ρΣt
ρ
1
= Vt − (Xt − Vt ) − (θbT − θ ∗ )(Xt − Vt )
θ
θ
θ∗
ρ
1 b
=
Vt − Xt − (θT − θ ∗ )(Xt − Vt )
θ
θ
θ
bT =
L
where
Thus,
where
Z
Z
T
0
Vbt2 dt = IT + (θbT − θ ∗ )(JT + (θbT − θ ∗ )KT )
IT
=
JT
=
KT
=
1 2
(ρ ST + (θ ∗ )2 LT − 2θ ∗ ρPT ),
θ2
1
(2ρST + 2θ ∗ LT − 2(θ + 2ρ)PT ),
θ2
1
(ST + LT − 2PT ).
θ2
b T (b
L
ρT − ρ∗ ) = ITV + (θbT − θ ∗ )(JTV + (θbT − θ ∗ )KTV )
Z TZ t
VbT2
∗
b
b
Vt δVt − ρ IT = −
= −
Ds Vbt (t − s)2H−2 dsdt − ρ∗ IT ,
+ αH
2
0
0
0
= −ρ∗ JT , and KTV = −ρ∗ KT .
ITV
JTV
T
On the other hand, using the formula (A-1) we obtain
RT
XT2
ρθ 2
1
H
ST = − 2θ∗ + θ∗ 0 RXs dBs − 2θ∗ Σ2T ,
V
T
PT = − XTθ∗VT + θ1∗ 0 Xs dBsH + 2θT∗ ,
L = − VT2 + 1 R T V dB H .
T
s
s
2ρ
ρ 0
Furthermore, using the relation between Young and Skorohod integrals,
RT
XT2
αH R T R t
1
H
2H−2 dsdt − ρθ Σ2 ,
ST = − 2θ∗ + θ∗ 0 RXs δBs + θ∗ 0 R 0 RDs Xt (t − s)
2θ ∗ T
V2
T
XT VT
αH T t
1
H
2H−2
PT = − θ∗ + θ∗ 0 Xs δBs + θ∗ 0 0 Ds Xt (t − s)
dsdt + 2θT∗ ,
L = − VT2 + 1 R T V δB H + αH R T R t D V (t − s)2H−2 dsdt.
T
2ρ
ρ
0
s
s
Setting
λT := αH
ρ
Z
0
T
0
Z
0
t
0
s t
e−(t−s) (t − s)2H−2 dsdt
18
(30)
we can write
λX
T
:=
αH
λVT :=
αH
λΣ
T :=
b
λVT
αH
:=
αH
Z
Z
Z
Z
T
0
T
0
T
0
T
0
= λX
T +
Z
Z
Z
Z
t
0
t
0
t
0
t
0
∗ Σ
θ λT
Ds Xt (t − s)2H−2 dsdt =
1
ρ2−2H − θ 2−2H λT
ρ−θ
Ds Vt (t − s)2H−2 dsdt = ρ1−2H λT
Ds Σt (t − s)2H−2 dsdt =
1 V
1
(λT − λX
θ 1−2H − ρ1−2H λT
T )=
θ
ρ−θ
b Σ
Ds Vbt (t − s)2H−2 dsdt = λX
T + θT λT + J1,T − J2,T
+ (θbT − θ ∗ )λΣ
T + J1,T − J2,T .
The last equality comes from the fact that Ds Vbt = Ds Xt + θbT Ds Σt + Σt Ds θbT .
Since
2
ρ X (θ ∗ )2 V
ρ∗
X
∗ Σ
λ
+
2ρλ
λ
−
−
λX
+
θ
λ
=
T
T
T
T
θ2
θ∗ T
ρ
we can write
ITV
where
VbT2
b
+ λVT − ρ∗ IT
2
∗
∗ )2 Z T
2 Z T
ρ
(θ
ρ
∗
Σ
H
H
= (θbT − θ )λT − 2 (−2ρ + ∗ )
Xs δBs +
Vs δBs + RT
θ
θ
ρ
0
0
= −
2
VT2
−VbT2
ρ∗ 2 XT2
ρθ 2
∗ 2 VT
∗ −XT VT
+ J1,T − J2,T − 2 ρ (− ∗ − ∗ ΣT − (θ )
− 2θ ρ(
+ ∗) .
RT =
2
θ
2θ
2θ
2ρ
θ∗
2θ
Combining previous estimations we obtain
where
Rθ
V
∗
V
b
bT (b
+ RT
L
ρT − ρ∗ ) = cρT I2 (fTρ ) + cθT I2 (fTθ ) + √T λΣ
T + JT + (θT − θ )KT
T
cρT
=
2
ρρ∗ (−2ρ + θρ∗ ) ρ∗ (θ ∗ )2
−ρλΣ
ρJTV
ρ(θbT − θ ∗ )KTV
T
−
−
−
−
(ρ − θ)ST
θ 2 (ρ − θ)
ρθ 2
(ρ − θ)ST
(ρ − θ)ST
2
ρρ∗ (−2ρ + θρ∗ ) ρ∗ (θ ∗ )2
ρλJ
−ρλΣ
−
−
−
−→ c =
(ρ − θ)η X
θ 2 (ρ − θ)
ρθ 2
(ρ − θ)η X
ρ
(31)
almost surely as T −→ ∞, and
cθT
=
2
θρ∗ (−2ρ + ρθ∗ ) ρ∗ (θ ∗ )2
θλΣ
θJTV
θ(θbT − θ ∗ )KTV
T
+
−
+
+
(ρ − θ)ST
θ 2 (ρ − θ)
ρθ 2
(ρ − θ)ST
(ρ − θ)ST
2
θρ∗ (−2ρ + θρ∗ ) ρ∗ (θ ∗ )2
θλJ
θλΣ
+
−
+
−→ c =
(ρ − θ)η X
θ 2 (ρ − θ)
ρθ 2
(ρ − θ)η X
θ
19
(32)
almost surely as T −→ ∞. These last two convergences come from the fact that
Σ
λΣ
T /T −→ λ =
and
2HΓ(2H) 1−2H
θ
− ρ1−2H
ρ−θ
2 ∗ X
θ η − HΓ(2H)θ ∗ ρ−2H
2
θ
almost surely as T −→ ∞, because λT /T −→ 2HΓ(2H) as T −→ ∞.
Thus
√1
cρT I2 (fTρ ) + cθT I2 (fTθ )
√
T
∗
T (b
ρT − ρ ) =
+ RTρ
b
JTV /T −→ λJ =
LT
T
where as T −→ ∞
RTρ
=
RTθ
V + (θ
bT − θ ∗ )K V
λΣ
+
J
T
T
T
bT
L
+
√
RT / T
bT
L
T
−→ 0
(33)
(34)
almost surely.
Finally, with the expressions (28) and (33) on hand, and the almost-sure negligibility of their
corresponding lower-order terms as proved in (29) and (34), we get
!
θ T
θ T
√
c
1
T
b
ρ−θ
S
ρ
T
LT
+ RTθ , RTρ
I2 (fTθ ), I2 (fT )
T θbT − θ ∗ , ρbT − ρ∗ = √
ρ T
ρ T
cT b
T
θ−ρ ST
LT
where as T −→ ∞
θ T
ρ−θ ST
ρ T
θ−ρ ST
T
bT
L
cρT bT
LT
cθT
!
−→ P :=
θ 1
ρ−θ ηX
ρ 1
θ−ρ ηX
cθ
b
ηL
cρ
b
ηL
!
(35)
almost surely. Now, applying Slutsky’s lemma and Theorem 22 combined with the above
convergences, the proof is complete.
4
Discrete observation
Assume that the process X is observed equidistantly in time with the step size ∆n : ti =
i∆n , i = 0, . . . , n, and Tn = n∆n denotes the length of the ‘observation window’. The goal
of this section is to construct two estimators θ̌n and ρ̌n of θ and ρ respectively based on the
sampling data Xti , i = 0, . . . , n, and study their strong consistency and asymptotic normality.
We also want to define estimators in such a way that
and normality results proved
consistency
b
in Section 3 for the continuous-data estimators θT , ρbT can be used to good effect in the
discrete case. The basic strategy for this is therefore to look for ways of discretizing the MLE
studied in Section 3. It turns out that the most efficient way of implementing this strategy
is to
defineseveral intermediate estimators, starting with ones where only the denominators
in θbT , ρbT are discretized, and then using an algebraic asymptotic interpretation of the
numerators to avoid a direct discretization of the corresponding Young or Skorohod integrals.
This method allows a rather direct use of the asymptotic normality Theorem 8 in Section
3, while for the strong consistency results, some of the almost-sure convergences proved in
20
Section 3 are used directly, and additional ones are newly established early on in this section.
See Section 1.2 for other details about the heuristics which explain the choices made below in
this Section.
For any given process Z, define
n
Qn (Z) :=
1X
(Zti−1 )2
n
i=1
The following well-known direct consequence of the Borel-Cantelli Lemma (see e.g. [18]), will
allows us to turn convergence rates in the p-th mean into pathwise convergence rates. This is
particularly efficient when working with sequences in Wiener chaos.
Lemma 9 Let γ > 0 and p0 ∈ N. Moreover let (Zn )n∈N be a sequence of random variables.
If for every p > p0 there exists a constant cp > 0 such that for all n ∈ N,
(E|Zn |p )1/p 6 cp · n−γ ,
then for all ε > 0 there exists a random variable ηε such that
|Zn | 6 ηε · n−γ+ε
almost surely
for all n ∈ N. Moreover, E|ηε |p < ∞ for all p > 1.
As before we assume that ∆n = tk+1 − tk is a function of n only. Of some importance,
particularly for the purpose of proving normal convergence theorems, is the case n−α with a
given α ∈ R. The case α > 0 implies that the observation frequency must increase even as the
horizon itself also increases. The case of α = 0 is of special importance because it corresponds
to a setup where the observation frequency is fixed ( ∆n = 1, no in-fill asymptotics, only
increasing horizon), which may be desirable in some applications. We will see that for some
almost-sure convergence results, we may even take a time step ∆n which grows with n. In
other words, this allows for very sparse observations. We will also see that most almost-sure
results are valid for the entire range H ∈ ( 12 , 1), while normal convergence results require
H ∈ ( 21 , 34 ). We begin by recording and proving some important technical estimates.
√
S
Lemma 10 Define δn (X) := Tn Qn (X) − TTnn . Then
E δn2 (X) 6 c(H, θ, ρ) min n∆n2H+1 ,
1
+ ∆H+1
+ ∆n4H−3
n
n∆n
In particular, if ∆n 6 nα for some α ∈ (−∞, 1/H), then
n
X
j=1
j 4H−4 .
Qn (X) −→ η X
almost surely as n → ∞.
Proof. The points 3) and 6) of Lemma 21 lead to
Z tj
n Z
i
h
2
1 X ti
E Xt2 − Xt2i−1 Xs2 − Xt2j−1 dsdt
E δn (X) =
Tn
ti−1 tj−1
i,j=1
6 c(H, θ, ρ)n∆n2H+1 .
21
(36)
(37)
On the other hand, we have
E δn2 (X)
Z tj
n Z
1 X ti
=
E Xt2 − Xt2i−1 E Xs2 − Xt2j−1 dsdt
Tn
i,j=1 ti−1 tj−1
Z tj
n Z
1 X ti
+
E Xt − Xti−1 Xs − Xtj−1 E Xt + Xti−1 Xs + Xtj−1 dsdt
Tn
i,j=1 ti−1 tj−1
Z tj
n Z
1 X ti
+
E Xt − Xti−1 Xs + Xtj−1 E Xt + Xti−1 Xs − Xtj−1 dsdt
Tn
ti−1 tj−1
:=
1
Tn
i,j=1
n
X
(D1 (i, j) + D2 (i, j) + D3 (i, j)) .
i,j=1
By using the points 2), 4) and 6) of Lemma 21 we obtain
#2
" n Z
n
1 X
1 X ti
D1 (i, j) =
E Xt2 − Xt2i−1 dt
Tn
Tn
i,j=1
i=1 ti−1
#2
"
n
X
c(H, θ, ρ)
−ti−1 /2
e
∆n
6
Tn
i=1
2
∆n
c(H, θ, ρ)
,
6
Tn
1 − e−∆n /2
=
n
1 X
(D2 (i, i) + D3 (i, i))
Tn
i=1
!
Z ti Z ti h
n
i
X
1
2(E(Xt Xs ))2 + 2(E(Xt2i−1 ))2 − 4(E(Xti−1 Xs ))2 dsdt
Tn
ti−1 ti−1
i=1
6 c(H, θ, ρ)∆H+1
,
n
and
n
1 X
(D2 (i, j) + D3 (i, j))
Tn
i6=j=1
=
n
2 X
Tn
i6=j=1
6
Z
ti
ti−1
Z
ti
ti−1
(E(Xt Xs ))2 − (E(Xti−1 Xs ))2 − (E(Xti−1 Xt ))2 + (E(Xti−1 Xtj−1 ))2 dsdt
n
c(H, θ, ρ) X 4H−2
∆n
|j − i − 1|4H−4
Tn
i<j=1
6
c(H, θ, ρ)∆n4H−3
n
X
j 4H−4 .
j=1
Thus (36) is obtained. Now, using (13), Lemma 9 and (18) we will be able to assert the
convergence (37), and thus the entire lemma, as soon as we can show that the right-hand side
22
!
of (36) divided by Tn converges to 0 as fast as some negative power of n. Thus we only need
to show that there exists ε > 0, such that as n → ∞
n
H
X
1
∆n
1 4H−4
qn := min ∆2H
j 4H−4 6 n−ε .
n ,
2 + n + n ∆n
(n∆n )
j=1
Let us concentrate first on the second part of the
defining qn . This is the sum of
Pnminimum
−1 ∆4H−4
4H−4 . The first of these three terms
/n,
and
n
the three terms (n∆n )−2 , ∆H
j
n
n
j=1
will tend to 0 like a negative power of n as soon as there exists ε1 > 0 such that ∆n > n−1+ε1 .
The second term will tend to 0 like a negative power of n as soon as there exists ε2 > 0 such
that ∆n 6 n1/H−ε2 . For the third term,
P we must separate the case H < 3/4 from the case
H > 3/4. When H < 3/4, the series nj=1 j 4H−4 is bounded, so the last term in the second
part of the min in qn will tend to 0 like a negative power of n as soon as there exists ε3 > 0
such that ∆n > n−1/(4−4H)+ε3 . When H > 3/4, the series is bounded above by a constant
times n4H−3 , yielding a contribution of (∆n /n)4H−4 ; so the last term in the second part of
the min in qn will tend to 0 like a negative power of n as soon as there exists ε4 > 0 such that
∆n > n−1+ε4 . The case H = 3/4 is done in the same fashion, with the same conclusion as
when H > 3/4. Thus we have proved that for each fixed n, if there exist ε1 , ε2 , ε3 > 0 such
that
1
(38)
max n−1+ε1 , n− 4−4H +ε3 6 ∆n 6 n1/H−ε2
then for some ε > 0,
qn 6 n−ε .
On the other hand notice that for every H ∈ (1/2, 1), there exist ε1 , ε3 > 0 such that
1
max n−1+ε1 , n− 4−4H +ε3 6 n−1/2 .
(39)
Thus for each fixed n, if we have
∆n 6 n−1/2 ,
(40)
using the first part of the min in the definition of qn , we get
qn 6 n−ε
with ε = H . To conclude, by (39), for each fixed n, we are either in the case (38) or (40), so
that qn 6 n−ε in all cases as soon as ∆n 6 n1/H−ε2 for some ε2 > 0. The proof of the lemma
is complete.
RT
√
Lemma 11 Define δn (Σ) := Tn Qn (Σ) − T1n 0 n Σ2t dt . Then
n
X
1
E δn2 (Σ) 6 c(H, θ, ρ) min n∆n2H+1 ,
+ ∆H+1
+ ∆n4H−3
j 4H−4 .
n
n∆n
(41)
Qn (Σ) −→ η Σ
(42)
j=1
In particular, if ∆n 6 nα for some α ∈ (−∞, 1/H), then
23
almost surely as n → ∞.
→ 0 for some α > 0,
On the other hand if n1+α ∆H+1
n
b − Qn (Σ) −→ 0
Qn (Σ)
almost surely as n → ∞, where
b n = ∆n
Σ
ti
In addition, if n3 ∆n2H+3 → 0,
p
in L2 (Ω) as n → ∞.
i
X
(43)
Xtj−1 .
j=1
b − Qn (Σ) −→ 0
Tn Qn (Σ)
(44)
Proof. By using same arguments as in the proof of Lemma 10, (21) and (19), we obtain
(41) and (42).
Now, we prove the convergence (43). We can write
b =
Qn (Σ) − Qn (Σ)
2 2 X
−1 X
b
bt
Σ
−
Σ
Σti−1 − Σ
Σ
+
ti−1
ti−1 .
ti−1
i−1
n
n
n
n
i=1
i=1
Using the point 6) of Lemma 21
i−1 X
i−1 Z
2
X
bt
Σti−1 − Σ
=
E
i−1
j=1 k=1
tj
tj−1
6 c(H, θ, ρ)
Z
tk
E[(Xs − Xtj−1 )(Xr − Xtk−1 )]drds
tk−1
i−1 Z
X
j=1
tj
tj−1
6 c(H, θ, ρ) n∆H+1
n
2
2
|s − tj−1 |H ds
Then, by Hölder inequality and the point 3) of Lemma 21 we obtain for every p > 1
h
p i1/p
b
E Qn (Σ) − Qn (Σ)
6 c(H, θ, ρ) n2 ∆n2H+2 + n∆H+1
.
n
Thus, by (13), Lemma 9 and that fact that n1+α ∆H+1
→ 0 for some α > 0 the convergence
n
(43) is obtained.
Furthermore, it is also easy to see that the convergence (44) is satisfied.
4.1
Auxiliary estimators θ̃ and ρ̃
The first step
in constructing a discrete-observation-based estimator for which the asymptotics
b
of θT , ρbT studied in Section 3 can be helpful, is to consider the following two auxiliary
estimators θen and ρen of θ ∗ and ρ∗ respectively, by leaving the numerators in θbT , ρbT alone,
and discretizing the denominators:
R Tn
1
Tn 0 Xt δXt
e
θn = −
Qn (X)
24
and
R Tn
Vbtn δVbtn
ρen (Σ) = −
,
Qn (X) + (θen )2 Qn (Σ)
1
Tn
where
0
Vbtn = Xt + θbn Σt ,
(45)
0 6 t 6 Tn .
and we recall that Qn (Z) is a notation for the Riemann-sum rectangle approximation
We also consider the version of ρen (Σ) based only on discrete observations of Σ:
b =−
ρen (Σ)
R Tn
1
n
Pn
2
i=1 (Zti−1 ) .
Vbtn δVbtn
.
b
Qn (X) + (θen )2 Qn (Σ)
1
Tn
0
Combining Lemma 3 and the almost-sure convergence (37) we deduce the strong consistency of θen .
Theorem 12 Assume H ∈ (1/2, 1). If ∆n 6 nα for some α ∈ (−∞, 1/H), then
θen −→ θ ∗
almost surely as n → ∞.
b
By Lemmas 6 and 11 it is easy also to deduce the strong consistency of ρen (Σ) and ρen (Σ).
Theorem 13 Assume H ∈ (1/2, 1). If ∆n 6 nα for some α ∈ (−∞, 1/H), then
ρen (Σ) −→ ρ∗
almost surely as n → ∞.
→ 0 for some α > 0,
In addition, if n1+α ∆H+1
n
b −→ ρ∗
ρen (Σ)
almost surely as n → ∞.
To establish the asymptotic normality of θen , ρen , we can write
p
∗
e
Tn θn − θ =
STn
Tn
Qn
p
Tn (θbTn − θ ) +
∗
Similarly,
p
∗
Tn (e
ρn − ρ ) =
bT
L
n
Tn
bn
Q
p
Tn (b
ρTn
√
S
θ ∗ Tn ( TTnn − Qn (X))
Qn (X)
√
b
L
bn )
ρ∗ Tn ( TTnn − Q
−ρ )+
.
bn
Q
.
∗
Theorem 8 provides the convergence of the last summands in each of the two lines above.
Combining this with the convergences we obtained in Lemmas 10 and 11, we obtain the
following result.
25
Theorem 14 Let H ∈ ( 21 , 34 ) and n∆H+1
→ 0. Then
n
p
law
Tn θen − θ ∗ , ρen (Σ) − ρ∗ −→ N 0,t P Γ P
In addition if n3 ∆n2H+3 → 0,
p
law
b − ρ∗ −→
Tn θen − θ ∗ , ρen (Σ)
N 0,t P Γ P
where P the matrix defined in (35).
4.2
X and Σ are observed
The problem with the auxiliary estimators θen and ρen is that they still contain Skorohod
integrals. In order to devise a further scheme that allows us to evaluate them, at least approximately, using discrete data only, we begin by using the discrete observations of X and
Σ, and recalling that, from Lemmas 3 and 6, we have
Z
1 T
Xt δXt −→ −ρθ(ρ + θ)η X ,
T 0
Z
1 T b b
Vt δVt −→ −ρθ(ρ + θ)η Σ
T 0
where η X and η Σ , which are also functions of H, θ, ρ, are given in Lemma 2. Since these
limits depend on the parameters we are trying to estimate, one strategy is to rewrite the
strong consistency results of Theorems 12 and 13 for θen and ρen (Σ) as implicit definitions of
new estimators, where the numerators in the definitions of θen and ρen (Σ) are replaced by their
limits recalled above, and each instance of θ and ρ therein are replaced by the new estimator
we are trying to define. The same substitution must be done with the expressions θ ∗ and ρ∗ ,
since these are the limits of θen and ρen . In other words we consider only that the denominators
in θen and ρen contain data, and replace all other instances of (θ, ρ) in the limits in Theorems
12 and 13 by the pair of estimators we are trying to define. After some minor
manipulations,
this leads to the following definition of a new pair of estimators θ̌n , ρ̌n as solution of the
system of the following two equations, if it exists:
HΓ(2H)[(ρ̌n )2−2H −(θ̌n )2−2H ]
θ̌n + ρ̌n =
(ρ̌n −θ̌n )Qn (X)
2
.
(46)
(θ̌n ) −(ρ̌n )2
HΓ(2H)
h
(θ̌ )2−2H −(ρ̌ )2−2H + ρ̌ +θ̌ 2 (ρ̌ )−2H −(θ̌ )−2H i = Qn (X)+(θ̌n +ρ̌n )2 Qn (Σ)
)
( n n) ( n
n
n
n
We emphasize that the above is an implicit definition of θ̌n , ρ̌n . It is also rather opaque.
The system (46) can be simplified slightly using more elementary manipulations. We find that
the definition of θ̌n , ρ̌n is equivalent to the following:
F θ̌n , ρ̌n = (Qn (X), Qn (Σ))
where F is a positive function of the variables (x, y) in (0, +∞)2 defined by: for every (x, y) ∈
(0, +∞)2
1
2−2H − x2−2H , x−2H − y −2H
if x 6= y
2 −x2 y
y
(47)
F (x, y) = HΓ(2H) ×
−2H
−2H−2
(1 − H)x
, Hx
if x = y.
26
Interestingly, this shows that a good candidate for the discrete version of the least-squares
estimator of (θ, ρ) is none other than a type of generalized method of moments estimator
RT
obtained via Lemma 2 after discretizing the expressions ST (X) := T −1 0 Xs2 ds and ST (Σ) :=
RT
T −1 0 Σ2s ds. We now consider the question whether System (46) has a unique solution
θ̌n2 , ρ̌2n , and how this may imply strong consistency for these estimators.
Since for every (x, y) ∈ (0, +∞)2 with x 6= y
(1−H)x1−2H
JF (x, y) = Γ(2H+1)
(x2 −y2 )−x(x2−2H −y2−2H )
(x2 −y 2 )2
Hx−2H−1 (x2 −y 2 )+x(x−2H −y −2H )
(x2 −y 2 )2
(1−H)y 1−2H (y 2 −x2 )−y (y 2−2H −x2−2H )
(x2 −y 2 )2
Hy −2H−1 (y 2 −x2 )+y (y −2H −x−2H )
(x2 −y 2 )2
the determinant of JF (x, y) is non-zero on in (0, +∞)2 . So, F is a diffeomorphism in (0, +∞)2
and its inverse G has a Jacobian
Hy−2H−1 y2 −x2 +y y−2H −x−2H
(1−H)y 1−2H (y 2 −x2 )−y (y 2−2H −x2−2H )
)
) (
(
−
Γ(2H + 1)
(x2 −y 2 )2
(x2 −y 2 )2
;
JG (a, b) =
1−2H x2 −y 2 −x x2−2H −y 2−2H
−2H−1 x2 −y 2 +x x−2H −y −2H
(1−H)x
Hx
)
) (
(
)
)
(
(
det JF (x, y) −
(x2 −y 2 )2
(x2 −y 2 )2
where (x, y) = G (a, b).
Hence, (37) and (42) lead to
θ̌n , ρ̌n = G (Qn (X), Qn (Σ)) −→ G η X , η Σ = (θ, ρ)
almost surely as n → ∞ as soon as ∆n 6 nα for some α ∈ (−∞, 1/H). Summarizing, we have
proved the following.
Theorem 15 Let H ∈ (1/2, 1) and assume that ∆n 6 nα for some α ∈ (−∞, 1/H). Then,
as n −→ ∞
θ̌n , ρ̌n −→ (θ, ρ)
almost surely.
We may now prove a normal convergence result for θ̌n , ρ̌n based on Theorem 14. Note
that the second part of Theorem 14 is not needed here because we rely on fully observed Σ in
this section.
Theorem 16 Suppose that H ∈ ( 12 , 43 ) and n∆H+1
→ 0. Then
n
p
law
Tn θ̌n − θ, ρ̌n − ρ −→ N (0,t M t P Γ P M )
where the matrices Γ, P and M are defined respectively in (A-5), (35) and (50).
Proof. We have
θen = −
and
ρen (Σ) = −
1
Tn
R Tn
0
Xt δXt
Qn (X)
R Tn
:=
Jnθ (X)
Qn (X)
Vbtn δVbtn
Jnρ (Vb )
:=
.
Qn (X) + (θen )2 Qn (Σ)
Qn (X) + (θen )2 Qn (Σ)
1
Tn
0
27
Then, we can write
p
Tn θ̌n − θ, ρ̌n − ρ)
!!
!
Jnθ (X) e −2 Jnρ (Vb ) Jnθ (X)
, (θn )
− (θ, ρ)
=
−
Tn G
ρen (Σ)
θen
θen
p
=
Tn GoL Jnθ (X), Jnρ (Vb ), θen , ρen (Σ) − (θ, ρ)
p
where L(r, s, u, v) = ur , us2 v − ur3 .
On the other hand for any ε ∈ (0, 1)
Z
1
T
T
0
Z
t
0
r 2H−2 e−r drdt = Γ(2H − 1) + o(
1
)
Tε
because
1
T 1−ε
Z
T
0
Z
∞
r
t
2H−2 −r
e
1
(1 − e−T /2 )
drdt 6
2T 1−ε
→ 0.
Z
∞
r 2H−2 e−r/2 dr
0
Combining this together with (23) and the point 5) of Lemma 21 we can write
1
(48)
Jnθ (X) = (ρ + θ)η X + o( √ )
Tn
√
) denotes a random variable such that Tn o( √1T ) converges to zero almost surely
where o( √1T
n
as Tn → ∞.
Similar argument leads to
n
1
Jnρ (Vb ) = ρθ(ρ + θ)η Σ + o( √ ).
Tn
Since
GoL (ρ + θ)η X , ρθ(ρ + θ)η Σ , θ ∗ , ρ∗ = (θ, ρ)
(49)
we can write
p
Tn θ̌n − θ, ρ̌n − ρ)
p
Tn GoL Jnθ (X), Jnρ (Vb ), θen , ρen (Σ) − GoL (ρ + θ)η X , ρθ(ρ + θ)η Σ , θ ∗ , ρ∗
=
p h
=
Tn GoL Jnθ (X), Jnρ (Vb ), θen , ρen (Σ) − GoL (ρ + θ)η X , ρθ(ρ + θ)η Σ , θen , ρen (Σ)
i
+ GoL (ρ + θ)η X , −ρθ(ρ + θ)η Σ , θen , ρen (Σ) − GoL (ρ + θ)η X , ρθ(ρ + θ)η Σ , θ ∗ , ρ∗
:=
s n + rn .
From (48) and (49) we obtain sn −→ 0 almost surely as n → ∞.
On the other hand, by Taylor’s formula
p
rn = Tn θen − θ ∗ , ρen (Σ) − ρ∗ M + dn
28
where
M=
∂h1
∗ ∗
∂u (θ , ρ )
∂h1 ∗ ∗
∂v (θ , ρ )
∂h2 ∗ ∗
∂u (θ , ρ )
∂h2 ∗ ∗
∂v (θ , ρ )
(50)
with
We can write
h(u, v) = (h1 , h2 )(u, v) = GoL (ρ + θ)η X , ρθ(ρ + θ)η Σ , u, v .
h(u, v) = (G1 , G2 ) og (u, v)
where
g(u, v) = (g1 , g2 )(u, v) = L (ρ + θ)η X , ρθ(ρ + θ)η Σ , u, v .
Moreover for i = 1, 2
∂hi
∂Gi
∂g1
∂Gi
∂g2
(u, v) =
(g(u, v))
(u, v) +
(g(u, v))
(u, v)
∂u
∂a
∂u
∂b
∂u
and
∂Gi
∂g1
∂Gi
∂g2
∂hi
(u, v) =
(g(u, v))
(u, v) +
(g(u, v))
(u, v).
∂v
∂a
∂v
∂b
∂v
On the other hand, dn converges in distribution to zero, because
p
kdn k 6 c(H, θ, ρ) Tn k(e
ρn (Σ) − θ ∗ , θen − ρ∗ )k2 .
It is elementary that if for any ω ∈ Ω there exists n0 (ω) ∈ N such that Xn (ω) = Yn (ω) for all
law
law
n > n0 (ω) and Xn −→ 0 as n → ∞, then Yn −→ 0 as n → ∞.
Combining this with Theorem 14 the proof is completed.
4.3
X is observed
In the previous section, we encountered theorems in which X and Σ are both assumed to be
fully observed in discrete time. Since Σ is the time-antiderivative of X, such an assumption
corresponds, for instance, to the physical situation where X is the velocity of a particle, and
Σ is its position.
In this section, we abandon such a framework, and assume instead
that only X is observed
in discrete time. Thus we consider the following pair of estimators θ̆n , ρ̆n :
b
θ̆n , ρ̆n = G Qn (X), Qn (Σ)
where the deterministic explicit function G was identified
in the previous section as the inverse
of the function F given in (47). Equivalently, θ̆n , ρ̆n is the solution of the system (46), or
b which relies only on observations
its equivalent form (47), with Σ replaced by the process Σ,
of X. Using same arguments as in Section 4.2 and Lemma 11, but relying now on the second
part of Theorem 13 (hence the stronger condition on ∆n for the strong consistency result)
and the second part of Theorem 14 (hence the stronger condition on ∆n for the convergence
in law result), we conclude the following.
29
Theorem 17 If n1+α ∆H+1
→ 0 for some α > 0,
n
θ̆n , ρ̆n −→ (θ, ρ)
almost surely as n → ∞.
Theorem 18 Let H ∈ ( 12 , 34 ). If n3 ∆n2H+3 → 0, then, as n → ∞
p
law
Tn θ̆n − θ, ρ̆n − ρ −→ N (0,t M t P Γ P M )
where the matrices Γ, P and M are defined respectively in (A-5), (35) and (50).
4.4
X and V are observed
When both X and V are observed, the estimator of θ ∗ based on continuous data is θbT given
in (9) but the estimator of ρ becomes the usual full-observation estimator of an fBm-driven
Ornstein-Uhlenbeck process as in [14], i.e.
RT
Vt δVt
ρT = − R0T
.
2
0 Vt dt
Following similar arguments as in the beginning of Section 4.2, the
natural candidate for the
estimator based on discrete data of X and V is the pair θn , ρn defined as the solution of
the following system:
HΓ(2H)
2−2H −(θ )2−2H ]
n
θ + ρ = (ρn −θn ) [(ρn )
n
Qn (X)
n
1
ρ = HΓ(2H) 2H .
Qn (V )
n
We see that ρn is defined explicitly autonomously via the discrete-data-based statistic Qn (V ).
With ρn now known, elementary manipulations yield that θn is precisely the solution of the
following simple equation
2−2H Q (X) 2
Qn (X)
n
2−2H
2
−
(θn )
−
ρn .
(θn ) = ρn
HΓ(2H)
HΓ(2H)
Define
F (x, y) = HΓ(2H) ×
(
y 2−2H −x2−2H
, y −2H
y 2 −x2
(1 − H)x−2H , x−2H
if x 6= y
if x = y.
Its Jacobian is given, for every (x, y) ∈ (0, +∞)2 such that x 6= y, by
JF (x, y) = Γ(2H+1)
(1−H)x1−2H (x2 −y 2 )−x(x2−2H −y 2−2H )
(x2 −y 2 )2
0
(1−H)y 1−2H (y 2 −x2 )−y (y 2−2H −x2−2H )
(x2 −y 2 )2
−y −2H−1
!
Thus the Jacobian of is inverse G is as follows
1−2H y 2 −x2 −y y 2−2H −x2−2H
)
) (
(
−2H−1 − (1−H)y
−y
Γ(2H + 1)
(x2 −y 2 )2
; (x, y) = G (a, b) .
JG (a, b) =
(1−H)x1−2H (x2 −y 2 )−x(x2−2H −y 2−2H )
det JF (x, y)
0
(x2 −y 2 )2
Using same arguments as in Section 4.2 we obtain
30
.
Theorem 19 Assume that ∆n 6 nα for some α ∈ (−∞, 1/H). Then, as n −→ ∞
θn , ρn −→ (θ, ρ)
almost surely.
Theorem 20 Suppose that H ∈ ( 12 , 43 ) and n∆H+1
→ 0. Then, as n → ∞
n
p
law
Tn θn − θ, ρn − ρ −→ N (0,t Q t P Γ P Q)
where Γ and P are defined respectively in (A-5) and (35), and where
Q=
∂f1
∂f2 ∗
(θ
,
ρ)
∂u
∂f2 ∗
∂v (θ , ρ)
∗
∂u (θ , ρ)
∂f1 ∗
∂v (θ , ρ)
such that
(f1 , f2 )(u, v) = Gol (u, v)
with
l(u, v) = L (ρ + θ)η X , ρ1−2H , u, v
and
L(r, s, u, v) :=
5
r s
.
,
u v
Appendix
In this appendix, we present some calculations used in the paper.
Fix T > 0. Let f, g : [0, T ] −→ R be Hölder continuous functions of orders α ∈ (0, 1) and
β ∈ (0, 1) respectively with α + β > 1. Young [29] proved that the Riemann-Stieltjes integral
RT
(so-called Young integral) 0 fs dgs exists. Moreover, if α = β ∈ ( 12 , 1) and φ : R2 −→ R is
R . ∂φ
R.
a function of class C 1 , the integrals 0 ∂φ
∂f (fu , gu )dfu and 0 ∂g (fu , gu )dgu exist in the Young
sense and the following chain rule holds:
Z t
Z t
∂φ
∂φ
(fu , gu )dfu +
(fu , gu )dgu , 0 6 t 6 T.
(A-1)
φ(ft , gt ) = φ(f0 , g0 ) +
∂f
0 ∂g
0
As a consequence, if H ∈ ( 12 , 1) and (ut , t ∈ [0, T ]) is a process with Hölder paths of order
RT
α ∈ (1 − H, 1), the integral 0 us dBsH is well-defined as a Young integral. Suppose moreover
that for any t ∈ [0, T ], ut ∈ D 1,2 , and
P
Z
T
0
Z
0
T
2H−2
|Ds ut ||t − s|
Then, by [1], u ∈ Domδ and for every t ∈ [0, T ],
Z
t
0
us dBsH
=
Z
t
0
us δBsH
+ H(2H − 1)
31
dsdt < ∞
Z tZ
0
0
t
= 1.
Ds ur |s − r|2H−2 drds.
(A-2)
In particular, when ϕ is a non-random Hölder continuous function of order α ∈ (1 − H, 1), we
obtain
Z T
Z T
ϕs δBsH = B H (ϕ).
(A-3)
ϕs dBsH =
0
0
In addition, for all ϕ, ψ ∈ |H|,
Z T
Z T
Z
H
H
ϕs dBs
ψs dBs = H(2H − 1)
E
0
0
T
0
Z
0
T
ϕ(u)ψ(v)|u − v|2H−2 dudv.
Lemma 21 Let m, m′ > 0 and let X m be the process defined in (15). Then,
R∞R∞
′
1−2H + m′ 1−2H ,
m
1) λ(m, m′ ) := H(2H −1) 0 0 e−ms e−m r |s−r|2H−2 drds = HΓ(2H)
m+m′
′
2) 0 6 λ(m, m′ ) − E(Xtm Xtm ) 6 c(H, m, m′ )e−t/2 ,
3) supt>0 E[|Xtm |p ] 6 c(H, m, p) < ∞,
′
4) 0 6 E(Xtm Xsm ) 6 c(H, m, m′ )|t − s|2H−2 ,
5) For every ε > 0,
XTm
Tε
→ 0 almost surely as T → ∞,
6) E(|Xtm − Xsm |p ) 6 c(H, m, p)|t − s|pH .
Proof. To prove equality 1), we just write
′
Z
∞Z ∞
′
e−ms e−m r |s − r|2H−2 drds
0
0
Z ∞
Z s
′
dse−ms
= H(2H − 1)
dre−m r |s − r|2H−2
0
0
Z ∞
Z ∞
′
−ms
dse
dre−m r |s − r|2H−2
+H(2H − 1)
λ(m, m ) = H(2H − 1)
0
=
s
HΓ(2H) 1−2H
′ 1−2H
.
m
+
m
m + m′
For the point 2) see [15]. For 3) and 6) we refer to [16], and for 4) and 5) see [14, Lemma 5.2
and Lemma 5.4]
Theorem 22 Let H ∈ ( 21 , 34 ). Define for m > 0
fTm(u, v) :=
Then, as T −→ ∞,
1 −m|u−v| ⊗2
e
11{[0,T ]} (u, v).
2
1
law
√
I2 (fTθ ), I2 (fTρ ) −→ N (0, Γ)
T
(A-4)
where Γ is a symmetric nonnegative definite matrix which has the following explicit expression
l1 l3
Γ = ηH
(A-5)
l3 l2
32
where l1 = θ 1−4H , l2 = ρ1−4H , l3 =
ηH
2ρθ
(4H−1)(ρ2 −θ 2 )
θ 1−4H − ρ1−4H and
Γ(2H)Γ(3 − 4H)Γ(4H − 1)
.
= H (4H − 1) Γ(2H) +
Γ(2 − 2H)
2
2
Proof. Notice that for (A-4) to hold it suffices that prove that for every a, b ∈ R,
GT := aI2 (fTθ ) + bI2 (fTρ )
converges in law to N 0, (a, b)Γt (a, b) as T −→ ∞.
Fix a, b ∈ R. Since GT is a multiple integral, by the isometry property of double stochastic
integral I2 , we get the variance of GT as follows
2
1
IT3
α2H
2 IT
2 IT
2
+b
+ 2ab
a
,
EGT =
2
T
T
T
where
IT1
IT2
IT3
=
=
=
Z
Z
[0,T ]4
e−θ|t−s| e−θ|u−v| |t − u|2H−2 |s − v|2H−2 dtdsdudv,
[0,T ]4
e−ρ|t−s| e−ρ|u−v| |t − u|2H−2 |s − v|2H−2 dtdsdudv,
[0,T ]4
e−ρ|t−s| e−θ|u−v| |t − u|2H−2 |s − v|2H−2 dtdsdudv.
Z
Using the same argument as in the proof of [14, Theorem 3.4], we have as T → ∞
limT →∞
α2H IT1
= ηH l1
2 T
limT →∞
α2H IT2
= ηH l2 .
2 T
and
Now, let us estimate IT3 . We have
#
"Z
dIT3
e−ρ(T −s) e−θ|u−v| (T − u)2H−2 |s − v|2H−2 dsdudv
= 2
dT
3
[0,T ]
#
"Z
−ρ|t−s| −θ(T −v)
2H−2
2H−2
e
e
(T − t)
|s − v|
dsdudv
+2
[0,T ]3
:=
AT (ρ, θ) + AT (θ, ρ).
Making the change of variables T − s = x , T − u = y and T − v = z
Z
e−ρx e−θ|y−z| y 2H−2 |z − x|2H−2 dxdydz.
AT (ρ, θ) = 2
[0,T ]3
This implies
limT →∞ AT (ρ, θ) = A∞ (ρ, θ) = 2
Z
[0,∞]3
33
e−ρx e−θ|y−z| y 2H−2 |z − x|2H−2 dxdydz.
Making the change of variables z − x = w, we obtain
"Z
Z
∞
A∞ (ρ, θ) = 2
= 2
[0,∞)2
"Z
[0,∞)2
+2
"Z
−x
Z
e−ρx e−θ|y−w−x|y 2H−2 |w|2H−2 dwdxdy
y−x
−x
(0,∞)2
Z
#
e−ρx e−θ(y−w−x) y 2H−2 |w|2H−2 dwdxdy
∞
y−x
#
#
e−ρx eθ(y−w−x) y 2H−2 |w|2H−2 dwdxdy .
Integrating in x we get
2
(−ρ + θ)
Z
Z
(e−ρ(y+w) − e−ρw e−θy )y 2H−2 w2H−2 dwdy
∞ Z +∞
(e(−ρ+θ)(y−w) − e(−ρ+θ)[(−w)∨0] )1[(y−w)−((−w)∨0)]+
0
−∞
−θ(y−w) 2H−2
2H−2
×e
y
|w|
dwdy
Z ∞ Z +∞
2
−
e(−θ−ρ)[(y−w)∨0] eθ(y−w) y 2H−2 |w|2H−2 dwdy
(−ρ − θ) 0
−∞
2
2
A1 (ρ, θ) +
A2 (ρ, θ).
=
θ−ρ
ρ+θ
A∞ (ρ, θ) =
Furthermore
1
A (ρ, θ) =
0
∞Z ∞
Z
0
∞Z y
(e−ρ(y−w) − e−θ(y−w) )y 2H−2 w2H−2 dwdy
Z ∞
2 Z ∞
2H−2 −ρw
2H−2 −θy
2H−2 −ρy
w
e
dw
y
e dy
y
e dy −
=
0
0
0
Z ∞Z ∞
+
(e−ρ(y−w) − e−θ(y−w) )y 2H−2 w2H−2 dydw
0
w
= Γ(2H − 1)2 ρ1−2H ρ1−2H − θ 1−2H
Z ∞Z ∞
+
(e−ρx − e−θx )(x + w)2H−2 w2H−2 dxdw.
+
0
Z ∞
0
Using (x + w)2H−2 =
0
0
1
Γ(2−2H)
R∞
0
ξ 1−2H e−ξ(w+x) dξ, the term A1 becomes
A1 (ρ, θ) = Γ(2H − 1)2 ρ1−2H ρ1−2H − θ 1−2H
Z Z
Γ(2H − 1) ∞ ∞ 2−4H −ξx −ρx
ξ
e (e
− e−θx )(dξdx
+
Γ(2 − 2H) 0
0
= Γ(2H − 1)2 ρ1−2H ρ1−2H − θ 1−2H
Γ(2H − 1)Γ(3 − 4H)Γ(4H − 2) 2−4H
+
ρ
− θ 2−4H .
Γ(2 − 2H)
34
Similarly, we obtain
A2 (ρ, θ) =
Z
0
+
+
∞Z ∞
Z
Z
0
e−ρ(y+w) y 2H−2 w2H−2 dwdy
0
∞Z y
e−ρ(y−w) y 2H−2 w2H−2 dwdy
0
∞Z ∞
0
2−4H
eθ(y−w) y 2H−2 w2H−2 dwdy
y
Γ(2H − 1)2
Γ(2H − 1)Γ(3 − 4H)Γ(4H − 2) 2−4H
+
ρ
+ θ 2−4H .
Γ(2 − 2H)
= ρ
Thus, A∞ (θ, ρ) is also obtained.
Consequently
limT →∞ α2H
IT3
T
= α2H (A∞ (ρ, θ) + A∞ (θ, ρ))
= 2ηH l3 .
Finally, combining the above convegences we deduce that as T → ∞
EG2T −→ (a, b) Γ t (a, b).
On the other hand
Ds GT
Z s
Z T
1
√
ae−θ(t−s) + be−ρ(t−s) δBt
ae−θ(s−t) + be−ρ(s−t) δBt +
T
s
0
1 a,b
a,b
√
Xs + Ys,T .
T
=
:=
Hence
kDGT k2H
=
=
:=
Z Z
αH T T a,b
a,b
a,b
|r − s|2H−2 dsdr
Xra,b + Yr,T
Xs + Ys,T
T 0 0
Z Z
αH T T a,b a,b
a,b
a,b a,b
Xs Xr + 2Xra,b Ys,T
+ Ys,T
Yr,T |r − s|2H−2 dsdr
T 0 0
αH a,b
AT + BTa,b + CTa,b .
T
Since Xsa,b belongs to the first Wiener chaos of B H ,
Z
2
a,b
a,b
E(Xsa,b Xra,b )E(Xua,b Xva,b )|u − r|2H−2 |v − s|2H−2 dsdrdudv.
=2
E AT − EAT
[0,T ]4
Using similar arguments as in [14, Lemma 5.4 of web-only Appendix],
Z
2
a,b
a,b
|s − r|2H−2 |v − u|2H−2 |u − r|2H−2 |v − s|2H−2 dsdrdudv
6 c(H, θ, ρ)
E AT − EAT
[0,T ]4
6
c(H, θ, ρ)
T 4−8H
Z
[0,1]4
|s − r|2H−2 |v − u|2H−2 |u − r|2H−2 |v − s|2H−2 dsdrdudv.
35
Using the same argument for BTa,b and CTa,b we conclude that
2
E kDGT k2H − EkDGT k2H
Z
c(H, θ, ρ)
6
|s − r|2H−2 |v − u|2H−2 |u − r|2H−2 |v − s|2H−2 dsdrdudv
T 6−8H [0,1]4
−→ 0
as T −→ ∞, because H < 34 . This completes the proof of Theorem 22.
References
[1] Alòs, E. and Nualart, D. (2003). Stochastic integration with respect to the fractional
Brownian motion. Stoch. stoch. reports 75, no. 3, 129-152.
[2] Azmoodeh, E. and Morlanes, I. (2013). Drift parameter estimation for fractional OrnsteinUhlenbeck process of the second kind. Statistics: A Journal of Theoretical and Applied
Statistics. DOI: 10.1080/02331888.2013.863888.
[3] Azmoodeh, E. and Viitasaari, L. (2014). Parameter estimation based on discrete observations of fractional Ornstein-Uhlenbeck process of the second kind. Statistical Inference
for Stochastic Processes. DOI 10.1007/s11203-014-9111-8.
[4] Baraboza Chinchilla, L. A. (2012). New methods of estimation of long-memory models
with an application in climatology. Ph.D. Thesis, Purdue University.
[5] Belfadli, R., Es-Sebaiy, K., and Ouknine, Y. (2011). Parameter Estimation for Fractional
Ornstein-Uhlenbeck Processes: Non-Ergodic Case. Frontiers in Science and Engineering
(An International Journal Edited by Hassan II Academy of Science and Technology). 1,
no. 1, 1-16.
[6] Bercu, B., Proia, F., and Savy, N. (2014). On Ornstein-Uhlenbeck driven by OrnsteinUhlenbeck processes. Statistics and Probability Letters, 85, pp. 36-44.
[7] Brouste, A., Iacus, S. M. (2012). Parameter estimation for the discretely observed fractional Ornstein-Uhlenbeck process and the Yuima R package. Computational Statistics.
DOI 10.1007/s00180-012-0365-6.
[8] Cénac, P. and Es-Sebaiy, K. (2012). Almost sure central limit theorems for random
ratios and applications to LSE for fractional Ornstein-Uhlenbeck processes. Preprint :
http://arxiv.org/abs/1209.0137.
[9] Chronopoulou, A., Viens, F. (2012). Stochastic volatility models with long-memory in
discrete and continuous time. Quantitative Finance, 12 no. 4, 635-649.
[10] Comte, F. and Renault, E. (1998). Long memory in continuous time stochastic volatility
models. Mathematical Finance, 8, 291-323.
[11] Es-Sebaiy, K. (2013). Berry-Esséen bounds for the least squares estimator for discretely
observed fractional Ornstein-Uhlenbeck processes. Statistics and Probability Letters 83,
2372-2385.
36
[12] Fouque, J.-P., Papanicolaou, G., Sircar, K.R. (2000). Derivatives in Financial Markets
with Stochastic Volatility. Cambridge U.P.
[13] Hairer, M.(2005). Ergodicity of Stochastic Differential Equations Driven by Fractional
Brownian Motion. The Annals of Probability, 33, No. 2, 703-758.
[14] Hu, Y. and Nualart, D. (2010). Parameter estimation for fractional Ornstein-Uhlenbeck
processes. Statist. Probab. Lett. 80, 1030-1038.
[15] Hu, Y. and Song, J. (2013). Parameter estimation for fractional Ornstein- Uhlenbeck processes with discrete observations. F. Viens et al (eds), Malliavin Calculus and Stochastic
Analysis:A Festschrift in Honor of David Nualart, Springer Proceedings in Mathematics
and Statistics 34, 427-442.
[16] Garrido-Atienza, M., Kloeden, P., and Neuenkirch, A. (2009): Discretization of stationary
solutions of stochastic systems driven by fractional Brownian motion. Appl. Math. Optim.
60, no. 2, 151-172.
[17] Kleptsyna, M. and Le Breton, A. (2002). Statistical analysis of the fractional OrnsteinUhlenbeck type process. Statistical Inference for Stochastic Processes 5, 229-241.
[18] Kloeden, P. and Neuenkirch, A. (2007). The pathwise convergence of approximation
schemes for stochastic differential equations. LMS J. Comp. Math. 10, 235-253.
[19] Liptser, R. S. and Shiryaev, A. N. (1978). Statistics of Random Processes. Springer-Verlag,
New York, 1978.
[20] Neuenkirch, A. Tindel, S. (2014). A least square-type procedure for parameter estimation
in stochastic differential equations with additive fractional noise. Statistical Inference for
Stochastic Processes 17, no. 1, 99-120.
[21] Newey, W.K. and McFadden, D. (1994). Large sample estimation and hypothesis testing.
In Handbook of Econometrics (Robert F Engle and Daniel L McFadden, eds.), 4, 21112245. Elsevier.
[22] Nourdin, I; Peccati, G. (2012). Normal Approximations with Malliavin Calculus: From
Stein’s Method to Universality. Cambridge UP.
[23] Nualart, D. (2006). The Malliavin calculus and related topics. Springer-Verlag, Berlin,
second edition.
[24] Nualart, D. and Ortiz-Latorre, S. Central limit theorems for multiple stochastic integrals
and Malliavin calculus. Stochastic Process. Appl. 118 (2008) 614–628.
[25] Pipiras, V. and Taqqu, M.S. (2000). Integration questions related to fractional Brownian
motion. Probab. Theory Rel. Fields. 118, no. 2, 251-291.
[26] Prakasa Rao, B. L. S. (2010). Statistical inference for fractional diffusion processes. Wiley
Series in Probability and Statistics, Chichester, John Wiley & Sons.
[27] Rosenbaum, M. (2008). Estimation of the volatility persistence in a discretely observed
diffusion model. Stochastic Processes and their Applications, 118 (8), 1434-1462
37
[28] Tudor, C. Viens, F. (2007). Statistical aspects of the fractional stochastic calculus. Ann.
Statist. 35, no. 3, 1183-1212.
[29] Young, L. C. (1936). An inequality of the Hölder type connected with Stieltjes integration.
Acta Math. 67, 251-282.
38
| 10math.ST
|
Compiling ER Specifications into Declarative
Programs⋆
arXiv:0711.0348v1 [cs.PL] 2 Nov 2007
Bernd Braßel
Michael Hanus
Marion Müller
Institut für Informatik, CAU Kiel, D-24098 Kiel, Germany.
{bbr|mh|mam}@informatik.uni-kiel.de
Abstract. This paper proposes an environment to support high-level
database programming in a declarative programming language. In order
to ensure safe database updates, all access and update operations related
to the database are generated from high-level descriptions in the entityrelationship (ER) model. We propose a representation of ER diagrams
in the declarative language Curry so that they can be constructed by
various tools and then translated into this representation. Furthermore,
we have implemented a compiler from this representation into a Curry
program that provides access and update operations based on a highlevel API for database programming.
1
Motivation
Many applications in the real world need databases to store the data they process. Thus, programming languages for such applications must also support some
mechanism to organize the access to databases. This can be done in a way that
is largely independent on the underlying programming language, e.g., by passing SQL statements as strings to some database connection. However, it is well
known that such a loose coupling is the source of security leaks, in particular, in
web applications [16]. Thus, a tight connection or amalgamation of the database
access into the programming language should be preferred.
In principle, logic programming provides a natural framework for connecting databases (e.g., see [5,7]) since relations stored in a relational database can
be considered as facts defining a predicate of a logic program. Unfortunately,
the well-developed theory in this area is not accompanied by practical implementations. For instance, distributions of Prolog implementations rarely come
with a standard interface to relational databases. An exception is Ciao Prolog
that has a persistence module [4] that allows the declaration of predicates where
the facts are persistently stored, e.g., in a relational database. This module supports a simple method to query the relational database, but updates are handled
by predicates with side effects and transactions are not explicitly supported. A
similar concept but with a clear separation between queries, updates, and transactions has been proposed in [11] for the multi-paradigm declarative language
⋆
This work was partially supported by the German Research Council (DFG) under
grant Ha 2457/5-2.
Curry [8,15]. This will be the basis for the current framework that provides an
environment for high-level programming with databases. The objectives of this
work are:
– The methods to access and update the database should be expressed by
language features rather than passing SQL strings around.
– Queries to the database should be clearly separated from updates that might
change the outcome of queries.
– Safe transactions, i.e., sequence of updates that keep some integrity constraints, should be supported.
– The necessary code for these operations should be derived from specifications
whenever possible in order to obtain more reliable applications.
For this purpose, we define an API for database programming in Curry that
abstracts from the concrete methods to access a given database by providing
abstract operations for this purpose. In particular, this API exploits the type
system of Curry in order to ensure a strict separation between queries and updates. This is described in detail in Section 2. To specify the logical structure of
the data to be stored in a database, we use the entity-relationship (ER) model
[3], which is well established for this purpose. In order to be largely independent of concrete specification tools, we define a representation of ER diagrams
in Curry so that concrete ER specification tools can be connected by defining
a translator from the format used in these tools into this Curry representation.
This representation is described in Section 3. Finally, we develop a compiler
that translates an ER specification into a Curry module that contains access
and update operations and operations to check integrity constraints according
to the ER specification. The generated code uses the database API described in
Section 2. The compilation method is sketched in Section 4. Finally, Section 5
contains our conclusions.
2
Database Programming in Curry
We assume basic familiarity with functional logic programming (see [13] for a
recent survey) and Curry [8,15] so that we give in the following only a short
sketch of the basic concepts relevant for this paper.
Functional logic languages integrate the most important features of functional and logic languages to provide a variety of programming concepts to the
programmer. For instance, the concepts of demand-driven evaluation, higherorder functions, and polymorphic typing from functional programming are combined with logic programming features like computing with partial information
(logic variables), constraint solving, and non-deterministic search for solutions.
This combination, supported by optimal evaluation strategies [1] and new design
patterns [2], leads to better abstractions in application programs such as implementing graphical user interfaces [9], programming dynamic web pages [10,12],
or access and manipulation of persistent data possibly stored in databases [6,11].
As a concrete functional logic language, we use Curry in our framework but
it should be possible to apply the same ideas also to other functional logic languages, e.g., TOY [17]. From a syntactic point of view, a Curry program is a
functional program extended by the possible inclusion of free (logic) variables in
conditions and right-hand sides of defining rules. Curry has a Haskell-like syntax
[18], i.e., (type) variables and function names usually start with lowercase letters and the names of type and data constructors start with an uppercase letter.
The application of f to e is denoted by juxtaposition (“f e”). A Curry program
consists of the definition of functions and data types on which the functions
operate. Functions are first-class citizens as in Haskell and are evaluated lazily.
To provide the full power of logic programming, functions can be called with
partially instantiated arguments and defined by conditional equations with constraints in the conditions. Function calls with free variables are evaluated by a
possibly nondeterministic instantiation of demanded arguments (i.e., arguments
whose values are necessary to decide the applicability of a rule) to the required
values in order to apply a rule.
Example 1. The following Curry program defines the data types of Boolean values, “possible” (maybe) values, union of two types, and polymorphic lists (first
four lines) and functions for computing the concatenation of lists and the last
element of a list:
data Bool
= True
| False
data Maybe a
= Nothing | Just a
data Either a b = Left a | Right b
data List a
= []
| a : List a
conc :: [a] -> [a] -> [a]
conc []
ys = ys
conc (x:xs) ys = x : conc xs ys
last :: [a] -> a
last xs | conc ys [x] =:= xs
= x
where x,ys free
The data type declarations define True and False as the Boolean constants,
Nothing and Just as the constructors for possible values (where Nothing is
considered as no value), Left and Right to inject values into a union (Either)
type, and [] (empty list) and : (non-empty list) as the constructors for polymorphic lists (a and b are type variables ranging over all types and the type
“List a” is usually written as [a] for conformity with Haskell).
Curry also offers other standard features of functional languages, like higherorder functions (e.g., “\x -> e” denotes an anonymous function that assigns to
each x the value of e), modules, or monadic I/O [19]. For instance, an operation of
type “IO t” is an I/O action, i.e., a computation that interacts with the “external
world” and returns a value of type t. Thus, purely declarative computations are
distinguished from I/O actions by their types so that they cannot be freely
mixed.
Logic programming is supported by admitting function calls with free variables (see “conc ys [x]” above) and constraints in the condition of a defining
rule. Conditional program rules have the form l | c = r specifying that l is
reducible to r if the condition c is satisfied (see the rule defining last above).
A constraint is any expression of the built-in type Success. For instance, the
trivial constraint success is an expression of type Success that denotes the
always satisfiable constraint. “c1 & c2 ” denotes the concurrent conjunction of the
constraints c1 and c2 , i.e., this expression is evaluated by proving both argument constraints concurrently. An equational constraint e1 =:= e2 is satisfiable
if both sides e1 and e2 are reducible to unifiable constructor terms. Specific
Curry systems also support more powerful constraint structures, like arithmetic
constraints on real numbers or finite domain constraints (e.g., PAKCS [14]).
Using functions instead of predicates has the advantage that the information
provided by functional dependencies can be used to reduce the search space and
evaluate goals in an optimal way (e.g., shortest derivation sequences, minimal
solution sets, see [1] for details). However, there are also situations where a
relational style is preferable, e.g., for database applications as considered in this
paper. This style is supported by considering predicates as functions with result
type Success. For instance, a predicate isPrime that is satisfied if the argument
(an integer number) is a prime can be modeled as a function with type
isPrime :: Int -> Success
The following rules define a few facts for this predicate:
isPrime 2 = success
isPrime 3 = success
isPrime 5 = success
isPrime 7 = success
Apart from syntactic differences, any pure logic program has a direct correspondence to a Curry program. For instance, a predicate isPrimePair that is
satisfied if the arguments are primes that differ by 2 can be defined as follows:
isPrimePair :: Int -> Int -> Success
isPrimePair x y = isPrime x & isPrime y & x+2 =:= y
In order to deal with information that is persistently stored outside the program
(e.g., in databases), [11] proposed the concept of dynamic predicates. A dynamic
predicate is a predicate where the defining facts (see isPrime) are not part of
the program but stored outside. Moreover, the defining facts can be modified
(similarly to dynamic predicates in Prolog). In order to distinguish between
definitions in a program (that do not change over time) and dynamic entities,
there is a distinguished type Dynamic for the latter.1 For instance, in order to
define a dynamic predicate prime to store prime numbers whenever we compute
them, we provide the following definition in our program:
prime :: Int -> Dynamic
prime dynamic
1
In contrast to Prolog, where dynamic declarations are often used for efficiency purposes, this separation is also necessary here due to the lazy evaluation strategy which
makes it difficult to estimate when a particular evaluation is performed. Thus, performing updates by implicit side effects is not a good choice.
If the prime numbers should be persistently stored, one has to replace the second
line by
prime persistent "store"
where store specifies the storage mechanism, e.g., a directory for a lightweight
file-based implementation or a database specification [6].
There are various primitives that deal with dynamic predicates. First, there
are combinators to construct complex queries from basic dynamic predicates.
For instance, the combinator
(<>) :: Dynamic -> Dynamic -> Dynamic
joins two dynamic predicates, and the combinators
(|>) :: Dynamic -> Bool
-> Dynamic
(|&>) :: Dynamic -> Success -> Dynamic
restrict a dynamic predicate with a Boolean condition or constraint, respectively.
Since the operator “<>” binds stronger then “|>”, the expression
prime x <> prime y |> x+2 == y
specifies numbers x and y that are prime pairs.2 On the one hand, such expressions can be translated into corresponding SQL statements [6] so that the
programmer is freed of dealing with details of SQL. On the other hand, one can
use all elements and libraries of a universal programming language for database
programming due to its conceptual embedding in the programming language.
Since the contents of dynamic predicates can change over time, one needs a
careful concept of evaluating dynamic predicates in order to keep the declarative
style of programming. For this purpose, we introduce a concept of queries that
are evaluated in the I/O monad, i.e., at particular points of time in a computation.3 Conceptually, a query is a method to compute solutions w.r.t. dynamic
predicates. Depending on the number of requested solutions, there are different
operations to construct queries, e.g.,
queryAll :: (a -> Dynamic) -> Query [a]
queryOne :: (a -> Dynamic) -> Query (Maybe a)
queryAll and queryOne construct queries to compute all and one (if possible)
solution to an abstraction over dynamic predicates, respectively. For instance,
qPrimePairs :: Query [(Int,Int)]
qPrimePairs = queryAll (\(x,y) -> prime x <> prime y |> x+2 == y)
is a query to compute all prime pairs. In order to access the currently stored
data, there is an operation runQ to execute a query as an I/O action:
runQ :: Query a -> IO a
2
3
Since the right argument of “|>” demands a Boolean value rather than a constraint,
we use the Boolean equality operator “==” rather than the equational constraint
“=:=” to compare the primes x and y.
Note that we only use the basic concept of dynamic predicates from [11]. The following interface to deal with queries and transactions is new and more abstract than
the concepts described in [11].
For instance, executing the main expression “runQ qPrimePairs” returns prime
pairs that can be derived from the prime numbers currently stored in the dynamic
predicate prime.
In order to change the data stored in dynamic predicates, there are operations
to add and delete knowledge about dynamic predicates:
addDB
:: Dynamic -> Transaction ()
deleteDB :: Dynamic -> Transaction ()
Typically, these operations are applied to single ground facts (since facts
with free variables cannot be persistently stored), like “addDB (prime 13)” or
“deleteDB (prime 4)”. In order to embed these update operations into safe
transactions, the result type is “Transaction ()” (in contrast to the proposal
in [11] where these updates are I/O actions). A transaction is basically a sequence of updates that is completely executed or ignored (following the ACID
principle in databases). Similarly to the monadic approach to I/O, transactions
also have a monadic structure so that transactions can be sequentially composed
by a monadic bind operator:
(|>>=) :: Transaction a -> (a -> Transaction b) -> Transaction b
Thus, “t1 |>>= \x -> t2” is a transaction that first executes transaction t1,
which returns some result value that is bound to the parameter x before executing transaction t2. If the result of the first transaction is not relevant, one can
also use the specialized sequential composition “|>>”:
(|>>) :: Transaction a -> Transaction b -> Transaction b
t1 |>> t2 = t1 |>>= \_ -> t2
A value can be mapped into a trivial transaction returning this value by the
usual return operator:
returnT :: a -> Transaction a
In order to define a transaction that depend on some data stored in a database,
one can also embed a query into a transaction:
getDB :: Query a -> Transaction a
For instance, the following expression exploits the standard higher-order functions map, foldr, and “.” (function composition) to define a transaction that
deletes all known primes that are smaller than 100:
getDB (queryAll (\i -> prime i |> i<100)) |>>=
foldr (|>>) (returnT ()) . map (deleteDB . prime)
Since such a sequential combination of transactions that are the result of mapping a list of values into a list of transactions frequently occurs, there is also a
single function for this combination:
mapT_ :: (a -> Transaction _) -> [a] -> Transaction ()
mapT_ f = foldr (|>>) (returnT ()) . map f
To apply a transaction to the current database, there is an operation runT that
executes a given transaction as an I/O action:
runT :: Transaction a -> IO (Either a TError)
runT returns either the value computed by the successful execution of the transaction or an error in case of a transaction failure. The type TError of possible
transaction errors contains constructors for various kinds of errors, i.e., it is
currently defined as
data TError = TError TErrorKind String
data TErrorKind = KeyNotExistsError | DuplicateKeyError
| UniqueError | MinError | MaxError | UserDefinedError
but this type might be extended according to future requirements (the string
argument is intended to provide some details about the reason of the error).
UserDefinedError is a general error that could be raised by the application
program whereas the other alternatives are typical errors due to unsatisfied
integrity constraints. An error is raised inside a transaction by the operation
errorT :: TError -> Transaction a
where the specialization
failT :: String -> Transaction a
failT s = errorT (TError UserDefinedError s)
is useful to raise user-defined transaction errors. If an error is raised inside a
transaction, the transaction is aborted, i.e., the transaction monad satisfies the
laws
errorT e |>>= t = errorT e
t |>>= \x -> errorT e = errorT e
runT (errorT e) = return (Right e)
Thus, the changes to the database performed in a transaction that raises an
error are not visible.
There are a few further useful operations on transactions that are not relevant
for this paper so that we omit them. We summarize the important features of
this abstract programming model for databases:
– Persistent data is represented in the application program as language entities
(i.e., dynamic predicates) so that one can use all features of the underlying
programming language (e.g., recursion, higher-order functions, deduction)
for programming with this data.
– There is a clear separation between the data access (i.e., queries) and updates
that can influence the results of accessing data. Thus, queries are purely
declarative and are applied to the actual state of the database when their
results are required.
– Transactions, i.e., database updates, can be constructed from a few primitive
elements by specific combinators. Transactions are conceptually executed as
an atomic action on the database. Transactions can be sequentially composed
but nested transactions are excluded due to the type system (this feature is
intended since nested transactions are usually not supported in databases).
All features for database programming are summarized in a specific Database
library4 so that they can be simply used in the application program by importing
4
http://www.informatik.uni-kiel.de/\char126pakcs/lib/CDOC/Database.html
Fig. 1. A simple entity-relationship diagram for university lectures
it. This will be the basis to generate higher-level code from entity-relationship
diagrams that are described in the following.
3
Entity-Relationship Diagrams
The entity-relationship model is a framework to specify the structure and specific constraints of data stored in a database. It uses a graphical notation, called
entity-relationship diagrams (ERDs) to visualize the conceptual model. In this
framework, the part of the world that is interesting for the application is modeled by entities that have attributes and relationships between the entities. The
relationships have cardinality constraints that must be satisfied in each valid
state of the database, e.g., after each transaction.
There are various tools to support the data modeling process with ERDs. In
our framework we want to use some tool to develop specific ERDs from which the
necessary program code based on the Database library described in the previous
section can be automatically generated. In order to become largely independent
of a concrete tool, we define a representation of ERDs in Curry so that a concrete
ERD tool can be applied in this framework by implementing a translator from
the tool format into our representation. In our concrete implementation, we have
used the free software tool Umbrello UML Modeller5 , a UML tool part of KDE
that also supports ERDs. Figure 1 shows an example ERD constructed with
this tool. The developed ERDs are stored in XML files in XMI (XML Metadata
Interchange) format, a format for the exchange of UML models. Thus, it is a
standard XML transformation task to translate the Umbrello format into our
ERD format.
The representation of ERDs as data types in Curry is straightforward. A
complete ERD consists of a name (that is later used as the module name for the
generated code) and lists of entities and relationships:
5
http://uml.sourceforge.net
data ERD = ERD String [Entity] [Relationship]
An entity has a name and a list of attributes, where each attribute has a name,
a domain, and specifications about its key and null value property:
data Entity = Entity String [Attribute]
data Attribute = Attribute String Domain Key Null
data Key = NoKey | PKey | Unique
type Null = Bool
data Domain = IntDom
(Maybe Int)
| FloatDom
(Maybe Float)
| CharDom
(Maybe Char)
| StringDom
(Maybe String)
| BoolDom
(Maybe Bool)
| DateDom
(Maybe ClockTime)
| UserDefined String (Maybe String)
| KeyDom String
-- later used for foreign keys
Thus, each attribute is part of a primary key (PKey), unique (Unique), or not
a key (NoKey). Furthermore, it is allowed that specific attributes can have null
values, i.e., can be undefined. The domain of each attribute is one of the standard domains or some user-defined type. In the latter case, the first argument
of the constructor UserDefined is the qualified type name used in the Curry
application program (note that the Database library is able to handle complex
types by mapping them into standard SQL types [6]). For each kind of domain,
one can also have a default value (modeled by the Maybe type in Curry). The
constructor KeyDom is not necessary to represent ERDs but will be later used to
transform ERDs into relational database schema.
Finally, each relationship has a name and a list of connections to entities
(REnd), where each connection has the name of the connected entity, the role
name of this connection, and its cardinality as arguments:
data Relationship = Relationship String [REnd]
data REnd = REnd String String Cardinality
data Cardinality = Exactly Int | Range Int (Maybe Int)
The cardinality is either a fixed integer or a range between two integers (where
Nothing as the upper bound represents an arbitrary cardinality). For instance,
the simple-complex (1:n) relationship Teaching in Figure 1 can be represented
by the term
Relationship "Teaching"
[REnd "Lecturer" "taught_by" (Exactly 1),
REnd "Lecture" "teaches" (Range 0 Nothing)]
4
Compiling ER Diagrams into Curry Programs
This section describes the transformation of ERDs into executable Curry code.
This transformation is done in the following order:
1. Translate an ERD into an ERD term.
2. Represent the relationships occurring in an ERD term as entities.
3. Map all entities into corresponding Curry code based on the Database library.
The first step depends on the format used in the ERD tool. As mentioned above,
we have implemented a translation from the XMI format used by the Umbrello
UML Modeller into ERD terms. This part is relatively easy thanks to the presence
of XML processing tools.
4.1
Transforming ERDs
The second step is necessary since the relational model supports only relations
(i.e., database tables). Thus, entities as well as relationships must be mapped
into relations. The mapping of entities into relations is straightforward by using the entity name as the name of the relation and the attribute names as
column names. The mapping of relationships is more subtle. In principle, each
relationship can be mapped into a corresponding relation. However, this simple
approach might cause the creation of many relations or database tables. In order
to reduce them, it is sometimes better to represent specific relations as foreign
keys, i.e., to store the key of entity e1 referred by a relationship between e1
and e2 in entity e2 . Whether or not this is possible depends on the kind of the
relation. The different cases will be discussed next. Note that the representation
of relationships as relations causes also various integrity constraints to be satisfied. For instance, if an entity has an attribute which contains a foreign key, the
value of this attribute must be either null or an existing key in the corresponding relation. Furthermore, the various cardinalities of each relationship must be
satisfied. Ideally, each transaction should ensure that the integrity constraints
are valid after finishing the transaction.
Now we discuss the representation of the various kinds of relationships in
the ER model. For the sake of simplicity, we assume that each relationship
contains two ends, i.e., two roles with cardinality ranges (min, max) so that we
can characterize each relationship by their related cardinalities (minA , maxA ) :
(minB , maxB ) between entities A and B (where maxi is either a natural number
greater than mini or ∞, i ∈ {A, B}).
Simple-simple (1:1) relations: This case covers all situations where each cardinality is at most one. In the case (0, 1) : (1, 1), the key of entity B is added
as an attribute to entity A containing a foreign key since there must be exactly one B entity for each A entity. Furthermore, this attribute is Unique
to ensure the uniqueness of the inverse relation. The case (0, 1) : (0, 1) can
be similarly treated except that null values are allowed for the foreign key.
Simple-complex (1:n) relations: In the case (0, 1) : (minB , maxB ), the key
of entity A is added as a foreign key (possibly null) to each B entity. If
minB > 0 or maxB 6= ∞, the integrity constraints for the right number
of occurrences must be checked by each database update. The case (1, 1) :
(0, maxB ) is similarly implemented except that null values for the foreign
key are not allowed.
Complex-complex (n:m) relations: In this case a new relation representing
this relationship is introduced. The new relation is connected to entities A
and B by two new relationships of the previous kinds.
Note that we have not considered relationships where both minimal cardinalities
are greater than zero. This case is excluded by our framework (and rarely occurs
in practical data models) since it causes difficulties when creating new entities
of type A or B. Since each entity requires a relation to an existing entity of the
other type and vice versa, it is not possible to create the new entities independently. Thus, both entities must be created and connected in one transaction
which requires specific complex transactions. Therefore, we do not support this
in our code generation. If such relations are required in an application (e.g.,
cyclic relationships), then the necessary code must be directly written with the
primitives of the Database library.
Based on this case distinction, the second step of our compiler maps an ERD
term into a new ERD term where foreign keys are added to entities and new
entities are introduced to represent complex-complex relations. Furthermore,
each entity is extended with an internal primary key to simplify the access to
each entity by a unique scheme.
4.2
Code Generation for ERDs
After the mapping of entities and relationships into relations as described above,
we can generate the concrete program code to organize the database access and
update. As already mentioned, we base the generated code on the functionality
provided by the library Database described in Section 2. The schemas for the
generated code are sketched in this section. We use the notation En for the name
of an entity (which starts by convention with an uppercase letter) and en for
the same name where the first letter is lowercase (this is necessary due to the
convention in Curry that data constructors and functions start with uppercase
and lowercase letters, respectively).
The first elements of the generated code are data types to represent relations.
For each entity En with attributes of types at1 , . . . , atn , we generate the following
two type definitions:
data En = En Key at1 ...atn
data EnKey = EnKey Key
Key is the type of all internal keys for entities. Currently, it is identical to Int.
Thus, each entity structure contains an internal key for its unique identification. The specific type EnKey is later used to distinguish the keys for different
entities by their types, i.e., to exploit the type system of Curry to avoid confusion between the various keys. For each relation that has been introduced for a
complex-complex relationship (see above), a similar type definition is introduced
except that it does not have an internal key but only the keys of the connected
entities as arguments. Note that only the names of the types are exported but
not their internal structure (i.e., they are abstract data types for the application program). This ensures that the application program cannot manipulate
the internal keys. The manipulation of attributes is possible by explicit getter
and setter functions that are described next.
In order to access or modify the attributes of an entity, we generate corresponding functions where we use the attribute names of the ERD for the names
of the functions. If entity En has an attribute Ai of type ati (i = 1, . . . , n), we
generate the following getter and setter functions and a function to access the
key of the entity:
enAi :: En -> ati
enAi (En _ ... xi ... _) = xi
setEnAi :: En -> ati -> En
setEnAi (En x1 ... _ ... xn ) xi = En x1 ... xi ... xn
enKey :: En -> EnKey
enKey (En k _ ... _) = EnKey k
As described in Section 2, data can be persistently stored by putting them into
a dynamic predicate. Thus, we define for each entity En a dynamic predicate
enEntry :: En -> Dynamic
enEntry persistent "..."
Since the manipulation of all persistent data should be done by safe operations,
this dynamic predicate is not exported. Instead, a dynamic predicate en is exported that associates a key with the data so that an access is only possible to
data with an existing key:
en :: EnKey -> En -> Dynamic
en key obj | key =:= enKey obj = enEntry obj
Although these operations seem to be standard functions, the use of a functional
logic language is important here. For instance, the access to an entity with a given
key k can be done by solving the goal “en k o” where o is a free variable that
will be bound to the concrete instance of the entity.
For each role with name rn specified in an ERD, we also generate a dynamic
predicate of type
rn :: En1 Key -> En2 Key -> Dynamic
where En1 and En2 are the entities related by this role. The implementation
of these predicates depend on the kind of relationship according to their implementation as discussed in Section 4.1. Since complex-complex relationships are
implemented as relations, i.e., persistent predicates (that are only internal and
not exported), the corresponding roles can be directly mapped to these. Simplesimple and simple-complex relationships are implemented by foreign keys in the
corresponding entities. Thus, their roles are implemented by accessing these keys.
We omit the code details that depend on the different cases already discussed in
Section 4.1.
Based on these basic implementations of entities and relationships, we also
generate code for transactions to manipulate the data and check the integrity
constraints specified by the relationships of an ERD. In order to access an entity
with a specific key, there is a generic function that delivers this entity in a
transaction or raises a transaction error if there is no entry with this key:
getEntry :: k -> (k -> a -> Dynamic) -> Transaction a
getEntry key pred =
getDB (queryOne (\info -> pred key info)) |>>=
maybe (errorT (KeyNotExistsError "no entry for..."))
returnT
This internal function is specialized to an exported function for each entity:
getEn :: EnKey -> Transaction En
getEn key = getEntry key en
In order to insert new entities, there is a “new” transaction for each entity. If
the ERD specifies no relationship for this entity with a minimum greater than
zero, there is no need to provide related entities so that the transaction has the
following structure (if En has attributes of types at1 , . . . , atn ):
newEn :: at1 -> · · · -> atn -> Transaction En
newEn a1 ... an = check1 |>> ... |>> checkn |>> newEntry ...
Here, checki are the various integrity checks (e.g., uniqueness checks for attributes specified as Unique) and newEntry is a generic operation (similarly to
getEntry) to insert a new entity. If attribute Ai has a default value or null values
are allowed for it, the type ati is replaced by Maybe ati in newEn. If there are
relationships for this entity with a minimum greater than zero, than the keys (in
general, a list of keys) must be also provided as parameters to newEn. The same
holds for the “new” operations generated for each complex-complex relationship.
For instance, the new operation for lectures according to the ERD in Figure 1
has the following type:
newLecture :: LecturerKey -> Int -> String -> Maybe Int
-> Transaction Lecture
The first argument is the key of the lecturer required by the relationship
Teaching, and the further arguments are the values of the Id, Title and Hours
attributes (where the attribute Hours has a default value so that the argument
is optional).
Similarly to newEn, we provide also operations to update existing entities.
These operations have the following structure:
updateEn :: En -> Transaction ()
updateEn e = check1 |>> ... |>> checkn |>> updateEntry ...
Again, the various integrity constraints must be checked before an update is finally performed. In order to get an impression of the kind of integrity constraints,
we discuss a few checks in the following.
For instance, if an attribute of an entity is Unique, this property must be
checked before a new instance of the entity is inserted. For this purpose, there
is a generic transaction
unique :: a -> (b -> a) -> (b -> Dynamic) -> Transaction ()
where the first argument is the attribute value, the second argument is a getter
function for this attribute, and the third argument is the dynamic predicate
representing this entity, i.e., a typical call to check the uniqueness of the new
value ai for attribute Ai of entity En is (unique ai enAi En). This transaction
raises a UniqueError if an instance with this attribute value already exists.
If an entity contains a foreign key, each update must check the existence of
this foreign key. This is the purpose of the generic transaction
existsDBKey :: k -> (a -> k) -> (a -> Dynamic) -> Transaction ()
where the arguments are the foreign key, a getter function (enKey) for the key
in the foreign entity and the dynamic predicate of the foreign entity. If the key
does not exist, a KeyNotExistsError is raised. Furthermore, there are generic
transactions to check minimum and maximum cardinalities for relationships and
lists of foreign keys that can raise the transaction errors MinError, MaxError,
or DuplicateKeyError. For each new and update operation generated by our
compiler, the necessary integrity checks are inserted based on the ER specification.
Our framework does not provide delete operations. The motivation for this
is that safe delete operations require the update of all other entities where this
entity could occur as a key. Thus, a simple delete could cause many implicit
changes that are difficult to overlook. It might be better to provide only the
deletion of single entities followed by a global consistency check (discussed below). A solution to this problem is left as future work.
Even if our generated transactions ensure the integrity of the affected relations, it is sometimes useful to provide a global consistency check that is regularly applied to all data. This could be necessary if unsafe delete operations
are performed, or the database is modified by programs that do not use the
safe interface but directly accesses the data. For this purpose, we also generate
a global consistency test that checks all persistent data w.r.t. the ER model. If
E1 , . . . , En are all entities (including the implicit entities for complex-complex
relations) derived from the given ERD, the global consistency test is defined by
checkAllData :: Transaction ()
checkAllData = checkE1 |>> ... |>> checkEn
The consistency test for each entity En is defined by
checkEn :: Transaction ()
checkEn = getDB (queryAll enEntry) |>>= mapT_ checkEnEntry
checkEnEntry :: En -> Transaction ()
checkEnEntry e = check1 |>> ... |>> checkn
where the tests checki are similar to the ones used in new and update operations
that raise transaction errors in case of unsatisfied integrity constraints.
5
Conclusions
We have presented a framework to compile conceptual data models specified
as entity-relationship diagrams into executable code for database programming
in Curry. This compilation is done in three phases: translate the specific ERD
format into a tool-independent representation, transform the relationships into
relations according to their complexity, and generate code for the safe access and
update of the data.
Due to the importance of ERDs to design conceptual data models, there are
also other tools with similar objectives. Most existing tools support only the generation of SQL code, like the free software tools DB-Main6 or DBDesigner47 . The
main motivation for our development was the seamless embedding of database
programming in a declarative programming language and the use of existing
specification methods like ERDs as the basis to generate most of the necessary
code required by the application programs. The advantages of our framework
are:
– The application programmer must only specify the data model in a high-level
format (ERDs) and all necessary code for dealing with data in this model is
generated.
– The interface used by the application programs is type safe, i.e., the types
specified in the ERD are mapped into types of the programming language
so that ill-typed data cannot be constructed.
– Updates to the database are supported as transactions that automatically
checks all integrity constraints specified in the ERD.
– Checks for all integrity constraints are derived from the ERD for individual
tables and the complete database so that they can be periodically applied
to verify the integrity of the current state of the database.
– The generated code is based on an abstract interface for database programming so that it is readable and well structured. Thus, it can be easily modified
and adapted to new requirements. For instance, integrity constraints not expressible in ERDs can be easily added to individual update operations, or
specific deletion operations can be inserted in the generated module.
For future work we intend to increase the functionality of our framework, e.g., to
extend ERDs by allowing the specification of more complex integrity constraints
or attributes for relations, which is supported by some ER tools, or to provide
also delete operations for particular entities. Finally, it could be also interesting
to generate access and update operations for existing databases by analyzing
their data model. Although this is an issue different from our framework, one
can reuse the API described in Section 2 and some other techniques of this paper
for such a purpose.
References
1. S. Antoy, R. Echahed, and M. Hanus. A Needed Narrowing Strategy. Journal of
the ACM, Vol. 47, No. 4, pp. 776–822, 2000.
6
7
http://www.db-main.be
http://www.fabforce.net/dbdesigner4
2. S. Antoy and M. Hanus. Functional Logic Design Patterns. In Proc. of the 6th
International Symposium on Functional and Logic Programming (FLOPS 2002),
pp. 67–87. Springer LNCS 2441, 2002.
3. P. P.-S. Chen. The Entity-Relationship Model—Toward a Unified View of Data.
ACM Transactions on Database Systems, Vol. 1, No. 1, pp. 9–36, 1976.
4. J. Correas, J.M. Gómez, M. Carro, D. Cabeza, and M. Hermenegildo. A Generic
Persistence Model for (C)LP Systems (and Two Useful Implementations). In Proc.
of the Sixth International Symposium on Practical Aspects of Declarative Languages
(PADL’04), pp. 104–119. Springer LNCS 3057, 2004.
5. S.K. Das. Deductive Databases and Logic Programming. Addison-Wesley, 1992.
6. S. Fischer. A Functional Logic Database Library. In Proc. of the ACM SIGPLAN
2005 Workshop on Curry and Functional Logic Programming (WCFLP 2005), pp.
54–59. ACM Press, 2005.
7. H. Gallaire and J. Minker, editors. Logic and Databases, New York, 1978. Plenum
Press.
8. M. Hanus. A Unified Computation Model for Functional and Logic Programming.
In Proc. of the 24th ACM Symposium on Principles of Programming Languages
(Paris), pp. 80–93, 1997.
9. M. Hanus. A Functional Logic Programming Approach to Graphical User Interfaces. In International Workshop on Practical Aspects of Declarative Languages
(PADL’00), pp. 47–62. Springer LNCS 1753, 2000.
10. M. Hanus. High-Level Server Side Web Scripting in Curry. In Proc. of the Third International Symposium on Practical Aspects of Declarative Languages (PADL’01),
pp. 76–92. Springer LNCS 1990, 2001.
11. M. Hanus. Dynamic Predicates in Functional Logic Programs. Journal of Functional and Logic Programming, Vol. 2004, No. 5, 2004.
12. M. Hanus. Type-Oriented Construction of Web User Interfaces. In Proceedings of
the 8th ACM SIGPLAN International Conference on Principles and Practice of
Declarative Programming (PPDP’06), pp. 27–38. ACM Press, 2006.
13. M. Hanus. Multi-paradigm Declarative Languages. In Proceedings of the International Conference on Logic Programming (ICLP 2007), pp. 45–75. Springer LNCS
4670, 2007.
14. M. Hanus, S. Antoy, B. Braßel, M. Engelke, K. Höppner, J. Koj, P. Niederau,
R. Sadre, and F. Steiner. PAKCS: The Portland Aachen Kiel Curry System.
Available at http://www.informatik.uni-kiel.de/~pakcs/, 2006.
15. M. Hanus (ed.). Curry: An Integrated Functional Logic Language (Vers. 0.8.2).
Available at http://www.informatik.uni-kiel.de/~curry, 2006.
16. S.H. Huseby. Innocent Code: A Security Wake-Up Call for Web Programmers.
Wiley, 2003.
17. F. López-Fraguas and J. Sánchez-Hernández. TOY: A Multiparadigm Declarative
System. In Proc. of RTA’99, pp. 244–247. Springer LNCS 1631, 1999.
18. S. Peyton Jones, editor. Haskell 98 Language and Libraries—The Revised Report.
Cambridge University Press, 2003.
19. P. Wadler. How to Declare an Imperative. ACM Computing Surveys, Vol. 29,
No. 3, pp. 240–263, 1997.
| 6cs.PL
|
arXiv:1706.05271v1 [cs.PL] 16 Jun 2017
A Coq-based synthesis of Scala programs
which are correct-by-construction
Youssef El Bakouny1,2 , Tristan Crolard1 , and Dani Mezher2
1
2
CEDRIC - CNAM - Paris - France
CIMTI - ESIB - Saint-Joseph University - Beirut - Lebanon
May 1, 2017
Abstract
The present paper introduces Scala-of-Coq, a new compiler that allows a
Coq-based synthesis of Scala programs which are “correct-by-construction”.
A typical workflow features a user implementing a Coq functional program, proving this program’s correctness with regards to its specification
and making use of Scala-of-Coq to synthesize a Scala program that can
seamlessly be integrated into an existing industrial Scala or Java application.
1 Introduction
In our modern world, software bugs are becoming increasingly detrimental
to the engineering industry. Since software components are interconnected, a
bug in one component is likely to affect others, causing a system-wide failure.
Additionally, software bugs have often created system vulnerabilities that are
exploitable by malicious users [10].
These bugs are essentially due to a lack of rigor in the software engineering
field where the correctness of a program is thoroughly tested but rarely proven.
In fact, current industrial software development practices rely solely on the
expensive and time-consuming development of tests in an attempt to hunt
down bugs. For non-trivial cases, this approach can never guarantee that a
software is correct with regards to its specification.
As a result, we have recently witnessed interesting initiatives aimed at
developing new science, technologies and tools capable of responding to the
aforementioned software engineering challenges. A remarkable example of
such an initiative is a U.S. National Science Foundation (NSF) expedition in
computing project called “the Science of Deep Specification (Deep Spec)” [10].
These initiatives make use of “formal methods”, potentially as a complement to software testing, with the goal of providing the means of developing
1
software that is correct-by-construction. These formal methods can be defined
as a set of techniques and mathematical theories allowing both the specification
of a software system in an unambiguous manner, and the rigorous verification
of the conformity of this software system with regards to its specification. This
verification can be done by proving that the program satisfies a set of lemmas
or theorems constituting its specification.
2 Proof assistants
Since the manual checking of realistic program proofs is impractical or, to say
the least, time-consuming; several proof assistants have been developed to
provide machine-checked proofs. Isabelle/HOL [8] and Coq [6] are currently
the world’s two leading proof assistants.
Both of these proof assistants allow the interactive construction of formal
proofs; Coq is based on a logical framework known as the calculus of inductive
constructions [11], while Isabelle/HOL is based on high order classical logic.
Coq supports code generation in Objective Caml, Haskell and Scheme [6], while
Isabelle/HOL supports code generation in Objective Caml, Haskell, SML and
Scala [3].
The code generation capabilities of these proof assistants enable the synthesis of programs which are correct-by-construction. However, they do not
address the verification needs of existing software programs. This need is covered by other initiatives such as Leon [1], Why3 [2] and, also, Sireum Logika1 .
Leon and Logika allow the automatic verification of Scala programs. Leon can
also resort to Isabelle/HOL machine-checked proofs when its automatic verification mechanism fails to give an answer. Similarly, Why3 allows the automatic
verification of Objective Caml programs with the option of resorting to Coq on
failure.
Coq has been successfully used to implement CompCert [5], the world’s
first and only formally verified C compiler. It is also widely used in several
research projects including the aforementioned DeepSpec project, hence our
interest in this particular proof assistant.
3 The Scala Programming Language
The Scala programming language [9] was developed by Martin Odersky and
his “École Polytechnique Fédérale de Lausanne” (EPFL) team in 2003. It is
currently of high interest to several industrial projects thanks to its practical
fusion of functional and object-oriented programming; as well as its seamless
interoperability with Java.
Scala is the implementation language of many important frameworks, including Apache Spark, Kafka, and Akka. It also provides the core infrastructure
for sites such as Twitter and Coursera.
1 http://logika.sireum.org
2
Given the importance of Scala in the industrial world, the Coq proof assistant
would highly benefit of a Scala code generation feature. For this purpose, the
“Scala-of-Coq” compiler is currently being implemented.
4 Compiling Coq to Scala
The main objective of the Scala-of-Coq compiler is to allow the synthesis of
Scala programs that are correct-by-construction. The compiler itself is written
in Scala and its underlying Coq parser is largely based on the use of parser combinators [4], easing the construction of complex parsers out of primitive ones.
The Scala library that supports these parser combinators [7] facilitated the implementation of the Coq parser without the need to resort to parser generators
such as Lex/Yacc or ANTLR. This is allowing us to create and manipulate Coq’s
Abstract Syntax Tree (AST) directly in Scala while taking full advantage of the
language’s fusion of object-oriented and functional paradigms.
Using Scala-of-Coq, the Coq program in listing 1, which computes the number of nodes in a binary tree, was successfully converted to the Scala program
in listing 2.
Listing 1: Source Coq Program
Inductive Tree : Type : =
Leaf : Tree
| Node(l r : Tree): Tree.
Fixpoint size ( t: Tree) : nat : =
match t with
Leaf => 1
| Node l r => 1 + ( size l) + ( size r)
end.
Listing 2: Generated Scala Program
object CertifiedTree {
sealed trait Tree
case object Leaf extends Tree
case class Node(l: Tree, r: Tree) extends Tree
def size(t: Tree): BigInt =
t match {
case Leaf => 1
case Node(l, r) => 1 + (size(l) + size(r))
}
}
It is important to note that the source Coq program can be formally verified
by proving the lemmas that define its specification. For example, the lemmas
in listing 3 were successfully proven using the definitions provided in listing 1.
3
Once the source Coq program is formally proven with regards to its specification (portrayed as Lemmas), the generated Scala program can also be considered, assuming a few reasonable hypotheses, formally verified with regards
to the same specification.
Lemma size_left: ∀ l r : Tree, size ( Node l r) > size l.
Lemma size_right: ∀ l r : Tree, size ( Node l r) > size r.
Listing 3: Coq Lemmas
5 Acknowledgments
The authors would like thank the National Council for Scientific Research in
Lebanon (CNRS-L)2 for their funding, as well as Murex S.A.S3 for providing
financial support.
6 Conclusion
In conclusion, the Scala-of-Coq compiler enables users of the Coq proof assistant
to generate Scala programs that are formally verified. An example of a practical
use case of Scala-of-Coq depicts a Coq user that generates a verified Scala
component and seamlessly integrates it into an existing industrial Scala or Java
application.
Future developments will focus on the verification of existing Scala programs using Coq. This verification should also include the support of imperative side effects and a restricted subset of parallel programs.
Finally, a framework for the support of additional languages, such as Swift,
should be implemented.
References
[1] Régis Blanc, Viktor Kuncak, Etienne Kneuss, and Philippe Suter. An
overview of the leon verification system: Verification by translation to
recursive functions. In Proceedings of the 4th Workshop on Scala, SCALA ’13,
pages 1:1–1:10, New York, NY, USA, 2013. ACM.
[2] Jean-Christophe Filliâtre and Andrei Paskevich. Why3 - where programs
meet provers. In Matthias Felleisen and Philippa Gardner, editors, Programming Languages and Systems - 22nd European Symposium on Programming, ESOP 2013, Held as Part of the European Joint Conferences on Theory and
Practice of Software, ETAPS 2013, Rome, Italy, March 16-24, 2013. Proceedings,
2 http://www.cnrs.edu.lb/
3 https://www.murex.com/
4
volume 7792 of Lecture Notes in Computer Science, pages 125–128. Springer,
2013.
[3] Florian Haftmann and Tobias Nipkow. Code generation via higher-order
rewrite systems. In Matthias Blume, Naoki Kobayashi, and Germán Vidal,
editors, Functional and Logic Programming: 10th International Symposium,
FLOPS 2010, Sendai, Japan, April 19-21, 2010. Proceedings, pages 103–117,
Berlin, Heidelberg, 2010. Springer.
[4] Graham Hutton and Erik Meijer. Monadic parsing in haskell. J. Funct.
Program., 8(4):437–444, 1998.
[5] Xavier Leroy. Mechanized semantics for compiler verification. In Chris
Hawblitzel and Dale Miller, editors, Certified Programs and Proofs - Second International Conference, CPP 2012, Kyoto, Japan, December 13-15, 2012.
Proceedings, volume 7679 of Lecture Notes in Computer Science, pages 4–6.
Springer, 2012.
[6] The Coq development team. The Coq proof assistant reference manual. LogiCal Project, 2004. Version 8.0.
[7] Adriaan Moors, Frank Piessens, and Martin Odersky. Parser combinators
in Scala. Technical report, K.U.Leuven, 2008.
[8] Tobias Nipkow, Lawrence C. Paulson, and Markus Wenzel. Isabelle/HOL
- A Proof Assistant for Higher-Order Logic, volume 2283 of Lecture Notes in
Computer Science. Springer, 2002.
[9] Martin Odersky and Tiark Rompf. Unifying functional and object-oriented
programming with scala. Commun. ACM, 57(4):76–86, 2014.
[10] Benjamin C. Pierce. The science of deep specification (keynote). In Eelco
Visser, editor, Companion Proceedings of the 2016 ACM SIGPLAN International Conference on Systems, Programming, Languages and Applications: Software for Humanity, SPLASH 2016, Amsterdam, Netherlands, October 30 November 4, 2016, page 1. ACM, 2016.
[11] Benjamin C. Pierce, Arthur Azevedo de Amorim, Chris Casinghino, Marco
Gaboardi, Michael Greenberg, Cǎtǎlin Hriţcu, Vilhelm Sjöberg, and Brent
Yorgey. Software Foundations. Electronic textbook, 2016.
5
| 6cs.PL
|
arXiv:1509.08278v2 [math.OA] 25 Feb 2016
A REMARK ON T-VALUED COHOMOLOGY GROUPS OF
ALGEBRAIC GROUP ACTIONS
YONGLE JIANG
Abstract. We prove that for a weakly mixing algebraic action σ : G y (X, ν),
the nth cohomology group H n (G y X; T), after quotienting out the natural
b as a natural subgroup for n = 1. If
subgroup H n (G, T), contains H n (G, X)
we further assume the diagonal actions σ2 , σ4 are T-cocycle superrigid and
b is torsion free as an abelian group, then the above also holds true
H 2 (G, X)
for n = 2. Applying it for principal algebraic actions when n = 1, we show
that H 2 (G, ZG) is torsion free as an abelian group when G has property (T)
as a direct corollary of Sorin Popa’s cocycle superrigidity theorem; we also use
it (when n = 2) to answer, negatively, a question of Sorin Popa on the 2nd
cohomology group of Bernoulli shift actions of property (T) groups.
1. Introduction
Let G be a countable infinite discrete group with integer group ring ZG. Recall
that an algebraic G-action is a homomorphism α : G → Aut(X) from G to the
group of (continuous) automorphisms of a compact metrizable abelian group X. By
b (see [20, Section 1]
duality, such an action of G is determined by a ZG-module X
ZG
b
for details). When X = ZGf for some f ∈ ZG, the corresponding algebraic action
is called a principal algebraic G-action. A central question on a principal
algebraic action is whether it is measurable conjugate to a Bernoulli shift under
natural assumptions. It is trivial to observe that the answer is yes when assuming
f = k ∈ ZG for some integer k. However, it is a highly non-trivial theorem saying
that the answer is yes when G = Zd for all positive integer d if we assume the action
has completely positive entropy [36]. When G is non-abelian, this question is still
wide open.
Towards this problem, it is natural to first investigate whether Bernoulli shifts
and principal algebraic actions share any common dynamical properties, such as
ergodicity, weakly mixing, mixing, freeness, completely positive entropy etc. However, it turns out that even these weaker questions are usually difficult to settle.
We refer the reader to [4, 12, 16] for research on these questions. Very recently, cohomological methods have been employed to study ergodicity of principal algebraic
actions in [20].
However, the entropy theory of principal algebraic actions has received much
attention in the past decade from amenable groups case in [6, 7, 19, 21, 38] to sofic
groups in [3, 4, 13] due to the emergence of sofic entropy in [2, 17].
2010 Mathematics Subject Classification. Primary 22D40, 46L10; Secondary 28D05, 20J06.
Key words and phrases. Algebraic actions; cohomology groups; cocycle superrigid actions;
random groups.
1
2
YONGLE JIANG
Our motivation behind this paper is to study the above questions from the
viewpoint of von Neumann algebras, following [22, 34, 35] in spirit.
In the theory of von Neumann algebras, especially II1 factors, a fundamental
theme is to study various rigidity phenomena such as finding T-cocycle superrigid
actions, where T denotes the unit circle. Recall that a free, ergodic, probability
measure preserving (p.m.p.) action of a countable discrete group G on a standard
probability space (X, ν) is said to be T-cocycle superrigid if any T-valued 1cocycle for the action G y (X, ν) must be cohomologous (see Definition 2.3) to a
homomorphism from G to T.
Denote
Q by G the class of groups G such that the standard Bernoulli shift action
G y g∈G (X0 , µ0 ) is T-cocycle superrigid, where (X0 , µ0 ) is any standard probability space. The following (non-amenable) groups are shown to be in G mainly
by Sorin Popa’s powerful deformation/rigidity theory [31] and Jesse Peterson’s L2 rigidity theory [28].
• [33] G is weakly rigid, i.e. there exists an infinite normal subgroup H <
G such that (G, H) has relative property (T); in particular, when G has
property (T).
• [29] When L(G) is a L2 -rigid II1 factor in the sense of Peterson [28]; for
example, when L(G) is a II1 factor which is either non-prime or has property
(Γ).
• [30] If G has an infinite rigid subgroup H such that H is wq-normal or
w-normal.
• [32] G = F2 × F2 , where F2 is the free group on two letters. For the general
version, see [32, Theorem 1.1, Corollary 1.2].
• [39] G is inner amenable but not amenable.
Therefore, one natural question to ask is whether a principal algebraic G-action
(or in general an algebraic action) is T-cocycle superrigid when G ∈ G under natural
assumptions.
Motivated by this question and inspired by [29, Corollary 4.2][37, Proof of Proposition 5.2], we prove the following theorem.
Theorem 1.1. Let G be a countable infinite discrete group, X be a compact metrizable abelian group with Haar measure ν and σ : G y (X, ν) be an algebraic action.
If σ is weakly mixing as a p.m.p. action, then we have an injective group homomorphism
1
b ֒→ H (G y X; T) .
H 1 (G, X)
H 1 (G, T)
Under further assumptions that the diagonal actions σ 2 : G y (X 2 , ν 2 ), σ 4 : G y
b is torsion free as an abelian group,
(X 4 , ν 4 ) are T-cocycle superrigid and H 2 (G, X)
we also have an injective group homomorphism
2
b ֒→ H (G y X; T) .
H 2 (G, X)
H 2 (G, T)
Remark 1.2. (a) Under the assumptions in the theorem, H n (G, T) is identified with
a subgroup of H n (G y X; T) for n = 1, 2 by the natural map described in Lemma
2.9 below.
b has
(b) In the proof of the second embedding, we only need to assume H 2 (G, X)
no non-trivial elements of order two (rather that being torsion free). Without the
A REMARK ON T-VALUED COHOMOLOGY GROUPS OF ALGEBRAIC GROUP ACTIONS 3
b is torsion free as an abelian group, a small modification
assumption that H 2 (G, X)
of our proof actually shows that the second embedding above still holds if we
quotient out corresponding torsion subgroups from both sides, which was pointed
out to us by Hanfeng Li.
The paper is organized as follows. In Section 2, we first recall the definitions
of weakly mixing actions, n-cocycles and nth cohomology groups in the setting of
group actions and modules, then we prove two lemmas; one is used to replace almost group homomorphisms with genuine ones; the other one justifies H n (G, T)
being a natural subgroup of H n (G y X; T) for n = 1, 2 under assumptions. In
Section 3, we prove Theorem 1.1. In Section 4, we give applications of Theorem 1.1
as mentioned in the abstract. We conclude with further questions related to this
paper in Section 5.
Acknowledgments. We are indebted to Hanfeng Li for constant support and pointing out errors in a previous version of this paper. We thank Christian Rosendal for
sharing his unpublished notes on almost group homomorphisms at an early stage
of this paper. We also thank Jesse Peterson for his interest in this work, and thank
Sorin Popa for helpful comments. Also thanks to Yves de Cornulier, Mike Mihalik,
Boris Okun and Ben Wieland for answering our questions on H 2 (G, ZG). Last
but not least, we are grateful to the referee for extremely helpful suggestions and
comments which improves the paper greatly.
2. Preliminaries
First, let us recall the well-known definition of weakly mixing actions.
Definition 2.1. A probability measure preserving action G y (X, µ) is said to be
weakly mixing if one of the following equivalent conditions is satisfied.
• The diagonal action G y (X × X, µ × µ) is ergodic.
• The diagonal action G y (X ×Y, µ×η) is ergodic for all ergodic G y (Y, η).
• The constant functions form the only finite-dimensional G-invariant subspace of L2 (X, µ).
Note that from the above definition, it follows that if G y (X, µ) is weakly mixing,
then the diagonal G-action on the product (X n , µn ) is ergodic, and weakly mixing
for all n ≥ 2.
Next, let us recall the definition of cocycles in the group action setting [8].
Definition 2.2. Let G y (X, µ) be a p.m.p. group action, we say a measurable
map c : Gn × X → T is a n-cocycle if for all g1 , g2 , · · · , gn+1 ∈ G, the equation
c(g2 , g3 , . . . , gn+1 , g1−1 x)c(g1 g2 , g3 , . . . , gn+1 , x)−1 c(g1 , g2 g3 , g4 , . . . , gn+1 , x) · · ·
n
· · · c(g1 , g2 , . . . , gn−1 , gn gn+1 , x)(−1) c(g1 , g2 , . . . , gn , x)(−1)
n+1
=1
holds true µ-almost every x ∈ X.
Since G is countable, in the above definition, we can further assume there exists
a conull subset Z of X such that the above equation holds true for all x ∈ Z and
all g1 , g2 , · · · , gn+1 ∈ G. Here, Z ⊂ (X, µ) is conull if µ(X \ Z) = 0.
4
YONGLE JIANG
Definition 2.3. Two n-cocycles c1 , c2 : Gn × X → T are called cohomologous if
there is a measurable map b : Gn−1 × X → T such that for all g1 , g2 , · · · , gn ∈ G,
c1 (g1 , . . . , gn , x)b(g2 , . . . , gn , g1−1 x)b(g1 g2 , g3 , . . . , gn , x)−1 b(g1 , g2 g3 , . . . , gn , x) · · ·
· · · b(g1 , . . . , gn−2 , gn−1 gn , x)(−1)
n−1
n
b(g1 , g2 , . . . , gn−1 , x)(−1) = c2 (g1 , . . . , gn , x)
holds true µ-almost every x ∈ X.
Definition 2.4. We say a n-cocycle is trivial if it is cohomologous to the cocycle
which takes the constant value 1 ∈ T.
The set of all n-cocycles for the action G y (X, µ) with values in T is denoted
by Z n (G y X; T), the set of trivial cocycles is denoted by B n (G y X; T), and
the set of equivalence class of cohomologous cocycles is denoted by H n (G y X; T).
Note that since T is abelian, Z n (G y X; T) is an abelian group under pointwise
multiplication, B n (G y X; T) is a subgroup and H n (G y X; T) is the quotient
group, which is called the nth cohomology group for the action G y (X, µ).
Let G be a group and M be a left G-module, the cohomology group H n (G, M )
for any natural number n can also be defined using cocycles [5, Page 59].
Definition 2.5. A map c : Gn → M is called a n-cocycle if the equation
g1 c(g2 , g3 , . . . , gn+1 ) − c(g1 g2 , g3 , . . . , gn+1 ) + c(g1 , g2 g3 , g4 , . . . , gn+1 ) − · · ·
· · · + (−1)n c(g1 , g2 , . . . , gn−1 , gn gn+1 ) + (−1)n+1 c(g1 , g2 , . . . , gn ) = 0
holds true for all g1 , · · · , gn+1 in G.
Definition 2.6. Two n-cocycles c1 , c2 : Gn → M are called cohomologous if there
is a map b : Gn−1 → M such that
c1 (g1 , · · · , gn ) + g1 b(g2 , · · · , gn ) − b(g1 g2 , g3 , . . . , gn ) + b(g1 , g2 g3 , . . . , gn ) − · · ·
· · · + (−1)n−1 b(g1 , . . . , gn−2 , gn−1 gn ) + (−1)n b(g1 , g2 , . . . , gn−1 ) = c2 (g1 , . . . , gn )
holds true for all g1 , g2 , · · · , gn ∈ G.
Definition 2.7. We say a n-cocycle is trivial if it is cohomologous to the cocycle
which takes the constant value 0 ∈ M .
In the same way, we define the nth cohomology group H n (G, M ) for a G-module
M as the quotient group Z n (G, M )/B n (G, M ). Especially, when M = ZG with
left G-multiplication and M = T with trivial G-action, we can define H n (G, ZG)
and H n (G, T) respectively.
A general version of the following lemma was communicated to us by Christian
Rosendal (see also [23, Theorem 3]). We are grateful to Hanfeng Li for the following
quick proof.
Lemma 2.8. Let X be a compact metrizable abelian group with Haar measure ν,
and let b : X → T be a measurable map. If b(x + y) = b(x)b(y) holds true for almost
every (x, y) ∈ X × X, then there is a continuous group homomorphism λ : X → T
such that λ(x) = b(x) for almost every x ∈ X.
Proof. Note that the translation action α : X y L2 (X) is a continuous group
action, and b ∈ L2 (X) with ||b||2 = 1. By assumption, we can find a conull subset
Z of X such that αz (b) = b(z)b holds true for all z ∈ Z. This implies b is an
eigenvector for αx for all x ∈ X, i.e. αx (b) = λ(x)b for all x ∈ X and some map
λ : X → T. In particular, λ(z) = b(z) for all z ∈ Z. It is easy to check that
λ : X → T satisfies the required properties.
A REMARK ON T-VALUED COHOMOLOGY GROUPS OF ALGEBRAIC GROUP ACTIONS 5
Lemma 2.9. For a p.m.p. action G y (X, ν), we have a natural group homomorphism Ψn : H n (G, T) → H n (G y X; T) for any natural number n. It is injective
for n = 1 if the action is weakly mixing; it is also injective for n = 2 if the diagonal
action σ 2 : G y (X 2 , ν 2 ) is weakly mixing and T-cocycle superrigid.
Proof. We define Ψn : H n (G, T) → H n (G y X; T) by Ψn ([c]) = [c′ ], where c′ :
Gn × X → T is defined by c′ (g1 , · · · , gn , x) = c(g1 , · · · , gn ).
It is easy to check that Ψn is a well-defined group homomorphism.
For n = 1, we claim Ψ1 is injective if the action is weakly mixing.
Note that H 1 (G, T) = Hom(G, T).
If [c] ∈ H 1 (G, T) and Ψ1 ([c]) = [0], then there exists a measurable map b : X → T
such that c(g) = b(g −1 x)b(x)−1 holds a.e. x ∈ X. Now, viewing b ∈ L2 (X), we
have gb = c(g)b, so Cb is an invariant subspace under the G-action. Now the action
is weakly mixing implies b(x) = λ ∈ T a.e. x ∈ X, so c(g) = 1 for all g ∈ G, and
[c] = [0].
For n = 2, we claim Ψ2 is injective if the diagonal action σ 2 : G y (X 2 , ν 2 ) is
weakly mixing and T-cocycle superrigid.
If [c] ∈ H 2 (G, T) and Ψ2 ([c]) = [0] ∈ H 2 (G y X; T), there exists a measurable
map b : G × X → T such that
c(g, h) = b(g, x)b(h, g −1 x)b(gh, x)−1
(2.1)
holds a.e. x ∈ X.
For brevity, we use bg (x) := b(g, x) below.
Defining b̂ : G×X 2 → T by b̂g (x, y) = bg (x)bg (y)−1 , then (2.1) implies b̂g σg2 (b̂h ) =
b̂gh ∈ L∞ (X 2 , T), i.e. b̂ is a 1-cocyle for the diagonal action σ 2 : G y X 2 . Since
it is a T-cocycle superrigid action, we can find a group homomorphism λ : G → T
and a measurable map t : X 2 → T such that
b̂g = λg tσg2 (t∗ )
or, equivalently:
bg (x) = bg (y)λg t(x, y)t(g −1 x, g −1 y)−1
(2.2)
holds a.e. (x, y) ∈ X 2 .
Now we claim t splits as a simple tensor, i.e. there exist measurable functions
φ, ψ : X → T such that t(x, y) = φ(x)ψ(y) holds a.e. (x, y) ∈ X 2 .
The proof is essentially the same as the proof of [9, Theorem 3.4](for the “absolute
case”). We outline it for completeness.
From (2.2), we also have
bg (z) = bg (y)λg t(z, y)t(g −1 z, g −1 y)−1 .
(2.3)
Substituting (2.2) in (2.3), we obtain the following identity which holds a.e.
(x, y, z) ∈ X 3 :
bg (z) = [t(z, y)t(x, y)−1 ]bg (x)[t(g −1 z, g −1y)t(g −1 x, g −1 y)−1 ]−1 .
def
Setting Φ(x, y, z) = t(z, y)t(x, y)−1 , the above takes the form
bg (z) = Φ(x, y, z)bg (x)Φ(g −1 x, g −1 y, g −1 z)−1
or, equivalently:
Φ(g −1 x, g −1 y, g −1 z) = bg (x)Φ(x, y, z)bg (z)−1 .
6
YONGLE JIANG
Next, observe that the morphism of the diagonal G-actions
q : (X 3 , ν 3 ) −→ (X 2 , ν 2 ), q(x, y, z) = (x, z)
is relative weakly mixing, as described in [9, Page 25].
Following the notations in [9], we observe that T ∈ Ufin ⊂ Ginv . Since the
“cocycle” assumption in the statement of [9, Lemma 3.2] is not used in its proof,
we can still apply [9, Lemma 3.2] for the measurable maps α, β : G × X 2 → T given
by α(g, x, z) := bg (x), β(g, x, z) := bg (z) to conclude that Φ(x, y, z) = f (x, z), for
some measurable map f : X 2 → T. Therefore
t(z, y)t(x, y)−1 = Φ(x, y, z) = f (x, z)
meaning that a.e. (x, y, z) ∈ X 3 , we have
t(z, y) = f (x, z)t(x, y).
Then there exists some z0 , such that
t(x, y) = t(z0 , y)f (x, z0 )−1
holds a.e. (x, y) ∈ X 2 .
After defining ψ(y) = t(z0 , y), φ(x) = f (x, z0 )−1 , we see that t(x, y) = φ(x)ψ(y)
holds a.e. (x, y) ∈ X 2 .
The definition of b̂ and (2.2) imply bg ⊗ b∗g = λg (φ ⊗ ψ)(σg (φ∗ ) ⊗ σg (ψ ∗ )) ∈
∞
L (X) ⊗ L∞ (X) or, equivalently:
(2.4)
1 ⊗ 1 = λg (b∗g φσg (φ∗ ) ⊗ bg ψσg (ψ ∗ )).
Since we can identify L∞ (X) ⊗ L∞ (X) with a subspace of L2 (X) ⊗ L2 (X), we may
think 1 ⊗ 1, b∗g φσg (φ∗ ) ⊗ bg ψσg (ψ ∗ ) as Hilbert-Schmidt operators on L2 (X). Then,
take the vector 1 ∈ L2 (X), and apply (2.4), we have
1 = h1, 1i1 = (1 ⊗ 1)1
= λg b∗g φσg (φ∗ ) ⊗ bg ψσg (ψ ∗ )(1)
= λg h1, b∗g ψ ∗ σg (ψ)ib∗g φσg (φ∗ ) ∈ L2 (X).
The above implies b∗g φσg (φ∗ ) is a constant function in L2 (X), say, b∗g φσg (φ∗ ) = µg
for some map µ : G → T; hence, bg = φσg (φ∗ )µ−1
g , substituting this in (2.1), we
−1
deduce c(g, h) = µ−1
µ
µ
,
so
[c]
=
[0].
gh
g
h
Remark 2.10. If G has property (T) and the action G y X is a Bernoulli shift, then
it is well-known that H 2 (G, T) is a subgroup of H 2 (G y X; T) by [33, Theorem
3.1].
Remark 2.11. If the diagonal action σ 2 : G y (X 2 , ν 2 ) is weakly mixing and Tcocycle superrigid, then σ : G y (X, ν) is also T-cocycle superrigid by [30, Theorem
3.1].
3. Proof of Theorem 1.1
Now, we are ready to give the proof of Theorem 1.1.
Proof. We may identify H n (G, T) with a subgroup of H n (G y X; T) for n = 1, 2
by Lemma 2.9. Then we can define a group homomorphism
n
b → H (G y X; T)
Φn : H n (G, X)
H n (G, T)
A REMARK ON T-VALUED COHOMOLOGY GROUPS OF ALGEBRAIC GROUP ACTIONS 7
by Φn ([c]) = [c̃] + H n (G, T), where c̃ : Gn × X → T is defined by c̃(g1 , · · · , gn , x) =
c(g1 , · · · , gn )(x) for all g1 , · · · , gn ∈ G and x ∈ X.
It is clear that Φn is a well-defined group homomorphism.
Case n = 1.
We claim Φ1 is injective if the action G y X is weakly mixing.
Note that H 1 (G, T) = Hom(G, T).
b and Φ1 ([c]) = [0] + H 1 (G, T), then there exist a measurable
If [c] ∈ H 1 (G, X)
map b : X → T and a group homomorphism φ : G → T such that
(3.1)
c(g)(x) = c̃(g, x) = b(g −1 x)b(x)−1 φ(g)
holds true for all g ∈ G and all x ∈ Z ⊂ X a conull set.
Then we prove that we can assume b is a continuous group homomorphism and
φ(g) = 1 ∈ T for all g ∈ G.
Note that the addition operation ∆ : X × X → X, ∆(x, y) := x + y, induces a
left invariant measure ∆∗ (µ × µ) on X and hence ∆∗ (µ × µ) = µ by the uniqueness
of Haar measure on X. So W := ∆−1 (Z) ∩ (Z × Z) ⊂ X × X is still a conull set
and we may assume it is G-invariant.
Since c(g)(x + y) = c(g)(x)c(g)(y) for all g ∈ G and all x, y ∈ X, by (3.1) we
deduce that
b(x + y)b(x)−1 b(y)−1 φ(g) = (gb)(x + y)[(gb)(x)]−1 [(gb)(y)]−1
holds true for all g ∈ G and all (x, y) ∈ W ⊂ X × X.
Now we consider the map b̂ : X 2 → T defined by b̂(x, y) = b(x + y)b(x)−1 b(y)−1 ,
the above equality implies
(3.2)
b̂(x, y)φ(g) = (g b̂)(x, y)
holds true for all g ∈ G and all (x, y) ∈ W ⊂ X × X.
Notice that the above implies the subspace Cb̂ is invariant under the action of G.
Since G y X 2 is weakly mixing, it follows that b̂ is a constant function in L2 (X 2 ).
Hence, b̂(x, y) ≡ λ a.e. for some constant λ ∈ T, i.e. b(x + y) = b(x)b(y)λ for all
(x, y) ∈ Z ′ ⊂ X × X a conull set.
From (3.2), we find φ(g) = 1 ∈ T for all g ∈ G.
After replacing b by bλ, which does not change the equality in (3.1), we may
assume b is an almost group homomorphism in the sense that b(x + y) = b(x)b(y)
for all (x, y) ∈ Z ′ ⊂ X × X a conull set.
Now, Lemma 2.8 shows that there is a continuous genuine group homomorphism
b̃ : X → T such that b̃(x) = b(x) for all x ∈ Y ⊂ X a conull set.
So, for any fixed g ∈ G, c(g)(x) = b̃(g −1 x)b̃(x)−1 holds for all x in a conull
subset of X. Since both sides are continuous functions on x, the equality holds
b c is a trivial 1-cocycle.
everywhere. Hence c(g) = (g − 1)b̃ ∈ X,
Case n = 2.
We claim Φ2 is injective if the diagonal actions G y X 2 , X 4 are weakly mixing
and T-cocycle superrigid.
b and Φ2 ([c]) = [0] + H 2 (G, T), then there is a measurable map
If [c] ∈ H 2 (G, X)
t : G × X → T and a 2-cocycle θ : G × G → T such that
(3.3)
c(g, h)(x) = θ(g, h)t(g, x)t(h, g −1 x)t(gh, x)−1
holds true for all g, h ∈ G and almost every x ∈ X.
For brevity, we use tg (x) := t(g, x) below.
8
YONGLE JIANG
b
Now we show that θ ∈ B 2 (G, T) and t can be replaced by a map from G to X.
b we have the following equality
Since c(g, h) ∈ X,
tgh (x + y)tgh (x)−1 tgh (y)−1 θ(g, h) = tg (x + y)tg (x)−1 tg (y)−1 th (g −1 (x + y))th (g −1 x)−1 th (g −1 y)−1
holds true for all g, h ∈ G and almost every (x, y) ∈ X × X.
Defining t̂ : G × X 2 → T by t̂g (x, y) = tg (x + y)tg (x)−1 tg (y)−1 , then the above
becomes
θ(g, h) = t̂g (x, y)t̂h (g −1 x, g −1 y)t̂gh (x, y)−1 .
Therefore, Ψ2 ([θ]) = [0] under the homomorphism Ψ2 : H 2 (G, T) → H 2 (G y
X 2 ; T). Now, since we assume the diagonal action G y X 4 is weakly mixing and
T-cocycle superrigid, Ψ2 is injective by Lemma 2.9, hence [θ] = [0] ∈ H 2 (G, T).
Now, we may assume θ(g, h) = 1 in (3.3) and find a new expression
(3.4)
c(g, h)(x) = tg (x)th (g −1 x)tgh (x)−1 ,
which holds true for all g, h ∈ G and almost every x ∈ X.
Repeating the above argument, we would conclude
t̂g (x, y)t̂h (g −1 x, g −1 y) = t̂gh (x, y).
Hence, t̂ ∈ Z 1 (G y X 2 ; T). Then we apply the assumption that G y X 2 is
T-cocycle superrigid to find a measurable map u : X 2 → T and a group homomorphism λ : G → T such that
t̂g = λg uσg2 (u∗ )
holds for all g ∈ G. This is equivalent to say
(3.5)
tg (x1 + x2 ) = tg (x1 )tg (x2 )λg u(x1 , x2 )u(g −1 x1 , g −1 x2 )−1 ,
holds for all (x1 , x2 ) ∈ E ⊂ X 2 , where E ⊂ X 2 is a conull set, and all g ∈ G.
Note that we may replace E by the smaller conull subset ∩τ ∈S2 τ (E), where
def
S2 denotes the permutation group of two elements and (x1 , x2 ) ∈ τ (E) ⇐⇒
(xτ (1) , xτ (2) ) ∈ E, so without loss of generality, we may assume E is symmetric.
Now we claim that we can assume u is a measurable symmetric 2-cocycle, i.e.
there exist conull sets F ′′ ⊂ X 3 and E ′′ ⊂ X 2 such that the following hold:
u(x1 , x2 )u(x1 + x2 , x3 ) =
u(x1 , x2 ) =
u(x2 , x3 )u(x1 , x2 + x3 ), for all (x1 , x2 , x3 ) ∈ F ′′ ;
u(x2 , x1 ), for all (x1 , x2 ) ∈ E ′′ .
First, we show u can be assumed to be symmetric.
Given any (x1 , x2 ) ∈ E, we have the following equality by (3.5)
tg (x1 + x2 ) = tg (x1 )tg (x2 )λg u(x2 , x1 )u(g −1 x2 , g −1 x1 )−1 .
Comparing it with (3.5), we deduce
u(x1 , x2 )u(x2 , x1 )−1 = u(g −1 x1 , g −1 x2 )u(g −1 x2 , g −1 x1 )−1 .
Since G y X 2 is weakly mixing, u(x1 , x2 ) = au(x2 , x1 ) for some constant a ∈ T
and for all (x1 , x2 ) ∈ E ′ ⊂ X 2 a conull set.
Defining E ′′ = ∩τ ∈S2 τ (E ∩ E ′ ), then E ′′ is still conull.
Note that E ′′ is symmetric; hence u(x1 , x2 ) = au(x2 , x1 ) = a2 u(x1 , x2 ), a = ±1.
If a = −1, u(x1 , x2 )2 = u(x2 , x1 )2 , and we may first study 2-cocycle c2 (and
hence t2g , λ2g , u2 respectively in (3.5)), and continue the proof below. In the end, we
b is torsion free, so
could show [c4 ] = [0], but this implies [c] = [0] since H 2 (G, X)
A REMARK ON T-VALUED COHOMOLOGY GROUPS OF ALGEBRAIC GROUP ACTIONS 9
without loss of generality, we may assume u(x1 , x2 ) = u(x2 , x1 ) for all (x1 , x2 ) ∈ E ′′ ,
i.e. u is symmetric.
Second, we show that we can assume u satisfies the 2-cocycle identity.
Defining F ⊂ X 3 by F = φ−1 (E ′′ ) ∩ ψ −1 (E ′′ ) ∩ p−1 (E ′′ ), where φ, ψ, p : X 3 →
2
X are defined by φ(x1 , x2 , x3 ) = (x1 , x2 + x3 ), ψ(x1 , x2 , x3 ) = (x2 + x3 , x1 ) and
p(x1 , x2 , x3 ) = (x2 , x3 ) respectively.
Note that F is a measurable conull subset of X 3 and we can further assume
F = τ (F ) for all τ ∈ S3 , i.e. (x1 , x2 , x3 ) ∈ F if and only if (xτ (1) , xτ (2) , xτ (3) ) ∈ F .
Given any (x1 , x2 , x3 ) ∈ F , applying (3.5) to (x1 + x2 , x3 ), we deduce the following identity
(3.6) tg (x1 +x2 +x3 ) = tg (x1 +x2 )tg (x3 )λg u(x1 +x2 , x3 )u(g −1 (x1 +x2 ), g −1 x3 )−1 .
Substituting tg (x1 + x2 ) in (3.6), via the formula (3.5), we further deduce
(3.7)
tg (x1 + x2 + x3 )
= u(x1 , x2 )u(g −1 x1 , g −1 x2 )−1 u(x1 +x2 , x3 )u(g −1 (x1 +x2 ), g −1 x3 )−1 .
tg (x1 )tg (x2 )tg (x3 )λ2g
Similarly, we could deduce the following
(3.8)
tg (x1 + x2 + x3 )
= u(x1 , x3 )u(g −1 x1 , g −1 x3 )−1 u(x1 +x3 , x2 )u(g −1 (x1 +x3 ), g −1 x2 )−1 .
tg (x1 )tg (x2 )tg (x3 )λ2g
Now, combining (3.7) and (3.8), we deduce the following equality,
u(x1 , x2 )u(x1 + x2 , x3 )
u(g −1 x1 , g −1 x2 )u(g −1 (x1 + x2 ), g −1 x3 )
=
.
u(x1 , x3 )u(x1 + x3 , x2 )
u(g −1 x1 , g −1 x3 )u(g −1 (x1 + x3 ), g −1 x2 )
Since G y X 3 is weakly mixing, we deduce that
(3.9)
u(x1 , x2 )u(x1 + x2 , x3 )
= a′
u(x1 , x3 )u(x1 + x3 , x2 )
for some constant a′ ∈ T and all (x1 , x2 , x3 ) ∈ F ′ ⊂ X 3 , where F ′ is a conull subset
of X 3 .
Defining F ′′ = ∩τ ∈S3 τ (F ∩ F ′ ), then F ′′ is still conull.
u(x1 ,x3 )u(x1 +x3 ,x2 )
Swapping x2 and x3 in (3.9), we also have u(x
= a′ , hence a′2 = 1,
1 ,x2 )u(x1 +x2 ,x3 )
a′ = ±1. If a′ = −1, we study u2 and c2 first as before, so without loss of generality,
we may assume a′ = 1;
Now, we have shown that u(x1 , x2 )u(x1 + x2 , x3 ) = u(x1 , x3 )u(x1 + x3 , x2 ) for
all (x1 , x2 , x3 ) ∈ F ′′ .
Since u is symmetric and X is abelian, we can also rewrite it as
u(x2 , x1 )u(x2 + x1 , x3 ) = u(x1 , x3 )u(x2 , x1 + x3 )
or, equivalently:
u(x1 , x2 )u(x1 + x2 , x3 ) = u(x2 , x3 )u(x1 , x2 + x3 )
for all (x1 , x2 , x3 ) ∈ F ′′ .
From above, we have shown that u can be assumed to be a measurable symmetric
2-cocycle. Since X is a compact separable abelian group, any such a 2-cocycle is
a (measurable) 2-coboundary (see e.g. [23, Theorem 10] and [24, page 39] or see
[25, line -4 on page 252]), i.e. there exists a measurable map v : X → T such that
u(x1 , x2 ) = v(x1 )v(x2 )v(x1 + x2 )−1 for a.e. (x1 , x2 ) ∈ X 2 .
10
YONGLE JIANG
Notice that (3.5) can be written as
tg (x1 +x2 ) = tg (x1 )tg (x2 )λg v(x1 )v(x2 )v(x1 +x2 )−1 [v(g −1 x1 )v(g −1 x2 )v(g −1 x1 +g −1 x2 )−1 ]−1
or, equivalently:
−1
[tg (x1 )v(x1 )v(g −1 x1 )−1 ][tg (x2 )v(x2 )v(g −1 x2 )−1 ] = λ−1
x1 + g −1 x2 )−1 ].
g [tg (x1 + x2 )v(x1 + x2 )v(g
Then after replacing tg by tg vσg (v ∗ )λg , which does not change the equality in (3.4),
b for all g ∈ G; hence (3.4) holds
and applying Lemma 2.8, we may assume tg ∈ X
b
for all x ∈ X, g, h ∈ G with tg ∈ X, [c] = [0].
Remark 3.1. The above proof (for n = 2) starting from (3.5) is similar to step 5 in
the proof of [14, Theorem 8.2].
4. Applications
Now, we give some applications of our theorem.
We denote by [k] a finite set with k points, and denote by νk the uniformly
distributed measure on [k].
Corollary 4.1. If G ∈ G, e.g. G has Property (T), then H 2 (G, ZG) is torsion free
as an abelian group.
ZG ∼
Proof. For any k ≥ 2, we take f = k ∈ ZG, then note that G y X = d
= [k]G
f
G
, νkG )
ZGf
is a Bernoulli shift, which is mixing. Notice that G y ([k]
is T-valued cocycle
superrigid implies H 1 (G y [k]G ; T) = H 1 (G, T), then applying Theorem 1.1, we
ZG
deduce H 1 (G, ZGf
) = 0.
ZG
→ 0 of
Note that we have a short exact sequence 0 → ZGk → ZG → ZGk
G-modules and ZGf ∼
= ZG whenever f ∈ ZG is not a right zero-divisor. By the
well-known fact in homological algebra [5, Chapter III, Proposition 6.1], we know
×k
the natural map H 2 (G, ZG) → H 2 (G, ZG) is injective for any positive integer k,
hence H 2 (G, ZG) is torsion free.
Remark 4.2. We do not know a direct proof of this corollary when G has Property
(T) without using Popa’s cocycle superrigidity theorem [30]. H 2 (G, ZG) is torsion
free for any finitely presented group G [10, Proposition 13.7.1], but a group G with
(T) might not be finitely presented, e.g. G = SL3 (Fp [X]) [1, Section 3.4].
Corollary 4.3. Considering the Bernoulli shift action G y (TG , ν G ) or G y
([k]G , νk G ) with G ∈ G, where ν is the Haar measure on T, the following are true.
(1) H n (G y TG ; T) = H n (G, T) implies H n (G, ZG) = 0 for n = 1, 2.
ZG
) = 0.
(2) H 1 (G y [k]G ; T) = H 1 (G, T) implies H 1 (G, ZGk
ZG
(3) H 2 (G, ZGk ) = 0 implies H 2 (G, ZG) is k-divisible, i.e. the multiplication by
k map on H 2 (G, ZG) is surjective.
ZG
) = 0 implies H 2 (G, ZG) has no (non-trivial) elements with order
(4) H 1 (G, ZGk
k.
b = ZG, by Corollary 4.1, we can apply Theorem 1.1 for n = 1, 2 to
Proof. Let X
see (1) holds.
b = ZG , we apply Theorem 1.1 for n = 1 to see (2) holds.
Let X
ZGk
For (3) and (4), the proof is similar to the proof of Corollary 4.1. In fact, (3)
and (4) hold without the assumption G ∈ G.
A REMARK ON T-VALUED COHOMOLOGY GROUPS OF ALGEBRAIC GROUP ACTIONS11
For a Bernoulli shift action G y X of a property (T) group G, Sorin Popa
asked whether H n (G y X; T) = H n (G, T) holds for n > 1 [30, Section 6.6];
unfortunately, this is not true for n = 2 in full generality.
Corollary 4.4. There exists a countable discrete group G with property (T) such
that H 2 (G y TG ; T) 6= H 2 (G, T).
Proof. By Corollary 4.3 (1) for n = 2, we only need to find a group G with (T)
and non-trivial H 2 (G, ZG). Then by [5, Chapter VIII, Proposition 6.7], it suffices
to find a group G with (T) such that it is of type F P and has cohomological
dimension cd(G) = 2. Such a group exists, see e.g. [26, Theorem 1][11, Section
9.B][27, Section I.2.b., Theorem 11; Section I.3.g., Theorem 27]. Indeed, we can
take G to be a random group in Gromov’s density model at density d ∈ ( 13 , 12 ).
It is with overwhelming probability torsion free, word-hyperbolic, of cohomological
dimension 2 [27, Section I.3.b], (hence of type F P by [15, Proposition 6.3(3)] and [5,
Chapter VIII, Proposition 6.1]) and has property (T) by Zuk [40] (further clarified
by Kotowski-Kotowski [18]).
Remark 4.5. The important problem of computing H 2 (G y X; T) explicitly for
a Bernoulli action of various examples of property (T) groups remains open. In
particular, it seems unclear whether H 2 (G y X; T) = H 2 (G, T) holds for G =
SL3 (Z).
5. Further Questions
We end this paper by discussing a few open questions which arose from our
investigation.
(1) Can we extend [29, Corollary 4.2] to second cohomology groups?
(2) Can we find a group G ∈ G and f ∈ ZG such that the principal algebraic
ZG
is weakly mixing and has nontrivial H 1 (G, ZG ), or in
action G y X := d
f
ZGf
ZGf
general, H 1 (G y Xf ; T) 6= H 1 (G, T)? Can we find such an example with f ∈
ZG ∩ ℓ1 (G)× ? If such an example exists, it would not be conjugate to a Bernoulli
ZG
) = 0 is equivalent to the right multiplication by f map
shift. Note that H 1 (G, ZGf
2
2
Rf : H (G, ZG) → H (G, ZG) is injective for G ∈ G and f ∈ ZG is not a right-zero
divisor. Also note that the right G-module structure on H 2 (G, ZG) might not be
trivial, which was pointed out to us by Boris Okun.
d
ZG
(3) Does there exist a weakly mixing principal algebraic action G y Xf := ZGf
ZG
)=
such that 0 6= H 2 (G, ZGf
H 2 (GyXf ;T)
?
H 2 (G,T)
This was suggested to us by the referee.
References
1. Bachir Bekka, Pierre de la Harpe, Alain Valette: Kazhdan’s property (T ), New Mathematical
Monographs, 11. Cambridge University Press, 2008.
2. Lewis Bowen: Measure conjugacy invariants for actions of countable sofic groups, J. Amer.
Math. Soc. 23 (2010), 217–245.
3. Lewis Bowen: Entropy for expansive algebraic actions of residually finite groups, Ergodic
Theory Dynam. Systems. 31 (2011), no. 3, 703–718.
4. Lewis Bowen, Hanfeng Li: Harmonic models and spanning forests of residually finite groups,
J. Funct. Anal. 263 (2012), no. 7, 1769–1808.
5. Ken Brown: Cohomology of groups, Graduate Texts in Mathematics, 87, Springer-Verlag
(1982)
12
YONGLE JIANG
6. Christopher Deninger: Fuglede-Kadison determinants and entropy for actions of discrete
amenable groups, J. Amer. Math. Soc. 19 (2006), 737–758.
7. Christopher Deninger, Klaus Schmidt: Expansive algebraic actions of discrete residually finite
amenable groups and their entropy, Ergodic Theory Dynam. Systems. 27 (2007), 769–786.
8. Jacob Feldman, Calvin C. Moore: Ergodic equivalence relations, cohomology, and von Neumann algebras I, II, Trans. Amer. Math. Soc. 234 (1977), 289–359.
9. Alex Furman: On Popa’s Cocycle Superrigidity Theorem, Inter. Math. Res. Notices IMRN,
2007 (2007), 1–46, Art. ID rnm073.
10. Ross Geoghegan: Topological methods in group theory, Graduate Texts in Mathematics, 243,
Springer, New York, 2008.
11. Mikhail Gromov: Asymptotic invariants of infinite groups, Geometric group theory, Vol.
2(Sussex, 1991), London Math. Soc. Lecture Note Ser., vol.182, Cambridge Uni. Press, Cambridge, 1993, pp.1–295.
12. Ben Hayes: Ergodicity of nilpotent group actions, Gauss’s lemma and mixing in the Heisenberg group, Senior Thesis, University of Washington, Seattle, 2009.
13. Ben Hayes: Fuglede-Kadison determinants and sofic Entropy, arxiv 1402.1135
14. Adrian Ioana, Sorin Popa, Stefaan Vaes: A class of superrigid group von Neumann algebras,
Ann. of Math. 178 (2013), 231–286.
15. Ilya Kapovich, Nadia Benakli: Boundaries of hyperbolic groups, Combinatorial and geometric
group theory (New York, 2000/Hoboken, NJ, 2001) 296 (2002), 39–93.
16. David Kerr: Bernoulli actions of sofic groups have completely positive entropy, Israel J. Math.
202 (2014), no. 1, 461–474.
17. David Kerr, Hanfeng Li: Entropy and the variational principle for actions of sofic groups,
Invent. Math. 186 (2011), no. 3, 501–558.
18. Marcin Kotowski, Michal Kotowski: Random groups and property (T): Żuk’s theorem revisited, J. London Math. Soc. 88 (2013), no. 2, 396–416.
19. Hanfeng Li: Compact group automorphisms, addition formulas and Fuglede-Kadison determinants, Ann. of Math. (2) 176 (2012), no. 1, 303–347.
20. Hanfeng Li, Jesse Peterson, Klaus Schmidt: Ergodicity of principal algebraic group actions,
Contemp. Math. 631 (2015), 201–210.
21. Hanfeng Li, Andreas Thom: Entropy, determinants, and L2 -torsion, J. Amer. Math. Soc. 27
(2014), no. 1, 239–292.
22. Niels Meesschaert, Sven Raum, Stefaan Vaes: Stable orbit equivalence of Bernoulli actions of
free groups and isomorphism of some of their factor actions, Exposition. Math. 31 (2013),
274–294.
23. Calvin C. Moore: Group extensions and cohomology for locally compact groups. III, Trans.
Amer. Math. Soc. 221 (1976), 1–33.
24. Calvin C. Moore: Group extensions and cohomology for locally compact groups. IV, Trans.
Amer. Math. Soc. 221 (1976), 35–58.
25. Sergey Neshveyev, Erling Størmer: Ergodic theory and maximal abelian subalgebras of the
hyperfinite factor, J. Funct. Anal. 195 (2002), 239–261.
26. Yann Ollivier: Sharp phase transition theorems for hyperbolicity of random groups, Geom.
Funct. Anal. 14 (2004), no. 3, 595–679.
27. Yann Ollivier: A January 2005 invitation to random groups, Ensaios Matemáticos [Mathematical Surveys], 10. Sociedade Brasileira de Matemática, Rio de Janeiro, 2005.
28. Jesse Peterson: L2 -rigidity in von Neumann algebras, Invent. Math. 175 (2009), no. 2, 417–
433.
29. Jesse Peterson, Thomas Sinclair: On cocycle superrigidity for Gaussian actions, Ergodic
Theory Dynam. Systems. 32 (2012), 249–272.
30. Sorin Popa: Cocycle and orbit equivalence superrigidity for malleable actions of w-rigid
groups, Invent. Math. 170 (2007), 243–295.
31. Sorin Popa: Deformation and rigidity for group actions and von Neumann algebras, in “Proceedings of the International Congress of Mathematicians” (Madrid 2006), Volume I, EMS
Publishing House, Zurich 2006/2007, pp. 445–479.
32. Sorin Popa: On the superrigidity of malleable actions with spectral gap, J. Amer. Math. Soc.
21 (2008), 981–1000.
33. Sorin Popa, Roman Sasyk: On the cohomology of Bernoulli actions, Ergodic Theory Dynam.
Systems. 27 (2007), 241–251.
A REMARK ON T-VALUED COHOMOLOGY GROUPS OF ALGEBRAIC GROUP ACTIONS13
34. Sorin Popa, Stefaan Vaes: Strong rigidity of generalized Bernoulli actions and computations
of their symmetry groups, Adv. Math. 217 (2008), 833–872.
35. Sorin Popa, Stefaan Vaes: Cocycle and orbit superrigidity for lattices in SL(n, R) acting on
homogeneous spaces, Geometry, rigidity and group actions, Eds. B. Farb and D. Fisher. The
University of Chicago Press, 2011, pp. 419–451
36. Daniel J. Rudolph, Klaus Schmidt: Almost block independence and Bernoullicity of Zd -actions
by automorphisms of compact abelian groups, Invent. Math. 120 (1995), 455–488.
37. Klaus Schmidt: The cohomology of higher-dimensional shifts of finite type, Pacific J. Math.
170 (1995), 237–270.
38. Klaus Schmidt: Dynamical Systems of Algebraic Origin, Progress in Mathematics, 128.
Birkhäuser Verlag, Basel, 1995.
39. Robin Tucker-Drob: Invariant means and the structure of inner amenable groups, arxiv:
1407.7474.
40. Andrzej Zuk: Property (T) and Kazhdan constants for discrete groups, Geom. Funct. Anal.
13 (2003), no. 3, 643–670.
Yongle Jiang, Department of Mathematics, SUNY at Buffalo, NY 14260-2900, U.S.A.
E-mail address: [email protected]
| 4math.GR
|
Rate Selection and Power Adaptation
using Maximal Ratio Combining
for the Random Access Gaussian Channel
Arman Hasanzadeh, Jean-Francois Chamberland, Krishna Narayanan
Department of Electrical and Computer Engineering, Texas A&M University
arXiv:1801.03557v1 [cs.IT] 10 Jan 2018
{armanihm,chmbrlnd,krn}@tamu.edu
Abstract—With the emergence of machine-driven communication, there is a renewed interest in the design of random
multiple access schemes for networks with large number of active
devices. Many of the recently proposed access paradigms are
enhancements to slotted ALOHA. One of the popular schemes,
irregular repetition slotted ALOHA (IRSA), is based on an
analogy between multiple access with successive interference
cancellation and message-passing decoding on bipartite graphs.
Most of the results on IRSA and its variants focus on the collision
channel and they ignore physical limitations such as transmit
power constraints and additive Gaussian noise at the physical
layer. As such, naive extensions of IRSA to the Gaussian multiple
access channel are not power efficient in the low signal-to-noiseratio regime. This work introduces a novel paradigm whereby
devices adapt their rates and/or transmit powers based on their
chosen repetition profiles. The receiver performs maximal ratio
combining over all the replicas prior to decoding a message.
Numerical results for finite number of users show that the
proposed scheme can provide substantial improvements in terms
of power efficiency and average rate over standard IRSA.
Index Terms—Slotted ALOHA, Interference cancellation,
Gaussian multiple access, Uncoordinated multiple access
I. I NTRODUCTION
Emerging wireless networks, such as the Internet of Things
(IoT) and vehicular networks, are characterized by massive
numbers of devices and unpredictable, bursty traffic. In such
networks, existing coordinated multiple access mechanisms
are known to consume significant resources to facilitate coordination among active devices. When the typical payloads
of data packets are small, the overhead associated with coordination can result in a substantial throughput penalty. This
situation invites the examination of alternate frameworks. In
particular, efficient random multiple access schemes form a
promising paradigm to manage networks with massive numbers of devices.
It is well known that traditional slotted ALOHA shows
poor performance when the number of users is large; indeed,
multiple access throughput is limited to 1/e ≈ 0.37 [1].
Fortunately, important improvements to slotted ALOHA have
been proposed recently. A key enhancement is produced when
collided packets are not dropped, but stored in a buffer and
subsequently decoded via successive interference cancellation
(SIC) which was first introduced in [2] and later in [3]. Two
important examples of such schemes are Contention Reso-
lution Diversity Slotted ALOHA (CRDSA) [4] and Irregular
Repetition Slotted ALOHA (IRSA) [5], [6], [7]. The latter
article draws a pivotal connection between SIC and messagepassing decoding on bipartite graphs in erasure channels. This
analogy has been leveraged to show that, in the absence of
noise and without any power constraints at the transmitter, the
soliton distribution is an optimal slot selection distribution in
the sense that throughput can be arbitrary close to one when
the number of users is asymptotically large [8].
For additive Gaussian noise multiple access channels with
power constraints, naive extensions of CRDSA or IRSA result
in poor power efficiency, particularly in the low signal-to-noise
ratio (SNR) regime. This is mainly due to the fact that only one
of the repetitions is used in the decoding process even though
a packet may be repeated multiple times. Thus, the energy
transmitted in other slots is not harnessed effectively. Madala
et al. address this shortcoming in [9] and propose a scheme
where each user picks a rate randomly according to a carefully
chosen probability distribution from a set of admissible rates.
It is shown that, asymptotically, the sum rate for all the users in
a frame converges to the GMAC capacity with an additive gap
of Θ(log log K) where K is the number of users. However,
they consider a different framework in which there are no
slots and all the users send at the same time. However, the
number of distinct rates is equal to the number of users K
which makes the receiver very complex and impractical even
for small number of users.
Kissling propose Contention Resolution ALOHA (CRA)
as extensions of CRDSA to the asynchronous (unslotted)
setting [10]. In CRA, users can have variable length packets
and the repetition rate is fixed. Clazzer et al. [11], [12] suggest
using all replicas of a packet for decoding through maximal ratio combining (MRC) or selection combining (SC) to improve
the performance of CRA in the presence of Gaussian noise.
We note that this modification can only improve performance.
Still, their algorithm, called Enhanced Contention Resolution
ALOHA (ECRA), applies to the asynchronous (unslotted)
setting and it is tailored to fixed repetition rates, like CRDSA.
Although their scheme shows good improvement over CRA, it
is very sensitive to rate, i.e., for some rates, performance can
be worse than CRDSA. Moreover, the transmission scheme
remains unaltered compared to CRDSA; only the operation of
the receiver is altered, leaving room for further improvements.
In this article, we capitalize on this opportunity and seek
to advance random multiple access schemes. We consider a
random multiple access channel with additive Gaussian noise
and assume that devices are subject to individual transmit
power constraints. We propose a scheme where devices pick
their repetition rates according to a prescribed distribution.
The devices then adapt their powers or select their rates as
functions of the number of times they repeat their messages
within a frame. More specifically, devices with higher degrees
transmit at lower power levels, a scheme which we call
power adaptation (PA)-IRSA. Alternatively, they can send
information at higher rates under rate selection (RS)-IRSA.
In addition, the receiver uses MRC to recover sent messages
rather than single slot decoding. We demonstrate, through extensive numerical simulations, that the proposed schemes yield
significant performance improvements, in terms of rate and
power consumption, compared to existing schemes, especially
in the low SNR regime.
slot 1
slot 2
slot 3
m4
m4
m3
m3
m2
m2
slot 5
m4
m3
m2
m1
device 1
device 2
device 3
device 4
Fig. 1. This picture offers a notional diagram for the envision scheme. In this
instance, K = 4 devices are transmitting multiple replicas of their messages
within a frame. The messages associated with device i are labeled mi . The
frame is partitioned into M = 5 slots. Each device randomly selects the
subset of slots during which they transmit their message.
Variables
II. S YSTEM M ODEL
In this section, we introduce the abstract framework for
the multiple access system we wish to study. Let K be the
number of active devices, each of which is trying to transmit
one short message to the receiver. This communication task
is enabled through a slotted structure, with every MAC frame
consisting of M slots. Individual slots are composed of L
channel uses. The focus is exclusively on the uplink scenario
where information flows from the wireless devices to a central
receiver. We assume that the access point is aware of K, the
total number of active devices, and it shares this information
with each of them. Based on the system parameters, every
device randomly chooses a repetition pattern according to a
prescribed distribution. Specifically, device k independently
draws repetition count Dk using distribution fD . This device
subsequently chooses Dk slots uniformly within the frame,
and then transmits its packet in all of the selected slots.
We note that this transmission patterns is established using
common randomness between a user and the access point.
As a consequence, all patterns are known at the receiver. A
notional diagram for the proposed access scheme appears in
Fig. 1.
The aggregate access schedule among the active devices can
be represented as a bipartite graph, with messages on one side
and slots on the other, as depicted in Fig. 2. When the receiver
employs successive interference cancellation, this bipartite
representation admits the application of powerful graphical
tools from coding theory [5]. To leverage pertinent concepts
from iterative decoding, we must first establish aPsuitable
notation. Following established literature, let L(x) = i Li xi ,
where Li = fD (i), be the left degree distribution
P polynomial
from the node perspective. Likewise, R(x) = i Ri xi is the
right degree distribution polynomial from the node perspective.
We emphasize that, once the variable node degree distribution
is specified, the check node degree distribution is implicitly
slot 4
Checks
+
c1 – slot 1
+
c2 – slot 2
+
c3 – slot 3
+
c4 – slot 4
+
c5 – slot 5
m1
m2
m3
m4
Fig. 2. The envisioned random access scheme naturally admits a bipartite
representation that can be put in correspondence with a Tanner graph. In this
analogy, the messages play the role of variable nodes, and the time slots act
as check nodes. Edges indicate the time slots during which a particular device
transmits a copy of its message. This relation provides a conceptual bridge
between standard tools in iterative decoding and the multiple access problem
at hand.
determined through the slot selection process. The average
variable node and check node degrees are given by
lavg =
X
i
iLi = L′ (1)
ravg =
K
lavg ,
M
where function L′ (1) represents the derivative of polynomial
L(x) evaluated at one.
Herein, we are particularly interested in decoding strategies
whereby the receiver uses maximal ratio combining, rather
than single-slot decoding, to recover the sent messages sequentially. Depending on the nature of the interference and
background noise, such strategies can enhance performance
drastically. Additionally, we let devices adapt their power
levels or select their code rates as functions of the degrees
of their variable nodes, which correspond to the number of
times a particular message is repeated within a frame. In our
envisioned implementation, users with higher degrees may
transmit at lower power levels (PA-IRSA), or they can send
information at higher rates (RS-IRSA). A design challenge
then is to find suitable left degree distributions for the nodes
under various transmit power constraints. We focus on low
SNR scenarios, an operating regime where the impacts of
the noise variance N0 cannot be neglected. This regime is
common in sensor networks and machine-driven wireless
communications.
III. R ATE S ELECTION
To establish a performance benchmark, we consider the
situation where all the devices employ a same power level.
The transmit energy per channel use, Es , is then uniform
across slots and messages. Moreover, the total expected energy
expended by all devices within a frame is Etot = Klavg LEs .
Suppose, in addition, that the decoding process relies on
standard SIC via message-passing. That is, MRC is not applied
at the receiver at this point; only one replica of a message
is used to recover the sent data. Incidentally, this is the
same procedure as the one proposed in IRSA [5]. Under
this decoding strategy and for a message with degree li ,
the amount of energy transmitted by a device that is not
employed during the decoding process is L(li − 1)Es . This
significantly reduces power efficiency, especially for degree
distributions with high lavg . Under the standard information
theoretic Gaussian approximation, this elementary scheme
yields the following maximum available rate per user
Es
L
.
RIRSA = log 1 +
2
N0
The sum of rates across all active devices per frame is then
limited by
SIRSA = M TIRSA RIRSA
Es
L
,
= M TIRSA log 1 +
2
N0
where TIRSA is the normalized throughput, i.e., the number of
decoded messages within a frame divided by the total number
of slots. Equivalently, one can think of M TIRSA as the total
number of decoded packets via iterative message passing under
standard SIC.
It is well known in the communication literature that MRC
can improve decoding performance, especially in the low SNR
regime [13]. This algorithmic opportunity therefore constitutes
a promising means to enhance decoding performance for
random multiple access schemes. To illustrate these potential
benefits, consider a particular realization of the frame. We
label the set of slots connected to user i by Ui . Then, when
we apply MRC at the receiver, the total SINR experienced by
user i at iteration q becomes
X
Es
,
(rj,q − 1)Es + N0
j∈Ui
where rj,q is the degree of slot j at decoding iteration q [14].
We emphasize that the effective degree of a slot can change at
every iteration of the decoding process because interference
is being peeled off from some slots every time a message
is recovered. Then, under the Gaussian approximation, the
maximum rate available to user i becomes
X
Es
L
i,q
.
Rmax
= log
2
(rj,q − 1)Es + N0
j∈Ui
While this argument is simple for a fixed realization, it
becomes more complicated for a general system. In the latter
case, parameters li and rj,q can be viewed as random variables,
and they are determined by the construction of the transmission graph. Performance improvements can accordingly
i,q be
. Yet,
assessed in terms of expected rate, R̄max = E Rmax
the rate selection task by every device is more ambiguous. In
the MRC case, the maximum rate for a device depends on
the realization of the bipartite graph. That is, for decodability,
i,q
we want every device to choose a rate slightly below Rmax
,
but this latter quantity depends on the actions of other users
and the decoding order adopted by the receiver. Unfortunately,
for uncoordinated random access, wireless devices do not have
access to the full transmission schedule and/or decoding order.
As such, they cannot select a maximum rate with certainty.
To circumvent this difficulty, we adopt a practical strategy
and analyze its performance. We assume that devices select
their rate according to an estimated interference level, with
the following formula
L
α(li − 1)Es
Es
i
RRS = log 1 +
.
+
2
N0
(βravg − 1)Es + N0
Parameters α and β are coefficients that depend on M , K,
Es , and the left degree distribution. Naively, user i attempts to
i,q
estimate Rmax
by assuming that one of the connected slots has
degree one and all other slots have degree equal to βravg . The
tuning variable α serves as a scaling parameter; it is chosen
i,q
such that the selected rate remains close to Rmax
. Finding
suitable values for α and β is accomplished through numerical
simulations. From this preliminary analysis, we gather that
wireless devices that transmit more frequently should pick
higher information rates. This is intuitively pleasing since
devices that are expending more transmit energy within the
context of a frame get larger rates.
A. Decoding Process and Analysis
We assume that the receiver has complete knowledge of the
bipartite graph at the onset of the decoding process. As such,
i
it can compute RRS
for any device and predict whether a
message can be successfully recovered at iteration q. Even
when packet recovery is enhanced by MRC, the graphical
portion of the decoding process follows a standard iterative
message-passing algorithm. That is, when there is a degreeone slot, the receiver attempts to decode the corresponding
packet while leveraging its copies in other slots through MRC.
While the receiver decodes message i, the effective SINR is
determined by the bipartite graph resulting from peeling at
this stage of the iterative process. The decoding of message i
succeeds and all replicas of this message are removed from the
i
graph through interference cancellation whenever RRS
is less
than the effective SINR at this particular stage. This sequential
process continues until there are no decodable degree-one
slot left. At this point, the receiver attempts to decode the
remaining messages, although they may not be connected to
degree-one slots. Here again, MRC is applied and, as such,
i
is below the maximum
decoding succeeds whenever RRS
achievable rate under the residual SINR. This cycle continues
until no additional messages can be recovered through MRC.
B. Efficiency Evaluation
To compare different degree distributions, we explore two
criteria. First, adopting the approach put forth in [5], we assess
the efficiency of MAC schemes by normalizing their total rate
over the sum capacity of the Gaussian MAC channel, Cref . To
make this notion clear, we go through an illustrative example.
Suppose that all the devices send packets in all the slots with
constant energy Ẽs . Then, the total energy transmitted by the
users is Eref = KM LẼs and the reference capacity per frame
can be written as
!
K Ẽs
LM
.
log 1 +
Cref =
2
N0
We define the efficiency η of a random access scheme as the
ratio of the sum rate per frame over the reference capacity.
The fact that the total energy expended in the random MAC
and reference schemes are equal, Etot = Eref , ensures a
meaningful assessment of efficiency. For our example, it is
straightforward to show this balance is reached with
lavg Es
.
M
One can then write the efficiency as
Ẽs =
η=
S
=
Cref
(1)
LM
2
log 1 +
i∈D
where D is the set of decoded users. Using the aforementioned
approximation, we get
α(lavg − 1)Es
Es
M TRS L
+
log 1 +
S RS ≈
2
N0
(βravg − 1)Es + N0
where TRS = |D|/M is the expected throughput of the
rate selection scheme. Altogether, the efficiency of the MRC
scheme with rate adaptation is adequately captured by
ηRS =
S RS
Cref
≈ TRS
log 1 +
Es
N0
+
log 1 +
α(lavg −1)Es
(βravg −1)Es +N0
Klavg Es
MN0
.
Conceptually, rate selection is a means to take advantage of
the benefits associated with MRC. Ideally, we would like
every device to select a rate matched to its effective SINR at
the corresponding decoding step. However and unfortunately,
this is impractical in the uncoordinated framework. Instead,
i
devices choose rates close to Rmax
as a proxy for good
i
performance. To see how well RRS
performs compared to
ideal rates, we introduce a maximum average efficiency,
"
#
X
i
S max = E
Rmax
i∈D
Likewise, we can introduce the notion of maximum efficiency
S
indistinguishable. With this in mind, the expected sum rate
for all the users within a frame is given by
"
#
X
α(li − 1)Es
Es
L
+
log 1 +
S RS = E
2
N0
(βravg − 1)Es + N0
Klavg Es
MN0
(2)
where S denotes the expected sum rate for all the devices
within a frame. Using (1) and (2), the efficiency associated
with IRSA can be expressed as
Es
log 1 + N
SIRSA
0
.
= TIRSA
ηIRSA =
Kl
E
Cref
log 1 + avg s
MN0
We can repeat these steps for a receiver that takes advantage
of MRC. For a prescribed degree distribution, let RRS be the
expected value of the rate selected by individual users, i.e.,
i
RRS = E[RRS
]
L
α(li − 1)Es
Es
= E log 1 +
.
+
2
N0
(βravg − 1)Es + N0
We note that we can derive a convenient upper bound for RRS
using Jensen’s inequality,
L
α(lavg − 1)Es
Es
RRS ≤ log 1 +
.
+
2
N0
(βravg − 1)Es + N0
It is also worth mentioning that, in the low SNR regime,
these two expressions are found empirically to be nearly
ηRS,max =
S max
.
Cref
This quantity is pertinent in that it showcases how close
the proposed framework can get to coordinated performance.
Furthermore, by comparing ηRS and ηRS,max , we can assess
the suitability of our SINR prediction strategy.
Our second performance criterion is related to spectral
efficiency. This popular quantity indicates how much information is received per slot, on average. For the messagepassing decoder without SIC, the average spectral efficiency
is proportional to
γIRSA = TIRSA RIRSA
Likewise, for MRC with rate adaptation, the average spectral
efficiency can be written as
"
#
X
Es
α(li − 1)Es
L
log 1 +
+
E
.
γRS =
2M
N0
(βravg − 1)Es + N0
i∈D
A good benchmark for the maximum spectral efficiency rate
selection with MRC is γRS,max = S max /M .
Having discussed the potential benefits of rate adaptation,
along with pertinent performance criteria, we turn to the
second class of uncoordinated schemes we wish to study.
Specifically, in the next section, we explore the potential gains
associated with varying transmit powers in the context of
random access with MRC and SIC at the receiver.
on Ěsi , the power used by device i, such that the resulting
SINR is equal to that of the idealized system. Using the
effective SINR associated with MRC, and by assuming that
every slot contains ravg messages, we get
IV. P OWER A DAPTATION
li Ěsi
Ês
=
.
(4)
N0
(ravg − 1)Ēs + N0
Isolating the parameters associated with device i, we obtain
For simplicity and fairness, assume that all the active devices adopt a common code rate to transmit data. Then, owing
to the random graph generation and the iterative decoding
process, it is foreseeable that not all active devices need to
employ a same transmit power. Thus, we explore the benefits
derived from power adaptation.
Let R̂ denote the nominal rate shared by all active devices.
In the absence of interference and under the Gaussian approximation, the energy needed to sustain this rate, which we denote
by Ês , is implicitly given by
!
Ês
L
.
R̂ = log 1 +
2
N0
In this section, we aim to design a power adaptation scheme
such that the effective SINR of every user, after applying MRC
and SIC at the receiver, is close to Ês /N0 .
We note that this task is slightly more involved than rate
selection. While successful decoding for rate adaptation is
dictated by the transmission graph, this information is not
sufficient to accurately determine decodability under power
adaptation. The empirical distribution of the transmit powers
associated with every slot is key in computing effective SINR.
Thus, we seek to estimate the right degree and the level of
interference within each slot.
To gain insight, assume that the transmission schedule
produces a regular bipartite graph. That is, all the messages
are repeated exactly lavg times and every slot has degree ravg .
Also, suppose that the energy transmitted per channel use is
Ēs for all the active devices. Then, the minimum possible
SINR experienced during the iterative decoding process is
lavg Ēs
.
(ravg − 1)Ēs + N0
To find the value of Ēs that would lead to a performance
comparable to our idealized system, we equate the expression
above to Ês /N0 ,
lavg Ēs
Ês
=
.
N0
(ravg − 1)Ēs + N0
Solving for Ēs , we obtain
Ēs =
Ês
Ês
+ lavg
(1 − ravg ) N
0
.
(3)
While unsophisticated, the argument above offers a starting
point for power adaptation.
Consider a more general transmission schedule, with a
possible irregular bipartite graph. For the sake of argument,
suppose that every device, other than device i, transmits at
nominal power Ēs as defined in (3). We wish to find conditions
Ês
(ravg − 1)Ēs + N0 .
N0
This equation does not provide an optional power level for
device i. Nevertheless, it offers a conceptual blueprint for a
pragmatic solution. In particular, in our candidate approach,
we assume that device i uses a transmit power Esi that is
proportional to the quantity above;
li Ěsi =
Esi = µĚsi ,
where µ is a tuning parameters that depends on M , K, R̂
and the left degree distribution. The lowest value for µ that
support decoding messages at a code rate equal to R̂, is found
through numerical computations.
Combining the expressions above, we gather that the total
energy transmitted by device i is equal to
li Esi L = µL
Ês
(ravg − 1)Ēs + N0
N0
= µLÊs
1+
(ravg − 1)Ês
(ravg − 1)Ês + N0 lavg
!
(5)
.
It is interesting to note that, irrespective of left degrees, the
total energy used by the devices is the same over a frame.
A. Decoding and Efficiency Evaluation
As before, the receiver has complete information about the
realization of the bipartite transmission graph. Consequently, it
can compute the transmit power selected by every device. With
this knowledge, the iterative decoding algorithm proceeds in
a manner akin to the standard process. The receiver initially
attempts to decode devices that are connected to degreeone slots. When the effective SINR of a message is greater
than Ês /N0 , decoding succeeds. The corresponding data is
recovered and the codeword can be peeled. Once this phase
terminates, the receiver tries to decode the remaining messages
based on the same SINR criteria, even when devices are not
connected to degree-one slots.
To evaluate the performance associated with different degree distributions, we explore the average energy per device
normalized by N0 . More precisely, we compare the average
normalized energy of the proposed power adaptation scheme
!
Ês
(ravg − 1)Ês N0
ΓPA = µ
N0 +
N0
(ravg − 1)Ês + N0 lavg
to the normalized energy associated with IRSA and the minimum level derived from the idealized central scheme,
ΓIRSA = lavg
Ês
N0
Γmin =
Ês
,
N0
Normalized Throughput (T)
1.5
-
L(1) (x) - ηRS
L(1) (x)-ηIRSA
L(1) (x) - ηRS,max
1.2
RS-IRSA
RS-IRSA
RS-IRSA
IRSA
IRSA
IRSA
L(3) (x) - ηRS
L(3) (x) - ηIRSA
L(3) (x) - ηRS,max
L(2) (x)-ηRS
L(2) (x) - ηIRSA
L(2) (x) - ηRS,max
1
Efficiency (η)
L(1) (x)
L(3) (x)
L(2) (x)
L(1) (x)
L(3) (x)
L(2) (x)
1
0.8
0.6
0.4
0.5
0.2
0
0
0.2
0.4
0.6
0.8
G
1
1.2
Fig. 3. Throughput, T , of RS-IRSA and IRSA with ideal soliton (L(1) (x))
with parameter M (number of slots), modified soliton (L(2) (x)) with paramẼs
eter 10 and L(3) (x) as left degree distributions. K = 300, N
= 0.0009.
o
respectively. Coefficient L is omitted from these definitions
because it is common to all of them and, hence, does not
influence the relative character of these values.
Finally, we need to establish a means to fairly compare the
performance of the proposed rate selection and power adaptation schemes. For a fixed ratio K/M , this is accomplished by
first identifying the maximum rate at which rate selection can
operate. We then adopt this rate as a target value for power
adaptation, and see if it can be attained using less energy.
V. N UMERIC R ESULTS
Since there are no closed-form expressions for throughput
as a function of α and β for rate selection, nor as a function of
µ for power adaptation, it is difficult to establish optimal left
degree distributions for these algorithms, both for finite K and
asymptotically. Instead, we seek to demonstrate the value of
the proposed schemes over existing ones. This is accomplished
through numerical computations with three expository degree
distributions. Note that, in all of them, almost half of the
variable nodes have degree equal to two. This conforms to the
intuition that, when properly accounting for radio resources,
excessive repetitions are costly. The candidate left degree
distributions are as follows:
•
The ideal soliton, which is deployed in [9], with parameter Y ,
(
1
,
i=1
(1)
Li = Y 1
2 ≤i ≤ Y;
i(i−1) ,
•
Our proposed distribution, the modified soliton with parameter Y ,
(2)
Li
=
1
1
+
,
i(i − 1) Y (Y − 1)
0
1.4
2≤i≤Y;
0
0.2
0.4
0.6
0.8
G
1
1.2
1.4
Fig. 4. Efficiency, η, of RS-IRSA and IRSA with ideal soliton (L(1) (x)) with
parameter M (number of slots), modified soliton (L(2) (x)) with parameter
Ẽs
= 0.0009.
10 and L(3) (x) as left degree distributions. K = 300, N
o
•
And, the numerically validated discrete distribution,
which is used in [5],
L(3) = 0.4977x2 + 0.2207x3 + 0.0381x4 + 0.0756x5
+ 0.0398x6 + 0.0009x7 + 0.0088x8 + 0.0068x9
+ 0.0030x11 + 0.0429x14 + 0.0081x15 + 0.0576x16 .
A. Rate Selection
Because we are primarily interested in performance at low
SNR, simulations are performed at
Ẽs
= 0.0009 = −30.46 dB.
N0
We note that this puts Es /N0 in the range of -12.78 dB to
1.97 dB. Figure 4 plots the efficiency η of RS-IRSA and
IRSA for different values of M , while K remains constant
at 300. Parameters α and β are derived numerically through
simulations as to ensure that throughput is almost equal to
G = K/M for the three expository distributions. The results
are averaged over 1000 independent trials for every operating
point. As mentioned above, one may anticipate that distributions with higher lavg will show lower performance because
of their misuse of radio resources. There is a natural tension
between a desire for a high left degree to boost performance
through MRC, and the fact that higher slot counts also produce
more interference.
For rate selection, the modified soliton shows the best
performance. This can be attributed to the fact that it features
(2)
the lowest average left degree lavg = 3.42 among the three
distributions. On the other hand, the ideal soliton offers
the lowest performance because of its high average degree,
(1)
lavg = 5.7. Our scheme is very close to the maximum possible
rate for G < 0.8, only losing 0.1 in terms of efficiency.
However, when G > 0.8, efficiency degrades because most
L(1) (x) - γRS
L(1) (x)-γIRSA
L(1) (x) - γRS,max
L(3) (x) - γRS
L(3) (x) - γIRSA
L(3) (x) - γRS,max
This is necessary because the number of interferers in each slot
also grows. Consequently, at larger values for G = K/M ,
performance gains are not as significant and ΓPA converges
to ΓIRSA . The ideal soliton features the largest marginal gain.
L(2) (x)-γRS
L(2) (x) - γIRSA
L(2) (x) - γRS,max
15
10
5
0
0
0.2
0.4
0.6
0.8
G
1
1.2
1.4
Fig. 5. Spectral Efficiency, γ, of RS-IRSA and IRSA with ideal soliton
(L(1) (x)) with parameter M (number of slots), modified soliton (L(2) (x))
Ẽs
with parameter 10 and L(3) (x) as left degree distributions. K = 300, N
=
o
0.0009.
slots have high right degrees and a user with a low left degree
cannot support a high rate in the presence of many interferers.
To maintain a good estimation for G > 0.8, we must decrease
α or increase β, which leads to a reduction in the sum capacity
of the frame. We stress that, in the aforementioned setup, RSIRSA can reach up to 80% of the capacity associated with the
best coordinated scheme, and it features a 50% improvement
in terms of capacity with respect to IRSA. Altogether, the
benefits of the proposed schemes are very significant.
Figure 3 shows that our scheme works even when G exceeds
one, with a throughput very close to G. This implies that the
receiver can decode almost all the messages, while preserving
good efficiency. Figures 4 and 5 show that both efficiency and
spectral efficiency improve as G increases, reaching their peak
value around G = 0.8. This may be a good target operating
point for G = K/M .
More importantly, our proposed rate adaptation scheme,
RS-IRSA, uniformly outperforms IRSA, with very significant
gains. This holds true for the three expository distributions
and different values of G = K/M . This points to the need to
carefully manage radio resources in practical implementations,
and to leverage existing techniques such as MRC within the
iterative decoding process.
B. Power Adaptation
Given our focus on system operation in the low SNR regime,
we fix rate R̂ at 10 and L = 100. The best value for µ
and the repetition distribution are optimized numerically such
that receiver can decode at least 90% of the messages for
K = 300. Figure 6 plots the energy per user normalized to
N0 . Figure 7 shows a comparison between the normalized
throughputs for PA-IRSA and IRSA. The results are averaged
over 1000 independent trials for every operating point. As G
grows, the throughput level is maintained by increasing energy.
Energy per User Normalized to N0 (dB)
Spectral Efficiency (γ bit/slot)
20
4
L(1) (x) - ΓP A
L(1) (x) - ΓIRSA
Γmin
L(2) (x) - ΓP A
L(2) (x) - ΓIRSA
L(3) (x) - ΓP A
L(3) (x) - ΓIRSA
2
0
−2
−4
−6
−8
−10
0
0.2
0.4
0.6
0.8
G
1
1.2
1.4
Fig. 6. Energy per uses normalized to N0 of PA-IRSA and IRSA with
ideal soliton (L(1) (x)) with parameter M (number of slots), modified
soliton (L(2) (x)) with parameter 10 and L(3) (x) as left degree distributions.
K = 300, R̂ = 10, L = 100.
Much like in the rate selection case, the modified soliton
produces the best performance among these three distributions.
In addition to significant power gains, the modified soliton
also shows better normalized throughput. Around G = 0.8,
where IRSA achieves its best normalized throughput, PAIRSA features comparable throughput; yet every device, on
average, employs at least 3 dB less energy than IRSA.
Our proposed scheme, PA-IRSA, works even when G exceeds one. For instance, at G = 1.5, the normalized throughput
corresponding to the modified soliton is 1.4, and it uses 1.5 dB
less energy than IRSA. We note that the throughput of IRSA
drops rapidly for G > 0.8, and it becomes zero at G = 1.5.
The data points for IRSA in Fig. 6 when G > 0.8 are present
solely to ensure an easy comparison between IRSA and PAIRSA.
In the case of power adaptation, again, the most important
lesson is to manage radio resources carefully and to take
advantage of MRC along with successive interference cancellation. Maximum ratio combining is an optimal means to
go beyond single-slot decoding, and it yields significant gains
in several instances. Once this algorithmic enhancement is
incorporated into the system, this opens up new possibilities
for power adaptation.
C. RS and PA Comparison
We compare alternate implementations under system parameter G = 0.8 because all the schemes, IRSA, PA-IRSA, and
L(1) (x)
L(3) (x)
L(2) (x)
L(1) (x)
L(3) (x)
L(2) (x)
1.2
1
-
Energy per User Normalized to N0 (dB)
Normalized Throughput (T)
1.4
PA-IRSA
PA-IRSA
PA-IRSA
IRSA
IRSA
IRSA
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
G
1
1.2
1.4
L(1) (x) - IRSA
L(2) (x) - IRSA
L(3) (x) - IRSA
5
L(1) (x) - PA-IRSA
L(2) (x) - PA-IRSA
L(3) (x) - PA-IRSA
L(1) (x) - RS-IRSA
L(2) (x) - RS-IRSA
L(3) (x) - RS-IRSA
0
−5
−10
−15
2
4
6
8
10
12
14
Average Rate of Users (bit/sec)
16
18
Fig. 7. Normalized throughput, T, of PA-IRSA and IRSA with ideal soliton
(L(1) (x)) with parameter M (number of slots), modified soliton (L(2) (x))
with parameter 10 and L(3) (x) as left degree distributions. K = 300, R̂ =
10, L = 100.
Fig. 8. RS-IRSA, PA-IRSA and IRSA comparison with ideal soliton
(L(1) (x)) with parameter M (number of slots), modified soliton (L(2) (x))
with parameter 10 and L(3) (x) as left degree distributions. K = 300, G =
0.8, L = 100.
RS-IRSA, reach their best performance around this operating
point. Figure 8 displays the respective performance of these
schemes. Various values of Es /N0 are fed to the RS-IRSA
scheme to find the best average rate per user, RRS , such
that normalized throughput is at least 0.78. Energy levels
are chosen as to remain in the low SNR regime. The y-axis
for RS-IRSA corresponds to lavg Es /N0 . The optimal rates
for rate adaptation, RRS , are selected as input to the PAIRSA scheme. Then, we find the minimum energy levels,
ΓPA , such that the normalized throughputs are at least 0.78
at the target rates. Finally, since IRSA only uses one replica
during the decode a message, the energy per channel use for
IRSA is derived directly from the definition of capacity. It
is subsequently multiplied by lavg to form a data point. The
results are averaged over 1000 independent trials for every
sample.
As the rate increases, the energy level also climbs. This
produces higher interference levels, which leads to lower
performances for RS-IRSA and PA-IRSA. Consequently, the
gains are not as pronounced at higher rates. RS-IRSA features
better performance than PA-IRSA. This can be explained by
the fact that, in PA-IRSA, two variables are estimated, i.e.,
number of interferers in a slot and their power levels; whereas,
only the number of interferers is estimated in RS-IRSA. At
low rates or, equivalently, very low SNRs, PA-IRSA and RSIRSA show similar performance. As anticipated, the modified
soliton produces the best performance among the candidate
distributions.
achievable rate per user and the maximum achievable SINR
per user, respectively, to adjust code rates or transmit powers.
The estimation process takes into account the repetition profile
of every device, and the fact that MRC and SIC are utilized at
the receiver. These schemes allow users with higher degrees
to increase their rates or lower their transmit powers, taking
advantage of their effective interference pattern.
Also, we show that, with a finite number of users and in
low SNR regime, the sum rate for all the users can reach
80% of the GMAC capacity with coordination. This, we feel,
is admirable given the information asymmetry between the
two problem formulations. In this same setting, RS-IRSA
showcases a 50% improvement over the state-of-the-art in
slotted random multiple access. Moreover, in this regime, PAIRSA yields a minimum gain of 3 dB in transmit power with
respect to IRSA. Altogether, our contributions demonstrate
that MRC and SIC should both be employed at the receiver.
Furthermore, once these enhancements are in place, it is
valuable to tailor the left degree distribution to this reality. This
results in important gains in terms of rates or power savings,
especially in the low SNR regime. These benefits hinge on
the ability to accurately predict effective interference levels
experienced during the decoding process.
There are several possible avenues for future inquiries.
While we introduce a new viewpoint for random access and
show excellent performance improvements, we do not present
an optimal family of distributions for slot selection. This
optimization problem appears very challenging, as it relies
on a multitude of factors including K, M , and N0 . Still,
providing better insights for the selection of good distributions
would be valuable. The framework studied in this paper
focuses on a single frame. Expanding our findings to scenarios
with multiple frames or with frame length adaptation forms
VI. C ONCLUSION
In this paper, we introduce two new random access protocols
for the slotted GMAC channel. Our proposed schemes, RSIRSA and PA-IRSA, employ estimates for the maximum
an interesting challenge. Finally, while the results are very
promising, many practical concerns should be address before
these techniques are pushed into standards. This too warrants
attention.
R EFERENCES
[1] R. Gallager, “A perspective on multiaccess channels,” IEEE Trans. Inform. Theory, vol. 31, no. 2, pp. 124–142, 1985.
[2] Y. Yu and G. Giannakis, “High-throughput random access using successive interference cancellation in a tree algorithm,” IEEE Trans. Inform. Theory, vol. 53, no. 12, pp. 4628–4639, 2007.
[3] S. Gollakota and D. Katabi, Zigzag decoding: Combating hidden terminals in wireless networks, vol. 38. ACM, 2008.
[4] E. Casini, R. De Gaudenzi, and O. Herrero, “Contention resolution diversity slotted ALOHA (CRDSA): An enhanced random access schemefor
satellite access packet networks,” IEEE Trans. on Wireless Commun.,
vol. 6, no. 4, pp. 1408–1419, 2007.
[5] G. Liva, “Graph-based analysis and optimization of contention resolution
diversity slotted ALOHA,” IEEE Trans. on Commun., vol. 59, no. 2,
pp. 477–487, 2011.
[6] E. Paolini, G. Liva, and M. Chiani, “High throughput random access via
codes on graphs: Coded slotted ALOHA,” in Proc. IEEE International
Conf. Commun., pp. 1–6, 2011.
[7] E. Paolini, C. Stefanovic, G. Liva, and P. Popovski, “Coded random
access: applying codes on graphs to design random access protocols,”
IEEE Communications Magazine, vol. 53, pp. 144–150, June 2015.
[8] K. R. Narayanan and H. D. Pfister, “Iterative collision resolution for
slotted ALOHA: An optimal uncoordinated transmission policy,” in
Proc. Intl. Symp. Turbo Codes and Rel. Topics., pp. 136–139, IEEE,
2012.
[9] S. Madala and K. Narayanan, “Uncoordinated rate selection: Approaching the capacity of Gaussian MAC without coordination,” in Proc. IEEE
International Conf. Commun., pp. 2057–2062, June 2015.
[10] C. Kissling, “Performance enhancements for asynchronous random
access protocols over satellite,” in 2011 IEEE International Conference
on Communications (ICC), pp. 1–6, June 2011.
[11] F. Clazzer, C. Kissling, and M. Marchese, “Exploiting combination techniques in random access mac protocols: Enhanced contention resolution
aloha,” arXiv preprint arXiv:1602.07636, 2016.
[12] F. Clazzer, F. Lazaro, G. Liva, and M. Marchese, “Detection and
combining techniques for asynchronous random access with time diversity,” in SCC 2017; 11th International ITG Conference on Systems,
Communications and Coding, pp. 1–6, Feb 2017.
[13] D. G. Brennan, “Linear diversity combining techniques,” Proceedings
of the IRE, vol. 47, pp. 1075–1102, June 1959.
[14] W. Jakes, Microwave mobile communications. IEEE Press classic
reissue, IEEE Press, 1974.
| 7cs.IT
|
Semi-supervised Question Retrieval with Gated Convolutions
Tao Lei
Hrishikesh Joshi Regina Barzilay
MIT CSAIL
Tommi Jaakkola
{taolei, hjoshi, regina, tommi}@csail.mit.edu
Katerina Tymoshenko
University of Trento
arXiv:1512.05726v2 [cs.CL] 4 Apr 2016
[email protected]
Alessandro Moschitti Lluı́s Màrquez
Qatar Computing Research Institute, HBKU
{amoschitti, lmarquez}@qf.org.qa
Abstract
Question answering forums are rapidly growing in size with no effective automated ability to refer to and reuse answers already available for previous posted questions. In this
paper, we develop a methodology for finding semantically related questions. The task
is difficult since 1) key pieces of information are often buried in extraneous details in
the question body and 2) available annotations on similar questions are scarce and fragmented. We design a recurrent and convolutional model (gated convolution) to effectively map questions to their semantic representations. The models are pre-trained within
an encoder-decoder framework (from body to
title) on the basis of the entire raw corpus,
and fine-tuned discriminatively from limited
annotations. Our evaluation demonstrates that
our model yields substantial gains over a standard IR baseline and various neural network
architectures (including CNNs, LSTMs and
GRUs).1
1
Introduction
Question answering (QA) forums such as Stack Exchange2 are rapidly expanding and already contain
millions of questions. The expanding scope and coverage of these forums often leads to many duplicate and interrelated questions, resulting in the same
questions being answered multiple times. By identifying similar questions, we can potentially reuse
1
Our code and data are available at https://github.
com/taolei87/rcnn
2
http://stackexchange.com/
Title: How can I boot Ubuntu from a USB?
Body: I bought a Compaq pc with Windows 8 a few
months ago and now I want to install Ubuntu but still
keep Windows 8. I tried Webi but when my pc restarts it
read ERROR 0x000007b. I know that Windows 8 has a
thing about not letting you have Ubuntu but I still want
to have both OS without actually losing all my data ...
Title: When I want to install Ubuntu on my laptop I’ll
have to erase all my data. “Alonge side windows” doesnt
appear
Body: I want to install Ubuntu from a Usb drive. It
says I have to erase all my data but I want to install it
along side Windows 8. The “Install alongside windows”
option doesn’t appear. What appear is, ...
Figure 1: A pair of similar questions.
existing answers, reducing response times and unnecessary repeated work. Unfortunately in most forums, the process of identifying and referring to existing similar questions is done manually by forum
participants with limited, scattered success.
The task of automatically retrieving similar questions to a given user’s question has recently attracted
significant attention and has become a testbed for
various representation learning approaches (Zhou et
al., 2015; dos Santos et al., 2015). However, the task
has proven to be quite challenging – for instance, dos
Santos et al. (2015) report a 22.3% classification accuracy, yielding a 4 percent gain over a simple word
matching baseline.
Several factors make the problem difficult. First,
submitted questions are often long and contain extraneous information irrelevant to the main question
being asked. For instance, the first question in Figure 1 pertains to booting Ubuntu using a USB stick.
A large portion of the body contains tangential de-
tails that are idiosyncratic to this user, such as references to Compaq pc, Webi and the error message.
Not surprisingly, these features are not repeated in
the second question in Figure 1 about a closely related topic. The extraneous detail can easily confuse
simple word-matching algorithms. Indeed, for this
reason, some existing methods for question retrieval
restrict attention to the question title only. While titles (when available) can succinctly summarize the
intent, they also sometimes lack crucial detail available in the question body. For example, the title
of the second question does not refer to installation
from a USB drive. The second challenge arises from
the noisy annotations. Indeed, the pairs of questions
marked as similar by forum participants are largely
incomplete. Our manual inspection of a sample set
of questions from AskUbuntu3 shows that only 5%
of similar pairs have been annotated by the users,
with a precision of around 79%.
In this paper, we design a neural network model
and an associated training paradigm to address these
challenges. On a high level, our model is used as
an encoder to map the title, body, or the combination to a vector representation. The resulting “question vector” representation is then compared to other
questions via cosine similarity. We introduce several departures from typical architectures on a finer
level. In particular, we incorporate adaptive gating
in non-consecutive CNNs (Lei et al., 2015) in order to focus temporal averaging in these models on
key pieces of the questions. Gating plays a similar
role in LSTMs (Hochreiter and Schmidhuber, 1997),
though LSTMs do not reach the same level of performance in our setting. Moreover, we counter the
scattered annotations available from user-driven associations by training the model largely based on
the entire unannotated corpus. The encoder is coupled with a decoder and trained to reproduce the title from the noisy question body. The methodology
is reminiscent of recent encoder-decoder networks
in machine translation and document summarization (Kalchbrenner and Blunsom, 2013; Sutskever
et al., 2014; Cho et al., 2014b; Rush et al., 2015).
The resulting encoder is subsequently fine-tuned
discriminatively on the basis of limited annotations
yielding an additional performance boost.
3
http://askubuntu.com/
We evaluate our model on the AskUbuntu corpus
from Stack Exchange used in prior work (dos Santos et al., 2015). During training, we directly utilize noisy pairs readily available in the forum, but
to have a realistic evaluation of the system performance, we manually annotate 8K pairs of questions.
This clean data is used in two splits, one for development and hyper parameter tuning and another
for testing. We evaluate our model and the baselines using standard information retrieval (IR) measures such as Mean Average Precision (MAP), Mean
Reciprocal Rank (MRR) and Precision at n (P@n).
Our full model achieves a MRR of 75.6% and P@1
of 62.0%, yielding 8% absolute improvement over
a standard IR baseline, and 4% over standard neural
network architectures (including CNNs, LSTMs and
GRUs).
2
Related Work
Given the growing popularity of community QA forums, question retrieval has emerged as an important
area of research (Nakov et al., 2015; Nakov et al.,
2016). Previous work on question retrieval has modeled this task using machine translation, topic modeling and knowledge graph-based approaches (Jeon
et al., 2005; Li and Manandhar, 2011; Duan et al.,
2008; Zhou et al., 2013). More recent work relies
on representation learning to go beyond word-based
methods. For instance, Zhou et al. (2015) learn
word embeddings using category-based metadata information for questions. They define each question
as a distribution which generates each word (embedding) independently, and subsequently use a Fisher
kernel to assess question similarities. Dos Santos
et al. (2015) propose an approach which combines
a convolutional neural network (CNN) and a bagof-words representation for comparing questions. In
contrast to (Zhou et al., 2015), our model treats each
question as a word sequence as opposed to a bag
of words, and we apply a recurrent convolutional
model as opposed to the traditional CNN model
used by dos Santos et al. (2015) to map questions
into meaning representations. Further, we propose
a training paradigm that utilizes the entire corpus of
unannotated questions in a semi-supervised manner.
Recent work on answer selection on community
QA forums, similar to our task of question retrieval,
(a) similarity
has also involved
the usemodel:
of neural network architectures (Severyn and Moschitti, 2015; Wang and
cosine similarity
Nyberg, 2015; Shen et al., 2015; Feng et al., 2015;
pooling
pooling
Tan et al., 2015). Compared
to our work,
these approaches focus encoder
on improving various
other aspects
encoder
of the model. For instance, Feng et al. (2015) explore different similarity measures beyond cosine
similarity, and Tan et al. (2015) adopt the neural atquestion 1
question 2
tention mechanism over RNNs to generate better answer representations given the questions as context.
(b) pre-training:
(a) similarity model:
cosine similarity
pooling
encoder
Question Retrieval Setup
encoder
…
decoder
encoder
question 1
question 2
(b) pre-training:
</s>
3
pooling
encoder
…
</s>
decoder
…
…
We begin by introducing the basic discriminative
<s>
setting for retrieving similar questions. Let q be
context (title/body)
title
a query question which generally consists of both
a title sentence and a body section. For efficiency
reasons, we do not compare q against all the other
queries in the data base. Instead, we retrieve first a
smallerQCRI/MIT-CSAIL
candidate set
of related
Q(q) using
Annual
Meeting
– questions
March
QCRI/MIT-CSAIL Annual
33
Meeting
– March
2014 2015
‹#› ‹#›
a standard
IR engine, and then we apply the more
sophisticated models only to this reduced set. Our
goal is to rank the candidate questions in Q(q) so
that all the similar questions to q are ranked above
the dissimilar ones. To do so, we define a similarity
score s(q, p; θ) with parameters θ, where the similarity measures how closely candidate p ∈ Q(q) is
related to question q. The method of comparison can
make use of the title and body of each question.
The scoring function s(·, ·; θ) can
be optimized on
−
the basis of annotated data D = (qi , p+
i , Qi ) ,
where p+
i is a question similar to question qi and
Q−
is
a
negative set of questions deemed not simii
lar to qi . During training, the correct pairs of similar
questions are obtained from available user-marked
pairs, while the negative set Q−
i is drawn randomly
from the entire corpus with the idea that the likelihood of a positive match is small given the size of
the corpus. The candidate set during training is just
−
Q(qi ) = {p+
i } ∪ Qi . During testing, the candidate
sets are retrieved by an IR engine and we evaluate
against explicit manual annotations.
In the purely discriminative setting, we use a maxmargin framework for learning (or fine-tuning) parameters θ. Specifically, in a context of a particular training example where qi is paired with p+
i , we
<s>
context (title/body)
title
Figure 2: Illustration of our model.
minimize the max-margin loss L(θ) defined as
+
max s(qi , p; θ) − s(qi , p+
i ; θ) + δ(p, pi ) ,
p∈Q(qi )
where δ(·, ·) denotes a non-negative margin. We set
+
δ(p, p+
i ) to be a small constant when p 6= pi and
0 otherwise. The parameters θ can be optimized
through sub-gradients ∂L/∂θ aggregated over small
batches of the training instances.
There are two key problems that remain. First,
we have to define and parameterize the scoring function s(q, p; θ). We design a recurrent neural network
model for this purpose and use it as an encoder to
map each question into its meaning representation.
The resulting similarity function s(q, p; θ) is just the
cosine similarity between the corresponding representations, as shown in Figure 2 (a). The parameters θ pertain to the neural network only. Second,
in order to offset the scarcity and limited coverage
of the training annotations, we pre-train the parameters θ on the basis of the much larger unannotated
corpus. The resulting parameters are subsequently
fine-tuned using the discriminative setup described
above.
4
4.1
Model
Non-consecutive Convolution
We describe here our encoder model, i.e., the
method for mapping the question title and body to
a vector representation. Our approach is inspired
by temporal convolutional neural networks (LeCun
et al., 1998) and, in particular, its recent refinement (Lei et al., 2015), tailored to capture longerrange, non-consecutive patterns in a weighted manner. Such models can be used to effectively summarize occurrences of patterns in text and aggregate them into a vector representation. However,
the summary produced is not selective since all pattern occurrences are counted, weighted by how cohesive (non-consecutive) they are. In our problem,
the question body tends to be very long and full of
irrelevant words and fragments. Thus, we believe
that interpreting the question body requires a more
selective approach to pattern extraction.
Our model successively reads tokens in the question title or body, denoted as {xi }li=1 , and transforms this sequence into a sequence of states
{hi }li=1 . The resulting state sequence is subsequently aggregated into a single final vector representation for each text as discussed below. Our approach builds on (Lei et al., 2015), thus we begin by
briefly outlining it. Let W1 and W2 denote filter matrices (as parameters) for pattern size n = 2. Lei et
al. (2015) generate a sequence of states in response
to tokens according to
ct0 ,t = W1 xt0 + W2 xt
X
0
ct =
λt−t −1 ct0 ,t
t0 <t
ht = tanh(ct + b)
where ct0 ,t represents a bigram pattern, ct accumulates a range of patterns and λ ∈ [0, 1) is a constant decay factor used to down-weight patterns with
longer spans. The operations can be cast in a “recurrent” manner and evaluated with dynamic programming. The problem with the approach for our
purposes is, however, that the weighting factor λ is
the same (constant) for all, not triggered by the state
ht−1 or the observed token xt .
Adaptive Gated Decay We refine this model by
learning context dependent weights. For example,
if the current input token provides no relevant information (e.g., symbols, functional words), the model
should ignore it by incorporating the token with a
vanishing weight. In contrast, strong semantic content words such as “ubuntu” or “windows” should be
included with much larger weights. To achieve this
effect we introduce neural gates similar to LSTMs
to specify when and how to average the observed
signals. The resulting architecture integrates recurrent networks with non-consecutive convolutional
models:
λt = σ(Wλ xt + Uλ ht−1 + bλ )
(1)
ct
(2)
ct
(1)
= λt
ct−1 + (1 − λt )
(W1 xt )
= λt
(2)
ct−1
(ct−1 + W2 xt )
+ (1 − λt )
(1)
···
(n)
ct
= λt
ht =
(n)
ct−1 + (1 − λt )
(n)
tanh(ct
(n−1)
(ct−1
+ Wn xt )
+ b)
where σ(·) is the sigmoid function and represents
(1)
(n)
the element-wise product. Here ct , · · · , ct are
accumulator vectors that store weighted averages of
1-gram to n-gram features. When the gate λt = 0
(vector) for all t, the model represents a traditional
(n)
CNN with filter width n. As λt > 0, however, ct
becomes the sum of an exponential number of terms,
enumerating all possible n-grams within x1 , · · · , xt
(seen by expanding the formulas). Note that the gate
λt (·) is parametrized and responds directly to the
previous state and the token in question. We refer
to this model as RCNN from here on.
Pooling In order to use the model as part of the
discriminative question retrieval framework outlined
earlier, we must condense the state sequence to a single vector. There are two simple alternative pooling
strategies that we have explored – either averaging
over the states4 or simply taking the last one as the
meaning representation. In addition, we apply the
encoder to both the question title and body, and the
final representation is computed as the average of the
two resulting vectors.
Once the aggregation is specified, the parameters
of the gate and the filter matrices can be learned in a
purely discriminative fashion. Given that the available annotations are limited and user-guided, we instead use the discriminative training only for fine
tuning an already trained model. The method of pretraining the model on the basis of the entire corpus
of questions is discussed next.
4
We also normalize state vectors before averaging, which
empirically gets better performance.
4.2
Pre-training Using the Entire Corpus
The number of questions in the AskUbuntu corpus
far exceeds user annotations of pairs of similar questions. We can make use of this larger raw corpus in
two different ways. First, since models take word
embeddings as input we can tailor the embeddings
to the specific vocabulary and expressions in this
corpus. To this end, we run word2vec (Mikolov
et al., 2013) on the raw corpus in addition to the
Wikipedia dump. Second, and more importantly,
we use individual questions as training examples
for an auto-encoder constructed by pairing the encoder model (RCNN) with an corresponding decoder (of the same type). As illustrated in Figure 2 (b), the resulting encoder-decoder architecture
is akin to those used in machine translation (Kalchbrenner and Blunsom, 2013; Sutskever et al., 2014;
Cho et al., 2014b) and summarization (Rush et al.,
2015).
Our encoder-decoder pair represents a conditional
language model P (title|context), where the context
can be any of (a) the original title itself, (b) the question body and (c) the title/body of a similar question. All possible (title, context) pairs are used during training to optimize the likelihood of the words
(and their order) in the titles. We use the question
title as the target for two reasons. The question body
contains more information than the title but also has
many irrelevant details. As a result, we can view the
title as a distilled summary of the noisy body, and
the encoder-decoder model is trained to act as a denoising auto-encoder. Moreover, training a decoder
for the title (rather than the body) is also much faster
since titles tend to be short (around 10 words).
The encoders pre-trained in this manner are subsequently fine-tuned according to the discriminative
criterion described already in Section 3.
5
Alternative models
For comparison, we also train three alternative
benchmark encoders (LSTMs, GRUs and CNNs) for
mapping questions to vector representations. LSTM
and GRU-based encoders can be pre-trained analogously to RCNNs, and fine-tuned discriminatively.
CNN encoders, on the other hand, are only trained
discriminatively. While plausible, neither alternative
reaches quite the same level of performance as our
pre-trained RCNN.
LSTMs LSTM cells (Hochreiter and Schmidhuber, 1997) have been used to capture semantic information across a wide range of applications, including machine translation and entailment recognition (Bahdanau et al., 2015; Bowman et al., 2015;
Rocktäschel et al., 2016). Their success can be attributed to neural gates that adaptively read or discard information to/from internal memory states.
Specifically, a LSTM network successively reads
the input token xt , internal state ct−1 , as well as
the visible state ht−1 , and generates the new states
c t , ht :
it = σ(Wi xt + Ui ht−1 + bi )
ft = σ(Wf xt + Uf ht−1 + bf )
ot = σ(Wo xt + Uo ht−1 + bo )
zt = tanh(Wz xt + Uz ht−1 + bz )
ct = i t
zt + ft
ht = ot
tanh(ct )
ct−1
where i, f and o are input, forget and output
gates, respectively. Given the visible state sequence
{hi }li=1 , we can aggregate it to a single vector exactly as with RCNNs. The LSTM encoder can be
pre-trained (and fine-tuned) in the similar way as
our RCNN model. For instance, Dai and Le (2015)
recently adopted pre-training for text classification
task.
GRUs A GRU is another comparable unit for sequence modeling (Cho et al., 2014a; Chung et al.,
2014). Similar to the LSTM unit, the GRU has two
neural gates that control the flow of information:
it = σ(Wi xt + Ui ht−1 + bi )
rt = σ(Wr xt + Ur ht−1 + br )
ct = tanh(Wxt + U(rt
ht = it
ct + (1 − it )
ht−1 ) + b)
ht−1
where i and r are input and reset gate respectively.
Again, the GRUs can be trained in the same way.
CNNs Convolutional neural networks (LeCun et
al., 1998) have also been successfully applied to various NLP tasks (Kalchbrenner et al., 2014; Kim,
2014; Kim et al., 2015; Zhang et al., 2015; Gao
et al., 2014). As models, they are different from
LSTMs since the temporal convolution operation
Corpus
Training
Dev
Test
# of unique questions
Avg length of title
Avg length of body
# of unique questions
# of user-marked pairs
# of query questions
# of annotated pairs
Avg # of positive pairs per query
# of query questions
# of annotated pairs
Avg # of positive pairs per query
167,765
6.7
59.7
12,584
16,391
200
200×20
5.8
200
200×20
5.5
Table 1: Various statistics from our Training, Dev, and Test
sets derived from the Sept. 2014 Stack Exchange AskUbuntu
dataset.
and associated filters map local chunks (windows) of
the input into a feature representation. Concretely, if
we let n denote the filter width, and W1 , · · · , Wn
the corresponding filter matrices, then the convolution operation is applied to each window of n consecutive words as follows:
ct = W1 xt−n+1 + W2 xt−n+2 + · · · + Wn xt
ht = tanh(ct + b)
The sets of output state vectors {ht } produced in
this case are typically referred to as feature maps.
Since each vector in the feature map only pertains
to local information, the last vector is not sufficient
to capture the meaning of the entire sequence. Instead, we consider max-pooling or average-pooling
to obtain the aggregate representation for the entire
sequence.
6
Experimental Setup
Dataset We use the Stack Exchange AskUbuntu
dataset used in prior work (dos Santos et al., 2015).
This dataset contains 167,765 unique questions,
each consisting of a title and a body5 , and a set of
user-marked similar question pairs. We provide various statistics from this dataset in Table 1.
Gold Standard for Evaluation User-marked similar question pairs on QA sites are often known
to be incomplete. In order to evaluate this in our
dataset, we took a sample set of questions paired
with 20 candidate questions retrieved by a search engine trained on the AskUbuntu data. The search engine used is the well-known BM25 model (Robert5
We truncate the body section at a maximum of 100 words.
son and Zaragoza, 2009). Our manual evaluation of
the candidates showed that only 5% of the similar
questions were marked by users, with a precision of
79%. Clearly, this low recall would not lead to a realistic evaluation if we used user marks as our gold
standard. Instead, we make use of expert annotations carried out on a subset of questions.
Training Set We use user-marked similar pairs as
positive pairs in training since the annotations have
high precision and do not require additional manual annotations. This allows us to use a much larger
training set. We use random questions from the corpus paired with each query question pi as negative
pairs in training. We randomly sample 20 questions
as negative examples for each pi during each epoch.
Development and Test Sets We re-constructed
the new dev and test sets consisting of the first 200
questions from the dev and test sets provided by dos
Santos et al. (2015). For each of the above questions, we retrieved the top 20 similar candidates using BM25 and manually annotated the resulting 8K
pairs as similar or non-similar.6
Baselines and Evaluation Metrics We evaluated
neural network models—including CNNs, LSTMs,
GRUs and RCNNs—by comparing them with the
following baselines:
• BM25, we used the BM25 similarity measure
provided by Apache Lucene.
• TF-IDF, we ranked questions using cosine
similarity based on a vector-based word representation for each question.
• SVM, we trained a re-ranker using SVM-Light
(Joachims, 2002) with a linear kernel incorporating several similarity measures from the
DKPro similarity package (Bär et al., 2013).
We evaluated the models based on the following IR
metrics: Mean Average Precision (MAP), Mean Reciprocal Rank (MRR), Precision at 1 (P@1), and
Precision at 5 (P@5).
6
The annotation task was initially carried out by two expert
annotators, independently. The initial set was refined by comparing the annotations and asking a third judge to make a final
decision on disagreements. After a consensus on the annotation
guidelines was reached (producing a Cohen’s kappa of 0.73),
the overall annotation was carried out by only one expert.
Method
Pooling
BM25
TF-IDF
SVM
CNNs
LSTMs
GRUs
RCNNs
LSTMs + pre-train
GRUs + pre-train
RCNNs + pre-train
mean
mean
mean
last
mean
last
last
MAP
52.0
54.1
53.5
58.5
58.4
59.1
59.9
58.3
59.3
61.3∗
Dev
MRR P@1
66.0
51.9
68.2
55.6
66.1
50.8
71.1
58.4
72.3
60.0
74.0
62.6
74.2
63.2
71.5
59.3
72.2
59.8
75.2
64.2
P@5
42.1
45.1
43.8
46.4
46.4
47.3
48.0
47.4
48.3
50.3∗
MAP
56.0
53.2
57.7
57.6
56.8
57.1
60.7
55.5
59.3
62.3∗
Test
MRR P@1
68.0
53.8
67.1
53.8
71.3
57.0
71.4
57.6
70.1
55.8
71.4
57.3
72.9
59.1
67.0
51.1
71.3
57.2
75.6∗ 62.0
P@5
42.5
39.7
43.3
43.2
43.2
43.6
45.0
43.4
44.3
47.1∗
Table 2: Comparative results of all methods on the question similarity task. Higher numbers are better. For neural network models,
we show the best average performance across 5 independent runs and the corresponding pooling strategy. Statistical significance
with p < 0.05 against other types of model is marked with ∗.
LSTMs
GRUs
CNNs
RCNNs
d
240
280
667
400
|θ|
423K
404K
401K
401K
n
3
2
Table 3: Configuration of neural models. d is the hidden dimension, |θ| is the number of parameters and n is the filter width.
Hyper-parameters We performed an extensive
hyper-parameter search to identify the best model
for the baselines and neural network models. For
the TF-IDF baseline, we tried n-gram feature order
n ∈ {1, 2, 3} with and without stop words pruning.
For the SVM baseline, we used the default SVMLight parameters whereas the dev data is only used
to increase the training set size when testing on the
test set. We also tried to give higher weight to dev
instances but this did not result in any improvement.
For all the neural network models, we used
Adam (Kingma and Ba, 2015) as the optimization method with the default setting suggested by
the authors. We optimized other hyper-parameters
with the following range of values: learning rate
∈ {1e − 3, 3e − 4}, dropout (Hinton et al., 2012)
probability ∈ {0.1, 0.2, 0.3}, CNN feature width
∈ {2, 3, 4}. We also tuned the pooling strategies
and ensured each model has a comparable number of
parameters. The default configurations of LSTMs,
GRUs, CNNs and RCNNs are shown in Table 3. We
used MRR to identify the best training epoch and
the model configuration. For the same model configuration, we report average performance across 5
independent runs.7
Word Vectors We ran word2vec (Mikolov et al.,
2013) to obtain 200-dimensional word embeddings
using all Stack Exchange data (excluding StackOverflow) and a large Wikipedia corpus. The word
vectors are fixed to avoid over-fitting across all experiments.
7
Results
Overall Performance Table 2 shows the performance of the baselines and the neural encoder models on the question retrieval task. The results
show that our full model, RCNNs with pre-training,
achieves the best performance across all metrics on
both the dev and test sets. For instance, the full
model gets a P@1 of 62.0% on the test set, outperforming the word matching-based method BM25 by
over 8 percent points. Further, our RCNN model
also outperforms the other neural encoder models and the baselines across all metrics. This superior performance indicates that the use of nonconsecutive filters and a varying decay is effective
in improving traditional neural network models.
Table 2 also demonstrates the performance gain
from pre-training the RCNN encoder. The RCNN
model when pre-trained on the entire corpus consistently gets better results across all the metrics.
7
For a fair comparison, we also pre-train 5 independent
models for each configuration and then fine tune these models. We use the same learning rate and dropout rate during pretraining and fine-tuning.
Method
CNNs, max-pooling
CNNs, mean-pooling
LSTMs + pre-train, mean-pooling
LSTMs + pre-train, last state
GRUs + pre-train, mean-pooling
GRUs + pre-train, last state
RCNNs + pre-train, mean-pooling
RCNNs + pre-train, last state
MAP
57.8
58.5
58.3
57.6
57.5
59.3
59.3
61.3
Dev
MRR P@1
69.9
56.6
71.1
58.4
71.5
59.3
71.0
58.1
69.9
57.1
72.2
59.8
73.6
61.7
75.2
64.2
P@5
47.7
46.4
47.4
47.3
46.2
48.3
48.6
50.3
Test
MRR P@1
73.1
59.6
71.4
57.6
67.0
51.1
69.8
55.2
67.3
52.4
71.3
57.2
72.3
57.3
75.6
62.0
MAP
59.6
57.6
55.5
57.6
55.5
59.3
58.9
62.3
P@5
45.4
43.2
43.4
43.7
42.8
44.3
45.3
47.1
Table 4: Choice of pooling strategies.
MRR
66.8
67.1
P@1
52.7
53.8
95
80
75
RCNN
GRU
LSTM 70
RCNNs, mean-pooling
title only
title + body
MAP
56.0
58.5
MRR
68.9
71.7
P@1
55.7
56.7
65
65
50
60
RCNNs, last state
title only
title + body
MAP
58.2
60.7
MRR
70.7
72.9
P@1
56.6
59.1
35
55
Table 5: Comparision between model variants on the test set
when question bodies are used or not used.
20
MRR
MAP
54.3
53.2
PPL
TF-IDF
title only
title + body
1
2
3
4
5
6
7
8
9
50
10
Epoch
Figure 3: Perplexity (dotted lines) on a heldout portion of the
corpus versus MRR on the dev set (solid lines) during pre-
Pooling Strategy We analyze the effect of various
pooling strategies for the neural network encoders.
As shown in Table 4, our RCNN model outperforms
other neural models regardless of the two pooling
strategies explored. We also observe that simply using the last hidden state as the final representation
achieves better results for the RCNN model.
Using Question Body Table 5 compares the performance of the TF-IDF baseline and the RCNN
model when using question titles only or when using
question titles along with question bodies. TF-IDF’s
performance changes very little when the question
bodies are included (MRR and P@1 are slightly better but MAP is slightly worse). However, we find
that the inclusion of the question bodies improves
the performance of the RCNN model, achieving a
1% to 3% improvement with both model variations.
The RCNN model’s greater improvement illustrates
the ability of the model to pick out components that
pertain most directly to the question being asked
from the long, descriptive question bodies.
Pre-training Note that, during pre-training, the
last hidden states generated by the neural encoder
are used by the decoder to reproduce the question titles. It would be interesting to see how such states
training. Variances across 5 runs are shown as vertical bars.
capture the meaning of questions. To this end, we
evaluate MRR on the dev set using the last hidden
states of the question titles. We also test how the encoder captures information from the question bodies
to produce the distilled summary, i.e. titles. To do
so, we evaluate the perplexity of the trained encoderdecoder model on a heldout set of the corpus, which
contains about 2000 questions.
As shown in Figure 3, the representations generated by the RCNN encoder perform quite well, resulting in a perplexity of 25 and over 68% MRR
without the subsequent fine-tuning. Interestingly,
the LSTM and GRU networks obtain similar perplexity on the heldout set, but achieve much worse
MRR for similar question retrieval. For instance, the
GRU encoder obtains only 63% MRR, 5% worse
than the RCNN model’s MRR performance. As a
result, the LSTM and GRU encoder do not benefit
clearly from pre-training, as suggested in Table 2.
The inconsistent performance difference may be
explained by two hypotheses. One is that the perplexity is not suitable for measuring the similarity
of the encoded text, thus the power of the encoder
is not illustrated in terms of perplexity. Another hy-
(a) how can i add guake terminal to the start-up applications
i
can
how
l
add guake rmina to
te
the
(b) banshee crashes with `` an unhandled exception was thrown : ''
s
p
rt-u cation
i
l
app
``
s h
e
she rashe wit
c
sta
ban
an
wn
led tion as
and xcep w thro
h
n
e
u
:
''
(c) i get the error message `` requires installation of untrusted packages every time i try to update after entering my password ...
i
get
the
or
err
ge ``
ssa
me
s
s y
n f
d
e
uire atio o truste ckage ever tim
req install
a
n
p
u
i
try
to
ate after tering my sword ...
en
pas
upd
(d) i recently bought samsung laptop and i facing hard time to boot my pen driver so that i can use ubuntu ...
i
g p d
ly ht
ent boug msun lapto an
a
s
rec
i
ing hard time
fac
to
boo
t
my
pen driver
so
i
t
tha
can
use
ubu
ntu ...
Figure 4: Visualizations of 1 − λt of our model on several question pieces from the data set. λt is set to a scalar value (instead of
400-dimension vector) to make the visualization simple. The corresponding model is a simplified variant, which is about 4% worse
than our full model.
pothesis is that the LSTM and GRU encoder may
learn non-linear representations therefore their semantic relatedness can not be directly accessed by
cosine similarity.
Adaptive Decay Finally, we analyze the gated
convolution of our model. Figure 5 demonstrates at
each word position t how much input information is
taken into the model by the adaptive weights 1 − λt .
The average of weights in the vector decreases as t
increments, suggesting that the information encoded
into the state vector saturates when more input are
processed. On the other hand, the largest value in
the weight vector remains high throughout the input,
indicating that at least some information has been
stored in ht and ct .
We also conduct a case study on analyzing the
neural gate. Since directly inspecting the 400dimensional decay vector is difficult, we train a
model that uses a scalar decay instead. As shown in
Figure 4, the model learns to assign higher weights
to application names and quoted error messages,
which intuitively are important pieces of a question
in the AskUbuntu domain.
8
Conclusion
In this paper, we employ gated (non-consecutive)
convolutions to map questions to their semantic
representations, and demonstrate their effectiveness
max
0.95
0.87
0.2
0.1
0
Figure 5:
20
40
t
60
mean
80
99
The maximum and mean value of the 400-
dimentional weight vector 1 − λt at each step (word position) t.
Values are averaged across all questions in the dev and test set.
on the task of question retrieval in community
QA forums. This architecture enables the model
to glean key pieces of information from lengthy,
detail-riddled user questions. Pre-training within an
encoder-decoder framework (from body to title) on
the basis of the entire raw corpus is integral to the
model’s success.
Acknowledgments
We thank Yu Zhang, Yoon Kim, Danqi Chen, the
MIT NLP group and the reviewers for their helpful comments. The work is developed in collaboration with the Arabic Language Technologies
(ALT) group at Qatar Computing Research Institute
(QCRI) within the I YAS project. Any opinions, findings, conclusions, or recommendations expressed in
this paper are those of the authors, and do not necessarily reflect the views of the funding organizations.
References
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly
learning to align and translate. In International Conference on Learning Representations.
Daniel Bär, Torsten Zesch, and Iryna Gurevych. 2013.
Dkpro similarity: An open source framework for text
similarity. In Proceedings of the 51st Annual Meeting
of the Association for Computational Linguistics: System Demonstrations, pages 121–126, Sofia, Bulgaria,
August. Association for Computational Linguistics.
Samuel R. Bowman, Gabor Angeli, Christopher Potts,
and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference.
In Proceedings of the 2015 Conference on Empirical
Methods in Natural Language Processing (EMNLP).
Association for Computational Linguistics.
Kyunghyun Cho, Bart van Merriënboer, Dzmitry Bahdanau, and Yoshua Bengio. 2014a. On the properties of neural machine translation: Encoder-decoder
approaches. arXiv preprint arXiv:1409.1259.
Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger
Schwenk, and Yoshua Bengio. 2014b. Learning
phrase representations using rnn encoder-decoder for
statistical machine translation. In Proceedings of the
2014 Conference on Empirical Methods in Natural
Language Processing (EMNLP).
Junyoung Chung, Caglar Gulcehre, KyungHyun Cho,
and Yoshua Bengio. 2014. Empirical evaluation of
gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555.
Andrew M Dai and Quoc V Le. 2015. Semi-supervised
sequence learning. In Advances in Neural Information
Processing Systems, pages 3061–3069.
Cicero dos Santos, Luciano Barbosa, Dasha Bogdanova,
and Bianca Zadrozny. 2015. Learning hybrid representations to retrieve semantically equivalent questions. In Proceedings of the 53rd Annual Meeting of
the Association for Computational Linguistics and the
7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages
694–699, Beijing, China, July. Association for Computational Linguistics.
Huizhong Duan, Yunbo Cao, Chin-Yew Lin, and Yong
Yu. 2008. Searching questions by identifying question
topic and question focus. In ACL, pages 156–164.
Minwei Feng, Bing Xiang, Michael R Glass, Lidan
Wang, and Bowen Zhou. 2015. Applying deep learning to answer selection: A study and an open task.
arXiv preprint arXiv:1508.01585.
Jianfeng Gao, Patrick Pantel, Michael Gamon, Xiaodong
He, Li Deng, and Yelong Shen. 2014. Modeling interestingness with deep neural networks. In Proceedings
of the 2013 Conference on Empirical Methods in Natural Language Processing.
Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky,
Ilya Sutskever, and Ruslan R Salakhutdinov. 2012.
Improving neural networks by preventing coadaptation of feature detectors.
arXiv preprint
arXiv:1207.0580.
Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long
short-term memory. Neural computation, 9(8):1735–
1780.
Jiwoon Jeon, W Bruce Croft, and Joon Ho Lee. 2005.
Finding similar questions in large question and answer
archives. In Proceedings of the 14th ACM international conference on Information and knowledge management, pages 84–90. ACM.
T. Joachims. 2002. Optimizing search engines using
clickthrough data. In ACM SIGKDD KDD.
Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent continuous translation models. In Proceedings of
the 2013 Conference on Empirical Methods in Natural Language Processing (EMNLP 2013), pages 1700–
1709.
Nal Kalchbrenner, Edward Grefenstette, and Phil Blunsom. 2014. A convolutional neural network for modelling sentences. In Proceedings of the 52th Annual
Meeting of the Association for Computational Linguistics.
Yoon Kim, Yacine Jernite, David Sontag, and Alexander M Rush. 2015. Character-aware neural language
models. Twenty-Ninth AAAI Conference on Artificial
Intelligence.
Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the Empiricial
Methods in Natural Language Processing (EMNLP
2014).
Diederik P Kingma and Jimmy Lei Ba. 2015. Adam: A
method for stochastic optimization. In International
Conference on Learning Representation.
Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. 1998.
Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324,
November.
Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2015.
Molding cnns for text: non-linear, non-consecutive
convolutions. In Proceedings of the 2015 Conference
on Empirical Methods in Natural Language Processing, pages 1565–1575, Lisbon, Portugal, September.
Association for Computational Linguistics.
Shuguang Li and Suresh Manandhar. 2011. Improving question recommendation by exploiting information need. In Proceedings of the 49th Annual Meeting
of the Association for Computational Linguistics: Human Language Technologies-Volume 1, pages 1425–
1434. Association for Computational Linguistics.
Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey
Dean. 2013. Efficient estimation of word representations in vector space. CoRR.
Preslav Nakov, Lluı́s Màrquez, Walid Magdy, Alessandro Moschitti, Jim Glass, and Bilal Randeree. 2015.
SemEval-2015 task 3: Answer selection in community question answering. In Proceedings of the 9th
International Workshop on Semantic Evaluation, SemEval ’15.
Preslav Nakov, Lluı́s Màrquez, Walid Magdy, Alessandro Moschitti, Jim Glass, and Bilal Randeree. 2016.
SemEval-2016 task 3: Community question answering. In Proceedings of the 10th International Workshop on Semantic Evaluation, SemEval ’16.
Stephen Robertson and Hugo Zaragoza. 2009. The
probabilistic relevance framework: BM25 and beyond.
Now Publishers Inc.
Tim Rocktäschel, Edward Grefenstette, Karl Moritz Hermann, Tomáš Kočiskỳ, and Phil Blunsom. 2016. Reasoning about entailment with neural attention. In International Conference on Learning Representations.
Alexander M Rush, Sumit Chopra, and Jason Weston.
2015. A neural attention model for abstractive sentence summarization. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language
Processing.
Aliaksei Severyn and Alessandro Moschitti. 2015.
Learning to rank short text pairs with convolutional
deep neural networks. In SIGIR.
Yikang Shen, Wenge Rong, Nan Jiang, Baolin Peng, Jie
Tang, and Zhang Xiong. 2015. Word embedding
based correlation model for question/answer matching. arXiv preprint arXiv:1511.04646.
Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. 2014.
Sequence to sequence learning with neural networks.
In Advances in neural information processing systems,
pages 3104–3112.
Ming Tan, Bing Xiang, and Bowen Zhou. 2015. Lstmbased deep learning models for non-factoid answer selection. arXiv preprint arXiv:1511.04108.
Di Wang and Eric Nyberg. 2015. A long short-term
memory model for answer sentence selection in question answering. In ACL, July.
Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015.
Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems, pages 649–657.
Guangyou Zhou, Yang Liu, Fang Liu, Daojian Zeng,
and Jun Zhao. 2013. Improving question retrieval
in community question answering using world knowledge. In Proceedings of the Twenty-Third international joint conference on Artificial Intelligence, pages
2239–2245. AAAI Press.
Guangyou Zhou, Tingting He, Jun Zhao, and Po Hu.
2015. Learning continuous word embedding with
metadata for question retrieval in community question
answering. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics
and the 7th International Joint Conference on Natural
Language Processing (Volume 1: Long Papers), pages
250–259, Beijing, China, July. Association for Computational Linguistics.
Note on Data Pre-processing Issue:
This paper is based on our early preprint version ”Denoising Bodies to Titles: Retrieving Similar
Questions with Recurrent Convolutional Models”.
We noticed, during the analysis of the models, that some question text in the AskUbuntu dataset
contain user-annotated information which should be removed for a more realistic evaluation (see
the figure below and the explanation).
We re-processed the dataset and updated the results accordingly. The performance of various
neural network models reported in this version may not match the numbers reported in our early
version exactly, but the relative observation (i.e. relative difference between models) does not
change.
AskUbuntu Dataset
user-marked duplicate question
A snapshot of one AskUbuntu question. The ”possible duplicate” section shows the title of a similar
question marked by the forum user. This section is included in the original XML dump of AskUbuntu, and
was not removed in our early experiments.
:
“Possible duplicate” was included in the original dump, and wasn’t
properly removed in our first experiments; same in (dos Santos et. al. 2015)
Annual
Meeting
– March
38 QCRI/MIT-CSAIL
QCRI/MIT-CSAIL
Annual
Meeting
– March
2014 2015
‹#› ‹#›
| 9cs.NE
|
Conic-sector-based analysis and control synthesis
for linear parameter varying systems
arXiv:1712.00722v2 [cs.SY] 30 Jan 2018
Sivaranjani S, James Richard Forbes, Peter Seiler and Vijay Gupta
Abstract— We present a conic sector theorem for linear
parameter varying (LPV) systems in which the traditional
definition of conicity is violated for certain values of the
parameter. We show that such LPV systems can be defined
to be conic in an average sense if the parameter trajectories
are restricted so that the system operates with such values
of the parameter sufficiently rarely. We then show that such
an average definition of conicity is useful in analyzing the
stability of the system when it is connected in feedback with
a conic system with appropriate conic properties. This can be
regarded as an extension of the classical conic sector theorem.
Based on this modified conic sector theorem, we design conic
controllers that allow the closed-loop system to operate in
nonconic parameter regions for brief periods of time. Due
to this extra degree of freedom, these controllers lead to less
conservative performance than traditional designs, in which the
controller parameters are chosen based on the largest cone
that the plant dynamics are contained in. We demonstrate
the effectiveness of the proposed design in stabilizing a power
grid with very high penetration of renewable energy while
minimizing power transmission losses.
I. I NTRODUCTION
The conic sector theorem, introduced by Zames [1], is
a powerful input-output stability result that is applicable to
both linear and nonlinear systems that are sector bounded.
Traditionally, conic-sector-based analysis has received relatively limited attention in comparison to small gain and
passivity-based analysis due to the difficulty involved in accurately characterizing conic bounds. However, recent results
in the efficient computation of conic bounds have made it
possible to consider the conic sector theorem as an attractive
tool for analysis and control design [2].
In this paper, we present a conic-sector-based stability
analysis and control synthesis for linear parameter varying
(LPV) systems. Typical control design approaches for LPV
systems involve designing and scheduling a bank of controllers to stabilize the system for all parameter values in its
operating trajectory [3]–[5]. Analysis and synthesis results
for LPV systems based on small gain theorems [6]–[8] and
dissipativity theory [9]–[11] are also available. Of particular
interest to this paper, controllers have been designed for
LPV systems [12] and specific subclasses like polytopic
systems [13] using the conic sector theorem. However, these
controllers are usually conservative since they are designed
Sivaranjani S and Vijay Gupta are with the Department of Electrical Engineering, University of Notre Dame, South Bend, IN.
{sseethar,vgupta2}@nd.edu. James Richard Forbes is with the Department
of Mechanical Engineering, McGill University, Montreal, QC, Canada.
[email protected]. Peter Seiler is with the Department of
Aerospace Engineering and Mechanics, University of Minnesota, Minneapolis, MN. [email protected].
for the worst-case sector bounds on the plant as a function
of the parameter. Furthermore, these results cannot be used
for LPV systems in which the system is non-conic for even
one value of the parameter. In this paper, we propose a less
conservative design approach based on a modification of the
conic sector theorem that does not constrain the designer to
consider the worst-case sector bound.
Recent results indicate that passivity in LPV systems can
be preserved on the average even when certain parameter
values lead to nonpassive operation, provided that the system
spends a sufficiently small amount of time operating with
these parameter values [14]. Here, we generalize these results
to obtain a modified conic sector theorem for intermittently
conic LPV systems, that is, LPV systems that may be
nonconic for certain parameter values. We then use this
theorem to design a conic controller that allows the closedloop LPV system to be intermittently conic. This approach
allows for less conservative control designs as the controller
is not designed for the worst case sector bound of the plant,
which, in fact, does not exist for intermittently conic plants.
The contributions of this paper are threefold. Firstly,
we derive conditions under which an LPV system with
intermittent conic behavior is conic in the average sense.
Secondly, we derive a modified conic sector theorem for
the feedback interconnection of an intermittently conic LPV
system and a conic system. Finally, we design a conic
controller based on this conic sector theorem that allows the
closed-loop system to be intermittently conic, leading to less
conservative designs than methods available in literature. We
also demonstrate an application of the proposed design in
stabilizing a power system with high penetration levels of
renewable energy while minimizing power losses.
Notation: R denotes the set of real numbers, Rp the set of
p-dimensional real vectors, R+ the set of non-negative real
numbers, N the set of natural numbers including zero and
Nn the set of natural numbers {1, · · · , n}, n ∈ N. Given two
sets A and B, A \ B denotes the set of all elements of A that
are not in B. For a vector v, [v]i denotes the i-th component.
I stands for the identity matrix with dimensions clear from
the context. A0 denotes the transpose of a matrix A.
II. P ROBLEM F ORMULATION
Linear Parameter Varying (LPV) system: Consider the
system described by
ẋ(t) = A(ρ(t))x(t) + B(ρ(t))u(t)
(1)
y(t) = C(ρ(t))x(t) + D(ρ(t))u(t)
where x(t) ∈ Rn is the process state, u(t) ∈ Rm is the
control input, y(t) ∈ Rm is the process output, and ρ ∈ Rp
is the parameter vector which is assumed to be a piecewise
continuously differentiable function of time. We will limit
consideration to the time interval [t0 , tn ] and consider t0 =
0 and tn → ∞ to define stability for this system. The
parameter trajectory from times t0 to tn (denoted by ρ(t)|ttn0
or simply ρ(t) when the interval [t0 , tn ] is clear from context)
refers to the set of parameter values that the system assumes
from time t0 to tn . The following assumptions will hold
throughout the paper.
Assumption 1: We assume that the input u(t) is norm
bounded so that ||u(t)||22 ∈ [ū21 , ū22 ] where ū1 > 0 and
ū2 < ∞ are positive constants.
LPV models are often constructed as linearizations of
nonlinear dynamics about a collection of different operating
points. Assumption 1 implies that u(t) = 0 for all t is not
allowed, that is, we disallow operation at an equilibrium for
such LPV models.
Assumption 2: The parameter trajectories satisfy the following range and rate bounds
dρ(t)
≤ ν̄, ∀1 ≤ i ≤ p. (2)
ρ ≤ [ρ(t)]i ≤ ρ̄, ν ≤
dt i
The special case when ν → −∞ and ν̄ → ∞ is called the
rate unbounded case.
Let all parameter trajectories that satisfy these constraints
be collected in the set A, which is the set of admissible
parameter trajectories. An admissible parameter trajectory
ρ(t)|ttn0 ∈ A where the parameter assumes constant values
ρ(i) indexed by i ∈ {1, 2, . . . , n}, n ∈ N, for time ti−1 <
t ≤ ti is defined as a discrete parameter trajectory over
the time interval [t0 , tn ]. We denote the set of all discrete
parameter trajectories by Ad , where Ad ⊂ A.
Conicity: The system (1) with a fixed value of the parameter is a linear time-invariant system and can be analyzed
for conicity [2]. As the parameter value varies, we use
the following definition of conicity for particular parameter
trajectories. The LPV system (1) with a given parameter
trajectory ρ(t)|ttn0 is said to be conic for that parameter
trajectory if there exist constants a and b such that 0 < a < b
and the inequalityZ t2
w(u(t), y(t))dt ≥ 0,
(3)
t1
holds for all times t1 and t2 such that t0 ≤ t1 < t2 ≤ tn
and all u(t) satisfying
1, where
Assumption
0
1
y(t) − 1b I
1 + ab I y(t)
2
w(u(t), y(t)) =
.
u(t) 12 1 + ab I
−aI u(t)
Further, if the system is conic for every parameter trajectory
in a set S ⊆ A, then it is said to be conic for that set
S. Note that the constants a and b will, in general, depend
on the parameter trajectory and are not unique. We assume
that the conic bounds represent the tightest bounds for the
set of parameter trajectories S ⊆ A and time interval being
considered, and can be characterized by solving the matrix
inequality (23) as described in Appendix I. We define the
interval [a, b] as the conic sector for the system (1) for the
set S, the value (b−a)/2 as the corresponding radius r of the
conic sector, and the quantity (b + a)/2 as the center c of the
conic sector. We also drop the dependence of w(u(t), y(t))
on time and use w(u, y) for brevity. Note that (3) is a finite
horizon integral quadratic constraint (IQC) on (u, y). The
most general IQC theory also allows dynamic mulitpliers,
with (3) being the special case where the multiplier is simply
the (static) identity matrix [15].
Intermittent conicity: Equation (3) constrains the parameter trajectory such that (3) holds for all choices of times
t1 and t2 where t0 ≤ t1 < t2 ≤ tn . We are interested
in trajectories that are only ‘intermittently conic’ in the
following sense. For a given parameter trajectory ρ(t)|ttn0 ,
identify times t0 ≤ t1 < t2 < · · · ≤ tn such that for every
i ∈ ic ⊂ Nn , the system is conic for the parameter trajectory
ρ(t)|ttii−1 , and for every i ∈ inc , where inc = Nn \ ic , the
system is not conic for the parameter trajectory ρ(t)|ttii−1 .
Note that the choice of these times is not unique. The conic
tn
parameter region of the trajectory
S ρ(t)|t0 is then defined
as the set Rc = {ρ(t) : t ∈
[ti−1 , ti ]} and the time
i∈ic
spent in this region is denoted by tc . The nonconic parameter
region of the trajectory ρ(t)|ttn0 is the set Rnc = {ρ(t) : t ∈
[t0 , tn ], t ∈
/ Rc } and the time spent in this region is denoted
by tnc . If the set Rc is not empty, the system is intermittently
conic for the parameter trajectory ρ(t)|ttn0 .
For intermittently conic systems, we extend the definition
of conicity proposed in (3) as follows. The LPV system (1)
with a given parameter trajectory ρ(t)|ttn0 is said to be conic
in the average sense for that parameter trajectory if there
exist constants a andZb with 0 < a < b such that
tn
w(u, y)dt ≥ 0
(4)
t0
holds for all u(t) satisfying Assumption 1. If the system is
conic in the average sense for every parameter trajectory in
a set S ⊆ A, then it is said to be conic in the average sense
for that set S. We will make the following assumptions on
the conic and nonconic parameter regions.
Assumption 3: For every i ∈ ic , there exist constants a
and b and a constant
i such that 0 < a < b and
Z
ti
w(u, y)dt ≥ i (ti − ti−1 ).
(5)
ti−1
Remark 1: Assumption 3 is similar to the notion of quasiQSR dissipativity with the supply function given by the left
hand side of (5) (see, e.g., [16, Chapter 4]). However, by
Assumption 1, we do not allow uniformly zero inputs over
any finite time horizon. Therefore, classical constraints on
the sign of i ([16, Chapter 4]) are not applicable in our
case. We note that Assumption 3 may be relaxed to allow
zero inputs over specific time intervals by imposing the sign
constraint i < 0 in such intervals. In this paper, we avoid
such a formulation due to increased conservatism.
Assumption 4: For every i ∈ inc , there exists a constant
αi > 0Z such that
Z
ti
ti
w(u, y)dt ≥ −αi (ti − ti−1 )
u2 (t)dt. (6)
ti−1
ti−1
R ti
Note that the ti−1
u2 (t)dt term in (6) ensures that no conic
parameter region satisfies (6). In the context of Assumption
4, a and b represent the smallest constants such that (6) holds
with 0 < a < b, rather than conventional sector bounds.
The constants i and αi depend on the parameter trajectory
ρ(t)|ttii−1 ; we drop this dependence for simplicity of notation.
In general, we jointly compute the tightest bounds for i and
αi with the sector bounds [a, b] (see Appendix I).
Problem statement: With these assumptions, the aim of
this paper is to
(i) derive conditions under which an LPV system (1) with
intermittently conic behavior for a parameter trajectory
ρ(t)|ttn0 is conic in the average sense for that trajectory,
(ii) derive a conic-sector theorem that ensures L2 stability
when system (1) with intermittent conic behavior for a
parameter trajectory is connected in negative feedback
with another conic system, and
(iii) design a conic controller using this conic sector theorem
to stabilize an intermittently conic LPV system while
guaranteeing some gain performance.
III. C ONICITY OF LPV S YSTEMS WITH I NTERMITTENT
C ONIC B EHAVIOR
We begin by deriving conditions for average system conicity in terms of the fractions of time spent in conic and
nonconic parameter regions.
Lemma 1 (Discrete Case): Let the parameter trajectories
of the intermittently conic system (1) be restricted to the set
of discrete parameter trajectories Ad ⊂ A. Then, (1) is conic
in the average sense if, for every ρ(i)|ttn0 ∈ Ad , we have
n
X
ti − ti−1
i , i ∈ ic
, ci =
. (7)
ci µi ≥ 0, µi =
2
−α
tn − t0
i ū1 , i ∈ inc
i=1
We can readily generalize Lemma 1 to the case where the
LPV system (1) can operate along any admissible parameter
trajectory as follows.
Theorem 1: The system (1) is conic in the average sense
if, for every ρ(t)|ttn0 ∈ A, we have
Z
Z
2
(ρ(t))dt ≥ ū1
α(ρ(t))dt,
(8)
t∈tc
t∈tnc
where (ρ(t)) and α(ρ(t)) are continuous approximations of
i and αi as defined in Appendix II.
The proofs of Lemma 1 and Theorem 1, along with the
formal definitions of (ρ(t)) and α(ρ(t)), can be found in
Appendix II. Lemma 1 and Theorem 1 formalize the idea
that the LPV system (1) is conic in the average sense if it
spends a sufficiently large fraction of time operating in conic
parameter regions.
IV. C ONIC S ECTOR T HEOREM FOR I NTERMITTENTLY
C ONIC S YSTEMS
Consider a negative feedback interconnection of LPV
systems G1 and G2 as shown in Fig. 1, where the dynamics
of G1 and G2 are described as follows.
G1 :
ẋp (t) = Ap (ρp (t))xp (t) + Bp (ρp (t))up (t)
yp (t) = Cp (ρp (t))xp (t) + Dp (ρp (t))up (t)
(9)
G2 : ẋc (t) = Ac (ρc (t))xc (t) + Bc (ρc (t))uc (t)
yc (t) = Cc (ρc (t))xc (t) + Dc (ρc (t))uc (t)
(10)
Fig. 1: Negative feedback interconnection of G1 and G2
Let Ap and Ac denote the admissible parameter trajectories of G1 and G2 respectively. Let G2 be conic with sector
bounds [ac , bc ] and G1 be intermittently conic with sector
bounds [ap , bp ]. We also assume ||up (t)||22 ∈ [ū2p,1 , ū2p,2 ] with
constants ūp,1 > 0 and ūp,2 < ∞. We now examine the
stability of the negative feedback interconnection of G1 and
G2 when G1 is intermittently conic as defined in Section II.
Theorem 2 (Modified Conic Sector Theorem): The negative feedback interconnection of G1 and G2 as shown in
Fig. 1 is L2 stable if, for every admissible ρp (t) ∈ Ap and
ρc (t)
Z ∈ Ac , tn → ∞,
Z
p (ρp (t)) dt ≥ ū2p,1
t∈tc
αp (ρp (t)) dt,
and,
(11)
t∈tnc
2
1
1 ac ap
+ac −
−
> 0.
bp
4 bc bp
(12)
Proof: Since G2 is conic, for all t1 and t2 such that t0 ≤
t1 < t2 ≤ tn , it satisfies
1
+ap > 0,
bc
Z
t2
t1
1
+ap
bc
0
1
I
yc (t) −
bc
1
uc (t)
1+
2
1
2
ac
bc
1+
I
I yc (t)
dt ≥ 0.
−ac I uc (t)
ac
bc
(13)
If (11) holds, then G1 is conic in the average sense, with
Z
t2
t1
0
1
I
yp (t) −
bp
1
up (t)
1+
2
1
2
ap
bp
1+
I
I yp (t)
dt ≥ 0.
−ap I up (t)
ap
bp
(14)
From Fig. 1, we have ec = uc + yp and ep = up − yc , which
when used in conjunction with (13) and (14) gives
0
Z tn
Y (t)
Q S Y (t)
dt ≥ 0,
(15)
E(t)
S 0 R E(t)
t0
0
0
0
0
where Y =[yc , yp ], E=[ec , ep ], and
" 1
#
a
− bc I−ap I bpp − abcc
−ac I 0
, R=
, and
Q=
1
0 −ap I
∗
− I−ac I
bp
1
1+ abcc I −ap I
2
.
S=
a
1
−ac I
1+ bpp I
2
(16)
Setting t0 = 0 and tn → ∞ in (15), the feedback interconnection is L2 stable if Q < 0, giving condition (12).
Remark 2: The modified conic sector theorem (Theorem
2) provides sufficient conditions for stability of the closedloop system based on the plant and feedback controller
parameters. This conic sector theorem can be graphically
interpreted as follows. The feedback interconnection of the
plant and the controller is stable if the two systems are conic
in the average sense (Equation 4) and the conic sectors they
lie in do not intersect. Note that Theorem 2 is an input-output
stability result; however, with additional assumptions like
zero state detectability and reachability, it can be extended
to prove Lyapunov stability along the lines of [17].
V. C ONIC S ECTOR T HEOREM BASED C ONTROL D ESIGN
In this section, we present a control design approach based
on the modified conic sector theorem derived in Section IV.
Key design idea: Consider an LPV plant that lies in
a range of conic sectors as depicted in Fig. 2. A conservative control design approach involves designing the
controller such that it lies in a cone that is complementary
to the largest conic sector that contains the plant dynamics
[12]. However, this strategy may result in very low gain
controllers. Therefore, we propose the following alternative
design. We choose a nominal plant cone that is smaller than
the largest plant cone, and design a controller that lies in
the conic sector complementary to this nominal cone, based
on Theorem 2. Now, the controller cone will intersect the
plant cone in certain parameter regions, making the closedloop system nonconic in these regions. Based on Theorem
1, we then bound the fraction of time that the closed-loop
system operates in these nonconic regions. Note that in
contrast to standard LPV design methods, we do not use
the parameter trajectory for gain scheduling. Rather, we use
the knowledge of the parameter trajectory to characterize
conic and nonconic parameter regions, which are then used
to design a single conic controller for a nominal operating
point.
Design Procedure: Let γ be the induced L2 gain of
the feedback interconnection of plant G1 and controller G2
shown in Fig. 1. We would like to choose a nominal plant
cone [a∗p , b∗p ] such that a controller G2 lying in the conic
sector [a∗c = − b1∗ + δ, b∗c = − a1∗ − δ], where δ > 0 is
p
p
a trivially small number, minimizes γ while keeping the
closed-loop system L2 stable. This design problem can be
Z
Z
formulated as
cl dt ≥
αcl dt,
(17)
P1 : min γ s.t.
ap ,bp
tc
tnc
where cl and αcl are the conicity indices as defined in (5)
and (6) respectively for the closed-loop system. We now have
the following result.
Proposition 1: If there exist ap = a∗p and bp = b∗p
satisfying matrix inequalities (18)-(20), where Acl , Bcl , Ccl
and Dcl are the closed loop system matrices, and Q, S and
R are defined in (16), then the conic controller G2 with
[ac , bc ] = [− b1∗ , − a1∗ ] renders the closed-loop system in
p
p
Fig. 1 L2 stable. Further, a∗p and b∗p solve problem P1 .
Proof: Let β > 0 be the smallest scalar such that
βI + R > 0 and (1/β + Q) < ζI, for some ζ >
0.
(15), with t0 = 0R and tn → ∞, we have
R ∞From
∞ 0
U (t)U (t)dt, where Y =
Y 0 (t)Y (t)dt ≥ − (β+λ(R))
ζ
0
0
0
0
0
0
[yc , yp ], U = [uc , up ] and λ(R) is the maximum eigenvalue
of R. Substituting ac = − b1p and bc = − a1p , the closedloop L2 gain can be written in terms of the plant parameters
as γ = (β + λ(R))/ζ. The numerator and denominator
of γ can be expressed as k1 (bp /ap − 1) and k2 (bp /ap )
respectively, where k1 and k2 are constants independent
of the optimization variables ap and bp . Therefore, the
optimization problem P1 is equivalent to
Z
Z
(bp − ap )
P2 : min rp =
, s.t.
cl dt ≥
αcl dt, (21)
ap ,bp
2
tc
tnc
that is, minimizing the conic radius of the nominal plant
cone. As rp is increased, the intersection of the plant and
controller cones increases. RTherefore, Rthe solution to problem
P2 is [a∗p , b∗p ] = {[ap , bp ] : tc cl dt − tnc αcl dt = 0}, which
is the smallest plant cone such that the closed-loop system
remains conic in the average sense defined in Theorem
1, implying (20). Equations (18) and (19) correspond to
computation of cl and αcl (see Appendix I).
VI. A PPLICATION E XAMPLE : P OWER S YSTEM WITH
V ERY H IGH P ENETRATION OF R ENEWABLE E NERGY
A class of systems where the proposed control design
finds application is power grids with a large penetration of
renewable energy resources. The dynamics of power grids
are described by nonlinear differential algebraic equations,
which are extremely difficult to analyze. Therefore, LPV
models of power grids have been proposed for the design of
wide-area controllers [18]. LPV models and switched system
approximations have also become relevant in the context of
time-varying generation from renewable energy sources [19].
Typical approaches to stabilize power grids with LPV models
involve the design of parameter varying controllers [18] that
can be difficult to design and implement. In this section,
we demonstrate the design of a conic controller to stabilize
power grids with very high penetration of renewable energy,
enhance their peak loading capacity and minimize losses.
Let the generation sources G1 , G2 , ..., GNg in the grid
be represented by LPV systems of the form (9) with the
parameters of the i-th generator being ρp,i = (Pi , Qi ),
representing the scheduled real and reactive power, and
the output being yi =[vi ], representing the voltage at the
i-th generator bus. The detailed nonlinear models of the
renewable generators and the electrical network can be found
in [20]. The system power flow constraint requires
Ng
Ng
X
X
PL =
Pi − Ploss , QL =
Qi − Qloss ,
(22)
i=1
Fig. 2: Conic-sector-based control design - key idea
i=1
where the total system load is PL + jQL
and the power
PN PN
2
losses are Ploss = 0.5
|i
|
r
and Qloss =
ij
ij
j=1
i=1
"
0
0
∂P (ρp )
+ Ccl QCcl
∂ρp
0
Ccl + Dcl QCcl
Acl P (ρp ) + P (ρp )Acl + ρ̇p
0
Bcl P (ρp )
−S
0
0
T
P (ρp )Bcl − Ccl
S + Ccl QDcl
0
Dcl QDcl
0
− S Dcl −
0
Dcl Scl
#
+ cl I ≤ 0
(18)
+R
0
0
0
0
∂ P̃ (ρp )
P̃ (ρp )B − Ccl S + Ccl QDcl
≤0
Acl P̃ (ρp ) + P̃ (ρp )Acl + ρ̇p ∂ρp + Ccl QCcl
0
0
0
0
0
0
Dcl QDcl − S Dcl − Dcl S + R − αcl I
Bcl P̃ (ρp ) − S Ccl + Dcl QCcl
Z
Z
cl =
αcl , P (ρp ) ≥ 0, P̃ (ρp ) ≥ 0
tc
(19)
(20)
tnc
TABLE I: Operating points, conic sector bounds and average
power losses.
PDF IG
(MW)
420
400
350
(Nominal)
300
Fig. 3: Modified IEEE 14-bus test system.
Fig. 4: Voltage profiles with conic and standard controllers.
P
N PN
2
0.5
|i
|
x
, with iij , rij and xij being the
ij
ij
j=1
i=1
line currents, resistances and reactances respectively. Our
objective is to minimize the total power loss in the system,
which is equivalent to minimizing the induced L2 norm from
the output vector yi to the input vector ui , subject to (22).
We consider a modified version of the IEEE 14-bus
standard test system as shown in Fig. 3 to illustrate the
proposed design. In this modified system, the generator at
bus 1 is replaced with an equivalent 600 MW doubly-fed induction generator (DFIG) wind farm (78% renewable energy
penetration). The system is then linearized around the power
flow solution at 4 operating points of (PDF IG , QDF IG )
(Table I) and a state space model with 48 states, 5 inputs
and 5 outputs is obtained. The five control inputs correspond
to the reference signals of the Automatic Voltage Regulator
(AVR) excitation controllers on the generators at buses 2,
3, 6 and 8, and the speed reference signal of the pitch
angle controller on the DFIG at bus 1. The conic sector
bounds for each operating point are listed in Table I.
We solve (18)-(20) over a grid of parameter values using
YALMIP/SeDuMi [21], [22] to obtain the nominal operating
point of (PDF IG , QDF IG ) = (350 MW,-28.2 MVAr) and the
QDF IG
MVAr
-34.00
-31.70
-28.20
ap
bp
7.501
7.201
7.575
Ploss
(MW)
46.68
37.34
29.43
Qloss
(MVAr)
157.90
121.32
90.27
Sloss
(MVA)
164.65
126.94
94.95
-3.499
-3.199
-2.425
-23.05
-2.899
6.901
21.93
60.72
64.55
corresponding nominal plant and controller conic bounds.
We then synthesize a conic controller with system matrices
(Ac , Bc , Cc ) of dimensions (48×48), (48×5) and (5×48)
respectively, satisfying these nominal conic bounds using
the procedure in [2, Section 5]. We also implement three
other controllers for comparison, namely, the classical AVR
and pitch angle controllers, referred to as the ‘standard
controller’, as described in [23], a conic controller designed
to lie in a sector complementary to the largest plant cone
(corresponding to the operating point PDF IG = 420 MW)
[12], and a dynamic gain scheduled controller designed to
stabilize the system for every operating point [18].
We test the performance of the closed-loop system when
all real and reactive loads are increased by 15% between t =
5s to t = 10s (Fig. 4). This test condition is chosen to compare the responses of the controllers to two major changes in
the system parameter (operating point) corresponding to the
increase in the total load at t = 5s and the decrease in the
system load at t = 10s. With the standard controller for the
14-bus system [23], the system loses stability for the 15%
load change, implying that the maximum loading capability
of the system is exceeded. With the conic controller designed
for the largest plant cone [12], the system remains stable but
exhibits poor performance, with the voltage at the DFIG bus
not settling at the nominal value. With the classical gain
scheduled controller, the system displays adequate transient
response to the first operating point change at t = 5s, but
responds poorly to the second change at t = 10s. On the
other hand, the system with the conic controller designed
for the nominal cone responds well to this load change and
maintains a voltage level close to the nominal value, which
is an advantage of the increased controller gain obtained
by allowing the intermittently conic operation. This result
indicates that the conic controller enhances the maximum
loading capability of the grid. We also observe from Table I
that the nominal conic controller has a significantly lower net
power loss (94.95 MVA) as compared to the conic controller
designed for the largest plant cone (164.65 MVA) and the
classical gain scheduled controller (116.36 MVA).
VII. C ONCLUSION
We derived a modified conic sector theorem for intermittently conic LPV systems, and employed this theorem to
synthesize conic controllers that are less conservative than
traditional designs. We also demonstrated the performance
and advantages of this design in stabilizing a power grid
with very high penetration of renewable energy.
A PPENDIX I
D ETERMINING C ONIC B OUNDS AND C ONICITY I NDICES
Theorem 3: The LPV system (1) is conic with sector
bounds [a, b] if and only if there exists a symmetric P̂ (ρ) > 0
such that for all ρ ∈ A, the following holds:
#
"
∂ P̂ (ρ)
C0 + C0D
A0 P̂ (ρ) + P̂ (ρ)A + ρ̇ ∂t + C 0 C P̂ (ρ)B − a+b
2
≤0.
∗
D0 D − a+b
(D + D0 ) + abI
2
(23)
The proof is along the lines of [12, Theorem 2.1] and is
omitted here. In conic parameter regions, ∃P (ρ)>0 such that
A0 P (ρ) + P (ρ)A < 0.
(24)
From (5), following standard literature [24], and defining
Q = −1/b, S = 12 (1 + ab ), and R = −aI, we can derive
0
∂P (ρ)
A P (ρ) + P (ρ)A + ρ̇ ∂t + C 0 QC P (ρ)B − C 0 S + C 0 QD + I≤0.
0
0
0
0
0
0
B P (ρ) − S C + D QC
D QD − S D − D S + R
(25)
This matrix inequality can then be solved with constraints
(2) and (24) to determine (ρ(t)). Similarly, from (6), we
can obtain the matrix inequality
#
"
∂ P̃ (ρ)
A0 P̃ (ρ) + P̃ (ρ)A + ρ̇ ∂t + C 0 QC P̃ (ρ)B − C 0 S + C 0 QD
≤0,
0
0
0
0
0
0
B P̃ (ρ) − S C + D QC
D QD − S D − D S + R − αI
(26)
where P̃ (ρ) > 0, which can be solved with bounds (2)
to determine α(ρ(t)). The matrix inequalities to compute
the conic bounds and conicity indices are infinite dimensional due to their dependence on ρ and can be solved by
discretization over a mesh on ρ and ρ̇. Further, the conic
bounds obtained are not unique; the computation of tight
conic bounds for each parameter is discussed in [2].
A PPENDIX II
Proof of Lemma1: We have
Ztn
ti
Zti
X Z
X
w(u, y)dt =
w(u, y)dt+
w(u, y)dt .
i∈ic
t0
i∈inc
ti−1
ti−1
(27)
From (27),Z(5), (6) and Assumption 1, we have
tn
w(u, y)dt ≥
t0
X
(i (ti − ti−1 ))
i∈ic
−
X
αi ū21 (ti − ti−1 ) .
(28)
i∈inc
Comparing (28) with (4), system (1) is conic in the average
sense if, for every ρ(t) ∈ Ad , (7) is satisfied.
Proof of Theorem 1: Consider the discrete parameter
trajectory
ρ(i) in Lemma
1. Then,the
I = [t0 ,tn ] =
Sn
S
S interval
S
i=1 [ti−1 , ti ] =
i∈ic [ti−1 , ti ]
i∈inc [ti−1 , ti ] . Define a partition of I to be P = {t0 , t1 , . . . , tn }
with norm |P | = max (ti − ti−1 ), i ∈ Nn . Let
R = {ρ(1), ρ(2), . . . , ρ(n)} such that for every i ∈ Nn ,
ρ(i) ∈ [ti−1 , ti ]. Define the tagged partition F(t, ρ) =
(P, R) to be the collection of the partition P and the
associated parameter set R. Define the Riemann sum Sr with
respect to tagged partition F(t, ρ) to be the RHS of (28).
In the limit |P | → 0, since ρ(t) is piecewise continuously
differentiable, Sr can be approximated by the Riemann
integral
Z
Z
(ρ(t))dt − ū21
lim Sr =
|P |→0
t∈tc
α(ρ(t))dt.
(29)
t∈tnc
From (28) and (29), (1) is conic in the average sense if (8)
is satisfied for every ρ(t) ∈ A, completing the proof.
R EFERENCES
[1] G. Zames, “On the input-output stability of time-varying nonlinear
feedback systems–parts i and ii,” IEEE Transactions on Automatic
Control, vol. 11, no. 2/3, 1966.
[2] L. J. Bridgeman and J. R. Forbes, “Conic-sector-based control to
circumvent passivity violations,” International Journal of Control,
vol. 87, no. 8, pp. 1467–1477, 2014.
[3] D. Leith and W. Leithead, “Gain-scheduled and nonlinear systems: dynamic analysis by velocity-based linearization families,” International
Journal of Control, vol. 70, no. 2, pp. 289–317, 1998.
[4] D. Lawrence and W. Rugh, “Gain scheduling dynamic linear controllers for a nonlinear plant,” Automatica, vol. 31, no. 3, pp. 381–390,
1995.
[5] J. Shamma and M. Athans, “Analysis of gain scheduled control for
nonlinear plants,” IEEE Transactions on Automatic Control, vol. 35,
no. 8, pp. 898–907, 1990.
[6] P. Apkarian and P. Gahinet, “A convex characterization of gainscheduled H∞ controllers,” IEEE Transactions on Automatic Control,
vol. 40, pp. 853–864, 1995.
[7] A. Packard, “Gain scheduling via linear fractional transformations,”
Systems and Control Letters, vol. 22, pp. 79–92, 1994.
[8] C. Scherer, “Gain-scheduled synthesis with dynamic positive real
multipliers,” in IEEE Conference on Decision and Control, 2012.
[9] P. Apkarian and R. Adams, “Advanced gain-scheduling techniques for
uncertain systems,” IEEE Transactions on Control Systems Technology, vol. 6, no. 1, pp. 21–32, 1998.
[10] A. Packard and G. Becker, “Quadratic stabilization of parametricallydependent linear systems using parametrically-dependent linear, dynamic feedback,” in ASME Advances in Robust and Nonlinear Control
Systems, vol. 43, 1992, pp. 29–36.
[11] F. Wu, “Control of linear parameter varying systems,” Ph.D. dissertation, University of California, Berkeley, 1995.
[12] S. Joshi and A. Kelkar, “Design of norm-bounded and sector-bounded
lqg controllers for uncertain systems,” Journal of Optimization Theory
and Applications, vol. 113, no. 2, pp. 269–282, 2002.
[13] A. Walsh, “Extremum-seeking guidance and conic-sector-based control of aerospace systems,” Ph.D. dissertation, University of Michigan,
2017.
[14] S. Sivaranjani, V. Gupta, and P. Seiler, “Passivity of linear parameter
varying systems with intermittent non-passive behavior,” in IEEE
Conference on Decision and Control. IEEE, 2015, pp. 753–758.
[15] A. Megretski and A. Rantzer, “System analysis via integral quadratic
constraints,” IEEE Transactions on Automatic Control, vol. 42, no. 6,
pp. 819–830, 1997.
[16] R. Lozano, B. Brogliato, O. Egeland, and B. Maschke, Dissipative
systems analysis and control: theory and applications.
Springer
Science & Business Media, 2013.
[17] D. Hill and P. Moylan, “The stability of nonlinear dissipative systems,”
IEEE Transactions on Automatic Control, vol. 21, no. 5, pp. 708–711,
1976.
[18] W. Qiu, V. Vittal, and M. Khammash, “Decentralized power system
stabilizer design using linear parameter varying approach,” IEEE
Transactions on Power Systems, vol. 19, no. 4, pp. 1951–1960, 2004.
[19] E. Agarwal, S. Sivaranjani, and P. J. Antsaklis, “Feedback passivation
of nonlinear switched systems using linear approximations,” in IEEE
Indian Control Conference (ICC), 2017, pp. 12–17.
[20] S. Sivaranjani and D. Thukaram, “Networked control of smart grids
with distributed generation,” in IEEE India Conference (INDICON),
2013, pp. 1–6.
[21] J. Löfberg, “Yalmip: A toolbox for modeling and optimization in
matlab,” in IEEE International Symposium on Computer Aided Control
Systems Design. IEEE, 2004, pp. 284–289.
[22] J. F. Sturm, “Using sedumi 1.02, a matlab toolbox for optimization
over symmetric cones,” Optimization methods and software, vol. 11,
no. 1-4, pp. 625–653, 1999.
[23] F. Milano, “An open source power system analysis toolbox,” IEEE
Transactions on Power Systems, vol. 20, no. 3, pp. 1199–1206, 2005.
[24] F. Wu, X. H. Yang, A. Packard, and G. Becker, “Induced L2 norm
control for LPV systems with bounded parameter variation rates,”
International Journal of Robust and Nonlinear Control, vol. 6, pp.
983–998, 1996.
| 3cs.SY
|
Dynamic Time-Dependent Route Planning in
Road Networks with User Preferences∗
Moritz Baum1 , Julian Dibbelt1 , Thomas Pajor2 and Dorothea Wagner1
1
Karlsruhe Institute of Technology, Germany, [email protected]
2
Sunnyvale, CA
December 30, 2015
Abstract
There has been tremendous progress in algorithmic methods for computing driving directions
on road networks. Most of that work focuses on time-independent route planning, where it is
assumed that the cost on each arc is constant per query. In practice, the current traffic situation
significantly influences the travel time on large parts of the road network, and it changes over
the day. One can distinguish between traffic congestion that can be predicted using historical
traffic data, and congestion due to unpredictable events, e. g., accidents. In this work, we study
the dynamic and time-dependent route planning problem, which takes both prediction (based
on historical data) and live traffic into account. To this end, we propose a practical algorithm
that, while robust to user preferences, is able to integrate global changes of the time-dependent
metric (e. g., due to traffic updates or user restrictions) faster than previous approaches, while
allowing subsequent queries that enable interactive applications.
1 Introduction
The development of practical algorithms for route planning in road networks has been a showpiece application of successful algorithm engineering during the past decade. While Dijkstra’s
algorithm [16] solves the problem in almost linear time, it is too slow for responsive applications on
real-world road networks of country or continental size. Thus, numerous speedup techniques have
been proposed that employ a preprocessing phase to compute auxiliary data, which accelerates
Dijkstra’s algorithm by several orders of magnitude [1].
A successful approach [3, 7, 15, 20, 24] exploits that road networks have small separators [8, 21, 23,
32, 33]. Since this is a rarely changing topological feature, which is independent of the routing cost,
preprocessing is split into two phases [7]: The first, metric-independent phase computes a nested
hierarchy of separators and the overlay graph [34] induced by it. The second, metric-dependent phase
(also called customization) computes distance-preserving weights for the predetermined overlay
based on a given cost metric. Given the weighted overlay, queries can then skip large parts of the
original graph during search. Customization turns out to be very fast: If the routing cost metric
can be expressed as a single, scalar value, customization times below a second have been achieved
on large continental networks [7, 15].
∗
Partially supported by EU grant 288094 (eCOMPASS) and EU grant 609026 (MOVESMART).
1
However, an important [13] aspect of route planning in realistic scenarios is the consideration of
traffic patterns and incidents. For dynamic, time-dependent route planning, routing costs are no
longer constant, scalar values per road segment. Rather, costs depend on the time of day at which
a road is traversed, i. e., they vary as a function of time [5, 18]. Moreover, while these functions
are typically precomputed from historic knowledge of traffic patterns [31], they have to be updated
to respect current traffic incidents or short-term traffic predictions [14]. Also, the complexity of
the functional description is known to strongly increase for long-distance routes [22], which poses a
difficulty when adapting preprocessing techniques based on (long-distance) shortcut edges [12].
In this work, we investigate and experimentally evaluate the challenges that arise when extending
a separator-based route planning approach to the dynamic time-dependent scenario.
Related Work. Dynamic shortest paths and time-dependency have been considered in previous
works, often as separate problems. Dijkstra’s algorithm [16] can be extended to compute timedependent shortest paths [5, 18] if the time-dependent cost functions have certain, reasonable
properties [30]. Two query variants are typically considered: (1) Given a source and target vertex
and a fixed departure time at the source, compute the earliest arrival time (EA) at the target; (2)
compute earliest arrival times for all departure times of a day (profile search). Time-dependent cost
functions are typically represented as piecewise-linear functions, mapping departure to travel time.
The complexity of such a function is expressed by its number of breakpoints. Foschini et al. [22]
show that the optimal path between two vertices can change nΘ(log n) times, when varying departure
time. In other words, for a given vertex pair, any representation of the functional dependency
between departure time and travel time needs super-polynomial size. While these are worst-case
bounds for arbitrary graph families, a very strong increase in functional complexity has also been
observed by experimental evaluation of profile searches on real-world road networks [12].
Several known speed-up techniques have been adapted to handle time-dependency. Some works [10,
11,29] use (scalar) lower bounds on the travel time functions to guide the graph search, circumventing
any problems with functional complexity increase. TD-CALT [10] yields reasonable EA query
times (at least for approximate solutions) and allows for fast dynamic traffic updates, but does
not enable profile search on large networks. TD-SHARC [6] offers profile search on a country-scale
network. Time-dependent Contraction Hierarchies (TCH) [2] enable both fast EA and profile
searches even on continental networks. The technique (greedily) computes a vertex order, building
a sequence of overlays by iteratively contracting “unimportant” vertices while inserting shortcuts
between the remaining neighbors; The weights of shortcuts are obtained using profile searches.
After preprocessing, piecewise-linear function approximation [25] can be used on the arc functions
to save space, dropping optimality of subsequent query results. Quite cleverly, a multi-phase
extension of the query algorithms (ATCH) restores exact results, despite approximated shortcuts [2].
Unfortunately, TCH and ATCH have not been shown to be robust against user preferences. In [28],
time-dependent shortest path oracles are considered that provide approximate time-dependent
distances in sublinear query time, after subquadratic preprocessing effort. The approach was also
evaluated in experimental settings [27], but preprocessing effort and space consumption are rather
high (even on a comparatively small, urban instance).
Other works consider unforeseen dynamic changes such as congestion after an accident. While this
can trivially be solved by repeating the whole preprocessing phase, less costly approaches [11, 19]
enable partial updates of preprocessed data. Customizable Route Planning (CRP) [7] goes even
further, offloading most preprocessing effort to an initial metric-independent, separator-based phase.
The preprocessed data is then customized to a given routing metric for the whole network within
seconds or below (for scalar, non-functional metrics) in an additional preprocessing phase. Not
2
only allows this fast dynamic changes, it also enables robust integration of user preferences and
extended costs models such as turn-costs: Small graph separators in road networks guarantee
efficient customization and query times for any scalar metric, since the induced overlay graphs are
small. Customizable Contraction Hierarchies (CCH) [15] follows a similar approach.
However, for time-dependent metrics, small separators do not guarantee fast customization and
query times: even having a low number of overlay arcs can be too expensive if the corresponding
functional arc weights increase rapidly in complexity. To the best of our knowledge, non-scalar
metrics for separator-based overlay approaches have so far only been investigated in the context of
energy-optimal routes for electric vehicles (EVCRP) [4], where energy consumption depends on the
battery state-of-charge, but functional complexity for long shortcuts grows very slowly.
On the other hand, it is important to note that the use of dynamic but scalar approaches
(such as CRP or CCH) for handling live traffic information yields inaccurate results for medium
and long-distance paths: Such methods will (wrongly) consider current traffic even at far away
destinations—traffic that might well have dispersed once reaching the destination. For most-realistic
results, a combination of dynamic and time-dependent (non-scalar, functional) route planning is
necessary that accounts for current traffic, short-term predictions, and long-term historic knowledge
about recurring traffic patterns.
Contribution and Outline. In this work, we carefully extend the Customizable Route Planning [7]
approach to handle time-dependent functions. As such, we are the first to evaluate partition-based
route planning on a challenging non-scalar metric. To this end, we integrate profile search into the
customization phase and compute time-dependent overlays. We observe that, unlike EVCRP and
TCH, a naïve implementation fails: Shortcuts on higher-level overlays become too expensive to
be kept in memory (and too expensive to evaluate during queries). In order to reduce functional
complexity, we propose to approximate overlay arcs after each level (in contrast to ATCH, which
uses a post-processing step). In fact, even slight approximation suffices to make our approach
practical, in accordance to theory [22]. The resulting algorithmic framework enables interactive
queries with low average and maximum error in a very realistic scenario consisting of live traffic,
short-term traffic predictions, and historic traffic patterns, while also supporting and being robust
to user preferences such as lower maximum driving speeds or the avoidance of left-turns.
The rest of this paper is organized as follows. Section 2 introduces necessary notation and
foundations of this work, while our approach is presented in Section 3. In Section 4, we provide an
extensive experimental evaluation. We finish with concluding remarks in Section 5.
2 Preliminaries
A road network is modeled as a directed graph G = (V, A) with n = |V | vertices and m = |A| arcs,
where vertices v ∈ V correspond to intersections and arcs (u, v) ∈ A to road segments. An s–tpath P (in G) is a sequence Ps,t = [v1 = s, v2 , . . . , vk = t] of vertices such that (vi , vi+1 ) ∈ A. If s
and t coincide, we call P a cycle. Every arc a has assigned a travel-time function fa : Π → R+ from
a given function space F, mapping departure time (within some period Π = [0, π]) to a positive,
real-valued travel time. Given a departure time τ at s, the (time-dependent) travel time τ[s,...,t]
of an s–t-path is obtained by consecutive function evaluation, i. e., τ[s,...,vi ] = f(vi−1 ,vi ) (τ[s,...,vi−1 ) ).
In what follows, we assume that F consists of periodic piecewise linear functions with period Π,
given by breakpoints. For a function f , we denote by |f | its number of breakpoints. Moreover,
we define f max as the maximum value of f , i.e., f max = maxτ ∈Π f (τ ). Analogously, f min denotes
the minimum value of f . A function f is constant if f ≡ c for some c ∈ Π. We presume that
3
functions fulfill the FIFO property, that is, for arbitrary (positive) real values σ ≤ τ ∈ Π, the
condition σ + f (σ) ≤ τ + f (τ ) holds. In other words, waiting at a vertex never pays off. Note that
the shortest path problem becomes N P-hard if this condition is not satisfied for all arcs, unless
waiting is allowed at vertices. On F we require binary link (composition) and merge operations.
Given two functions f, g ∈ F, the link operation is defined as link(f, g) := f + g ◦ (id +f ), whereas
we define merging f and g by merge(f, g) := min(f, g). Given two piecewise linear functions f, g,
the result of the link operation link(f, g) is piecewise linear again, with at most |f | + |g| breakpoints
(namely, at departure times of breakpoints of f and backward projections of departure times of
points of g). Similarly, the result of merging piecewise linear functions is piecewise linear, and the
number of breakpoints is in O(|f | + |g|) (containing breakpoints of the two original functions and at
most one intersection per linear segment). Both operations can be implemented by (coordinated)
linear sweeps over the breakpoints of the corresponding functions.
Given a path P = [v1 , . . . , vk ], its (travel-time) profile is defined as the function fP : Π → R+ that
maps departure time τ (at v1 ) to travel time on P . Starting from f[v1 ,v2] = f(v1 ,v2 ) , we obtain the
desired profile by consecutively applying the link operation, i. e., f[v1 ,...,vi ] = link(f[v1 ,...,vi−1 ] , f(vi−1 ,vi ) ).
Given a set of s–t-paths P, the corresponding s–t-profile is defined by fP (τ ) = minP ∈P fP (τ ) for
τ ∈ Π, i. e., the minimum profile over all paths in P. In other words, the profile s–t-profile
maps departure time to minimum travel time for the given paths. Clearly, it can be obtained by
(iteratively) merging the respective paths.
S
A partition of the vertices V is a set C = {C1 , . . . , Ck } of disjoint sets such that ki=1 = V . More
generally, a nested multi-level partition consists of sets {C 1 , . . . , C L } such that C ` is a partition of
V for all ` ∈ {1, . . . , L}, and additionally for each cell Ci in partition C ` of level ` < L, there is a
partition C `+1 of level ` + 1 containing a cell Cj such that Ci ⊆ Cj . We call Cj the supercell of Ci .
For the sake of consistency, we define C 0 = {{v} | v ∈ V } and C L+1 = {V }. An arc (u, v) ∈ A for
which u and v are in different cells of C ` is called a boundary arc on level `. The vertices u and v
are called boundary vertices (of level `).
Query Variants and Algorithms. Given a departure time τ and vertices s and t, an earliestarrival (EA) query asks for the minimum travel time (from s to t) when departing at time τ .
Similarly, a latest-departure (LD) query asks (given vertices s, t, and an arrival time τ ) for the
minimum travel time of an s–t-path reaching t at time τ . A profile query for given source s and
target t asks for the minimum travel time at every possible departure time τ , i. e., a profile fs,t
from s to t (over all s–t-paths in G). EA queries can be handled by a time-dependent variant of
Dijkstra’s algorithm [18]. It maintains (scalar) arrival time labels d(·) for each vertex, initially
set to τ for the source s (∞ for all other vertices). In each step, a vertex u with minimum d(u)
is extracted from a priority queue (initialized with s). Then, the algorithm relaxes all outgoing
arcs (u, v): if d(u) + f(u,v) (d(u)) improves d(v), it updates d(v) accordingly and adds (or updates) v
in the priority queue. Analogously, we can answer LD queries, by running the algorithm from t and
relaxing incoming instead of outgoing arcs, while maintaining departure time labels. We refer to
this algorithm as TD-Dijkstra.
Profile queries can be solved by a label-correcting approach based on link and merge operations [12].
It generalizes Dijkstra’s algorithm, storing (tentative) s–v profiles fv at each vertex v ∈ V . Initially, we set fs ≡ 0, and fv ≡ ∞ for all other vertices. The algorithm continues along the lines of
TD-Dijkstra, with a queue operating on scalar values fvmin as key of a vertex v with profile fv . For
extracted vertices u, arc relaxations propagate profiles rather than travel times. Hence, for each
arc (u, v), we compute g := link(fu , f(u,v) ) and set fv := merge(fv , g). If the profile fv changed,
we update its key in the queue accordingly. We call this algorithm Profile-Dijkstra. As shown by
4
Foschini et al. [22], the number of breakpoints of the profile of an s–v-paths can be superpolynomial,
and hence, so is the space consumption per vertex label and the running time of Profile-Dijkstra
in the worst case. Accordingly, Profile-Dijkstra is not feasible for large-scale instances, even in
practice [12].
Customizable Route Planning (CRP). To allow fast integration of user-dependent routing preferences, CRP [7] introduces a three-phase workflow. The (metric-independent) preprocessing step of
CRP consists of the computation of a multi-level partition of the input graph, for a given number of
levels L. Several graph partition algorithms tailored to road networks exist, that provide partitions
with balanced cell sizes and small cuts [8,23,32,33]. The multi-level partition is computed in an offline
step and is therefore allowed to be (relatively) costly. For each level ` ∈ {1, . . . , L}, the resulting
partition C ` induces an overlay graph H ` , containing all boundary vertices and boundary arcs in C `
and shortcut arcs between boundary vertices of each cell Ci` ∈ C ` . We define C 0 = {{v} | v ∈ V }
and H 0 := G for consistency. The topology (and the memory layout) of the overlay graphs (of each
level) is constructed during preprocessing. This includes the allocation of data structures to store
the metric-dependent data.
In the metric customization phase, costs of all shortcuts (added to the overlay graphs during
preprocessing) are computed. This is achieved by running, for each boundary vertex of a given cell,
Dijkstra’s algorithm on the respective cell-induced subgraph. On higher levels, previously computed
overlays are used for fast computation of shortcuts.
Finally, Multilevel Dijkstra [9, 24, 26] is used during the online phase to speedup queries. Given a
source vertex s and a target vertex t, the search graph consists of the overlay graph induced by
the top-level partition C L , all overlays of cells of lower levels containing s or t, and the level-0 cells
(in the input graph G) that contain s or t. Then, the query consists of a bidirectional variant of
Dijkstra’s algorithm, executed on this search graph.
3 Our Approach
In this section, we present Time-Dependent CRP (TDCRP), a speedup technique for time-dependent
queries that still allows fast integration of dynamic or user-dependent metric changes. In particular,
TDCRP is able to incorporate long-term traffic patterns, live traffic, as well as short-term traffic
predictions, while supporting preferences of users or vehicle restrictions (e. g., avoiding highways,
maximum preferred driving speed, avoiding left-turns, vehicle height restrictions). Instead of scalar
arc weights, we use functions of departure time at arcs, taking into account historic knowledge
of traffic patterns.Additionally, we allow current and/or predicted traffic updates with a limited
departure time horizon (i. e., accounting for the fact that underlying traffic situations resolve over
time). This conceptual change has important consequences: While for (plain) CRP the topology
of the overlays is fixed after preprocessing (enabling several micro-optimizations with significant
impact on customization and query [7]), in our case the functional complexity is metric-dependent
(influenced by, e. g., user preferences) and has to be handled dynamically during customization.
This renders adaptation to dynamic time-dependent scenarios highly nontrivial, as we require new
data structures and algorithmic changes during customization. Below, we describe our extensions to
handle time-dependent scenarios in TDCRP. We incorporate profile queries into the customization
phase to obtain time-dependent shortcuts. Moreover, we adapt the query phase to efficiently compute
time-dependent shortest routes.
Regarding (metric-independent) preprocessing, on the other hand, the computation of a (multilevel)
partition remains unchanged (compared to plain CRP). Building the overlay, we use the clique
5
matrix representation, storing cliques of boundary vertices in matrices of contiguous memory [7].
However, instead of (scalar) arc weights, matrix entries now represent pointers to functions (whose
complexity is not known until customization). To improve locality, all functions are stored in a
single array, such that profiles corresponding to outgoing arcs of a certain boundary vertex are in
contiguous memory. This dynamic data structure rules out several optimizations for (plain) CRP,
such as microcode instructions, that require preallocated ranges of memory for the metric [7].
3.1 Customization
In the metric customization phase, we run profile searches to obtain costs of all time-dependent
shortcuts. In particular, we require, for each boundary vertex u (in some cell C ` at level ` ≥ 1), the
time-dependent distances for all τ ∈ Π to all boundary vertices v ∈ C ` . To this end, we run a profile
query on the overlay H `−1 . By design, this query is restricted to subcells of C ` , i. e., cells C `−1 on
level ` − 1 for which C `−1 ⊆ C ` holds. This yields profiles for all outgoing (shortcut) arcs (u, v) in
C ` from u. Unfortunately, profile queries are expensive in terms of both running time and space
consumption. Below, we describe improvements to remedy these effects, mostly by tuning the profile
searches.
Improvements. The main bottleneck of the profile search is performing link and merge operations,
which require linear time in the function size (cf. Section 2). To avoid unnecessary operations,
we explicitly compute and store the minimum f min and the maximum f max of each profile f in
its corresponding label (and in shortcuts of overlays). These values can be used for early pruning
(while avoiding costly link and merge operations): before relaxing an arc (u, v), we check whether
min > f max , i. e., the minimum of the linked profile exceeds the maximum of the label at
fumin + f(u,v)
v
vertex v. If this is the case, the arc (u, v) does not need to be relaxed. Otherwise, the functions are
linked. We distinguish four cases, depending on whether the first and second function are constant,
respectively. If both functions are constant, linking becomes trivial (summing up two integers). If
one of them is constant, simple shift operations suffice (note that we need to distinguish two cases,
depending on which of the two functions is constant). Only if none of the functions is constant, we
apply the actual link operation with running time linear in |fu | + |f(u,v) |.
After linking f(u,v) to fu , we obtain a tentative label f˜v together with its minimum f˜vmin and its
maximum f˜vmax . We perform additional checks before merging the profiles fv and f˜v in order to
avoid unnecessary expensive merge operations. First, we perform bound checks: if f˜vmin > fvmax , the
function fv remains unchanged, and no merge is necessary (note that this situation may occur even
though we checked the bounds before linking). Conversely, if fvmax < fvmin , we can simply replace fv
by f˜v . Second, if both checks fail and the functions fv and f˜v are both nonconstant, we perform an
additional check to determine whether merging is necessary (still, one function might dominate the
other). This check consists of a simulated merge operation (thus, a coordinated linear-time sweep
over the breakpoints of each function), which omits (computationally involving and numerically
unstable) line segment intersections.
Additionally, we can maintain a special value per vertex, indicating its unique parent (i. e., the
distinct vertex contributing to the profile at some vertex) if it exists. (Recall that for different
points in time, the optimal path to a vertex, and thus its parent, may vary.) We use this value to
perform hopping reduction [17]. If at some point, before relaxing the arc (v, u), u is the unique
parent of v (i. e., the label at v has not been updated since this arc was relaxed), we can safely
omit linking and merging (because a shortest path cannot contain a cycle). Alternatively, we can
use more powerful clique flags: For each vertex of the overlay graph, we add a flag to its label
that is true if all its parents belong to the same cell (on the current level). This flag is set to true
6
whenever the corresponding vertex label fv is set to the tentative function f˜v after relaxing a clique
arc (u, v) (i. e., the label is set for the first time or the former label fv is dominated by the tentative
function f˜v ). It is set to false if the vertex label is (partially) improved after relaxing a boundary
arc. For flagged vertices, we do not relax outgoing clique arcs, as this cannot possibly improve labels
within the same cell (due to the triangle inequality and the fact that we use full cliques).
Parallelization. Since cells on a given level are processed independently, CRP can be parallelized
in a straightforward manner, assigning cells to different threads [7]. In our scenario, however,
the workload may differ significantly between cells, as it strongly correlates with the number of
time-dependent arcs in the search graph. (Conversely, cell sizes are the bounding factor in plain CRP,
which are balanced by the partition.) As an extreme example, a cell that contains no time-dependent
arcs at all reduces profile searches to simple Dijkstra runs. In realistic data sets, the distribution of
time-dependent arcs is clearly not uniform: it depends on the road type (e. g., highways vs. side
roads) and the area (rural vs. urban). To balance the load per thread, we no longer parallelize per
cell, but per boundary vertex.
In contrast to plain CRP, shortcut profiles are written to dynamic containers, since the number
of breakpoints is not known in advance. Therefore, we have to prohibit parallel (writing) access to
the respective data structure. One way to solve this is to make use of locks. However, this can be
expensive if many threads try to write profiles at the same time. Alternatively, we can use threadlocal profile containers (i. e., each thread uses its own container to store profiles). After customization
of each level, we synchronize data by copying profiles to the global container sequentially. For better
locality during queries, we maintain the relative order of profiles (wrt. the matrix layout, so profiles
of adjacent vertices are likely to be contiguous in memory). Since the relative order of profiles
in thread-local containers is ensured by running queries accordingly, running merge sort on these
containers is sufficient to achieve the global order.
Approximations. On higher levels of the partition, shortcuts represent larger parts of the graph
and contain more breakpoints. Accordingly, shortcuts on higher levels are more expensive (wrt.
memory consumption). This makes profile searches fail on large graphs due to insufficient memory,
even on modern hardware. Moreover, running time is strongly correlated to the complexity of
profiles. In order to save space and time, we can simplify functions during customization. To this
end, we use the algorithm by Imai and Iri [25] that, given a maximum (relative or absolute) error ε,
computes an approximation of a given piecewise linear function with a provably minimum number
of breakpoints. In [2], this technique is applied after preprocessing to reduce space consumption.
Instead, we use this algorithm to simplify profiles after computing all shortcuts of a certain level.
Therefore, searches on higher levels use approximated functions from lower levels, leading to (slightly)
less accurate profiles but faster customization. The parameter ε is a tuning parameter: larger
values allow faster customization, but decrease quality. Also, approximation is not necessarily
applied on all levels, but can be restricted to the higher ones. Note that, when using clique flags on
approximated shortcuts, the triangle inequality may no longer hold (due to the fact that bounds
were simplified after computing shortcuts). Therefore, clique flags yield faster (profile) queries and
(slightly) decreased quality (additional arc relaxations improve shortcut bounds).
3.2 Live Traffic and Short-Term Traffic Predictions
Updates due to, e. g., live traffic, require that we rerun the customization phase. Clearly, we only
have to perform customization only for affected cells, i. e., cells that contain at least one arc for
which an update is made. However, we can do better if we exploit the fact that live traffic updates
7
are not periodic. Also, short-term updates only affect a limited time horizon (rather than the
complete period). Thus, we do not have to propagate updates to boundary vertices that cannot
reach the affected arc before the end of the time horizon.
We assume that short-term updates are given as partial functions f : [π 0 , π 00 ] → R+ , where π 0 ∈ Π
and π 00 ∈ Π are the beginning and end of the time horizon, respectively. Let a1 = (u1 , v1 ), . . . , ak =
(uk , vk ) denote the updated arcs inside some cell C ` at level `, and let f1 , . . . , fk be the corresponding
partial functions representing time horizons. Moreover, let τ be the current point in time. To
update C ` we run, on its induced subgraph, a multi-target latest-departure query from the tails
of all updated arcs. In other words, we initially insert the vertices u1 , . . . , uk into the queue. The
departure time label of ui (for i ∈ {1, . . . , k}) is set to πi00 (i. e., the end of the time horizon [πi0 , πi00 ]
of the partial function fi ). Consequently, the LD query computes, for each vertex of the cell C ` , the
latest possible departure time such that some affected arc is reached on a shortest path (before the
end of its corresponding time horizon). Whenever the search reaches a boundary vertex of the cell,
it is marked as affected by the update. We stop the search as soon as the departure time label of the
current vertex is below τ (recall that the LD query visits vertices in decreasing order of departure
time). Thereby, we ensure that only such boundary vertices are marked from which an updated arc
can be reached in time.
Next, we run the actual profile searches for C ` (similar to regular customization), but only from
affected vertices. For all shortcut profiles obtained during these searches, we test whether they
improve the corresponding stored shortcut. If this is the case, we add the interval of the profile for
which a change occurs to the set of time horizons for the next level. If shortcuts are approximations,
we test whether the change is significant (e. g., if the maximum difference between the profiles
exceeds a certain bound). We continue the update process on the next level accordingly.
3.3 Queries
The query algorithm makes use of shortcuts computed during customization to reduce the search
space. Similar to plain CRP, the search graph for a given source s and target t consists of the overlay
graph induced by the top-level partition C L and subgraphs induced by cells containing s or t. This
search graph does not have to be constructed explicitly, but can be obtained implicitly by obtaining,
at each vertex v, the highest uncommon level of v and both s and t in the multi-level partition:
We compute the highest levels `s,v and `v,t of the partition such that v and s or t, respectively, are
in distinct cells of the partition (0 if v is in the same level-1 cell as s or t). Then, we relax only
outgoing edges of v at level min{`s,v , `v,t } (recall that H 0 = G). Since the search graph is known at
query time, we can simply run (unidirectional) TD-Dijkstra to answer earliest arrival queries.
min stored at arcs during EA queries: If, at some
We also make use of the minimum values f(u,v)
min , does not yield an improvement of d(v), we do not have to relax
point during the query, d(u) + f(u,v)
the arc (u, v). Thereby, we avoid costly function evaluation. For approximated queries (i. e., if we
use approximated shortcuts), we observe that clique flags improve running times particularly when
activated on higher levels (where cliques are larger). On the other hand, we observe quality issues
with activated clique flags mostly on the lower levels (since errors propagate to higher levels). Hence,
we activate clique flags only on the top-k levels of the partition (where k is a tuning parameter).
3.4 Implementation Details
Implementation of the functional operations (evaluation, link, merge, and subroutines) requires
considerable effort and is a great source for numerical imprecision. We tested several implementation
variants and settled on using fixed-point arithmetics at milliseconds resolution. While specifics are
8
beyond the scope of this paper, we note that we found it beneficial to use half-plane tests (which
require no division) to determine whether an intersection between line segments occurs, before
computing the actual intersection.
We reorder vertices of the graph, moving boundary vertices to the front, and ordering vertices of
the same level by cell IDs [7]. This improves locality of subsequent memory accesses. In a naive
implementation, each vertex requires its own vertex label. Since the search graph is known in
advance, we can reduce the number of vertex labels to save space. Instead of explicitly extracting
the search graph, we compute the ranges of IDs of vertices of the current cell for each level (note that
there is at most one range per level, due to vertex reordering). Then, we can remap the ranges of
each level (for source and target cell, respectively) to a smaller, global range of vertex indices. The
size of the global range depends on the maximum cell size, which is known after preprocessing. The
following (mixed) variant worked best in our experiments: We only remap bottom level inner-cell
indices (the majority of vertices), while keeping a distinct vertex label for every boundary vertex
in the graph. Thereby, we save a significant amount of space, improve locality, but keep vertex
mapping overhead limited during queries.
To reset labels between queries, rather than using standard approaches (like timestamps), we
again exploit the vertex reordering: We store, for every cell, its vertex ranges, and reset only the (at
most two) cells per level touched during a query, along with all top level vertices. With labels being
on adjacent ranges, this can be done efficiently in practice.
Finally, we can save some space by storing cell IDs only at boundary vertices. Before running the
actual query algorithm, the source and target cell can be retrieved by running an additional DFS in
their respective (bottom-level) cells at negligible overhead.
4 Experiments
We implemented all algorithms in C++ using g++ 4.8 (flag -O3) as compiler. Experiments were
conducted on a dual 8-core Intel Xeon E5-2670 clocked at 2.6 GHz, with 64 GiB of DDR3-1600
RAM, 20 MiB of L3 and 256 KiB of L2 cache. We ran customization in parallel (using all 16 threads)
and queries sequentially.
Input Data and Methodology. Our test instances are based on the road network of Germany
(n = 4.7 million, m = 10.8 million) and Western Europe (n = 18 million, m = 42.2 million), kindly
provided by PTV AG, and the road network of Berlin/Brandenburg (n = 443 k, m = 988 k), kindly
provided by TomTom. All inputs contain time-dependent data. For the Germany and Berlin
instance, data stems from historical traffic patterns. For both instances, we extracted the 24 hour
Table 1. Network properties. We report the number of vertices and arcs of the routing graph, and
as a measure of time-dependency, the total amount of break points in the whole network as well as
the average time-dependent arc complexity.
# Break points
Network
# Vertices
# Arcs
Time-dep. arcs
Total
Avg.
Berlin
Germany
Europe
443 191
4 692 091
18 010 173
988 493
10 805 429
42 188 664
271 688 (27.5%)
777 984 (7.2%)
2 609 651 (6.2%)
3 464 241
12 714 370
29 362 693
12.8
16.3
11.3
9
Table 2. Custom performance on Europe instance for varying approximation errors (ε). We report,
per level, the number of breakpoints (bps, in millions) in the resulting overlay, the percentage
of clique arcs that are time-dependent (td.clq.arcs), the average complexity of time-dependent
arcs (td.arc.cplx), as well as the customization time for that level. Without approximation, level 5
and 6 cannot be computed as they do not fit into main memory.
ε
Lvl1
Lvl2
Lvl3
Lvl4
Lvl5
Lvl6
Total
—
[106 ]
bps
td.clq.arcs [%]
td.arc.cplx
time [s]
99.1
17.0
21.0
11.4
398.4
52.6
68.9
52.0
816.4
76.0
189.0
152.9
1 363.4
84.2
509.3
206.2
—
—
—
—
—
—
—
—
2 677.4
—
—
375.7
0.01%
bps [106 ]
td.clq.arcs [%]
td.arc.cplx
time [s]
75.6
16.9
16.0
4.5
182.7
52.6
31.6
18.2
244.6
75.9
56.6
33.1
240.8
84.1
89.9
83.4
149.3
85.1
108.6
152.2
59.2
82.4
108.0
153.3
952.2
—
—
444.7
0.1%
bps [106 ]
td.clq.arcs [%]
td.arc.cplx
time [s]
60.7
17.0
12.9
4.2
107.5
52.7
18.6
16.1
111.5
76.0
25.8
21.6
87.9
84.2
32.8
41.2
47.9
85.2
34.8
63.2
17.5
82.5
32.1
55.7
432.9
—
—
202.0
1.0%
bps [106 ]
td.clq.arcs [%]
td.arc.cplx
time [s]
45.6
17.0
9.7
4.1
58.0
52.7
10.0
14.1
45.6
76.0
10.6
14.9
29.2
84.2
10.9
22.85
14.7
85.2
10.7
30.0
5.3
82.5
9.8
24.3
198.5
—
—
110.2
profile of a Tuesday. For Western Europe, travel time functions were generated synthetically [29].
See Table 1 for details on inputs. For partitioning, we used PUNCH [8], which is explicitly developed
for road networks and aims at minimizing the number of boundary arcs. For Germany, we use a
5-level partition, with maximum cell sizes of 2[4:8:12:15:18] . For Europe, we use a 6-level partition,
with maximum cell sizes 2[4:8:11:14:17:20] . Finally, we use a 5-level partition for Berlin, with cell size
restricted to 2[4:8:11:14:17] . Compared to plain CRP, we use partitions with more levels, to allow
fine-grained approximation. Computing the partition took 20 seconds for Berlin, 5 minutes for
Germany, and 23 minutes for Europe. Given that road topology changes rarely, this is sufficiently
fast in practice.
Evaluating Customization. Table 2 shows details on customization for different approximation
parameters ε on the Europe instance. We report, for several choices of ε (and for each level of
the partition), the total number of breakpoints, the ratio of time-dependent (i. e., non-constant)
shortcuts, the average number of breakpoints per time-dependent shortcuts, and the (parallelized)
customization time. The first block shows figures for exact profile computation. Customization
had to be aborted after the fourth level, because the 64 GiB of main memory were not sufficient
to store all vertex (profile) labels. For remaining levels, we clearly see the strong increase of total
breakpoints per level. Also, the (relative) amount of time-dependent arcs rises with each level (since
shortcuts become longer). Finally, customization time clearly correlates with profile complexity,
from 10 seconds on the lowest level, to more then three minutes on the fourth. When approximating,
we clearly see that customization becomes faster for larger values of ε. Recall that higher levels work
10
Table 3. Single-core query performance on Europe instance as a trade-off of customization effort
and approximation. Since we employ approximation per level, the resulting query errors can be
higher than the input parameter.
Customization
Query
Time[s]
ε
# Vertices
# Arcs
# Bps
Time [ms]
Avg. err. [%]
Max. err. [%]
444.7
202.0
110.2
0.01%
0.1%
1.0%
3 499
3 499
3 499
12 231
12 248
12 259
347 559
78 404
54 784
6.28
4.02
3.65
<0.01
0.04
0.54
0.08
0.81
3.21
on approximated shortcuts of previous levels, so ε is not an approximation bound for all shortcuts.
Still, even a very small value (0.01 %) yields a massive drop of profile complexity (more than a
factor 5 at level 4), and immediately allows complete customization. For reasonably small values
(ε = 0.1 %, ε = 1.0 %), we see that customization becomes much faster (less than two minutes for
ε = 1.0 %) and in particular fast enough for real-time updates. Even for larger values of ε, the
higher levels are far more expensive (due to the large amount of time-dependent arcs).
Evaluating Queries. Table 3 details query performance for different values of the approximation
parameter ε on the Europe. For each, we report the number of settled vertices, arc relaxations,
and total number of evaluated breakpoints during the query. Furthermore, we report timings,
average, and maximum error for 100 000 point-to-point queries. In each query, source and target
vertex, as well as departure time were chosen uniformly at random. The data shows that (similar to
customization), query times decrease with higher approximation ratio. Again, this is due to the
smaller number of breakpoints in profiles. As expected, the average and maximum error increase
with ε. However, we see that even for the largest value ε = 1.0 %, the maximum error is very
low (about 3 %, which should be acceptable in most reasonable scenarios). Moreover, query times
are quite practical for all values of ε, ranging from less than 4 ms to below 7 ms. In summary,
our approach allows query times that are fast enough for interactive applications, if a reasonably
low error is allowed. Given that the input functions are based on statistical input with inherent
inaccuracy, this should be more than acceptable for most applications.
Comparison with Related Work. Finally, Table 4 provides an overview of our approach compared
to the most relevant previous work on time-dependent routing. For each technique, we report
performance of metric-dependent preprocessing and EA queries. We evaluated our approach on
all benchmark instances, namely, Berlin, Germany, and Europe (each for the fastest variant, i. e.,
ε = 1.0). For every approach (and instance), the table shows (metric-dependent) preprocessing time
(and the number of cores), as well as space consumption. For EA queries, we report vertex scans,
query time, and (average and maximum) errors. We see that our approach competes with all related
approaches, but has much lower customization time. Accounting for differences in hardware and
implementation, TDCALT may possibly be preprocessed much faster (also, running preprocessing
in parallel should be possible). This makes TDCALT the only candidate that might compete
with TDCRP in terms of preprocessing effort and resulting query performance. However, without
TDCALT allows no profile searches, making it a less versatile approach. Also, the support of
advanced features, such as short-term traffic predictions, has not beed discussed for TDCALT. All
in all, we see that TDCRP is a competitive approach that clearly broadens the state-of-the-art of
time-dependent route planning, as it handles a wider range of practical requirements.
11
Table 4. Overview of related work. For Germany and Europe, we present figures (as-is) for
TDCALT [10] (evaluated on an Opteron 2218 at 2.67 GHz), SHARC [6] (Opteron 2218, 2.67 GHz),
TCH and ATCH [2] (2×4×Core-i7, 2.67 GHz). For Berlin, we report measures for FLAT [28]
(machine not specified). Also, instances are slightly different for Berlin. For TCH and ATCH,
preprocessing times distinguish vertex ordering and contraction.
Metric-dep. Prepro.
Algorithm
Instance
TDCALT
TDCALT-K1.15
eco L-SHARC
heu SHARC
TCH
ATCH (1.0)
inex. TCH (0.1)
inex. TCH (1.0)
inex. TCH (2.5)
inex. TCH (10.0)
TDCRP (1.0)
Germany
Germany
Germany
Germany
Germany
Germany
Germany
Germany
Germany
Germany
Germany
TDCALT
TDCALT-K1.05
TDCALT-K1.10
TDCALT-K1.15
eco L-SHARC
heu SHARC
TCH
ATCH (1.0)
inex. TCH (0.1)
inex. TCH (1.0)
inex. TCH (2.5)
inex. TCH (10.0)
TDCRP (1.0)
Europe
Europe
Europe
Europe
Europe
Europe
Europe
Europe
Europe
Europe
Europe
Europe
Europe
FLAT K2000
FLAT SR2000
TDCRP (1.0)
Berlin*
Berlin*
Berlin
[h:m:s] (# Thr)
EA Queries
Space
[B/n]
# Vert.
Time [ms]
Err. [%] Err. [%]
avg.
max.
9:00
9:00
1:18:00
3:26:00
5:04 + 1:14
5:04 + 1:14
5:04 + 1:14
5:04 + 1:14
5:04 + 1:14
5:04 + 1:14
0:08
(1)
(1)
(1)
(1)
(8)
(8)
(8)
(8)
(8)
(8)
(16)
50
50
219
137
995
239
286
214
172
113
77
3 190
1 593
2 776
818
520
588
642
654
668
898
2 152
5.36
1.87
6.31
0.69
0.75
1.24
0.70
0.69
0.72
1.06
1.17
—
0.05
—
n/a
—
—
0.02
0.27
0.79
3.84
0.68
—
13.84
—
0.61
—
—
0.10
1.01
2.44
9.75
3.60
1:00:00
1:00:00
1:00:00
1:00:00
6:49:00
22:12:00
37:42 + 8:02
37:42 + 8:02
37:42 + 8:02
37:42 + 8:02
37:42 + 8:02
37:42 + 8:02
1:50
(1)
(1)
(1)
(1)
(1)
(1)
(8)
(8)
(8)
(8)
(8)
(8)
(16)
61
61
61
61
198
127
599
208
239
195
175
144
133
60 961
32 405
12 777
6 365
18 289
5 031
1 021
1 223
1 722
1 782
1 875
1 801
3 499
121.4
62.5
21.9
9.2
38.29
2.94
2.11
2.89
2.70
2.76
2.94
2.92
3.65
—
0.01
0.09
0.26
—
n/a
—
—
0.02
0.20
0.48
2.88
0.54
—
3.94
7.88
8.69
—
0.61
—
—
0.15
1.50
3.37
16.21
3.21
4:30:00 (6)
4:30:00 (6)
0:01 (16)
58 000
58 000
67
1 522
120
854
0.55
0.07
0.28
0.24
0.73
1.47
n/a
n/a
2.69
12
5 Conclusion
In this work, we introduced TDCRP, an extension of CRP to time-dependent query scenario.
We showed how integrating profile searches into the customization phase of CRP allows us to
compute time-dependent overlays. Thereby, we enable integration of dynamic or user-dependent
metric changes. Additionally, TDCRP also efficiently incorporates live traffic and short-term traffic
predictions. Our experimental analysis on a continental road network shows that for higher-level
overlays (i. e., where overlay vertices are far apart), the increase in functional complexity is too high
to be kept in memory (both in contrast to the known techniques EVCRP and TCH). To reduce
memory consumption, we approximate the overlay arcs at each level, accelerating the customization
and query times. As a result, we obtain an approach that enables fast near-optimal queries, while
being able to quickly integrate user preferences, live traffic updates and traffic predictions.
There are several aspects of future work. First, we would like to further investigate TDCALT,
which appears quite competitive in scenarios where profile searches are not a requirement. For
example, TDCALT should be easy to parallelize, so a re-evaluation might be fruitful (e. g., exploiting
insights from [19]). Additionally, we are interested in alternative customization approaches (avoiding
expensive profile searches). This could be achieved, e. g., by making use of kinetic data structures [22],
or balanced contraction within cells. As demonstrated for ATCH, one could also aim at exact
queries based on approximated shortcuts, by employing additional query phases.
Acknowledgements. We thank Gernot Veit Batz, Daniel Delling, Felix König, Spyros Kontogiannis,
and Ben Strasser for interesting conversations.
13
References
[1] Hannah Bast, Daniel Delling, Andrew V. Goldberg, Matthias Müller–Hannemann, Thomas Pajor,
Peter Sanders, Dorothea Wagner, and Renato F. Werneck. Route planning in transportation networks.
Technical Report abs/1504.05140, ArXiv e-prints, 2015.
[2] Gernot Veit Batz, Robert Geisberger, Peter Sanders, and Christian Vetter. Minimum time-dependent
travel times with contraction hierarchies. ACM Journal of Experimental Algorithmics, 18(1.4):1–43,
April 2013.
[3] Reinhard Bauer, Tobias Columbus, Ignaz Rutter, and Dorothea Wagner. Search-space size in contraction hierarchies. In Proceedings of the 40th International Colloquium on Automata, Languages, and
Programming (ICALP’13), volume 7965 of Lecture Notes in Computer Science, pages 93–104. Springer,
2013.
[4] Moritz Baum, Julian Dibbelt, Thomas Pajor, and Dorothea Wagner. Energy-optimal routes for electric
vehicles. In Proceedings of the 21st ACM SIGSPATIAL International Conference on Advances in
Geographic Information Systems, pages 54–63. ACM Press, 2013.
[5] K. Cooke and E. Halsey. The shortest route through a network with time-dependent internodal transit
times. Journal of Mathematical Analysis and Applications, 14(3):493–498, 1966.
[6] Daniel Delling. Time-dependent SHARC-routing. Algorithmica, 60(1):60–94, May 2011.
[7] Daniel Delling, Andrew V. Goldberg, Thomas Pajor, and Renato F. Werneck. Customizable route
planning in road networks. Transportation Science, 2015.
[8] Daniel Delling, Andrew V. Goldberg, Ilya Razenshteyn, and Renato F. Werneck. Graph partitioning
with natural cuts. In 25th International Parallel and Distributed Processing Symposium (IPDPS’11),
pages 1135–1146. IEEE Computer Society, 2011.
[9] Daniel Delling, Martin Holzer, Kirill Müller, Frank Schulz, and Dorothea Wagner. High-performance
multi-level routing. In The Shortest Path Problem: Ninth DIMACS Implementation Challenge, volume 74
of DIMACS Book, pages 73–92. American Mathematical Society, 2009.
[10] Daniel Delling and Giacomo Nannicini. Core routing on dynamic time-dependent road networks. Informs
Journal on Computing, 24(2):187–201, 2012.
[11] Daniel Delling and Dorothea Wagner. Landmark-based routing in dynamic graphs. In Proceedings of
the 6th Workshop on Experimental Algorithms (WEA’07), volume 4525 of Lecture Notes in Computer
Science, pages 52–65. Springer, June 2007.
[12] Daniel Delling and Dorothea Wagner. Time-dependent route planning. In Robust and Online Large-Scale
Optimization, volume 5868 of Lecture Notes in Computer Science, pages 207–230. Springer, 2009.
[13] Ugur Demiryurek, Farnoush Banaei-Kashani, and Cyrus Shahabi. A case for time-dependent shortest
path computation in spatial networks. In Proceedings of the 18th ACM SIGSPATIAL International
Conference on Advances in Geographic Information Systems (GIS’10), pages 474–477, 2010.
[14] Themistoklis Diamantopoulos, Dionysios Kehagias, Felix König, and Dimitrios Tzovaras. Investigating
the effect of global metrics in travel time forecasting. In Proceedings of the 16th International IEEE
Conference on Intelligent Transportation Systems (ITSC), pages 412–417. IEEE, 2013.
[15] Julian Dibbelt, Ben Strasser, and Dorothea Wagner. Customizable contraction hierarchies. In Proceedings
of the 13th International Symposium on Experimental Algorithms (SEA’14), volume 8504 of Lecture
Notes in Computer Science, pages 271–282. Springer, 2014.
[16] Edsger W. Dijkstra. A note on two problems in connexion with graphs. Numerische Mathematik,
1:269–271, 1959.
[17] Yann Disser, Matthias Müller–Hannemann, and Mathias Schnee. Multi-criteria shortest paths in timedependent train networks. In Proceedings of the 7th Workshop on Experimental Algorithms (WEA’08),
volume 5038 of Lecture Notes in Computer Science, pages 347–361. Springer, June 2008.
14
[18] Stuart E. Dreyfus. An appraisal of some shortest-path algorithms. Operations Research, 17(3):395–412,
1969.
[19] Alexandros Efentakis and Dieter Pfoser. Optimizing landmark-based routing and preprocessing. In
Proceedings of the 6th ACM SIGSPATIAL International Workshop on Computational Transportation
Science, pages 25:25–25:30. ACM Press, November 2013.
[20] Alexandros Efentakis, Dieter Pfoser, and Yannis Vassiliou. Salt. a unified framework for all shortest-path
query variants on road networks. In Proceedings of the 14th International Symposium on Experimental
Algorithms (SEA’15), Lecture Notes in Computer Science, pages 298–311. Springer, 2015.
[21] David Eppstein and Michael T. Goodrich. Studying (non-planar) road networks through an algorithmic
lens. In Proceedings of the 16th ACM SIGSPATIAL international conference on Advances in geographic
information systems (GIS ’08), pages 1–10. ACM Press, 2008.
[22] Luca Foschini, John Hershberger, and Subhash Suri. On the complexity of time-dependent shortest
paths. Algorithmica, 68(4):1075–1097, April 2014.
[23] Michael Hamann and Ben Strasser. Graph bisection with pareto-optimization. Technical report, arXiv,
2015.
[24] Martin Holzer, Frank Schulz, and Dorothea Wagner. Engineering multilevel overlay graphs for shortestpath queries. ACM Journal of Experimental Algorithmics, 13(2.5):1–26, December 2008.
[25] H. Imai and Masao Iri. An optimal algorithm for approximating a piecewise linear function. Journal of
Information Processing, 9(3):159–162, 1987.
[26] Sungwon Jung and Sakti Pramanik. An efficient path computation model for hierarchically structured
topographical road maps. IEEE Transactions on Knowledge and Data Engineering, 14(5):1029–1046,
September 2002.
[27] Spyros Kontogiannis, George Michalopoulos, Georgia Papastavrou, Andreas Paraskevopoulos, Dorothea
Wagner, and Christos Zaroliagis. Analysis and experimental evaluation of time-dependent distance
oracles. In Proceedings of the 17th Meeting on Algorithm Engineering and Experiments (ALENEX’15),
pages 147–158. SIAM, 2015.
[28] Spyros Kontogiannis, Dorothea Wagner, and Christos Zaroliagis. Hierarchical oracles for time-dependent
networks. Technical report, 2015. Technical Report on arXiv.
[29] Giacomo Nannicini, Daniel Delling, Leo Liberti, and Dominik Schultes. Bidirectional A* search on
time-dependent road networks. Networks, 59:240–251, 2012. Best Paper Award.
[30] Ariel Orda and Raphael Rom. Shortest-path and minimum delay algorithms in networks with timedependent edge-length. Journal of the ACM, 37(3):607–625, 1990.
[31] Dieter Pfoser, Sotiris Brakatsoulas, Petra Brosch, Martina Umlauft, Nektaria Tryfona, and Giorgos
Tsironis. Dynamic travel time provision for road networks. In Proceedings of the 16th ACM SIGSPATIAL
international conference on Advances in geographic information systems (GIS ’08). ACM Press, 2008.
[32] Peter Sanders and Christian Schulz. Distributed evolutionary graph partitioning. In Proceedings of the
14th Meeting on Algorithm Engineering and Experiments (ALENEX’12), pages 16–29. SIAM, 2012.
[33] Aaron Schild and Christian Sommer. On balanced separators in road networks. In Proceedings of
the 14th International Symposium on Experimental Algorithms (SEA’15), Lecture Notes in Computer
Science. Springer, 2015.
[34] Frank Schulz, Dorothea Wagner, and Karsten Weihe. Dijkstra’s algorithm on-line: An empirical case
study from public railroad transport. ACM Journal of Experimental Algorithmics, 5(12):1–23, 2000.
15
| 8cs.DS
|
Deriving Quests from Open World Mechanics
Ryan Alexander
Chris Martens
North Carolina State University
[email protected]
arXiv:1705.00341v1 [cs.MM] 30 Apr 2017
ABSTRACT
Open world games present players with more freedom than games
with linear progression structures. However, without clearly-defined
objectives, they often leave players without a sense of purpose. Most
of the time, quests and objectives are hand-authored and overlaid
atop an open world’s mechanics. But what if they could be generated organically from the gameplay itself? The goal of our project
was to develop a model of the mechanics in Minecraft that could
be used to determine the ideal placement of objectives in an open
world setting. We formalized the game logic of Minecraft in terms
of logical rules that can be manipulated in two ways: they may be
executed to generate graphs representative of the player experience
when playing an open world game with little developer direction;
and they may be statically analyzed to determine dependency orderings, feedback loops, and bottlenecks. These analyses may then be
used to place achievements on gameplay actions algorithmically.
KEYWORDS
procedural content generation, quests, Minecraft, game modeling
ACM Reference format:
Ryan Alexander and Chris Martens. . Deriving Quests from Open World
Mechanics. In Proceedings of Foundations of Digital Games Conference, Cape
Cod, MA, August 2017 (FDG2017), 6 pages.
DOI:
1
INTRODUCTION
In an open world game, players enjoy a great deal of autonomy in
selecting from a complex system of mechanics, explorable spaces,
and goals. Rather than experiencing a prescribed linear (or partiallyordered) progression of challenges and plot points orchestrated by
a game designer, players are free to experiment with the consequences of the game’s mechanics and exhibit more creativity in
deciding what to do, frequently devising their own goals. In the
framework of Jesper Juul [6], open world games emphasize emergence over progression: they are more interested in opening a wide
space of explorable consequences of the game’s mechanics than in
delivering a specific sequential experience.
Minecraft is a popular example of such a game, whose open world
mechanics are supported by procedurally generated terrain: players
can mine the terrain for natural resources, which they then craft into
new tools necessary for survival, exploration, and additional mining.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
FDG2017, Cape Cod, MA
© ACM. . . . $15.00
DOI:
North Carolina State University
[email protected]
Minecraft also comes with a hand-authored achievement tree (see
Figure 1) that serves to overlay some progression structure on top
of the open world mechanics: the player is notified Achievement
Unlocked when they reach certain points in their exploration of the
game’s mechanics, and eventually they may be led down any of
several branches to find special items and characters.
Clearly, although open worlds without prescriptive goals have
wide appeal, progression structures such as quests provide something that players are missing: a sense of purpose or progress
as they navigate the game’s mechanics. However, once players
complete Minecraft’s hand-authored achievement tree and special
quests, they are back to where they started in a directionless open
world. This quandary leads us to our problem statement: could
we procedurally generate quests based purely on an open world
game’s mechanics?
Examination of Minecraft’s hand-authored achievement tree
suggests that much of it could be derived algorithmically. The
tool crafting portion of the tree is a prime example of this, due to
the cyclical nature of the crafting system. Making new tools and
using new materials for tools are rewarded with achievements. By
examining these gameplay elements we can determine why they
were chosen to have achievements and how to apply that decision
making to other elements.
We present a formal computational model of Minecraft’s mining and crafting rules that is amenable to algorithmic treatment
analogous to a topological sort of these mechanics. The results
align naturally to progression structures in the game and suggest
locations in game traces where achievements might be placed. Furthermore, we show how two distinct analyses of the rules provide
distinct progression structures.
Our results constitute a novel approach to quest generation,
based not on a model of story distinct from a game’s mechanics, but
in fact derived from them and directly leveraging their emergent
behavior. As a secondary contribution, we illustrate the methodological value in using a formal modeling tool based on logic to
gain insight into a game’s mechanics.
2
RELATED WORK
On a broad level, the context for this work is the field of procedural
content generation (PCG). Within this field, we observe a shift in
interest from algorithms that can create artbitrary, varied content
to those that can generate to fit a specfication, such as the argument
for generating with ASP [9], a framework whose primary edge is
its rich constraint specification capabilities.
Others have also observed that, while PCG traditionally has
centered on crafting content to fit with the emergent dynamics of a
game (such as level generation), generating progression structures
such as narrative and quests poses novel and important challenges.
For instance, the Grail framework [10] addresses the problem of
constructing quests whose goals may have emergent solutions
FDG2017, August 2017, Cape Cod, MA
based on the game’s mechanics, rather than scripted solutions
based on the author’s intended path for the player. Likewise, the
Symon project [5] procedurally generates fetchquest-style narrative
puzzles from a flexible definition of object interactions. Compared
to our work, the goals of these projects were different: rather than
starting with an open world and building quest content atop it, they
aim to allow an author to specify only enough world rules for the
generator to create a flexible range of playable story content.
The “mission/space” dichotomy [3] explicitly formulates the integration of generating spatial content (e.g. levels) with progression
content (e.g. missions). First author Dormans also provides a more
thorough theoretical account of integrating emergence and progression [2]. Dormans et al.’s work closely relates to ours in that
it explicitly addresses co-generation of an open, explorable world,
and an ordered, finite progression, including consideration of how
to generate a space to fit a provided mission, or an emergent structure from a progression. Our work explores the opposite direction,
generating progression structures from emergent ones, and further
considers richer world mechanics (crafting, mining) than the lockand-key-based level exploration primarily explored by Dormans et
al.
Recent interest in general level generation [7] poses a problem
quite similar to the one we are addressing: how can a program,
given a game’s mechanics (specified in VGDL), generate a level
(including a goal condition) for the game? In this framework, the
level serves as the progression structure (an initial condition and a
goal) whereas the games rules, specified in terms of which entities
exist, their properties, and how they interact with other properties,
create emergence. This problem aligns very well with the case we
examine of taking Minecraft’s open world rules and generating a
beginning and end condition for a player to interact with them.
Finally, others have considered the problem of the quest in a
generated world [1], which is similar to the idea of generating quests
for an open world in that what a player can do within the world
is not known at the outset. They formulate a quest in terms of
spatial progression, and their system considers randomly generated
levels and generates lock-and-key puzzles for them. In this work,
the actions that make up the quest (collecting keys and unlocking
doors) are layered atop a generated world without a pre-existing set
of emergent rules. Our work seeks to construct quests not by adding
affordances (keys, locks) to a generated world but by making use of
existing game world rules, which permits us to directly integrate
the actions a player takes to complete a quest with the mechanics
that permit open play.
3 BACKGROUND
3.1 Emergence and Progression in Minecraft
Emergence and progression are terms used by Jesper Juul [6] to
describe two different modes of gameplay. Progression structures
engage the player through a structured set of objectives or quests.
Emergent play revolves around the naturally occurring events that
stem from a small number of simple rules. There are drawbacks to
both methodologies used in isolation: games of pure progression
may feel “railroaded” to the player, and furthermore demand more
time from developers to hand craft the desired experience, while
games of pure emergence risk overwhelming players with the vast
Ryan Alexander and Chris Martens
Figure 1: The achievement menu. Hovering the cursor over
the icon for a visible achievement shows a brief description.
space of possibilities available to them, or alternatively boring them
with no clear direction for advancement. Thus, many games, such
as Minecraft, utilize elements of both: they provide a simple set
of open-ended mechanics for exploratory play while also adding
quests and achievements to give the player a sense of progression.
Minecraft is an open world survival game, where players can
traverse an infinite terrain of blocks. These blocks come in a variety
of materials and can be collected by the player then used to build
structures or make items. Due to the open world setting, there are
no restrictions on where the player can go, which lends more freedom to the playing experience. Alongside this lack of boundaries,
also comes a lack of explanation. Upon spawning into the game,
players have “no clear idea of what he or she can do within the
world” [4]. The only hint is a message informing players that they
have unlocked an achievement upon opening their inventory. These
achievements are the main source of progression in this mostly
emergent game.
In Minecraft, achievements are optional objectives that the player
can choose to accomplish. The achievement menu displays them
as a tree where one achievement will serve as a requirement for
completing another. There is no in-game reward for completing
an achievement, other than unlocking the ability to see the next
achievements in the tree. As shown in Figure 1, each achievement
has a square icon linked to other achievements by a line. The colors
of the icons and lines represent the status of the achievement. Bright
icons indicate that the achievement has been unlocked, darker icons
mean that the achievement is not unlocked but can be unlocked and
its details are visible to the player. The darkest icons signal that an
achievement cannot be unlocked yet and the player does not have
access to information about the achievement. Similarly, gray lines
link unlocked achievements, green join unlocked achievements
to those that can be unlocked, while black lines connect locked
objectives.
3.2
Quests, Achievements, and Objectives
So far, we have been using the terms quest, achievement system,
and objective placement relatively interchangeably. In this work,
we adopt a simplified notion of quest meaning, essentially, a subset
of the nodes in a gameplay trace that are recognized by the game
Deriving Quests from Open World Mechanics
system as progress (e.g. with “achievement” messages). An objective in this framework is simply some communication of the fact
that the player should try to reach achievement nodes. We do not
closely examine these communication mechanisms (e.g. Minecraft’s
achievement tree conveys a great deal of information about achievements’ inter-dependency, whereas in some games, how to complete
achievements is left completely opaque).
3.3
Ceptre
Designing quests from the rules of a game requires formalizing
its mechanics in a way that is analyzable by an algorithm. We
created our formal model of Minecraft’s mechanics in Ceptre [8],
which allows us to represent game logic in a high-level, quicklyprototypable, yet rigorous and formal way.
Ceptre is a rule-based specification language describing how
program states (e.g. game states) may evolve. Each Ceptre rule
consists of a left-hand side (LHS), its inputs, and a right-hand side
(RHS), its outputs. A collection of rules is interpreted by multiset
rewriting: an initial state is supplied, a multiset of predicates. Then,
each step of execution involves selecting a rule that may fire in the
current state, and firing it, which means replacing the elements
in the multiset that match the LHS of the rule with new elements
matching the RHS.
Collections of rules within Ceptre programs are called stages,
and each stage may be designated as autonomous or interactive.
When a stage runs manually, Ceptre displays a list of possible
state transitions to the user from which she may choose to step
the program forward. If the program is run autonomously, it will
randomly pick possible transitions until no more remain.
Regardless of which method is used, it is possible to view a
visual representation of the rules that were executed by generating
the trace graph such as the one shown in Figure 5. These traces
provide information about how rules fit together to help establish
where any progression occurs. Being able to analyze how rules
interoperate this way is the main reason Ceptre was chosen for this
project. Running Ceptre code autonomously multiple times creates
a random sampling of play trace space and generates corresponding
trace graphs that visually demonstrate how the rules interact, while
manually running the code allows the user to see each step of the
process.
4
A FORMAL MODEL OF MINECRAFT
The game mechanics we chose to look at in Minecraft fall into two
groups, rules for gathering materials and rules to craft tools from
those materials. In this section, we discuss the Minecraft rules we
chose to represent and how they are formalized in Ceptre.
First, we encoded various facts about the world, such as the
quantity of materials carried by the player, which tools she has
built out of which materials, and so on, as predicates in Ceptre.
The mapping from predicate to Minecraft state is shown in Table 1.
Note that predicates with arguments are written in Ceptre with
adjacency, e.g. pred Arg, rather than with parentheses.
4.1
Rules
4.1.1 Rules for Gathering. The first category of rules is based
on taking resources found in the world of Minecraft and refining
FDG2017, August 2017, Cape Cod, MA
Predicate
Meaning
tree N
There are N trees in the environment
wood_block N
Player has N wood blocks
cT
Player has crafting table
plank N
Player has N wood planks
stick N
Player has N sticks
cobble N
Player has N cobblestone pieces
…
…
pickaxe M
Player has a pickaxe of material M
sword M
Player has a sword of material M
…
…
Table 1: Predicates used in our model. Predicates used in
the paper are shown; additional predicates included in the
formal model are indicate with ellipses.
craftTable : plank (N + 4) -o cT * plank N.
Figure 2: Crafting a crafting table using the inventory menu
and its formalization in Ceptre.
them into materials for crafting items and tools. For the purpose
of this project, a distinction has been made between gathering and
mining. While both involve collecting resources, mining refers to
the materials that can only be collected when using a pickaxe, one
of the tools in the game. Due to this limitation, the gathering rules
only include the recipes needed to make a wood pickaxe and begin
mining.
An example of a gathering rule is one that will allow the player
to gather wood from a tree:
tree : wood (N+1) -o wood_block 1 * wood N.
How to parse this rule: the rule is named tree. The name is
separated from the rule itself with a colon (:). The rule’s LHS is
wood (N+1), meaning that for the rule to fire, there must be some
number of trees representable as N + 1 for some N . After the rule
fires, this predicate will be consumed, meaning there will no longer
be N + 1 trees. The rule’s RHS is wood_block 1 * wood N, which
is two predicates conjoined with tensor (*). The first predicate,
wood_block 1, means that firing the rule results in the player
having 1 wooden block. The second predicate wood N means that
after the rule fires, there will be N trees, which is 1 fewer than what
we started with.
At an early stage in the game, the player has access to a crafting
interface that is a 2x2 grid serving as the input for materials, and
a single square on the right for the recipe’s output. By filling the
grid with wooden planks, the player is able to produce a crafting
table block. Interacting with the crafting table is what provides
FDG2017, August 2017, Cape Cod, MA
access to the 3x3 crafting menu needed for more complex recipes,
such as tools. The rule for creating a crafting table is shown in both
Minecraft’s crafting interface and as a rule in Ceptre in Figure 2.
4.1.2 Rules for Tools and Mining. After creating a crafting table
and gaining access to the 3x3 interface, the program can move
on to tools and the materials needed to make them. The tools in
Minecraft are shovels, swords, axes, hoes, and pickaxes. All of these
recipes involve a combination of materials and sticks. The material
used to create a tool dictates the strength and durability of a tool.
Some materials may only be collected by tools of a certain strength,
so upgrading to a higher tier opens up new crafting opportunities
for the player. While Minecraft has a system in place where the
condition of a tool decreases after prolonged use until it breaks
and disappears, for the purposes of this project that feature is not
represented in the rules. If the program creates one of a tool at a
certain strength level, it does not need to make another of that same
material. The strength levels go from wood as the weakest, then
stone, iron/gold, and diamond is the strongest. Gold is a special
case in that it requires an iron pickaxe to be collected, but is not
stronger than iron. All of these materials must be mined using
a pickaxe of the tier directly below it, except for wood which is
already covered under gathering.
Ryan Alexander and Chris Martens
cobblestone when broken with a pickaxe. Cobblestone is used for
crafting stone tools, while stone cannot. Also, breaking stone without a pickaxe results in no cobble being dropped, which is why
wooden pickaxes must be crafted beforehand. However, a pickaxe
of any material can be used to mine stone, so the “M” signifies that
any pickaxe may be used as a predicate.
craftStonePickaxe : $cT * stick (N + 2) * cobble (F + 3)
-o pickaxe s * (stick N) * (cobble F).
Figure 4: Crafting a stone pickaxe in Minecraft and in Ceptre.
Once the player has stone, they are able to use it to make tools
like in Figure 4 where a stone pickaxe is being Crafted in Minecraft
and in Ceptre. The similarities between Figure 3 and Figure 4 are
evident both in the game and in the code. Any wooden planks in
the recipe were replaced by cobblestone. This basic pattern became
a key component in determining quest placement.
5
craftWoodPickaxe : $cT * stick (N + 2) * plank (F + 3)
-o pickaxe wood * (stick N) * (plank F).
TRACE ANALYSIS
In order to understand how the rules operate in relationship to
each other, the program was run both autonomously and interactively. These procedures provided insight into how the rules were
connected, allowing us to analyze the progression elements in the
gameplay. When generating these models, the main patterns we
Figure 3: The recipe for crafting a wooden pickaxe in
Minecraft and the Ceptre formalization of this action.
Pickaxes in Minecraft are crafted by placing three of a material
in the top row and two sticks in the remaining spots of the middle
column, as shown in Figure 3. The Ceptre representation of this
recipe is shown underneath the game screenshot. The rule consumes a crafting table, at least two sticks, and at least three wooden
planks, and it produces a wooden pickaxe and any remaining sticks
and planks, as well as the crafting table. The crafting table is not
destroyed during this process, so the symbol “$” is used to indicate
that the predicate appears on both sides of the rule. The term wood
is the material argument to the pickaxe predicate.
The player can then use a wooden pickaxe to mine stone, codified
by the following rule:
mineStone : $pickaxe M * stone (N + 1) * cobble C -o
cobble (C + 1) * stone N.
When given a pickaxe and at least one block of stone, this
mine/stone rule can be called to return the same pickaxe, one more
piece of cobblestone, and any remaining stone. The difference between cobblestone (represented by the “cobble” predicate) and stone
in Minecraft is that stone is found more commonly, but becomes
Figure 5: A portion of a trace graph. Predicate nodes are
shown as blue ellipses, while transition nodes are green rectangles. An edge from a predicate to a transition indicates
that the transition consumes the predicate, while an edge
from a transition to a predicate suggests that the transition
produces the predicate.
Deriving Quests from Open World Mechanics
FDG2017, August 2017, Cape Cod, MA
looked for were bottlenecks and feedback loops. The gameplay of
Minecraft, specifically the mining aspect, tends to be cyclical with
players repetitively making stronger tools and finding stronger
materials. Running the Ceptre model helped us identify key rules
that stratify this process into stages.
As more rules fire, the number of possible choices in the system
increases. Each run begins with the same rules being necessary,
specifically the rules for gathering wood and making the crafting
table and sticks—the crafting table recipe is an important node in
all play traces, because no tools may be crafted until the player has
made it. Such a node represents a bottleneck in the mechanics.
Once the program has these materials, rules for wooden tools
can be run, which opens access to the mining rules. The mining
rules must fire multiple times in order to have sufficient resources
to create a tool. The way that these rules are repeated hints towards
their significance in the player experience. This interactive model
represents the player experience through text, but is similarly timeconsuming. After the initial discovery of how important the mining
rules were and how the crafting table expands the player’s options,
we moved on to studying the trace graphs to generate larger batches
of data to compare.
Trace graphs are a visual representation of all rules selected
during a run of the program—see Figure 5. This trace graph shows
that the tree rule is dependent on itself, since having some number
of trees in the environment is both a requirement for the rule and
an outcome. We can also see that the numeric argument to the
rule decreases on each subsequent firing of the rule, and that the
other resources generated (the wood blocks) are used in other,
unrelated rules. These graphs allowed us to compare multiple
possible outcomes when the program is given the same initial
resources. These comparisons reinforced the results found through
the interactive testing and highlighted the significance of the mining
rules. The high likelihood that a player will perform these actions
multiple times during a playthrough indicates that objectives such
as “Mine X amount of stone” might be worthwhile. While these
quests tend to be viewed as unoriginal and tedious, if the value
for “X” is chosen based on the expected amount of times a player
will perform this action through the course of playing the game
naturally, it can potentially make a task that is tedious in general
more rewarding. Creating multiples of these trace graphs and
analyzing how many times certain rules are run can help determine
what an appropriate “X” value is for that rule.
6
STATIC ANALYSIS
We have identified two different methods for organizing rules that
separate segments of a trace into different phases base on static
analysis of the LHS for each rule. We generated a static dependency
graph between rules, where an edge between R 1 and R 2 indicates
that R 2 consumes predicates that R 1 produces. Then, we collapsed
rules with similar LHS together into a single node with a list of
possible alternate arguments. We then interpret the bottleneck
points and actions immediately following them as game events that
should be rewarded with achievements. The two factors that were
used to determine which LHSes were similar were the materials
represented as arguments to predicates and the quantities of raw
materials that a rule required, which corresponds to its tool type.
Figure 6: A visualization of the tool rules organized based
on the type of materials used. (A) marks where Minecraft
currently has achievements.
Material-Based Sorting. One way to sort these rules is based
on the type of material they require. All tools in Minecraft are a
combination of sticks and the material that determines the tools
strength. Figure 6 groups rules that have the same inputs together.
The wood section is all of the rules that take some amount of
sticks and wood as inputs, the stone rules have inputs of sticks
and stone and so on. Objectives based on this type of organization
FDG2017, August 2017, Cape Cod, MA
Ryan Alexander and Chris Martens
Figure 7: The tool rules categorized based on the type of tool that is generated. (A) marks where Minecraft currently has
achievements.
could reward players for unlocking improved versions of items
or abilities that were previously accessible. An example of this
is the Minecraft achievement ”Getting an Upgrade” that rewards
the player for acquiring a stone pickaxe, which in Figure 6 would
be the seen as moving from the wood group to the stone group.
These types of goals reward the player by enhancing an aspect
of the gameplay that they are already familiar with. In Minecraft,
the stone pickaxe performs the same basic mining function as the
wooden one, but does it faster and on resources that are too strong
to be collected by the wood pickaxe. This method of organizing is
useful for finding objectives that make the player stronger.
Quantity-Based Sorting. A different approach to organizing rules
is based on the number of materials that the rule takes. In Minecraft,
pickaxes are made with two sticks and three of another material,
while swords are one stick and two of a material. Focusing on the
format of the recipe instead of on the specific ingredients leads
to grouping tools that have similar functionality , but differing
strength. This method is shown in Figure 7 where rules are grouped
by the type of tool that is produced. There are achievements for
the player crafting their first hoe, sword, and pickaxe, which are all
marked in the first node of their respective groups. These achievements all represent the player unlocking a new element of gameplay.
Hoes, swords, and pickaxes are the critical tools for farming, combat,
and mining respectively. Categorizing rules based on the quantities
of inputs aids in the formation of objectives that reward the player
with new ways to play the game.
7
CONCLUSION
Through our formalization and analysis of the rules of Minecraft
in Ceptre, we have shown that it is possible to re-discover handauthored progression structures algorithmically from the openworld mechanics of the game. This process can help explain and
identify ideal achievement placement in open world settings. We
carried out two analyses: an informal examination of the play
traces resulting from autonomous execution of the rules, and a
formal analysis of the static, syntactic structure of the encoded rules
themselves. The algorithmic discoveries we have made about the
system dynamics match up with achievements that were manually
authored for Minecraft.
In future work, we hope to apply this process of deriving progression structures from mechanics to other games, particularly
those that do not already have manually-authored quests or achievements. We also intend to devise alternative achievement placement
strategies and compare them to the two sorting-based strategies.
The ultimate goal of such a research programme is to improve
the player’s sense of purpose in open world games by providing
more affordances for potential intermittent objectives, where those
objectives stem organically from the gameplay itself.
REFERENCES
[1] Calvin Ashmore and Michael Nitsche. 2007. The quest in a generated world.
In Proc. 2007 Digital Games Research Assoc.(DiGRA) Conference: Situated Play.
503–509.
[2] Joris Dormans. 2011. Integrating emergence and progression. In Think Design
Play: Proceedings of the 2011 Digital Games Research Association Conference,
Hilversum the Netherlands.
[3] Joris Dormans and Sander Bakkes. 2011. Generating missions and spaces for
adaptable play experiences. IEEE Transactions on Computational Intelligence and
AI in Games 3, 3 (2011), 216–228.
[4] Sean Duncan. 2011. Minecraft, beyond construction and survival. Well Played: a
journal on video games, value and meaning (2011).
[5] Clara Fernández-Vara and Alec Thomson. 2012. Procedural generation of narrative puzzles in adventure games: The puzzle-dice system. In Proceedings of the
The third workshop on Procedural Content Generation in Games. ACM, 12.
[6] Jesper Juul. 2005. Half-Real: Video Games between Real Rules and Fictional Worlds.
61–87 pages.
[7] Ahmed Khalifa, Diego Perez-Liebana, Simon M Lucas, and Julian Togelius. 2016.
General video game level generation. In Proceedings of the 2016 on Genetic and
Evolutionary Computation Conference. ACM, 253–259.
[8] Chris Martens. 2015. Ceptre: A Language for Modeling Generative Interactive
Systems. In Artificial Intelligence and Interactive Digital Entertainment.
[9] Adam M Smith, Erik Andersen, Michael Mateas, and Zoran Popović. 2012. A case
study of expressively constrainable level design automation tools for a puzzle
game. In Proceedings of the International Conference on the Foundations of Digital
Games. ACM, 156–163.
[10] Anne Sullivan, Michael Mateas, and Noah Wardrip-Fruin. 2012. Making quests
playable: Choices, CRPGs, and the Grail framework. Leonardo Electronic Almanac
17, 2 (2012).
| 2cs.AI
|
arXiv:1501.04094v2 [math.AG] 30 Sep 2015
ON THE EFFECTIVE CONE OF Pn BLOWN-UP AT n + 3
POINTS
MARIA CHIARA BRAMBILLA, OLIVIA DUMITRESCU, AND ELISA POSTINGHEL
Abstract. We compute the facets of the effective and movable cones of divisors on the blow-up of Pn at n + 3 points in general position. Given any
linear system of hypersurfaces of Pn based at n + 3 multiple points in general
position, we prove that the secant varieties to the rational normal curve of degree n passing through the points, as well as their joins with linear subspaces
spanned by some of the points, are cycles of the base locus and we compute
their multiplicity. We conjecture that a linear system with n + 3 points is
linearly special only if it contains such subvarieties in the base locus and we
give a new formula for the expected dimension.
1. Introduction
Let L = Ln,d (m1 , . . . , ms ) denote the linear system of hypersurfaces of degree d
in Pn passing through a collection of s points in general position with multiplicities
at least m1 , . . . , ms . A classical question is to compute the dimension of L. A
parameter count provides a lower bound: the (affine) virtual dimension of L is
denoted by
X
s
n + mi − 1
n+d
(1.1)
vdim(L) =
−
n
n
i=1
and the (affine) expected dimension of L is edim(L) = max(vdim(L), 0). If the
dimension of L is strictly greater that the expected dimension we say that L is
special.
The dimensionality problem, that is the classification of all special linear systems,
is still open in general, in spite of intensive investigation by many authors. In particular, the study of linear systems requires information on the effective cone Eff R (X)
of divisors on the blow-up X of Pn at the given points. Also the computation of
the effective cone is in general a difficult task.
Let us overview now some known results. In the planar case, the SegreHarbourne-Gimigliano-Hirschowitz Conjecture describes all effective special linear
systems. It predicts that a linear system in P2 is special only if it contains in its
base locus (−1)-curves. In particular, it conjectures the effective cone of divisors
on P2 blown-up at points (see [12, 23, 24, 25]). On the negative side, we mention
Nagata’s Conjecture that predicts the nef cone of linear systems in the blown-up
plane at general points. Even for the case of dimension two in spite of many partial
results, both conjectures are open in general (see [13]).
2010 Mathematics Subject Classification. Primary: 14C20. Secondary: 14J70, 14J26, 13D40.
The first author is partially supported by MIUR and INDAM. The second author is a member
of the Simion Stoilow Institute of Mathematics of the Romanian Academy. The third author is
supported by the Research Foundation - Flanders (FWO).
1
2
MARIA CHIARA BRAMBILLA, OLIVIA DUMITRESCU, AND ELISA POSTINGHEL
In the case of P3 , Laface and Ugaglia Conjecture states that a Cremona reduced
special linear system in P3 contains in its base locus either lines or the unique
quadric surface determined by nine general points (see e.g. [28, 29]). The base
locus lemma for the quadric in P3 is difficult; it is related to Nagata’s Conjecture
for ten points in P2 (see [6]). The degeneration technique introduced by Ciliberto
and Miranda (see e.g. [6, 14]) is a successful method in the study of interpolation
problems in higher dimensions.
In the case of Pn general results are rare and few things are known. The wellknown Alexander-Hirschowitz Theorem [1] classifies completely the case of double
points (see [7, 16, 17, 32] for more recent and simplified proofs). In general, besides
the computation of some sporadic examples and the formulation of conjectures
about the speciality (see e.g. [3, 4]), very little is known.
The important feature of special linear systems interpolating double points in Pn
is that every element is singular along a positive dimensional subvariety containing
the points. As conjectured by Ciliberto and Miranda (see [12, Conjecture 6.4]), it
is expected that this is the case also for higher multiplicities. So it is natural to
look at the base locus of special linear systems and try to understand the possible
connection with the speciality. We call obstructions the subvarieties that whenever
contained with multiplicity in the base locus of a linear system force it to be special.
In [5, 21] the authors started a systematic study of special linear systems from
this point of view and considered in particular the case of linear obstructions.
Taking into account the contribution of all the linear cycles of the base locus,
a new notion of expected dimension can be given, the linear expected dimension
ldim(L) (see Definition 2.2). We say that a system L is linearly special (resp.
linearly non-special ) if its dimension differs from (resp. equals) the linear expected
dimension. In other words a linear system is linearly special if its speciality cannot
be explained completely by linear obstructions. Any linear system with s ≤ n + 2
base points is linearly non-special. This was proved in [5, 21] by means of a complete
cohomological classification of strict transforms in subsequently blown-up spaces of
linear systems, see Section 2 for an account.
The first instance of a non-linear cycle of the base locus of a special linear
system is the rational normal curve of degree n through n + 3 general points of Pn .
The well-known Veronese Theorem (often referred to as the Castelnuovo Theorem)
tells us that there exists exactly one such a curve. In P2 an instance of this is
the unique conic through five points. In this article we focus on special linear
system obstructed by rational normal curves and related varieties. We prove first a
base locus lemma (Lemma 4.1) for linear systems with arbitrary number of general
points which describes the following non-linear cycles of the base locus: the rational
normal curve, its secant varieties and the cones over them. For instance, the fixed
cubic hypersurface of P4 interpolating 7 double points, that appears as one of the
exceptions in the Alexander-Hirschowitz theorem, is the variety of secant lines to
the rational normal curves given by the seven points.
We expect that when the multiplicity of containment in the base locus is high
enough with respect to the degree, those cycles forces the linear system to be
special. More precisely, we give a conjectural formula (Definition 6.1, Conjecture
6.4) for the dimension of linear systems based at n + 3 points. The formula in
Definition 6.1 takes into account the contribution of the linear cycles and also that
of the normal curves and related cycles in the base locus. In Section 6.2 we prove
ON THE EFFECTIVE CONE OF Pn BLOWN-UP AT n + 3 POINTS
3
that this conjecture holds for n = 2, 3 and for general n in a number of interesting
families of homogeneous linear systems.
As a consequence of our analysis, we deduce the main result of this paper, that
is an explicit description of all effective divisors in X, the blown-up Pn at n + 3
points. We give in Theorem 5.1 a list of inequalities that define the effective cone
of X, and we also describe the movable cone of X, Theorem 5.3.
We mention that a new approach to the dimensionality problem for s = n + 3
points was introduced in [35] and their analysis relies on sagbi bases. For s ≤ n + 3,
in [10, 31] it was proved that the blow-up of Pn at s points in general position
is a Mori dream space. In particular Castravet and Tevelev [10] gave the rays
of the effective cone, see Section 5.1 for more details. What is interesting is the
fact that Castravet and Tevelev’s extremal rays can be formulated in terms of
hypersurfaces that are either secant varieties to the rational normal curve through
the n + 3 points or their joins with linear subspaces spanned by the points, see
Section 3. In this paper we show that in fact both the effective cone (Theorem 5.1)
and movable cone (Theorem 5.3) of the blown-up space, and the dimensionality
problem, depend exclusively on these secant varieties seen as cycles of arbitrary
codimension in Pn . Moreover there is a bijection between the 2n+2 weights of the
half-spin representations of so2(n+3) and the generators of the Cox ring of the blowup of Pn at n + 3 points in general position (see [20, 10, 34]). In [34] in particular
this bijection interprets the latter space as a spinor variety. It would be interesting
to establish a dictionary that translates the language of secant varieties into that
of spinor varieties.
The article is organized as follows. In Section 2 we give an account on the notion
of linear speciality [5, 21].
In Section 3 we give a geometric description of the rational normal curves and
(cones over) their secants and we give an interpretation in terms of divisors of those
among them that are of codimension 1, by means of Cremona transformations of Pn .
In particular, these divisors are the Castravet-Tevelev rays generating the effective
cone.
In Section 4 we prove the base locus lemma for rational normal curves and related
cycles, Lemma 4.1.
In Section 5 we describe the effective and movable cones of X, Theorem 5.1 and
Theorem 5.1.
In Section 6 we introduce the new notion of expected dimension, σldim (Definition 6.1), and state our Conjecture 6.4, exhibiting a list of evidences in Section
6.2.
Acknowledgements. The authors would like to thank the Research Center FBKCIRM Trento for the hospitality and financial support during the stay for the
Summer School “An interdisciplinary approach to tensor decomposition” (Summer
2014) and during their one month “Research in Pairs” program (Winter 2015). We
are grateful to Ana-Maria Castravet for pointing out general aspects of this work
and further research directions and to Cinzia Casagrande for pointing out some
inaccuracies in a previous version of this paper. We thank the referee for his/her
helpful comments.
4
MARIA CHIARA BRAMBILLA, OLIVIA DUMITRESCU, AND ELISA POSTINGHEL
2. Linear speciality of linear systems
Given a linear system L, we say that its base locus contains a subvariety L with
multiplicity k if any hypersurface in L has multiplicity at least k along each point
of L.
Let L = Ln,d (m1 , . . . , ms ) be a non-empty linear system, let I(r) ⊆ {1, . . . , s}
be any multi-index of length |I(r)| = r + 1, for 0 ≤ r ≤ min(n, s) − 1 and denote
by LI(r) the unique r-linear cycle through the points pi , for i ∈ I(r). Set
X
mi − rd, 0 .
(2.1)
kI(r) = max
i∈I(r)
It is an easy consequence of Bézout’s Theorem that if kI(r) > 0 then all elements
of L vanish along LI(r) .
In [5] a (sharp) base locus lemma for linear cycles, that we will refer to as linear
base locus lemma, for linear systems with at most n + 2 points was proved and
later generalized in [21] to linear systems with arbitrary numbers of points. We
summarize the content of the two above mentioned results in the following
Lemma 2.1 ([21, Proposition 4.2]). For any non-empty linear system L with arbitrary number of points and for any 0 ≤ r ≤ n − 1, the multiplicity of containment
of the cycle LI(r) in the base locus of L is kI(r) .
When the order of vanishing is high, precisely when kI(r) > r, then LI(r) provides
obstruction to the non-speciality. This observation yields the following definition
of expected dimension.
Definition 2.2 ([5, Definition 3.2]). The (affine) linear virtual dimension of L is
the number
s−1
X
X
n + kI(r) − r − 1
.
(−1)r+1
(2.2)
n
r=−1
I(r)⊆{1,...,s}
where we set I(−1) = ∅ and kI(−1) = d.
The (affine) linear expected dimension of L, denoted by ldim(L), is 0 if L is
contained in a linear system whose linear virtual dimension is negative, otherwise
is the maximum between the linear virtual dimension of L and 0.
−r−1
In (2.2), the number (−1)r+1 n+kI(r)
computes the contribution of the
n
linear cycle LI(r) ∼
= Pr spanned by the points pij , ij ∈ I(r). If all the numbers kI(r)
are zero, the linear virtual dimension (2.2) equals the virtual dimension (1.1) of L.
Asking whether the dimension of a given linear system equals its linear expected
dimension is a refinement of the classical question of asking whether the dimension
equals the expected dimension. We say that a linear system L is linearly special if
dim(L) 6= ldim(L). On the other hand a linear system is called linearly non-special
(or only linearly obstructed) if its dimension equals the linear expected dimension.
We recall here, for the reader convenience, the following results on linearly speciality and effectiveness.
Theorem 2.3 ([5, Corollary 4.8, Theorem 5.3]). All non-empty linear systems with
s ≤ n + 2 points are linearly non-special.
ON THE EFFECTIVE CONE OF Pn BLOWN-UP AT n + 3 POINTS
5
Moreover, for s ≥ n + 3 let s(d) ≥ 0 is the number of points of multiplicity d. If
s
X
mi ≤ nd + min(n − s(d), s − n − 2), 1 ≤ mi ≤ d,
(2.3)
i=1
then L is linearly non-special.
Theorem 2.4 ([5, 8, 10]). If s ≤ n + 2, then L is non-empty if and only if
s
X
mi ≤ nd.
(2.4)
mi ≤ d, ∀i = 1 . . . , s,
i=1
Moreover if s ≥ n + 3 and (2.3) is satisfied, then L is non-empty.
Since Theorem 2.4 gives a statement about linear systems we assumed that
the coefficients d and mi are positive. In order to translate the statement into
the language of divisors on the blow-up of Pn at points we relax the positivity
assumption on the coefficients mi ’s and we obtain the following
Corollary 2.5. The effective cone of Pn blown-up at s ≤ n + 2 points in general
position is described by (2.4) and the inequalities
X
(2.5)
d ≥ 0,
mi ≤ nd, ∀I ⊆ {1, . . . , s}, |I| = n + 1.
i∈I
We remark that (2.3) is also sufficient condition for the base locus of L to not
contain any multiple rational normal curve. In Section 4 we will give a sharp base
locus lemma for the rational normal curve for all linear systems based at n+3 general
points. Moreover in Section 5 we will give necessary and sufficient conditions for a
linear system in Pn based at n + 3 general points to be non-empty.
2.1. Connection to the Fröberg-Iarrobino Conjecture. The problem of determining the dimension of linear systems with assigned multiple points is related
to the Fröberg-Iarrobino Weak and Strong Conjectures [22, 27], which give a predicted value for the Hilbert series of an ideal generated by s general powers of linear
forms in the polynomial ring with n + 1 variables. Such an ideal corresponds, via
apolarity, to the ideal of a collection of fat points, therefore it is possible to give a
geometric interpretation of this conjecture, as Chandler pointed out [18]. See also
[5, Sect. 6.1] for more details.
In terms of our Definition 2.2 the Weak Conjecture can be stated as follows: the
dimension of a homogeneous linear system, i.e. one for which all points have the
same multiplicity, is bounded below by its linear expected dimension.
Conjecture 2.6 (Weak Fröberg-Iarrobino Conjecture). The linear system L =
Ln,d (ms ) satisfies dim(L) ≥ ldim(L).
Moreover, the Strong Conjecture states that a homogeneous linear system is
always linearly non-special besides a list of exceptions.
Conjecture 2.7 (Strong Fröberg-Iarrobino Conjecture). The linear system L =
Ln,d (ms ) satisfies dim(L) = ldim(L) except perhaps when one of the following
conditions holds:
(i) s = n + 3;
(ii) s = n + 4;
(iii) n = 2 and s = 7 or s = 8;
6
MARIA CHIARA BRAMBILLA, OLIVIA DUMITRESCU, AND ELISA POSTINGHEL
(iv) n = 3, s = 9 and d ≥ 2m;
(v) n = 4, s = 14 and d = 2m, m = 2 or 3.
Notice that in [18] the formulation of case (iv) requires the condition d = 2m,
anyway it is known (see for example [6]) that for any degree d ≥ 2m the linear
system Ln,d (ms ) is linearly special for m high enough.
In this paper we conjecture that linear systems with n + 3 points are linearly
special only if they contain in their base locus the rational normal curve given by
the n + 3 points, its secant varieties or their joins with linear subspaces spanned
by subsets of the set of the n + 3 points (Conjecture 6.4). Moreover, we give a new
definition of expected dimension, the secant linear dimension σldim (see Definition
6.1), that provides a correction term for ldim. In particular, in the homogeneous
case this completes the Strong Fröberg-Iarrobino Conjecture.
Remark 2.8. It would be interesting to extract the Hilbert series of ideals generated
by n + 3 powers of linear forms from our formula of σldim, Definition 6.1.
2.2. General vision. We can interpret the base locus lemma for rational normal
curves and related cycles −that we prove in Section 4− and the definition of σldim
both as extensions of the results contained in our previous works [5, 21]: Lemma
2.1 and Definition 2.2.
We expect rational normal curves and related cycles to appear as obstruction
to the non-speciality also in the case of linear systems with s ≥ n + 3 base points.
A natural generalization of Conjecture 2.6 would be that σldim provides a lower
bound for the dimension of any general non-homogeneous linear system. We plan
to further investigate this problem. We chose to dedicate this work to the case of
n + 3 points because it is the first case where non-linear obstructions appear and
was not understood before for the general case of Pn .
3. Secant varieties to rational normal curves and Cremona
transformations
In this section we collect a series of well-known geometric aspects of secant
varieties to rational normal curves. The first important point is the following.
Theorem 3.1 (Veronese). There exists a unique rational normal curve of degree
n passing through n + 3 points in general position in Pn .
This theorem is classically known and its first proof is due to Veronese [37],
although it is often attributed to Castelnuovo.
In this section and throughout this paper we will adopt the following notation.
Let p1 , . . . , pn+3 ∈ Pn general points, let C be the rational normal curve of degree
n interpolating them and, for every t ≥ 1, let σt := σt (C) ⊂ Pn be the variety of
t-secant Pt−1 ’s to C. In this notation we have σ1 = C.
Rational normal curves are never secant defective; in particular we have the
following formula for the secant dimension:
dim(σt ) = min(n, 2t − 1).
Moreover rational normal curves are of minimal secant degree if 2t − 1 < n, see
[15]:
n−t+1
deg(σt ) =
.
t
ON THE EFFECTIVE CONE OF Pn BLOWN-UP AT n + 3 POINTS
7
Secant varieties are highly singular, in particular for t ≥ 2, 2t − 1 < n, we have
σt−1 ⊂ Sing(σt ). Moreover the multiplicity of σt along στ , for all 1 ≤ τ < t, satisfies
the following (see e.g. [15])):
n−t−τ +1
n−t
.
multC (σt ) =
,
multστ (σt ) ≥
t−τ
t−1
3.1. Cones over the secant varieties to the rational normal curve. In this
section, we consider cones over the σt with vertex spanned by a subset of the base
points. Let I ⊂ {1, . . . , n + 3} with |I| = r + 1. We use the conventions |∅| = 0 and
σ0 = ∅. Let us denote by
(3.1)
J(LI , σt )
the join of LI and σt .
Recall that σt = J(σt−1 , C) = J(σt−2 , σ2 ) etc. Notice also that J(LI , σt ) ⊂ σ|I|+t .
The dimensions of such joins can be easily computed:
(3.2)
rI,σt := dim(J(LI , σt )) = dim(LI ) + dim(σt ) + 1 = |I| + 2t − 1.
3.2. Divisorial cones. When J(LI , σt ) is a hypersurface, namely when rI,σt =
n − 1 that is I is such that |I| = n − 2t, we can characterize these cones as the
unique section of a certain linear system of hypersurfaces of Pn interpolating points
p1 , . . . , pn+3 with multiplicity.
We will denote by Ln,d (m1 , . . . , ms ) the linear system of degree−d hypersurfaces
of Pn interpolating the n + 3 points with multiplicity m1 , . . . , ms respectively.
We first discuss the case when σt is a hypersurface. Precisely, when n = 2t,
I = ∅, we have that σt is a degree (t + 1) hypersurface with multiplicity t along
C and in particular at the fixed points p1 , . . . , pn+3 . In this notation we have that
σt belongs to the the linear system L2t,t+1 (t2t+3 ). Moreover one can prove that
it is the only element satisfying the interpolation condition, see also Section 6.2.3
(Proposition 6.12). For instance for t = 1 one obtains the plane conic through five
points, L2,2 (15 ), for t = 2 one obtains L4,3 (27 ).
Remark 3.2. In Section 4 (Corollary 4.3), we will show that, when n = 2t, σt has
multiplicity exactly t − τ + 1 on στ , for all 1 ≤ τ < t.
Assume now that σt has higher codimension in Pn . Fix I such that |I| = n−2t ≥
1 and consider πI : Pn 99K P2t the projection from the linear subspace LI . Denote
by C ′ := πI (C) the projection of C and σt′ := πI (σt (C)) the projection of its tsecant variety. Then C ′ is a rational normal curve of degree 2t and σt′ = σt (C ′ ) is
the t-secant variety to C ′ . Hence the hypersurface J(LI , σt ) is the cone with vertex
the linear subspace LI over the secant variety σt′ .
We conclude that for any I such that |I| ≥ 0, the following formula holds:
(3.3)
J(LI , σt ) = Ln,t+1 ((t + 1)n−2t , t2t+3 ).
We mention that in [10, Theorem 2.7] the authors prove that divisors of the form
(3.3) are the rays of the effective cone Eff R (X) (see also Section 5.1).
3.3. The standard Cremona transformation. We recall that the standard Cremona transformation of Pn is the birational transformation defined by the following
rational map:
−1
Cr : (x0 : · · · : xn ) → (x−1
0 : · · · : xn ),
8
MARIA CHIARA BRAMBILLA, OLIVIA DUMITRESCU, AND ELISA POSTINGHEL
see e.g. [20]. Let L = Ln,d (m1 , . . . , ms ) be a linear system based on s points in
general position; we can assume, without loss of generality, that the first n + 1 are
the coordinate points. The map Cr can be seen as the morphism associated to the
linear system Ln,n ((n − 1)n+1 ). This induces an automorphism of the Picard group
of the n-dimensional space blown-up at s points by sending the strict transform of
Ln,d (m1 , . . . , ms ) into the strict transform of
Cr(L) := Ln,d−c (m1 − c, . . . , mn+1 − c, mn+2 , . . . , ms )
where
c := m1 + · · · + mn+1 − (n − 1)d.
We have the following equality
(3.4)
dim(L) = dim(Cr(L)).
If c ≤ 0 we will say that the linear system L is Cremona reduced.
Remark 3.3. One can check that the join divisor J(LI , σt ) is in the orbit of the Weyl
group of an exceptional divisor. To see this, order the multiplicities decreasingly
and apply the Cremona action to the first n + 1 points t + 1 times. Indeed,
c(Ln,t+1 ((t + 1)n−2t , t2t+3 )) = 1
therefore
Cr(Ln,t+1 ((t + 1)n−2t , t2t+3 )) = Ln,t (tn−2t+2 , (t − 1)2t+1 ).
This proves the claim since one can recursively replace t by t − 1 until t = 0.
4. Base locus lemma
In this section we give a sharp base locus lemma for the rational normal curve,
and (cones over) its secant varieties, that generalizes Lemma 2.1 from the case of
at most n + 2 points to the case of arbitrary number of points s.
If s ≥ n + 3, as in Section 3, we denote by C the unique rational normal curve
through any subset of n+3 points, say p1 , . . . , pn+3 , and by σt its t-th secant variety.
We denote by J(I, σt ) the join between any index set I ⊂ {1, . . . , s} and σt , and by
rI,σt its dimension. For s = n + 3, this notions coincide with the ones introduced
in (3.1) and (3.2).
To a linear system Ln,d (m1 , . . . , mn+3 , . . . , ms ) we associate the following integers:
(4.1)
kC :=
n+3
X
mi − nd,
i=1
(4.2)
kI,σt :=
X
mi + tkC − (|I| + t − 1)d.
i∈I
Notice that, by setting
M :=
n+3
X
mi ,
i=1
one can write
(4.3)
kI,σt = tM +
X
i∈I
mi − ((n + 1)t + |I| − 1)d.
ON THE EFFECTIVE CONE OF Pn BLOWN-UP AT n + 3 POINTS
9
Moreover, if in (4.2) we replace t = 0 we obtain
X
kI :=
mi − (|I| − 1)d
i∈I
(cfr. (2.1)); if |I| = 0 and t = 1 we obtain kC := k∅,σ1 = M − nd; if |I| = 0 we
obtain kσt := tkC − (t − 1)d.
The number kσt is the multiplicity of containment of a t-secant Pt−1 to C in
the base locus of L. This is a straightforward consequence of the linear base locus
lemma, knowing that kC is the multiplicity of containment of C. In the next
lemma we prove that in fact the whole σt is contained in the base locus with that
multiplicity.
Lemma 4.1 (Base locus lemma). Let L be an effective linear system with s base
points. In the same notation as above, let C be the rational normal curve given by
n + 3 of them, fix any I ⊂ {1, . . . , s} and t ≥ 0 such that rI,σt ≤ n − 1.
If kI,σt ≥ 1, then the cone J(LI , σt ) is contained in the base locus with exact
multiplicity kI,σt .
Proof. Since all of the results used in this proof hold for arbitrary number of points
s, it is enough to prove that statement for s = n + 3 and for the corresponding C.
If t = 0 then J(LI , σt ) = LI and the statement follows from Lemma 2.1.
Assume that I = ∅ and t = 1. Then J(LI , σt ) = C is the rational normal curve
through the n + 3 points. In [9, Theorem 4.1], the authors prove that performing
the Cremona transformation based at the first n + 1 base points of L, then C is
mapped to the line through the last two points pn+2 and pn+3 , that we may denote
by LI(1) . Let KC be the multiplicity of containment of C in L: one has KC ≥ kC by
Bézout’s Theorem. Observe that KC is also the multiplicity of containment of the
line LI(1) in Cr(L). We conclude by noticing that by the linear base locus lemma,
this is given by
!
n+1
X
mi = M − nd = kC .
KC = kI(1) = mn+2 + mn+3 − nd −
i=1
Assume that I = ∅ and t ≥ 2, 2t − 1 < n. The above parts imply that any secant
(t − 1)-plane spanned by t distinct points of C is contained in the base locus of L
with multiplicity exactly kσt . Moreover, since the multiplicity is semi-continuous,
it follows that all limits of t-secant (t − 1)-planes are contained in the base locus
with multiplicity at least kσt . Hence the secant variety σt has multiplicity kσt .
Finally, the case I 6= ∅, t ≥ 1 follows from the above. Indeed every line L in
J(LI , σt ) connecting a point of the vertex LI and a point of the base σt , is contained
in the base locus of L with multiplicity kI + kσt − d.
Remark 4.2. Notice that the effectivity of L impliesP
the following inequality kC ≤
mi ≤ d, for all i. Indeed if kC > mi for some i, then j6=i mj > nd, a contradiction
by Theorem 5.1. This in particular implies kσ|I|+t ≤ kI,σt ≤ kI . Moreover the
P
obvious equality kI,σt = i∈I mi − |I|d + kσt and the effectivity condition mi ≤ d
imply kI,σt ≤ kσt .
Because of the containment relations LI , σt ⊆ J(LI , σt ) ⊆ σ|I|+t , the above
inequalities read as: if LI or σt is not in the base locus of L, neither is J(LI , σt )
nor σ|I|+t ; if J(LI , σt ) is not contained in the base locus, neither is σ|I|+t .
10
MARIA CHIARA BRAMBILLA, OLIVIA DUMITRESCU, AND ELISA POSTINGHEL
4.1. Geometric consequences of the base locus lemma. An immediate consequence of the base locus lemma is a description of the singularities of the secant
variety, whenever this is a hypersurfaces. Indeed since σt is the unique element of
the linear system L2t,t+1 (t2t+3 ), one can compute the multiplicity along the lower
order secant varieties.
Corollary 4.3. Let n = 2t and 1 ≤ τ ≤ t. Then σt is singular with multiplicity
t − τ + 1 on στ \ στ −1
Another consequence of Lemma 4.1 is the following result that in particular
implies that Cremona reduced linear systems are movable.
Corollary 4.4. Let L be an effective linear system with arbitrary number of points.
Assume that L is Cremona reduced. Then L does not contain any divisorial component of type J(I(n − 2t − 1), σt ) in its base locus.
Proof. Write n = 2l + ǫ, with ǫ ∈ {0, 1}. By Lemma 4.1, it is enough to prove that
kI(n−2t−1),σt ≤ 0 for all 0 ≤ t ≤ l + ǫ.
Since L is Cremona reduced, the hyperplane spanned by the collection of points
parametrized by I(n− 1) is not contained in the base locus, for any I(n− 1). Indeed
if I(n − 1) ⊂ I(n), for some I(n), we have
X
mi − (n − 1)d ≤ 0.
kI(n−1) <
i∈I(n)
This proves the statement for t = 0.
Assume 1 ≤ t ≤ l + ǫ. For any fixed index set I := I(n − 2t − 1) of cardinality
n−2t, choose 2t distinct indices in its complement: {i1 , . . . , i2t } ⊂ {1, . . . , n+3}\I.
We have
X
kI,σt = tM +
mi − (t + 1)(n − 1)d
i∈I
=
s
X
j=1
M − mij − mij+t
2t
X
X
mij +
− (n − 1)d +
mi − (n − 1)d < 0
j=1
i∈I
The first t terms are negative by assumption, the last is strictly negative because
of the hyperplane case t = 0.
Remark 4.5. In Section 5.1 we will see that effective divisors in the blown-up Pn at
n + 3 general points without fixed components of type J(I(n − 2t − 1), σt ), namely
those satisfying kI(n−2t−1),σt ≤ 0, are movable. Hence the cone of Cremona reduced
effective divisors, that is polyhedral since defined by inequalities, is contained in
the movable cone, that is in turn contained in the effective cone.
5. Effective and movable cones
In this section we will give necessary and sufficient conditions for linear systems
Ln,d (m1 , . . . , mn+3 ) in Pn with n + 3 base points in general position to have at
least one section. This is equivalent to an effectivity theorem for divisors on X, the
blown-up Pn at n + 3 general points, and provides a generalization of Theorem 2.4
and Corollary 2.5.
Throughout this section, we will use the same notation introduced in Section
4. Moreover we let N1 (X) denote the Neron-Severi group of X with coordinates
ON THE EFFECTIVE CONE OF Pn BLOWN-UP AT n + 3 POINTS
11
(d, m1 , . . . , mn+3 ) corresponding to the hyperplane class and the classes of the exceptional divisors.
Theorem 5.1 (Effectivity Theorem).
(I) For n ≥ 2, a linear system L =
Ln,d (m1 , . . . , mn+3 ) is non-empty if and only if
mi ≤ d,
(An )
∀i = 1, . . . , n + 3,
M − mi ≤ nd,
(Bn )
∀i = 1, . . . , n + 3,
kI,σt ≤ 0,
(Cn,t )
∀|I| = n − 2t + 1, 1 ≤ t ≤ l + ǫ,
where n = 2l + ǫ, ǫ ∈ {0, 1}.
(II) The facets of the effective cone of divisors Eff R (X) are given by the equalities in (An ), (Bn ), and (Cn,t ), with −1 ≤ t ≤ l + ǫ.
Remark 5.2. In Section 4 the numbers kI,σt (4.2) appeared as the multiplicities of
containment in the base locus, of cycles of codimension at least one, namely for
1 ≤ t ≤ l + ǫ and 0 ≤ |I| ≤ n − 2t − 1. In this section, the numbers kI(n−2t),σt
appearing in (Cn,t ) (Theorem 5.1) are formal generalizations of the above to the zero
n
codimensional case: J(I(n− 2t),
Psσt ) = P . In analogy with the cases s = n+ 1, n+ 2
points, where the condition i=1 mi ≤ nd corresponds to asking that the linear
span of all of the points − that is the whole space Pn − is not in the base locus (see
[21, Theorem 1.6]), here we ask that no n-dimensional virtual cycle J(I(n − 2t), σt )
is in the base locus of linear systems with n + 3 points.
The proof of Theorem 5.1 is by induction on n. For this reason we think it is
convenient to treat the initial case, n = 2, separately. Here the conditions of part
(I) read as follows.
(A2 )
mi ≤ d,
∀i = 1, . . . , 5,
(B2 )
M − mi ≤ 2d,
∀i = 1, . . . , 5,
(C2,1 )
M + mi ≤ 3d,
∀i = 1, . . . , 5,
Proof of Theorem 5.1, part (I), case n = 2. Without loss of generality we may assume m1 ≥ m2 ≥ · · · ≥ m5 ≥ 1. It is enough to prove that L is non-empty if and
only if
m1 ≤ d,
m1 + m2 + m3 + m4 ≤ 2d,
2m1 + m2 + m3 + m4 + m5 ≤ 3d.
If L is non-empty, then obviously m1 ≤ d. Moreover since L2,d (m1 , . . . , m4 ) is
non-empty, then m1 + m2 + m3 + m4 ≤ 2d.
P
To prove the third inequality assume first that kC = j=1 mj − 2d ≤ 0. Then
2m1 + m2 + m3 + m4 + m5 − 3d = kC + m1 − d ≤ 0. If kC ≥ 1, then by Lemma
′
4.1, the conic C isP
a fixed component of L and the residual
P5part has degree d :=
5
′
d − 2kC = 5d − 2 j=1 mj and multiplicities mi = 2d − j=1 mj + mi at the 5
points. Notice that m′i ≥ 0 by the second inequality. Effectivity implies d′ ≥ m′1
and this is equivalent to the third inequality.
P5
We now prove the other implication. If kC =
j=1 mj − 2d ≤ 0 then L is
non-empty by Theorem 2.4. Assume kC ≥ 1. By Lemma 4.1 the conic through
the five points is contained in the base locus with multiplicity kC . Notice that
m1 + m2 + m3 + m4 ≤ 2d implies kC ≤ m5 . The residual is L′ with d′ = d − 2kC =
12
MARIA CHIARA BRAMBILLA, OLIVIA DUMITRESCU, AND ELISA POSTINGHEL
P5
P5
5d − 2 j=1 mj and m′i = mi − kC = 2d − j=1 mj + mi ≥ 0, for all i = 1, . . . , 5.
P5
Obviously j=1 m′j − 2d′ = 0. We claim d′ ≥ m′i , for all i = 1, . . . , 5. Hence L′ is
P5
effective. To prove the claim for i = 1, notice that d′ − m′1 = 5d − 2 j=1 mj −
P5
P5
2d + j=1 mj − m1 = 3d − j=1 mj − m1 ≥ 0.
We now complete the proof of the first part of effectivity theorem for n ≥ 3.
Proof of Theorem 5.1, part (I), n ≥ 3. Without loss of generality, we may reorder
the points so that m1 ≥ · · · ≥ mn+3 . If mn+3 = 0, the set of conditions (Bn )
Pn+2
becomes just j=1 mj ≤ nd and the third set of conditions, (Cn,t ) is redundant.
In this case the result was proved in [5, Lemma 2.2], see Theorem 2.4. Hence we
will assume mn+3 ≥ 1.
“Only if ” implication:
If L is effective then (An ) and (Bn ) trivially hold.
The expanded expressions of condition (Cn,t ) is
X
kI,σt = tM +
mi − ((t + 1)n − t)d ≤ 0,
i∈I
n−1
for all 1 ≤ t ≤ 2 = l + ǫ and all multi-index I = I(n − 2t), see (4.3). Fix such
t and I = I(n − 2t). Take any j ∈ I(n − 2t) and denote by I \ {j} its complement
in I. Write
X
mi − ((t + 1)n − (t + 1))d.
kI\{j},σt = tM +
i∈I\{j}
In order to prove the inequality (Cn,t ), we consider the following cases.
Case (1). Assume that kI\{j},σt ≤ 0. Since kI,σt = kI\{j},σt + (mj − d), we
conclude by (An ).
Case (2). Assume that kI\{j},σt ≥ 1. By Lemma 4.1, kI\{j},σt is the
multiplicity of containment of the cone J(I \ {j}, σt ), that is the hypersurface
Ln,t+1 ((t + 1)n−2t , t2t+3 ), see (3.3). The residual of L after its removal is still effective by assumption. Let us denote by d′ and by m′j the degree and the multiplicity at
the point pj of the residual, that is d′ = d−(t+1)kI\{j},σt and m′j = mj −tkI\{j},σt .
Effectivity implies that d′ ≥ m′j ≥ 0. We conclude by noticing that d′ − m′j ≥ 0 is
equivalent to kI,σt ≤ 0.
“If ” implication:
The proof is by induction on n, with initial step the case n = 2 for which the
statement is already proved to hold. Assume the statement true for n − 1.
We will construct recursively an element that belongs to L, hence proving nonemptiness. We will treat the following cases and subcases separately.
(0) kC ≤ 0.
(1) kC ≥ 1 and m1 = d.
(2) kC ≥ 1 and m1 ≤ d − 1,
(2.a) L is Cremona reduced,
(2.b) L is not Cremona reduced.
ON THE EFFECTIVE CONE OF Pn BLOWN-UP AT n + 3 POINTS
13
Case (0). In this case L is effective by Theorem 2.4.
Case (1). Notice that the elements of L, cones with vertex at p1 , are in bijection
with the elements of a linear system L′ = Ln−1,d (m2 , . . . , mn+3 ). One can check
easily that L′ satisfies conditions (An−1 ), (Bn−1 ), being those implied by (An ),
(Bn ) respectively. Moreover for 1 ≤ t ≤ l and any index set I = I(n − 2t) such that
1 ∈ I, condition (Cn,t ) implies condition (Cn−1,t ), for the index set I ′ = I \ {1} =
I(n − 1 − 2t). Indeed we have
X
kI ′ ,σt = tM ′ +
m′i − ((t + 1)(n − 1) − t)d′
i∈I ′
= tM +
X
mi − (t + 1)m1 − ((t + 1)n − t)d + (t + 1)d
i∈I
= kI,σt ≤ 0.
′
Since L is effective, then L is.
Case (2.a). Notice that in this case mn+3 ≥ 2. Set I := {1, . . . , n − 2} and
consider the cone J(I, C) over the rational normal curve C with vertex the linear
subspaces LI spanned by the first n−2 points. As in (3.3), J(I, C) can be interpreted
as the fixed divisor Ln,2 (2n−2 , 15 ). Let us denote by L′ the kernel of the restriction
map L → L|J(I,C) . We can write
L′ = Ln,d′ (m′1 , . . . , m′n+3 ) := Ln,d−2 (m1 −2, . . . , mn−2 −2, mn−1 −1, . . . , mn+3 −1),
and the inequality dim(L) ≥ dim(L′ ) is satisfied.
Notice that if mn−2 = 2, i.e. m′n−2 = 0, then L′ is based on at most n + 2 points.
In this case we have
n+3
X
m′i − nd′ = (M − 2s − 1) − n(d − 2)
i=1
= (M − m1 − mn−2 − (n − 1)d) + (mn−2 + m1 − d − 1) ≤ 0.
The inequality follows by the fact that L is Cremona reduced and that m1 ≤ d − 1.
One concludes by noticing that L′ falls into case (0).
Otherwise, if mn−2 ≥ 3, since we also have mn+3 ≥ 2, i.e. m′n+3 ≥ 1, then L′
is based on n + 3 points. We claim that such a L′ satisfies conditions (An ), (Bn ),
′
(Cn,t ). Moreover kC
= kC − 1, namely the multiplicity of containment of C in the
′
′
base locus of L has decreased by one. If kC
= 0 we conclude by case (0), otherwise
we proceed with cases (1) or (2).
We are now left with showing the claim. One can easily check that the first two
conditions are satisfies, because of the assumption mi ≤ d − 1. In order to prove
that the third set of conditions, (Cn,t ), is also satisfied for any set I = I(n − 2t),
n ≥ 2t − 1, notice that
X
X
m′i ≤
mi − (n − 2t + 1 + f ),
i∈I
i∈I
where f is the cardinality of the index set I ∩ {1, . . . , n − 2}. From this, it follows
that
X
X
tM ′ +
m′i − ((t + 1)n − t)d′ ≤ tM +
mi − ((t + 1)n − t)d + (n − t − f − 1).
i∈I
i∈I
14
MARIA CHIARA BRAMBILLA, OLIVIA DUMITRESCU, AND ELISA POSTINGHEL
Now, choose 2t distinct indices {i1 , . . . , i2t } ⊂ {1, . . . , n + 3} \ I; the right hand side
of the above expression equals
t
X
j=1
where
α :=
M − mij − mij+t − (n − 1)d + α,
2t
X
mii +
j=1
X
i∈I
mi − nd + (n − t − f − 1).
Here we introduce the integer α for the sake of simplicity as we will treat different
cases in what follows. Notice that because of the assumption that L is Cremona
reduced, in order to conclude it is enough to prove that α ≤ 0.
Assume d ≥ n − t − 1. We have
2t
X
X
α=
mij +
mi − (n − 1)d + (n − t − f − 1 − d) ≤ 0,
j=1
i∈I
where the inequality follows from the fact that L is Cremona reduced.
If d ≤ n − t − 2, using mi ≤ d − 1 we obtain
α ≤ (n + 1)(d − 1) − nd + (n − t − f − 1) = n − 2t − 4 − f ≤ 0,
where the last inequality is implied by the fact that f ≥ min{0, n − 2t − 4}.
Case (2.b). Assume that L is not Cremona reduced, namely that
c :=
n+1
X
mi − (n − 1)d ≥ 1
i=1
and write
L′ := Cr(L) = Ln,d′ (m′1 , . . . , m′n+3 ).
We have dim(L) = dim(L′ ) by (3.4). We claim that L′ satisfies conditions (An ),
(Bn ), (Cn,t ). Hence we can reiterate the entire procedure for L′ , hence reducing the
proof of the effectivity of L to the proof of the effectivity of its Cremona transform
L′ .
We now prove the claim. We refer to conditions (An ), (Bn ) and (Cn,t ) for L′ as
(An )′ , (Bn )′ and (Cn,t )′ .
Notice that m′i ≤ d′ if and only if mi ≤ d, for all i ≤ n + 1. Moreover m′i ≤ d′ is
equivalent to (Bn ), for i = n + 2, n + 3.
One can easily check that (Bn ) implies (Bn )′ .
We now prove that (Cn,t )′ is satisfied for any t and any index set I = I(n − 2t),
with n ≥ 2t − 1. The expanded expression of condition (Cn,t )′ is
X
tM ′ +
m′i − ((t + 1)n − t)d′ ≤ 0.
i∈I
Assume that I ⊂ {1, . . . , n + 1}. The left-hand side of the above expression equals
X
tM +
mi − ((t + 1)n − t)d − c
i∈I
that is negative because c ≥ 1 and (Cn,t ) is satisfied.
Assume that |I \ {1, . . . , n + 1}| = 1, that occurs only if n ≥ 2t. One can easily
verify that (Cn,t )′ is equivalent to (Cn,t ).
ON THE EFFECTIVE CONE OF Pn BLOWN-UP AT n + 3 POINTS
15
Assume that |I \ {1, . . . , n + 1}| = 2, that occurs only when n ≥ 2t + 1. The
left-hand side of the expanded expression of condition (Cn,t )′ equals
X
X
mi − ((t + 2)n − (t + 1))d
tM +
mi − ((t + 1)n − t)d + c = (t + 1)M +
i∈I
i∈I(n−2t−2)
and this is bounded above by zero by (Cn,t+1 ).
Proof of Theorem 5.1, part (II). Notice that the arithmetic condition (Cn,−1 ) corresponds to d ≥ 0 and (Cn,0 ) corresponds to M −mi −mj ≤ nd. When multiplicities
mi are positive, condition (Cn,0 ) is redundant and the statement was proved in part
(I).
We assume now m1 ≥ · · · ≥ mn+3 and mn+3 < 0. Then conditions (Cn,−1 ) and
(Cn,0 ) imply (Cn,t ) for t ≥ 1. Moreover the non-emptiness of L is equivalent to the
non-emptiness of Ln,d (m1 , . . . , mn+2 ). This holds because every effective divisor D
in L decomposes as D = −mn+3 En+3 + (D + mn+3 En+3 ). The statement follows
now from the description of the effective cone of divisors on the blown-up Pn at
s ≤ n + 2 points, that is given by (An ), (Bn ), (Cn,−1 ) and (Cn,0 ), see Corollary
2.5.
5.1. Movable Cone of Divisors. Mori dream spaces were introduced by Hu and
Keel ([26], see also [2]), we give now an alternative definition. Let X be a normal
Q-factorial variety whose Picard group, Pic(X), is a lattice. Define the Cox ring of
X as
M
H 0 (X, D)
Cox(X) :=
D∈Pic(X)
with multiplicative structure defined by a choice of divisors whose classes form a
basis for the Picard group Pic(X). We say that X is a Mori dream space if the Cox
ring, Cox(X), is finitely generated.
We define the movable cone of a variety, Mov(X), to be the cone generated
by divisors without divisorial base locus. For a Mori dream space, the movable
cone and the effective cone of divisors are polyhedral. Moreover, the movable cone
decomposes into disjoint union of nef chambers that are the nef cones of all small
Q-factorial modifications.
Let X now denote the blow-up of the projective space Pn at s ≤ n + 3 points in
general position. We first recall that X is a Mori dream space [10] (see also [31]).
Moreover results of [8, 10] imply that the effective cone Eff R (X) is generated as a
cone and semigroup by divisors in the Weyl orbit W · Ei . Here, W represents the
Weyl group of X and W · Ei the orbit with respect to its action on an exceptional
divisor. Recall that every element of W corresponds to a birational map of Pn
lying in the group generated by projective automorphisms and standard Cremona
transformations of Pn . Also,
Pic(X) = hH, E1 , E2 , . . . , Es i.
Following [30] we introduce a symmetric bilinear on Pic(X) acting on its generators
as:
hEi , Ej i = −δi,j , hEi , Hi = 0, hH, Hi = n − 1.
(5.1)
∨
Let Eff R (X) denote the dual cone of the cone of effective divisors with real
coefficients. Namely, the dual cone Eff R (X)∨ consists of divisors D such that
16
MARIA CHIARA BRAMBILLA, OLIVIA DUMITRESCU, AND ELISA POSTINGHEL
hD, F i ≥ 0 with all F ∈ Eff R (X). One can define the degree of a divisor D ∈ Pic(X)
as follows (see [10]):
1
hD, −KX i,
deg(D) :=
n−1
where KX denotes the canonical divisor of the blown-up projective space, X. For
s ≤ n + 3 the movable cone can be described as the intersection between the cone
of effective divisors with real coefficients and its dual
Mov(X) = Eff R (X) ∩ Eff R (X)∨ ,
(5.2)
(see [8, Theorem 4.7]).
The facets of the effective cone Eff R (X) for s ≤ n + 2 are given in Corollary 2.5.
As an application of this result and using (5.2), one can easily describe the facets
of the movable cone Mov(X) for s ≤ n + 2 (see [8]).
We will extend this description to the case with s = n + 3 points. The facets of
the effective cone Eff R (X) of the blown-up Pn at n + 3 points in general position
are computed in Theorem 5.1. Moreover the generators of the effective cone are
described in [10] by the classes of divisors in the set
)
(
X
X
Ei : |I| = n − 2t, −1 ≤ t ≤ l + ǫ ,
(5.3) A = (t + 1)H − (t + 1)
Ei − t
i∈I
i∈I
/
where n = 2l + ǫ, ǫ ∈ {0, 1}. Notice that the divisors in A are the only divisors
on X of degree 1 and are the strict transforms of the one-section linear systems
described in (3.3); these are precisely the divisors that will appear in Conjecture
6.4.
We can now describe the movable cone of divisors on the blown-up Pn at the
collection of n + 3 points. They are the effective divisors on X for which the
corresponding linear system has no fixed divisorial component of type A.
Theorem 5.3. For n ≥ 2, letting (d, m1 , . . . , mn+3 ) be the coordinates of the
Neron-Severi group, N1 (X), then the movable cone Mov(X) is generated by the
inequalities (An ) and (Bn ), of Theorem 5.1 and
(Dn,t )
kI,σt ≤ 0,
∀|I| = n − 2t, −1 ≤ t ≤ l + ǫ.
Proof. It follows from Theorem 5.1, [8, Theorem 4.7] and [10, Theorem 2.7]. Indeed
a divisor in N1 (X) of the form
D = dH −
n+3
X
mi Ei
i=1
with d, mi ≥ 0 lies in Eff R (X)∨ if and only if it has non-negative intersection number
(5.1) with all elements of the generating set A described in (5.3). We leave it to
the reader to verify that these conditions are equivalent to the set of inequalities
(Dn,t ). Notice also that the conditions (Cn,t ) in Theorem 5.1 are redundant. One
concludes the proof by using (5.2).
5.2. Faces of the movable cone and contractions. From Mori theory it follows
that the faces of the movable cone are in one to one correspondence with classes of
divisorial and fibre type contractions from small Q-factorial modifications of X to
normal projective varieties.
ON THE EFFECTIVE CONE OF Pn BLOWN-UP AT n + 3 POINTS
17
In particular, contractions given by divisors in the boundary of the effective cone,
corresponding to the first three sets of equalities, namely (An ) and (Bn ), are of fibre
type contractions (i.e. projections to lower dimensional Mori dream spaces), while
contractions associated to the last set of equalities, namely (Dn,t ), corresponding
to the boundary of the dual effective cone, are divisorial contractions.
6. A new notion of expected dimension
Secant varieties to the rational normal curve and cones over them are a natural
generalization of the linear obstructions. In this section we introduce a new notion
of expected dimension for linear systems with n + 3 points in general position, Definition 6.1, that takes into account their contributions. Furthermore we conjecture
that those are the only non-linear obstructions, see Conjecture 6.4.
In Section 6.2 we prove this conjecture for n ≤ 3 and for some homogeneous
linear systems in families.
We adopt the same notation as in the previous sections (3.2) and (4.2). We
recall here that the join J(I, σt ) has dimension rI,σt ≤ n − 1 whenever 0 ≤ t ≤ l + ǫ,
n = 2l + ǫ and 0 ≤ |I| ≤ n − 2t.
Definition 6.1. Let L = Ln,d (m1 , . . . , mn+3 ) be a linear system. The (affine)
secant linear virtual dimension of L is the number
X
|I| n + kI,σt − rI,σt − 1
(6.1)
,
(−1)
n
I,σt
where the sum ranges over all indexes I ⊂ {1, . . . , n + 3} and t such that 0 ≤
t ≤ l + ǫ, n = 2l + ǫ and 0 ≤ |I| ≤ n − 2t.
The (affine) secant linear expected dimension of L, denoted by σldim(L) is defined as follows: if the linear system L is contained in a linear system whose secant
linear virtual dimension is negative, then we set σldim(L) = 0, otherwise we define
σldim(L) to be the maximum between the secant linear virtual dimension of L and
0.
Remark 6.2. Using the base locus lemma (Lemma 4), one may generalise formula
(6.1) for arbitrary number of points, by taking into account all of the rational
normal curves of degree n given by sets of n + 3 points (and related cycles).
Remark 6.3. One can easily verify that kI,σt ≤ rI,σt , so its corresponding Newton
binomial in (6.1) is zero. In particular one can check that this inequality is satisfied
for all I and t when kC ≤ 1, namely when the rational normal curve C is contained
in the base locus of L at most simply. In all of these cases σldim(L) = ldim(L).
Conjecture 6.4. Let L be a non-empty linear system of Pn with n + 3 base points
in general position and let C be the rational normal curve through the base points.
Then L is special only if its base locus contains either linear cycles, or cones over
the secant varieties σt of C. Moreover, we have dim(L) = σldim(L).
We illustrate this idea in the following examples.
Example 6.5. The linear system L = L6,8 (69 ) is linearly special, since dim(L) = 1
and ldim(L) = −147. The rational curve C, given by the 9 base points, is contained
in the singular locus of the fixed hypersurface L with multiplicity kC = 6. Moreover,
for each of the 9 base points, say p, the cone J(p, C) as well as σ2 are contained with
18
MARIA CHIARA BRAMBILLA, OLIVIA DUMITRESCU, AND ELISA POSTINGHEL
multiplicity 4 in the singular locus of L, by Lemma 4.1. Hence one can compute
σldim = 1.
Example 6.6. Consider the linear system L4,10 (9, 73 , 53 ). The rational normal
curve is contained 5 times and the cone J(p1 , C) is contained with multiplicity 4
We leave it to the reader to verify that σldim(L) = 2 and that dim(L) = 2, the last
equality following a series of Cremona transformations (see Section 3.3).
6.1. Properties of σldim. In this section we prove two technical lemma which
will be useful in the sequel.
Recall that a linear system L = Ln,d (d, m2 , . . . , ms ) has the same dimension
of the linear system Ln−1,d (m2 , . . . , ms ). We will call the second system the cone
reduction of L and we denote it by Cone(L).
Lemma 6.7. The secant linear expected dimension of a linear system L =
Ln,d (d, m2 , . . . , mn+3 ) is invariant under cone reduction:
σldim(L) = σldim(Cone(L)).
Proof. Let Cone(L) = Ln−1,d (m2 , . . . , mn+3 ) be the cone reduction of L.
We write the formula (6.1) for σldim(L) as follows:
X
σldim(L) =
BL (J(LI , σt ))
I,t
denoting by BL (J(LI , σt )) the contribution in the sum given by the cycle J(LI , σt )
that is
n + kI,σt − rI,σt − 1
.
BL (J(LI , σt )) := (−1)|I|
n
Now, recalling the formula ab = a−1
+ a−1
b
b−1 , it is easy to check that, for any
I ⊆ {2, . . . , n + 3}, one has
BL (J(LI , σt )) + BL (J(LI∪{1} , σt )) = BCone(L) (J(LI , σt )).
Lemma 6.8. The secant linear expected dimension of a linear system L =
Ln,d (m1 , . . . , mn+3 ) is invariant under Cremona transformations:
σldim(L) = σldim(Cr(L)).
Proof. Assume that
c=
n+1
X
mi − (n − 1)d ≥ 1
i=1
and let
Cr(L) = Ln,d−c (m1 − c, . . . , mn+1 − c, mn+2 , mn+3 ).
be the system obtained after the Cremona transformation, see Section 3.3.
First of all consider the linear system obtained from L forgetting the last two
e = Ln,d−c (m1 − c, . . . , mn+1 − c) be
points: Le = Ln,d (m1 , . . . , mn+1 ) and let Cr(L)
e are linearly non-special
the corresponding Cremona transform. Since Le and Cr(L)
by Theorem 2.3 and a Cremona transformation preserves the dimension of a linear
system (see (3.4)) we have:
e = dim(L)
e = dim(Cr(L))
e = ldim(Cr(L)).
e
ldim(L)
ON THE EFFECTIVE CONE OF Pn BLOWN-UP AT n + 3 POINTS
19
Using the same notation as in the proof of Lemma 6.7, one can split the sum as
follows:
X
e +
BL (J(LI , σt ))
σldim(L) = ldim(L)
|I∩{n+2,n+3}|=1
X
+
BL (J(LI , σt ))
|I∩{n+2,n+3}|=2
X
+
BL (J(LI , σt )),
t≥1,I∩{n+2,n+3}=∅
and similarly
e +
σldim(Cr(L)) = ldim(Cr(L))
+
X
BCr(L) (J(LI , σt ))
|I∩{n+2,n+3}|=1
X
BCr(L) (J(LI , σt ))
|I∩{n+2,n+3}|=2
+
X
BCr(L) (J(LI , σt )).
t≥1,I∩{n+2,n+3}=∅
Now it is not difficult to check that σldim(Cr(L)) = σldim(L). Indeed first of all
we note that
(6.2)
c
kC
=
n+3
X
mi − c(n + 1) − n(d − c) = kC − c = mn+2 + mn+3 − d,
i=1
c
where we denote by kC (resp. kC
) the multiplicity of containment of C in the base
locus of L (resp. of Cr(L)).
c
Let us also denote by kI,σt (resp. kI,σ
) the multiplicity of containment of
t
J(LI , σt ) in the base locus of L (resp. of Cr(L)). We leave it to the reader to
check by using (6.2) that the following holds.
c
If |I ∩ {n + 2, n + 3}| = 1, then kI,σ
= kI,σt , hence BCr(L) (J(LI , σt )) =
t
BL (J(LI , σt )).
c
If |I ∩{n+2, n+3}| = 2, then kI,σ
= kI\{n+2,n+3},t+1 , hence BCr(L) (J(LI , σt )) =
t
BL (J(LI\{n+2,n+3} , σt+1 )).
c
If t ≥ 1 and I ∩ {n + 2, n + 3} = ∅, then kI,σ
= kI∪{n+2,n+3},t−1 , hence
t
BCr(L) (J(LI , σt )) = BL (J(LI∪{n+2,n+3} , σt−1 )).
6.2. Cases where Conjecture 6.4 holds. In this section we provide a list of
evidences to Conjecture 6.4.
6.2.1. Conjecture 6.4 holds for Cremona transforms of only linearly obstructed linear systems.
Proposition 6.9. Let L be linear system with n + 3 base points for which kC ≤ 1.
Any linear system L′ that can be Cremona reduced to L satisfies Conjecture 6.4.
Proof. We have σldim(L′ ) = σldim(L) = ldim(L) = dim(L). The first equality
follows from Lemma 6.8, the second follows from Definition 6.1 (see also Remark
6.3), the last inequality follows from Theorem 2.3.
20
MARIA CHIARA BRAMBILLA, OLIVIA DUMITRESCU, AND ELISA POSTINGHEL
6.2.2. Conjecture 6.4 is true for n ≤ 3.
Proposition 6.10. Conjecture 6.4 holds for n = 2.
Proof. Set L = L2,d (m1 , . . . , m5 ). It is a well-known fact that the Segre-HarbourneGimigliano-Hirschowitz Conjecture holds for five points. Moreover, from RiemannRoch Theorem on the blow-up projective plane it follows that
X
X
5
mi + mj − d
kC
mi + 1
d+2
+
+
dim(L) =
−
2
2
2
2
1
i,j
and one can easily verify that the right-hand side of the above is σldim(L).
Proposition 6.11. Conjecture 6.4 holds true in n = 3.
Proof. Let L be a linear system in P3 . If L is Cremona reduced, then it is linearly
non-special by [19, Theorem 5.3], that is dim(L) = ldim(L). On the other hand
since kC ≤ 0 by Remark 6.3 we have σldim(L) = ldim(L) and this concludes the
proof in this case.
Assume now that L is not Cremona reduced and denote by Cr(L) the corresponding Cremona reduced linear system. We have σldim(L) = σldim(L′ ) = dim(L′ ) =
dim(L). The first equality follows from Lemma 6.8, the second follows from the
previous case and the last one from (3.4).
6.2.3. Families of homogeneous linear systems that satisfy Conjecture 6.4. Consider
the following family of linear systems
L(t, a) := L2t,a(t+1) ((at)2t+3 ),
for all t, a ≥ 1. Notice that in the case a = 1, the linear system has one section
that is σt , see Section 3.2.
Proposition 6.12. For any t, a ≥ 1, L = L(t, a) has one element, aσt . In particular it satisfies Conjecture 6.4.
Proof. The hypersurface aσt belongs to L(t, a) because it has degree a(t + 1) and
multiplicity at along the rational normal curve given by the 2t + 3 points, see
discussion in Section 3.
We prove by induction on t that aσt is the unique element of L(t, a) and that
σldim(L) = 1. If t = 1, then the system L = L2,2a (a5 ) has one section that
consists of the multiple conic aσ1 ⊂ P2 . Furthermore it is easy to compute that
σldim(L) = 1.
Now assume that t ≥ 2. First, by means of a Cremona transformation
we reduce to Cr(L(a, t)) = L2t,at ((a(t − 1))2t+1 , (at)2 ) and σldim(L(a, t)) =
σldim(Cr(L(a, t))), by Lemma 6.8. Second, we observe that Cone(Cr(L(a, t))) =
L(t − 1, a). We conclude by induction and by Lemma 6.7.
For every b ≥ 1, let us consider the following linear system
L(b) := Ln,b(n+2) ((bn)n+3 ).
Proposition 6.13. Let n ≥ 2 and b ≥ 1. The linear system L(b) has one element
if n is even and empty otherwise. In particular L(b) satisfies Conjecture 6.4.
ON THE EFFECTIVE CONE OF Pn BLOWN-UP AT n + 3 POINTS
21
Proof. The proof is by induction on n ≥ 1. If n = 1, one has dim(L1,3b (b4 )) = 0.
If n = 2, one has dim(L2,4b ((2b)5 )) = 1.
Now assume n ≥ 3. Notice that Cone(Cr(L(b))) = Ln−2,bn ((b(n − 2))n+1 ). We
conclude by induction on n, using Lemma 6.7 and Lemma 6.8.
Consider the family
L = Ln,d (nn+3 ).
Proposition 6.14. The linear system L satisfies Conjecture 6.4 for any n ≥ 2 and
d ≥ 1.
Proof. If d ≤ n + 1, then the system is empty by Theorem 2.4. If d = n + 2 we
conclude by applying Proposition 6.13 in the case b = 1. If d ≥ n + 3, the statement
follows from Theorem 2.3 and Remark 6.3.
References
[1] J. Alexander, A. Hirschowitz, Polynomial interpolation in several variables, J. Algebraic
Geom. 4 (1995), no. 2, 201–222.
[2] I. Arzhantsev, U. Derenthal, J. Hausen, A. Laface, Cox rings, Cambridge Studies in Advanced
Mathematics 144. Cambridge University Press viii, 530 p. (2015).
[3] C. Bocci, Special effect varieties in higher dimension, Collectanea Mathematica (2005) Volume: 56, Issue: 3, page 299-326; ISSN: 0010-0757
[4] C. Bocci, Special Effect Varieties and (−1)-Curves, Rocky Mountain J. Math. Volume 40,
Number 2 (2010), 397-419
[5] M.C. Brambilla, O. Dumitrescu and E. Postinghel, On a notion of speciality of linear systems
in Pn , Trans. Am. Math. Soc. 367 (2015), 5447-5473
[6] M.C. Brambilla, O. Dumitrescu and E. Postinghel, On linear systems of P3 with nine base
points, Ann. Mat. Pura Appl., DOI: 10.1007/s10231-015-0528-5
[7] M.C. Brambilla and G. Ottaviani, On the Alexander-Hirschowitz theorem, J. Pure Appl.
Algebra 212 (2008), no. 5, 1229–1251.
[8] S. Cacciola, M. Donten-Bury, O. Dumitrescu, A. Lo Giudice, J. Park, Cones of divisors of
blow-ups of projective spaces, Matematiche (Catania) 66 (2011), no. 2, 153–187.
[9] E. Carlini and M.V. Catalisano, Existence results for rational normal curves, J. London
Math. Soc. (2) 76 (2007), 73-86.
[10] A.M. Castravet and J. Tevelev, Hilbert’s 14th problem and Cox rings, Compos. Math. 142
(2006), no. 6, 1479–1498.
[11] M.V. Catalisano, A.V. Geramita, A. Gimigliano, Higher secant varieties of Segre-Veronese
varieties, in Projective varieties with unexpected properties (2005),81–107.
[12] C. Ciliberto, Geometrical aspects of polynomial interpolation in more variables and of Waring’s problem, European Congress of Mathematics, Vol. I (Barcelona, 2000), 289–316, Progr.
Math., 201, Birkhäuser, Basel (2001)
[13] C. Ciliberto, B. Harbourne, R. Miranda, J. Roé, Variations on Nagata’s conjecture, Clay
Math. Proc. 18, 185–203, Amer. Math. Soc., (2013)
[14] C. Ciliberto and R. Miranda, Degenerations of Planar Linear Systems, J.Reine Angew. Math.
501, 191–220, (1998)
[15] C. Ciliberto and F. Russo. Varieties with minimal secant degree and linear systems of maximal dimension on surfaces, Adv. Math., 200 (2006), no. 1, 1–50.
[16] K. Chandler, A brief proof of a maximal rank theorem for generic double points in projective
space, Trans. Amer. Math. Soc. 353 (2001), no. 5, 1907–1920.
[17] K. Chandler, Linear systems of cubics singular at general points of projective space, Compositio Mathematica 134 (2002), 269–282.
[18] K. Chandler, The geometric interpretation of Fröberg-Iarrobino conjectures on infinitesimal
neighbourhoods of points in projective space, J. Algebra 286 (2005), no. 2, 421–455.
[19] C. De Volder and A. Laface, On linear systems of P3 through multiple points, J. Algebra 310
(2007), no. 1, 207–217.
22
MARIA CHIARA BRAMBILLA, OLIVIA DUMITRESCU, AND ELISA POSTINGHEL
[20] I. Dolgachev, Weyl groups and Cremona transformations, Singularities, Part 1 (Arcata,
Calif., 1981), 283–294, Proc. Sympos. Pure Math., 40, Amer. Math. Soc., Providence, RI,
1983.
[21] O. Dumitrescu and E. Postinghel, Vanishing theorems for linearly obstructed divisors,
arXiv:1403.6852 (2014).
[22] R. Fröberg, An inequality for Hilbert series of graded algebras, Math. Scand. 56 (1985), no.
2, 117–144.
[23] A. Gimigliano, On linear systems of plane curves, Ph.D. Thesis, Queen’s University, Canada
(1987)
[24] B. Harbourne, The geometry of rational surfaces and Hilbert functions of points in the plane,
Can. Math. Soc. Conf. Proc. 6, 95–111 (1986)
[25] A. Hirschowitz, Une conjecture pour la cohomologie des diviseurs sur les surfaces rationnelles
génériques, J. Reine Angew. Math. 397 (1989), 208–213.
[26] Y. Hu and S. Keel, Mori Dream Spaces and GIT, Michigan Math. J. 48 (2000), 331–348.
[27] A. Iarrobino, Inverse system of symbolic power III. Thin algebras and fat points, Compositio
Math. 108 (1997), no. 3, 319–356.
[28] A. Laface and L. Ugaglia, On a class of special linear systems on P3 , Trans. Amer. Math.
Soc. 358 (2006), no. 12, 5485–5500 (electronic).
[29] A. Laface and L. Ugaglia, Elementary (−1) curves of P3 , Comm. Algebra 35 (2007), 313–324
[30] S. Mukai, Geometric realization of T-Shaped root systems and counterexamples to Hilbert’s
fourteenth problem, Algebraic Transformation Groups and Algebraic Varieties Encyclopaedia
of Mathematical Sciences Volume 132, 2004, pp 123-129.
[31] S. Mukai, Finite generation of the Nagata invariant rings in A-D-E cases, 2005, RIMS
Preprint 1502.
[32] E. Postinghel, A new proof of the Alexander-Hirschowitz interpolation theorem, Ann. Mat.
Pura Appl. (4) 191 (2012), no. 1, 77-94.
[33] Z. Ran, Enumerative geometry of singular plane curves, Inventiones Math. 97 (1989), 447–
465.
[34] B. Sturmfels, M. Velasco. Blow-ups of Pn−3 at n points and spinor varieties, J. Commut.
Algebra, 2 (2010), no. 2, 223-244.
[35] B. Sturmfels and Z. Xu, Sagbi bases of Cox-Nagata rings, J. Eur. Math. Soc. (JEMS) 12
(2010), no. 2, 429-459
[36] A. Van Tuyl, An appendix to a paper of M. V. Catalisano, A. V. Geramita and A.
Gimigliano. The Hilbert function of generic sets of 2-fat points in P1 × P1 : Higher secant varieties of Segre-Veronese varieties in Projective varieties with unexpected properties
(2005), 109–112.
[37] G. Veronese, Behandlung der projectivischen Verhältnisse der Räume von verschiedenen
Dimensionen durch das Princip des Projicirens und Schneidens, (German), Math. Ann. 19
(1881), no. 2, 161–234.
E-mail address: [email protected]
Università Politecnica delle Marche, via Brecce Bianche, I-60131 Ancona, Italy
E-mail address: [email protected]
Institut für Algebraische Geometrie GRK 1463, Welfengarten 1, 30167 Hannover,
Germany
E-mail address: [email protected]
KU Leuven, Department of Mathematics, Celestijnenlaan 200B, 3001 Heverlee, Belgium
| 0math.AC
|
THIS PAPER IS A PREPRINT OF A PAPER SUBMITTED TO IET COMPUTER VISION. IF ACCEPTED, THE COPY OF RECORD WILL BE AVAILABLE AT THE IET DIGITAL LIBRARY1
Visual Multiple-Object Tracking for Unknown
Clutter Rate
arXiv:1701.02273v3 [cs.CV] 30 Nov 2017
Du Yong Kim
Abstract—In multi-object tracking applications, model parameter tuning is a prerequisite for reliable performance. In
particular, it is difficult to know statistics of false measurements
due to various sensing conditions and changes in the field of
views. In this paper we are interested in designing a multi-object
tracking algorithm that handles unknown false measurement
rate. Recently proposed robust multi-Bernoulli filter is employed
for clutter estimation while generalized labeled multi-Bernoulli
filter is considered for target tracking. Performance evaluation
with real videos demonstrates the effectiveness of the tracking
algorithm for real-world scenarios.
Index Terms—multi-object tracking, random finite set, multiBernoulli filtering
I. I NTRODUCTION
Multi-object tracking is one of fundamental problems in
many applications. There are abundant research works, however, it is still far from practical use. The overwhelming
majority of multi-target tracking algorithms are built on the
assumption that multi-object system model parameters are
known a priori, which is generally not the case in practice
[1], [4]. While tracking performance is generally tolerant to
mismatches in the dynamic and measurement noise, the same
cannot be said about missed detections and false detections. In
particular, mismatches in the specification of missed detection
and false detection model parameters such as detection profile
and clutter intensity can lead to a significant bias or even
erroneous estimates [25].
Unfortunately, except for a few application areas, exact
knowledge of model parameters is not available. This is especially true in visual tracking, in which the missed detection
and false detection processes vary with the detection methods.
The detection profile and clutter intensity are obtained by trial
and error. A major problem is the time-varying nature of the
missed detection and false detection processes. Consequently,
there is no guarantee that the model parameters chosen from
training data will be sufficient for the multi-object filter at
subsequent frames.
In radar target tracking applications, stochastic multi-object
tracking algorithms based on Kalman filtering or Sequential
Monte Carlo (SMC) method have been widely used [4],
[16]. This approach also has been used in visual multiobject tracking research [7], [8], [23]. On the other hand,
deterministic approach such as network flow [5], continuous
energy optimisation [10], has become a popular method for
multi-object tracking problem in visual tracking application.
Du Yong Kim is with the Department of Electrical and Computer
Engineering, Curtin University, Bentley, 6102 Australia e-mail: [email protected].
This approach is known to be free from tuning parameters,
however, it is useful only when reliable object detection is
available.
Unknown observation model parameters (i.e., clutter rate,
detection profile) in online multi-object filtering was recently
formulated in a joint estimation framework using random finite
set (RFS) approach [2], [3]. Recently, Mahler [25] showed that
clever use of the CPHD filter can accommodate unknown clutter rate and detection profile. In [26] it was demonstrated that
by bootstrapping clutter estimator from [25] to the Gaussian
mixture CPHD filter [17], performed very close to the case
with known clutter parameter can be achieved. [27] extended it
to multi-Bernoulli filter with SMC implementation. The multiBernoulli filter was used for visual multi-object tracking in
[11]. While the solution for filtering with unknown clutter
rate exists, these filters do not provide tracks that identify
different objects. In particular, the conference version of this
work [11] is seriously extended as a new algorithm that is
able to provides track identities with completely new structure
and evaluated using challenging pedestrian tracking and cell
migration experiments To the best of our knowledge this paper
is the first attempt for handling unknown false measurement
information in online tracking. The main contribution of this
paper is to design a multi-object tracker that also produces
trajectories and estimates unknown clutter rate on the fly.
II. P ROBLEM F ORMULATION
nx
Let X = R denote the space of the target kinematic state,
and {0, 1} denote the discrete space of labels for clutter model
and actual targets. Then, the augmented state space is given
by
X̆ = X × {0, 1}
(1)
where × denotes a Cartesian product. Consequently, the state
variable contains the kinematic state, and target/clutter indicator. We follow the convention from [27] that the label u = 0
will be used as a subscript to denote the clutter generators and
the label u = 1 for actual targets.
Suppose that there are Tk target and clutter object, and we
have Ok observations (i.e., detections). In the RFS framework,
the collections of targets (including clutter objects) and measurements can be described as finite subsets of the state and
observation spaces, respectively as (2)
k
X̆k = {x̆k,i }Ti=1
⊂ X̆,
k
Zk = {zk,j }O
j=1 ⊂ Z,
(2)
where x̆k,i represent either the kinematic state of actual target
or clutter target; zk,j is a measurement, and Z is the space
of measurement, respectively. Considering the dynamic of the
THIS PAPER IS A PREPRINT OF A PAPER SUBMITTED TO IET COMPUTER VISION. IF ACCEPTED, THE COPY OF RECORD WILL BE AVAILABLE AT THE IET DIGITAL LIBRARY2
state, the RFS model of the multi-target state at time k consists
of surviving targets and new targets entering in the scene. This
new set is represented as the union
[
[
X̆k =
Sk|k−1 (x̆k−1 )
Γk ,
(3)
x̆k−1 ∈X̆k−1
pu,k−1 ,pS,u,k
where Γk is a set of spontaneous birth objects (actual target
or clutter targets) and Sk|k−1 (·) is the set of survived object
states at time k with survival probability pS (x) < 1.
The set of observations given the multi-target state is
expressed as
S
Zk = ZT,0,k ZT,1,k ,
(4)
where ZT,0,k and ZT,1,k are, respectively, sets of clutter
and target-originated observations with unknown detection
probability pD (x) < 1.
With the RFS multi-target dynamic and measurement
model, the multi-object filtering problem amounts to propagating multi-target posterior density recursively forward in time
via the Bayes recursion. Note that in the classical solution
to this filtering problem such as PHD [2], CPHD [3], and
multi-Bernoulli filters [18], [22], [23], instead of clutter target
measurement set, the Poisson clutter intensity is given and the
detection profile pD (x) is also known a priori [1].
III. M ULTI - OBJECT TRACKER WITH UNKNOWN CLUTTER
RATE
A. Robust Multi-Bernoulli Filter
The multi- Bernoulli filter parametrizes the multi-object
posterior density by using a set of pair, i.e., Bernoulli pa(i)
rameter, {(r(i) , p(i)) }M
and p(i) represent the
i=1 where r
existence probability and the density of the state among M
Bernoulli components. In the following the predicted and
updated densities are represented by propagating a set of
Bernoulli parameters. The multi-Bernoulli filter recursion for
extended state space called RMB filter [27] is summarized to
make the paper self-contained.
If the posterior multi-object density of the multi-Bernoulli
form at time k − 1 is given as
(i)
M
k−1
{(rk−1 , pu,k−1 )}i=1
.
(5)
Then, the predicted intensity is approximated by the following multi-Bernoulli
(i)
(i)
M
k|k−1
{(rk|k−1 , pu,k|k−1 )}i=1
.
(6)
A set of predicted Bernoulli components is a union of birth
MΓ,k
(i)
(i)
components {(rΓ,k , pΓ,u,k )}i=1
and surviving components
(i)
(i)
M
where x is the kinematic state, pS,u,k is the survival probability
to time k and fu,k|k−1 (x|·) is the state transition density
specified by either for actual target f1,k|k−1 (x|·) or for clutter
target f0,k|k−1 (x|·).
If at time k, the predicted multi-target density is multiBernoulli of the form (6), then the updated multi-Bernoulli
density approximation is composed of the legacy components
with the subscript L and the measurement updated components
with the subscript U as follows (8)
(i)
k|k−1
{(rk|k−1 , pu,k|k−1 )}i=1
. The birth Bernoulli components
are chosen a priori by considering the entrance region of the
M
(i)
k|k−1
∪ {(rU,k (z), pU,u,k (·; z))}z∈Zk .
{(rL,k , pL,u,k )}i=1
(8)
The legacy and measurements updated components are calculated by a series of equations (9) as follows.
P
(i)
(i)
rL,k = u=0,1 rL,u,k
,
D
E
(i)
(i)
rL,u,k =
rk|k−1 pu,k|k−1 ,1−pD,u,k
D
E,
P
(i)
(i)
1−rk|k−1 u0 =0,1 pu0 ,k|k−1 ,pD,u0 ,k
(i)
pL,u,k (x)
1−pD,u,k pu,k|k−1 (x)
D
E,
P
(i)
u0 =0,1 pu0 ,k|k−1 ,1−pD,u0 ,k
(i)
rU,k (z) =
The aim of this paper is to propose a new online multi-object
tracker that is able to accommodate unknown clutter rate. For
this purpose, the Robust Multi-Bernoulli (RMB) filter [27] is
employed to adapt unknown clutter rate. Then, the estimated
clutter rate is plugged into the Generalized Multi-Bernoulli
(GLMB) tracker [19] to boost the performance in real-world
scenarios.
(i)
visual scene, e.g., image border. The surviving components
are calculated by
(i)
(i) P
(i)
rP,k|k−1 = rk−1 u=0,1 hpu,k−1 , pS,u,k i,
D
E
(i)
(7)
fu,k|k−1 (x|·,)pu,k−1 ,pS,u,k
(i)
D
E
,
pu,k|k−1 (x) =
(i)
(i)
=
P
u=0,1 rU,u,k (z),
PMk|k−1
r
(i)
k|k−1
1−r
(i)
k|k−1
i=1
rU,u,k (z) =
,g
(z,|·)pD,u,k i
hp(i)
u,k|k−1 u,k
2
(i)
(i)
1−r
p 0
,p
u0 =0,1
k|k−1
u =0,1 D,u0 ,k
P
(i)
(i)
PMk|k−1 rk|k−1 u0 =0,1 pu,k|k−1 ,gu0 ,k (z|·)pD,u0 ,k
P
i=1
(i)
(i)
1−r
p 0
,p
u0 =0,1
k|k−1
u ,k|k−1 D,u0 ,k
P
h
i
pU,u,k (x; z) =
(i)
(i)
k|k−1
pu,k|k−1 (x)gu,k (z|x)·pD,u,k
(i)
1−r
k|k−1
(i)
E
P
PMk|k−1 rk|k−1 D (i)
pu0 ,k|k−1 ,gu0 ,k (z|·)pD,u0 ,k
u0 =0,1
i=1
(i)
1−r
k|k−1
PMk|k−1
r
i=1
(9)
where pD,u,k is the state dependent detection probability,
gu,k (z|x) is the measurement likelihood function that will be
defined in the following section. Note that the SMC implementation of summarized equations (6)-(10) can be found in
[27].
B. Boosted Generalized labeled Multi-Bernoulli Filter
The generalized labeled multi-Bernoulli (GLMB) filter provides a solution of multi-object Bayes filter with unique labels.
In this paper, the GLMB filter is used as a tracker that
returns trajectories of multi-object given the estimated clutter
rate from the RMB. As shown in Fig. 1, GLMB and RMB
filters are interconnected by sharing tracking parameters and
facilitate feedback mechanism in order for robust tracking
against time-varying clutter background. Note that one step
RMB filter is used for the estimation of clutter rate, thus, it is
not a parallel implementation of independent filter.
We call the proposed tracker as Boosted GLMB tracker.
In multi-object tracking with labels, formally, the state of an
object at time k is defined as xk = (xk , `k ) ∈ X × Lk , where
,
THIS PAPER IS A PREPRINT OF A PAPER SUBMITTED TO IET COMPUTER VISION. IF ACCEPTED, THE COPY OF RECORD WILL BE AVAILABLE AT THE IET DIGITAL LIBRARY3
Fig. 1. Schematic diagram of the proposed tracker
Lk denotes the label space for objects at time k (including
those born prior to k). Note that Lk is given by Bk ∪ Lk−1 ,
where Bk denotes the label space for objects born at time k
(and is disjoint from Lk−1 ) and we do not consider clutter
generator in designing GLMB, thus, the label u is omitted.
Suppose that there are Nk objects at time k as (2), but only
consider actual target with label xk,1 , ..., xk,Nk , in the context
of multi-object tracking,
Xk = {xk,1 , ..., xk,Nk } ∈ F(X × Lk ),
(10)
where F(X × Lk ) denotes the space of finite subsets of
X × Lk . We denote cardinality (number of elements) of X
by |X| and the set of labels of X, {` : (x, `) ∈ X}, by LX .
Note that since the label is unique, no two objects have the
same label, i.e. δ|X| (|LX |) = 1. Hence ∆(X) , δ|X| (|LX | is
called the distinct label indicator.
In the GLMB the posterior density takes the form of a
generalized labeled multi-Bernoulli
iX
h
X (c)
(ξ)
πk−1 (X) = ∆(X)
ωk−1 (LX ) pk−1 .
(11)
Fig. 2. Comparison of OSPA distance (Boosted GLMB)
where Θk is the space of mappings θ : Lk → {0, 1, ..., |Zk |},
such that θ(i) = θ(i0 ) > 0 implies i = i0 , and
(c,θ)
(c)
(c,θ)
ωZk (L) ∝ δθ−1 ({0:|Zk |}) (L)ωk|k−1 (L)[ηZk ]L ,
(c,θ)
λˆc =
c∈C
Given the posterior multi-object density of the form (11), the
predicted multi-object density to time k is given by
h
iX
X (c)
(c)
πk|k−1 (X) = ∆(X)
ωk|k−1 (LX ) pk|k−1
(12)
c∈C
p
(c)
(x,`)ψZ (x,`;θ)
pk|k (x, `|Zk ) = k|k−1 (c,θ) k
,
ηZ
(`)
k
D
E
(c,θ)
(c)
ηZk (`) = pk|k−1 (·, `), ψZk (·, `; θ) ,
ψZk (x, `; θ) = δ0 (θ(`))(1 − pD,k (x, `))
p
(x,`)g (z
|x,`)
+ (1 − δ0 (θ(`))) D,k κk (zkθ(`)θ(`)
)
where κk ∼ λˆc U(Z) denotes the clutter density. λˆc is the
estimated clutter rate from the RMB filter. Specifically, the
extraction of clutter rate can be simply obtained by the EAP
estimate of clutter target number as
Mk
X
(i)
r0,k pD,0,k
(14)
i=1
(i)
where r0,k is the existence probability of clutter target introduced in the previous section, pD,0,k is the probability of
detection for clutter targets, U(Z) is a uniform density on the
observation region Z.
where
(c)
(c)
ωk|k−1 (L) = wB,k (L ∩ Bk )ωS,k (L ∩ Lk−1 ),
(c)
(c)
pk|k−1 (x, `) = 1Lk−1 (`)pS,k (x, `) + (1 − 1Lk−1 (`))pB,k (x, `),
D
E
(c)
pS,k−1 (·,`)fk|k−1 (x|·,`),pk−1 (·,`)
(c)
,
pS,k (x, `) =
(c)
ηS (`)
D
E
R
(c)
(c)
ηS,k (`) =
pS,k−1 (·, `)fk|k−1 (x|·, `), pk−1 (·, `) dx,
P
(c)
(c)
(c)
(c)
ωS,k (J) = [ηS,k ]Lk−1 I⊆Lk−1 1I (J)[qS ]I−J ωk−1 ,
D
E
(c)
(c)
qS (`) = 1 − pS,k−1 (·, `), pk−1 (·, `) ,
where c is the index for track hypothesis, L is an instance of
label set, I is track labels from previous time step.
Moreover, the updated multi-object density is given by
h
iX
X X (c,θ)
(c,θ)
πk|k (X|Zk ) = ∆(X)
ωZk (LX ) pk|k (·|Zk )
c∈C θ∈Θk
(13)
IV. E XPERIMENTAL RESULTS
In this section, two types of experimental results are given.
A nonlinear multi-object tracking example is tested in order
to show the performance of the proposed tracker with respect
to the standard performance metric, i.e., OSPA distance [28].
In addition, the proposed tracker is also evaluated for visual
multi-object tracking datasets [13], [14], [31].
A. Object motion model and basic parameters
The target dynamic described as a coordinated turn model
as (15)
f1,k|k−1 = N (xk ; mx,1,k|k−1 (xk−1 ), Px,1,k|k−1 ),
(15)
THIS PAPER IS A PREPRINT OF A PAPER SUBMITTED TO IET COMPUTER VISION. IF ACCEPTED, THE COPY OF RECORD WILL BE AVAILABLE AT THE IET DIGITAL LIBRARY4
Dataset
PETS09-S2L1
TUD-Stadtmitte
ETH
BAHNHOF and
SUNNYDAY
Method
Boosted GLMB
GLMB [19]
RMOT [15]
Boosted GLMB
GLMB [19]
RMOT [15]
Boosted GLMB
GLMB [19]
RMOT [15]
Recall
90.2 %
82.6 %
80.6 %
83.4 %
80.0 %
82.9 %
73.1 %
71.5 %
71.5 %
Precision
89.5 %
81.4 %
85.4 %
85.6 %
83.0 %
86.6 %
82.6 %
76.3 %
76.3 %
FPF
0.03
0.16
0.25
0.10
0.16
0.19
0.78
0.88
0.98
GT
19
19
19
10
10
10
124
124
124
MT
90 %
82.7 %
84.7 %
80 %
78.0 %
80 %
60.4 %
58.7 %
57.7 %
PT
10 %
17.3 %
15.3 %
20 %
22.0 %
20 %
34.6 %
27.4 %
37.4 %
ML
0.0 %
0.0 %
0.0 %
0.0 %
0.0 %
0.0 %
5.0 %
13.9 %
4.8 %
Frag
23
23
20
12
23
10
110
112
68
IDS
10
12
11
16
12
16
20
40
40
TABLE I
C OMPARISON WITH THE STATE - OF - THE - ART TRACKERS
where mx,1,k|k−1 (xk−1 )
=
[F (ωk−1 )xk−1 , ωk−1 ]T ,
2
T
2
Px,1,k|k−1 = diag([σw GG , σω ]),
T2
sin ωT
ωT
1
0 − 1−cos
0
ω
ω
2
0 cos ωT 0 −sin ωT
0
,G = T
F (ω) =
sin ωT
0 1−cos ωT 1
0 T2 ,
ω
ω
2
0 sin ωT
0
cos ωT
0
T
(16)
where T is the sampling time, σw is the standard deviation of
the process noise, σω is the standard deviation of the turn rate
noise. These standard deviation values are determined by the
maximum allowable object motion with respect to the image
frame rate. For clutter targets, the transition density f0,k|k−1 ,
is given as a random walk to describe arbitrary motion [27].
B. Numerical example
The proposed algorithm is tested with a nonlinear multitarget tracking scenario in [19], [27]. The actual target is
observed from noisy bearing and range information zk =
[θk , rk ]T and its likelihood function is given by
gk (zk |xk ) = N (zk ; mz,1,k (xk ), Pz,1,k ),
(17)
q
where mz,1,k (xk ) = [arctan(px,k /py,k ), p2x,k + p2y,k ] and
Pz,1,k = diag([σθ2 , σr2 ]). For RMB implementation, we follow
the same parameter setting as given in [27]. Performance
comparison between the GLMB tracker with known clutter
rate and the proposed tracker (Boosted GLMB tracker) is
studied. As can be seen in Fig. 1, OSPA distances for both
trackers are similar. This result verifies that the Boosted
GLMB shows reliable performance even when the clutter rate
is unknown.
C. Pedestrian tracking in vision
For the evaluation in real-world data, we are interested in
tracking of multiple pedestrians. To detect pedestrians, we
apply the state-of-the-art pedestrian detector proposed by Piotr
et. al, called ACF detector [29]. The detector used in the
experiment integrates a set of image channels (normalized
gradient magnitude, histogram of oriented gradients, and LUV
color channels) to extract various types of features in order to
discriminate objects from the background.
Assuming that the object state xk = [px,k ṗx,k , py,k ṗy,k ]T
(x-, y- positions and velocities) is observed with additive
Fig. 3. Pedestrian detection results with clutter measurements
Gaussian noise, the measurement likelihood function is given
by
gk (zk |xk ) = N (zk ; Hxk , Σ),
(18)
where N (z; m, P ) denotes a normal distribution with mean m
and covariance P , zk is the response from designated detector;
H = [1 0 0 0; 1 0 0 0] i,e., x-, y- position is observed
by the detector, Σ is the covariance matrix of the observation
noise.
Sample detection results in Fig. 2 contain false positive
detections from other types of object with similar shapes as
pedestrians. Based on our experiences ACF detector is robust
to partial occlusions, however, there are more false positive
detections than other single-model based detectors [30]. Thus,
it is relatively difficult to remove false positive detections by
hard thresholding when it is used for visual scenes with timevarying imaging conditions or moving camera. In particular, in
visual scene from autonomous vehicles, the average number of
clutters (i.e., clutter rate) is varying with respect to the change
in the field of view due to the vehicle pose change. The basic
assumption behind the existing visual multi-object tracking is
that the offline-designed detector, e.g., HOG detector [29], [30]
gives reasonably clean detections. Thus, direct data association
algorithms such as [5], and [10] show reasonable performance
with minor number of clutter measurements. However, in
practice, there are false positive detections which make data
association results inaccurate and computationally intensive.
”S2.L1” sequence from the popular PETS’09 dataset [31],
”TUD-Stadtmitte” sequence from TUD dataset [13], and
”Bahnhof” and ”Sunnyday” sequences from ETHZ dataset
[14] are tested in the experiment where maximum 8-15 targets
are moving in the scene. The number of targets varies in time
due to births and deaths, and the measurement set includes
target-originated detections and clutter. In our experiments,
unlike previous works, we use the ACF detector with low
threshold for nonmaximum suppression so as to have less
THIS PAPER IS A PREPRINT OF A PAPER SUBMITTED TO IET COMPUTER VISION. IF ACCEPTED, THE COPY OF RECORD WILL BE AVAILABLE AT THE IET DIGITAL LIBRARY5
TABLE II
C OMPARISON OF AVERAGED OSPA DISTANCE
Method
Boosted GLMB (Ours)
MHT [9]
Average OSPA
5
8.5
D. Cell migration analysis in microscopy image
Fig. 4. A snapshot of microscopy image of stem cells
Fig. 5. Reconstructed cell trajectories (left: MHT [9], right: Boosted GLMB)
number of miss-detections but increased false alarms with
time-varying rate. It is more realistic setting especially in
ETHZ dataset that is recorded with frequent camera view
changes. The Boosted GLMB, is compared with the original
GLMB (with fixed clutter rate) [19], and state-of-the-art online
Bayesian multi-object tracker called RMOT [15]. Quantitative
experiment results are summarized in Table 1 using wellknown performance indexes given in [6]. In Table 1, Boosted
GLMB shows superior performance compared to the GLMB
in all indexes and compatible with the recent online tracker,
RMOT. The proposed Boosted GLMB outperforms other
trackers with respect to FPF where tracker is able to effectively
reject clutters with estimated clutter rate. On the hand, inferior
performance in Fragmentation and ID switches are observed
compared with RMOT because of the lack of relative motion
model.
In summary, it is verified from the experiment that the
Boosted GLMB filter is effective when the clutter rate is
not known a priori which is often required in real-world
applications. We make experiments using unoptimized MATLAB code with Intel 2.53GHz CPU laptop. The computation
time per one image frame of size 768×586 is 0.2s which is
reasonably suitable for real-time visual tracking application.
Further improvements can be made by code optimization to
speed up.
The proposed algorithm is also tested with live-cell microscopy image data for cell migration analysis. The proposed
GLMB tracking method is tested on a real stem cell migration
sequence as illustrated in Fig. 4. The image sequence is
recorded for 3 days, i.e., 4320 min and each image is taken
in every 16 min.
Performance comparison is conducted with the state-of-theart Multiple Hypothesis Tracker (MHT) [9]. The same motion
and measurement models are used as in the first experiments
and spot detection in [9] is applied for the fair comparison. As
shown in Fig. 5, the Boosted GLMB provides reliable tracking
results compared to the MHT. The MHT is tuned to obtain the
best tracking results. The Boosted GLMB tracker produces
significantly less false tracks and alleviate fragmented tracks
because the tracker efficiently manages time-varying clutter
information and keep confident tracks. Quantitatively, as can
be seen in Table II, time averaged OSPA distances [28] for
both trackers verify that the Boosted GLMB shows reliable
performance even when the clutter rate is unknown.
V. C ONCLUSION
In this paper, we propose a new multi-object tracking algorithm for unknown clutter rate based on two interconnected
random finite set filters. Unknown clutter rate is estimated
using one step robust Bernoulli filter [27]. Then, trajectories
of objects are estimated using [19] with estimated clutter rate
online. Two filters are sharing tracking parameters so that
there is no need for tuning of clutter parameters. Comparison
results via a synthesized nonlinear multi-object tracking and
visual tracking datasets (visual surveillance and biomedical)
with state-of-the-art online trackers illustrate that the proposed
multi-object tracker shows reliable performance. Interesting
future research direction would be the extension of the tracking
algorithm to adaptive survival probability and handling of
missed-detections for further improvement.
R EFERENCES
[1] R. Mahler, Statistical Multisource-Multitarget Information Fusion, Norwood, MA: Artech House, 2007.
[2] R. Mahler, “Multi-target Bayes Filtering via First-order Multi-target
Moments,” IEEE Trans. Aerosp. Electron. Syst, 39 (4): 1152-1178, 2003.
[3] R. Mahler, “PHD Filters of Higher Order in Target Number,” IEEE Trans.
Aerosp. Electron. Syst, 43(3): 1523-1543, 2007.
[4] Y.Bar-Shalom and T.E. Fortmann, Tracking and Data Association, Academic Press, San Diego, 1998.
[5] L. Zhang, Y. Li and R. Nevatia, “Global data association for multi-object
tracking using network flows,” In CVPR 2008.
[6] C.-H. Kuo and R. Nevatia, “How Does Person Identity Recognition Help
Multi-Person Tracking?,” Proc. IEEE Conf. Computer Vision and Pattern
Recognition, 2011.
THIS PAPER IS A PREPRINT OF A PAPER SUBMITTED TO IET COMPUTER VISION. IF ACCEPTED, THE COPY OF RECORD WILL BE AVAILABLE AT THE IET DIGITAL LIBRARY6
[7] M.D. Breitenstein, F. Reichlin, B.Leibe, E. K.-Meier, and L.V. Gool,
”Online multiperson tracking-by-detection from a single, uncalibrated
camera,” IEEE Trans. Pattern Analysis and Machine Intelligence, 33 (9):
1820-1833, 2011.
[8] T.D. Laet, H. Bruyninckx, J.D. Schutter, “Shape-based online multitarget
tracking and detection for targets causing multiple measurements: Variational Bayesian clustering and lossless data association,” IEEE Trans.
Pattern Analysis and Machine Intelligence, 33 (12): 2477-2491, 2011.
[9] N. Chenouard, I. Bloch, and J.-C. Olivo-Marin, “Multiple Hypothesis
Tracking for Cluttered Biological Image Sequences,” IEEE Trans. Pattern
Analysis and Machine Intelligence, 35 (11): 2736-2750, 2013.
[10] A. Milan, S. Roth, and K. Schinlder, “Continuous energy minimization
for multitarget tracking,” IEEE Trans. Pattern Analysis and Machine
Intelligence, 36 (1): 58-72, 2014.
[11] Du Yong Kim, and Moongu Jeon, “Robust Multi-Bernoulli Filtering for
Visual Tracking,” Proc. IEEE Conf. Control Automation and Information
Sciences, 2014.
[12] I. Smal, M. Loog, W.J. Niessen, and E. Meijering, ”Quantitative Comparison of Spot Detection Methods in Fluorescence Microscopy,” IEEE
Trans. on Medical Imaging, 29 (2): 282-301, 2010.
[13] M. Andriluka, S. Roth, and B. Schiele “People-Tracking-by-Detection
and People-Detection-by-Tracking,” Proc. IEEE Conf. Computer Vision
and Pattern Recognition, 2008.
[14] A. Ess, B. Leibe, and L.V. Gool “Depth and Appearance for Mobile
Scene Analysis,” Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2007.
[15] J. H. Yoon, M.-H. Yang, J. Lim, and K.-J. Yoon, ”Bayesian MultiObject Tracking using Motion Context from Multiple Objects,” Proc.
IEEE Winter Conference on Applications of Computer Vision (WACV),
2015.
[16] D. Reid, “An algorithm for tracking multiple targets,” IEEE Trans. Aut.
Control, 24 (6): 843–854, 1979.
[17] B.-T. Vo, B.-N. Vo, and A. Cantoni, “Analytic implementations of the
cardinalized probability hypothesis density filter,” IEEE Trans. Signal
Processing, 55 (7): 3553–3567, 2007.
[18] B.-T. Vo, B.-N. Vo, and A. Cantoni, “The Cardinality Balanced MultiTarget Multi-Bernoulli Filter and Its Implementations,” IEEE Trans.
Signal Processing, 57 (2): 409-423, 2009.
[19] B.-T. Vo, and B.-N. Vo, “Labeled random finite sets and multi-object
conjugate priors,” IEEE Trans. Signal Processing, 61 (13): 3460–3475,
2013.
[20] S. Reuter, B.-T. Vo, B.-N. Vo, and K. Dietmayer, ”The labelled multiBernoulli filter,” IEEE Trans. Signal Processing, 62 (12): 3246-3260,
2014.
[21] B.-N. Vo, B.-T. Vo, N.-T. Pham and D. Suter, “Joint detection and
estimation of multiple objects from image observations,” IEEE Trans.
Signal Processing, 58 (10): 5129–5241, 2010.
[22] R. Hoseinnezhad, B.-N.Vo, B.-T. Vo, and D. Suter, “Visual tracking
of numerous targets via multi-Bernoulli filtering of image data,” Pattern
Recognition, 45 (10): 3625-3635, 2012.
[23] R. Hoseinnezhad, B.-N. Vo and B.-T. Vo, “Visual tracking in background
subtracted image sequences via multi-Bernoulli filtering”, IEEE Trans.
Signal Processing, 61 (2): 392–397, 2013.
[24] D.Y. Kim, B.-T. Vo and B.-N. Vo, “Data Fusion in 3D Vision Using
a RGB-D Data Via Switching Observation Model and Its Application
to People Tracking,” Int. Conf. Control, Aut. & Info. Sciences, Vietnam,
November 2013.
[25] R. Mahler, B.-T. Vo, and B.-N. Vo, “CPHD Filtering with Unknown
Clutter Rate and Detection Profile,” IEEE Trans. Signal Processing, 59
(8): 3497-3513, 2011.
[26] M. Beard, B.-T. Vo, and B.-N. Vo, ”Multi-target Filtering with Unknown
Clutter Density using a Bootstrap GM-CPHD Filter”, IEEE Signal
Processing Letters, 20 (4): 323-326, 2013.
[27] B.-T. Vo, B.-N. Vo, R. Hoseinnezhad, and R.P.S. Mahler, “Robust MultiBernoulli Filtering,” IEEE Jour. of Select. Topics in Signal Processing, 7
(3):399-409, 2013.
[28] D. Schuhmacher, B.-T. Vo, and B.-N. Vo, “A Consistent metric for
performance evaluation of multi-object filters ” IEEE Trans. Signal
Processing, 56(8), pp. 3447-3457, 2008
[29] P. Dollár and Z. Tu and P. Perona and S. Belongie, “Integral Channel
Features,” In BMVC, 2009.
[30] N. Dalal and B. Triggs, “Histogram of oriented gradients for human
detection,” In CVPR, 2005.
[31] J. Ferryman, IEEE Workshop Performance Evaluation of Tracking and
Surveillance, 2009.
| 1cs.CV
|
Dynamic Watermarking for General LTI Systems
arXiv:1703.07760v2 [math.OC] 25 Sep 2017
Pedro Hespanhol, Matthew Porter, Ram Vasudevan, and Anil Aswani
Abstract— Detecting attacks in control systems is an important aspect of designing secure and resilient control systems.
Recently, a dynamic watermarking approach was proposed
for detecting malicious sensor attacks for SISO LTI systems
with partial state observations and MIMO LTI systems with
a full rank input matrix and full state observations; however,
these previous approaches cannot be applied to general LTI
systems that are MIMO and have partial state observations.
This paper designs a dynamic watermarking approach for
detecting malicious sensor attacks for general LTI systems, and
we provide a new set of asymptotic and statistical tests. We
prove these tests can detect attacks that follow a specified attack
model (more general than replay attacks), and we also show
that these tests simplify to existing tests when the system is
SISO or has full rank input matrix and full state observations.
The benefit of our approach is demonstrated with a simulation
analysis of detecting sensor attacks in autonomous vehicles.
Our approach can distinguish between sensor attacks and wind
disturbance (through an internal model principle framework),
whereas improperly designed tests cannot distinguish between
sensor attacks and wind disturbance.
I. INTRODUCTION
Secure and resilient control requires the development of
mechanisms to allow safe operation in the face of malicious attacks or external interferences. This is particularly
challenging for cyber-physical systems (CPS) that feature
interconnection between physical sensors and actuators with
the communication and computation capabilities of routers,
servers, etc. Such concerns are motivated by real-world
instances of attacks on CPS, including: the Maroochy-Shire
incident [1], the Stuxnet worm [2], and other incidents [3].
For control systems, two possible modes of attacks are
either an attacker inserting faulty measurements into the
output sensor signal or an attacker inserting malicious values
into the actuator input for the control system. Cybersecurity
techniques [4], [5], [6], [7] are an important component
of designing resilient CPS. However, CPS frequently has a
decentralized structure; and so approaches that detect attacks
by decoupling different sensing and actuating components of
the system are particularly useful for ensuring safe operation.
This paper designs a dynamic watermarking approach for
detecting malicious sensor attacks for general LTI systems,
and has two main contributions: First, we generalize the watermarking approach developed in [8] for SISO LTI systems
This work was supported by the UC Berkeley Center for Long-Term
Cybersecurity, and by Ford Motor Company.
Pedro Hespanhol and Anil Aswani are with the Department of Industrial Engineering and Operations Research, University of California,
Berkeley, CA 94720, USA [email protected],
[email protected]
Matthew Porter and Ram Vasudevan are with the Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI 48109, USA
[email protected], [email protected]
with partial state observations and MIMO LTI systems with
a full rank input matrix and full state observations under an
arbitrary attack, and our generalization applies to general LTI
systems under a specific attack model that is more general
than replay attacks [9]. Second, we show that modeling
is important for designing watermarking techniques: For
instance, dynamic watermarking was used to detect sensor
attacks in an intelligent transportation system [10]; however,
here we show that persistent disturbances such as those
from wind can invalidate watermarking approaches, and we
propose an approach based on the internal model principle to
compensate for persistent disturbances. This second contribution motivates our generalization of dynamic watermarking
to general MIMO LTI systems with partial observations,
since internal model states are never directly observed.
A. Watermarking for CPS
Defense and security for CPS is classified into either
detection and identification [3], [11], [12], and a number of
“passsive” techniques have been proposed. State estimation
algorithms [13], [14] have been suggested in order to handle
attacks on the physical plants within a CPS. Another related
approach [15] provides a metric to characterize the resilience
of a system facing stealthy attacks on the actuators.
More recently, “active defense” based on watermarking
has been developed for detecting sensor attacks [8], [10],
[16], [17], [18], [9], [19]. The idea is that honest (i.e.,
not compromised by an attacker) actuators superimpose a
random signal onto the control input to ensure security in
face of sensor attacks. One set of approaches [16], [17],
[18], [9], [19] develops statistical hypothesis tests that detect
attacks with a certain error rate, while dynamic watermarking
approaches [8], [10] develop a test to ensure that only
attacks which add a zero-average-power signal to the sensor
measurements can remain undetected. The first set of techniques applies to general LTI systems under specific attack
models, but cannot ensure the zero-average-power property
for attacks; while the second set of techniques applies to
specific LTI systems under general attack models. Our first
contribution in this paper is to partially bridge the gap
between these two techniques by developing a method that
applies to general LTI systems under specific attack models
and that ensures the zero-average-power property for attacks.
B. Security for Intelligent Transportation Systems
The design of intelligent transportation systems (ITS) is
receiving increased attention [10], [20], [21], [22], [23], [24],
[25], and one significant area for further study is the design
of methods to ensure the safe and resilient operation of
ITS. One recent work [10] considered the use of dynamic
watermarking to detect sensor attacks in a network of autonomous vehicles coordinated by a supervisory controller;
the watermarking approach was successfully able to detect
attacks. However, large-scale deployments of ITS must be
resilient in the face of persistent disturbances from environmental and human factors. Wind is an example of such a
persistent disturbance. A second contribution of this work is
from the perspective of modeling: We show that persistent
disturbances such as those from wind can invalidate watermarking approaches, and we propose an internal model
principle-based approach to handle persistent disturbances.
This motivates our generalization of dynamic watermarking
to general MIMO LTI systems with partial observations,
since internal model states are not directly observed.
C. Outline
Section II reviews the general LTI system model (i.e.,
MIMO systems with partial observations) and specifies our
attack model, and Sect. III provides intuition on why existing
dynamic watermarking approaches cannot by used on a
general LTI system. We construct a detection consistent
dynamic watermarking approach for general LTI systems
under our attack model in Sect. IV, and our term detection
consistent test is used to refer to a test that ensures the
zero-average-power property (described above) for attacks.
Next, Sect. V describes how our asymptotic tests can be
converted into statistical tests, and Sect. VI shows how our
tests are special cases of those in [8] for the SISO case or
the MIMO case with full rank input matrix and full state
observations. We conclude with Sect. VII, which conducts
simulations of an autonomous vehicle: Our tests are able
to distinguish between sensor attacks and wind disturbances
when including wind disturbance in the system dynamics using the internal model principal, while improperly designed
tests cannot distinguish between attacks and wind.
II. LTI S YSTEM AND ATTACK M ODEL
Let [r] = {1, . . . , r}, and consider a MIMO LTI system
xn+1 = Axn + Bun + wn with partial observations yn =
Cxn + zn + vn , where x ∈ Rp , u ∈ Rq , and y, z, v ∈ Rm .
The vn should be interpreted as an additive measurement
disturbance added by an attacker, while wn represents zero
mean i.i.d. process noise with a jointly Gaussian distribution
and covariance ΣW , and zn represents zero mean i.i.d.
measurement noise with a jointly Gaussian distribution and
covariance ΣZ . We further assume the process noise is
independent of the measurement noise, that is wn for n ≥ 0
is independent of zn for n ≥ 0.
If (A, B) is stabilizable and (A, C) is detectable, then a
stabilizing output-feedback controller can be designed when
vn ≡ 0 using an observer and the separation principle. Let K
be a constant state-feedback gain matrix such that A + BK
is Schur stable, and let L be a constant observer gain matrix
such that A + LC is Schur stable. The idea of dynamic
watermarking in this context will be to superimpose a private
(and random) excitation signal en known in value to the
controller but unknown in value to the attacker. As a result,
we will apply the control input un = K x̂n + en , where x̂n is
the observer-estimated state and en are i.i.d. Gaussian with
zero mean and
constant
by Tthe controller.
T
variance ΣE fixed
T
T
T
B
BT , C =
x
x̂
=
Let
x̃
=
,
and
define
B
T
T
C 0 , D = I 0 , L = 0 −LT , and
A
BK
A=
(1)
−LC A + BK + LC
Then the closed-loop system with private excitation is given
by x̃n+1 = Ax̃n + Ben + Dwn + L(zn + vn ). If we define
the observation error
δ =
x̂ − x, then with the change of
variables x̌T = xT δ T we have the dynamics x̌n+1 =
Ax̌n + Ben + Dwn + L(zn + vn ), where B T = B T 0 ,
DT = I −I , L = L, and
A + BK
BK
.
(2)
A=
0
A + LC
Recall that A is Schur stable whenever A+BK and A+LC
are both Schur stable.
Since the controller is fixed, we can suppose the attacker
chooses vn = α(Cxn + zn ) + Cξn + ζn for some fixed
α ∈ R, where ξn+1 = (A + BK)ξn + ωn , ζn are i.i.d.
Gaussian with zero mean and constant variance ΣS fixed
by the attacker, and ωn are i.i.d. Gaussian with zero mean
and constant variance ΣO fixed by the attacker. The idea
underlying this attack model is that the attacker allows some
fraction of the true output Cxn + zn to be measured by
the controller, and at the same time also incorporates the
measurement of a false state ξn that evolves according the
dynamics that would be expected under the controller.
III. I NTUITION FOR D ESIGNING A N EW T EST
To better understand how to design a new test, it is
instructive to apply existing dynamic watermarking schemes
and the associated tests [8] to particular LTI systems. Such
an exercise provides intuition that we use to design new tests.
Our main example is an LTI system with
1 1
0
A=
, B=
, and C = 1 0 .
(3)
0 1
1
Suppose the attacker chooses vn = −(Cxn + zn ) + Cξn +
ζn with ΣS = ΣZ and ΣO = ΣW , meaning the output
measurement yn = Cξn + ζn has no component from the
actual system. This is a SISO (i.e., m = q = 1) system with
partial state measurement, and the tests in [8] pass for this
example, even though the sensor has been compromised by
an attacker. The problem in this example is that the test
PN −1
as-limN N1 n=0 L(C x̂n − yn )eTn−1 = 0
(4)
from [8] correlates the innovations process L(C x̂n − yn )
with the private excitation only one step back in time en−1 ;
however, it takes two time steps for the control input to enter
into the output in this example. And so when designing a new
test for general LTI systems, we need to take into consideration that there is generally some delay between when some
private excitation is applied to when it is observed.
IV. D ETECTION C ONSISTENT T EST
Now let ΣX be the positive semidefinite matrix that solves
the following
ΣX = AΣX AT + BΣE B T + DΣW DT + LΣZ LT . (5)
PN −1
Note that ΣX = as-limN N1 n=0 x̌n x̌Tn . Similarly let Σ∆
be the positive semidefinite matrix that solves the following
Σ∆ = (A + LC)Σ∆ (A + LC)T + ΣW + LΣZ LT . (6)
PN −1
T
1
T
Note Σ∆ = as-lim
N N n=0 δn δn and Σ∆ = M ΣX M ,
where M = 0 I . Recall that ΣX and Σ∆ exist because
the above are Lyapunov equations with matrices A, (A+LC)
that are Schur stable.
Lemma 1: We have that
(A + BK)r B
r
A B=
(7)
(A + BK)r B
for all r ≥ 0
Proof: The result holds for r = 0 since A0 = I and
(A + BK)0 = I. Now suppose the result holds for r: We
prove that it holds for r + 1. In particular, note that
(A + BK)r B
(A + BK)r+1 B
=
, (8)
Ar+1 B = A
(A + BK)r B
(A + BK)r+1 B
where the first equality holds by the inductive hypothesis,
and the second equality follows by calculation of the matrix
multiplication. Hence the result follows by induction.
Proposition 1: Let A(α) = A + αH with
0
0
,
(9)
H=
−LC 0
and define k 0 = min{k ≥ 0 | C(A + BK)k B 6= 0}. Then
we have that A(α)k B = Ak B for 0 ≤ k ≤ k 0 .
Proof: If k 0 = 0, then the result holds trivially. So
assume k 0 ≥ 1. We have that A(α)0 B = A0 B = B since
A(α)0 = A0 = I. Now suppose A(α)k B = Ak B for 0 ≤
k ≤ k 0 − 1. But using Lemma 1 implies that
(A + BK)k B
k+1
k+1
A(α)
B=A
B + αH
=
(A + BK)k B
0
= Ak+1 B, (10)
Ak+1 B + α
−LC(A + BK)k B
where we have used that LC(A+BK)k B = 0 since k < k 0 .
And so the result follows by induction.
Now let k 0 = min{k ≥ 0 | C(A + BK)k B 6= 0}, and
consider the following tests
PN −1
as-limN N1 n=0 (C x̂n − yn )(C x̂n − yn )T =
CΣ∆ C T + ΣZ
as-limN
1
N
PN −1
n=0
(C x̂n − yn )eTn−k0 −1 = 0.
(11)
(12)
Theorem 1: Suppose (A, B) is stabilizable, (A, C) is detectable, ΣE is full rank, and k 0 = min{k ≥ 0 | C(A +
BK)k B 6= 0} exists. If the test (11)–(12) holds, then
PN −1
as-limN N1 n=0 vnT vn = 0,
(13)
meaning that vn asymptotically has zero power.
Proof: Observe that the dynamics for x̃ are given by
x̃n+1 = A(α) · x̃n + Ben + Dwn + L((1 + α)zn + Cξn + ζn ),
where A(α) = A + αH with H given in (9). Next note that
a basic calculation gives
Pk−1
0
x̃n = A(α)k x̃n−k + k0 =0 A(α)k−k −1 Ben+k0 −k +
Dwn+k0 −k + (1 + α) · Lzn+k0 −k +
LCξn+k0 −k + Lζn+k0 −k . (14)
If we define C = −C C , then C x̂n − yn = C x̃n − α ·
C x̃n − (1 + α) · zn − Cξn − ζn , and so for k ∈ [p] we have
PN −1
1
T
n=0 E (C x̂n − yn )en−k =
N
(C − α · C) · A(α)k−1 BΣE . (15)
Note that k 0 ≤ p − 1 by the Cayley-Hamilton theorem. So
combining Proposition 1 with (15) implies
PN −1
k0
1
T
n=0 E (C x̂n − yn )en−k0 −1 = (C − α · C) · A BΣE
N
0
= −α · C · Ak BΣE
(16)
where the second equality holds by by Lemma 1 and the
definition of C. Because the test (12) holds, the quantity
(16) should equal 0. But since ΣE is full rank
by assumption,
0
Sylvester’s rank inequality implies C · Ak BΣE 6= 0 since
0
k0
0
C ·A B =
6= 0,
(17)
C(A + BK)k B
where the first equality holds by Lemma 1 and the definition
of C. Thus we must have α = 0.
Next consider the expression
PN −1
1
T
n=0 (C x̂n − yn )(C x̂n − yn ) =
N
PN −1
1
n=0 (C x̂n − (1 + α) · (Cxn + zn ) − Cξn − ζn )×
N
(C x̂n − (1 + α) · (Cxn + zn ) − Cξn − ζn )T . (18)
We showed above that α = 0, and so the expectation of the
above expression is
PN −1
CΣ∆ C T + ΣZ + ΣS + N1 n=0 E Cξn ξnT C T +
PN −1
1
N −1
x0 (C(A + BK)N −1 ξ0 )T . (19)
n=0 C(A + BK)
N
Since (A + BK) is Schur stable, the associated property of
exponential stability implies
PN −1
limN N1 n=0 C(A+BK)N −1 x0 (C(A+BK)N −1 ξ0 )T = 0
(20)
by combining the Cauchy-Schwartz inequality with the exponential stability. However from the test (11), the expectation
must equal CΣ∆ C T + ΣZ in the limit. Since all the terms in
the above expectation (19) are positive semidefinite or have
zero limit, this implies
PN −1
ΣS + as-limN N1 n=0 E Cξn ξnT C T = 0.
(21)
Finally, consider the expression
PN −1
PN −1
1
1
T
n=0 vn vn = N
n=0 (α(Cxn +zn )+Cξn +ζn ×
N
α(Cxn + zn ) + Cξn + ζn )T . (22)
Since α = 0, the expectation of the above expression is
PN −1
ΣS + N1 n=0 E Cξn ξnT C T +
PN −1
1
N −1
x0 (C(A + BK)N −1 ξ0 )T . (23)
n=0 C(A + BK)
N
PN −1
Combining (20)–(23) implies as-limN N1 n=0 vn vnT = 0.
However, vnT vn equals the sum P
of the diagonal entries of
N −1
vn vnT . Thus we have as-limN N1 n=0 vnT vn = 0.
Existence of k 0 = min{k ≥ 0 | C(A + BK)k B 6= 0} is
easy to verify because Cayley-Hamilton implies k 0 ≤ p − 1
or it does not exist, but we also give sufficient conditions.
Corollary 1: Suppose (A, B) is controllable, (A, C) is
observable, and ΣE is full rank. If the test (11)–(12) holds,
then
PN −1
(24)
as-limN N1 n=0 vnT vn = 0,
meaning that vn asymptotically has zero power.
Proof: We claim that, under the conditions stated, k 0 =
min{k ≥ 0 | C(A + BK)k B 6= 0} ≤ p − 1 exists. Indeed,
since (A, B) is controllable we have that: (A + BK, B) is
controllable, and the controllability matrix
C = B (A + BK)B . . . (A + BK)p−1 B
(25)
has rank(C) = p. And so by Sylvester’s rank inequality, we
have rank(CC) ≥ rank(C) + rank(C) − p = rank(C). But
(A, C) is observable, and so the observability matrix
C
I
CA
A
O = . = diag(C, . . . , C) .
(26)
..
..
CAp−1
Ap−1
has rank(O) = p. Again applying Sylvester’s rank inequality
implies p rank(C) ≥ rank(O) = p, or equivalently that
rank(C) ≥ 1. Combining this with the earlier inequality
gives rank(CC) ≥ 1, and so CC 6= 0. This means k 0 ≤ p−1
exists since CC is a block matrix consisting of the blocks
C(A + BK)k B. Thus the result follows by Theorem 1.
V. S TATISTICAL V ERSION OF T EST
For the purpose of implementation, we can also construct
a statistical version of our test (11)–(12). Our approach is
similar to [8] in that we construct a hypothesis test by
thresholding the negative log-likelihood. Before defining the
test, we make the following useful observation:
Proposition 2: Let ψnT = (C x̂n − yn )T eTn−k0 −1 . The
test (11)–(12) holds if and only if the following test holds:
PN −1
CΣ∆ C T + ΣZ
0
. (27)
as-limN N1 n=0 ψn ψnT =
0
ΣE
Moreover, if the test (11)–(12) holds or equivalently the test
(27) holds, then we have that as-limn E(ψn ) = 0.
Proof: The equivalence between (11)–(12) and (27)
follows from the definition of ψn and of the tests. Next
suppose either (equivalent) test holds: Using the dynamics
on x̌ we have E(x̌n+1 ) = AE(x̌n ) + LE(vn ). But we have
Pn−1
vn = C(A + BK)n ξ0 + C k=0 (A + BK)n−k−1 ωk + ζn
since α = 0 as shown in the proof of Theorem 1, and
so E(vn ) = C(A + BK)n ξ0 . Since (A + BK) is Schur
stable, we have limn E(vn ) = 0 and hence limn E(x̌n ) =
A limn E(x̌n ). This means that limn E(x̌n ) = 0 since I − A
is full rank (which can be seen by recalling that A is Schur
stable, so cannot have any eigenvalue of exactly one,
andthus
det(sI − A) 6= 0 for s = 1). Since C x̂n − yn = 0 C x̌n ,
we have that E(C x̂n − yn ) = 0. This implies E(ψn ) = 0
since E(en−k0 −1 ) = 0 by construction.
This result
Pn+`implies that asymptotically the summation
Sn = 1` n+1 ψn ψnT with ` ≥ m + q has a Wishart
distribution with ` degrees of freedom and a scale matrix
that matches (27), and we use this observation to define a
statistical test. In particular, we check if the negative loglikelihood
L(Sn ) = (m + q + 1 − `) · log det Sn +
(CΣ∆ C T + ΣZ )−1
0
trace
× Sn
(28)
0
Σ−1
E
corresponding to this Wishart distribution and the summation
Sn is large by conducting the hypothesis test
(
reject,
if L(Sn ) > τ (α)
(29)
accept, if L(Sn ) ≤ τ (α)
where τ (α) is a threshold that controls the false error rate
α. A rejection corresponds to the detection of an attack,
while an acceptance corresponds to the lack of detection of
an attack. This notation emphasizes the fact that achieving
a specified false error rate α (a false error in our context
corresponds to detecting an attack when there is no attack
occurring) requires changing the threshold τ (α).
VI. R ELATIONSHIP TO E XISTING T ESTS
It is interesting to compare our test (11)–(12) to those
designed in [8]. More specifically, [8] designed a related
sequence of tests adapted to different (and less complex)
assumptions about the model dynamics. We will show that
our test is closely related to (and generalizes) these previous
tests developed under assumptions of less complex dynamics.
The simplest test in [8] was designed for systems with
direct state measurement (i.e., C = I), no measurement error
(i.e, zn ≡ 0), and full rank input matrix (i.e, rank(B) =
p). The SISO (i.e., m = p = q = 1) and MIMO cases
were considered separately in [8], though the SISO case is a
special case of the MIMO case. If we choose L = −A, then
we have that: yn = xn + vn , xn+1 = Axn + BK x̂n + Ben +
wn , x̂n+1 = Ayn + BK x̂n + Ben , Σ∆ = ΣW , and k 0 = 1
since rank(CB) = p. So our test (11)–(12) simplifies to
PN −1
as-limN N1 n=0 (yn+1 − Ayn − BK x̂n − Ben )×
(yn+1 − Ayn − BK x̂n − Ben )T = ΣW
as-limN
1
N
PN −1
n=0
(30)
(yn+1 − Ayn − BK x̂n − Ben )×
eTn = 0. (31)
This exactly matches the test designed in [8] for LTI systems
with the above described properties.
10
2
-3
Test (11)
10-3
5
Un-Attacked
Attacked
1
Test (12)
0
Un-Attacked
Attacked
-5
0
-10
-1
-15
0
10
20
30
40
50
0
10
20
10-4
2
30
40
50
Time(s)
Time(s)
Test (11) with Wind Model
10-3
5
Test (12) with Wind Model
Un-Attacked
Attacked
1
0
0
Un-Attacked
Attacked
-1
-2
-5
0
10
20
30
40
50
0
Fig. 1.
Deviation of (11) in Simulation of Autonomous Vehicle
A more complex test in [8] was designed for SISO (i.e.,
m = q = 1) systems with partial state measurement. In our
notation, the tests in [8] for this case simplify to
PN −1
as-limN N1 n=0 L(C x̂n − yn )(C x̂n − yn )T LT =
L CΣ∆ C T + ΣZ LT (32)
PN −1
as-limN N1 n=0 L(C x̂n − yn )eTn−1 = 0.
(33)
But k 0 = 1 since B is a nonzero vector and rank(CB) = 1 in
this case. So the test (32)–(33) from [8] essentially matches
our test (11)–(12), but with the difference that the test in [8]
considers quantities with L(C x̂n −yn ), while our test directly
considers quantities with C x̂n − yn ; this is a negligible
difference since C x̂n − yn is a scalar in this SISO case.
VII. S IMULATIONS : AUTONOMOUS V EHICLE
A standard model [26] for error
kinematics
of lane keeping
and speed control has xT = ψ y s γ v and uT = r a ,
where ψ is heading error, y is lateral error, s is trajectory distance, γ is vehicle angle, v is vehicle velocity, r is steering,
and a is acceleration. Linearizing about a straight trajectory
and constant velocity v0 = 10, and then performing exact
discretization with sampling period ts = 0.05 yields
1
1
0
0
1 0 0 10
400
1 1 0 1 0
1
0
40
2
2400
1
1
A = 0 0 1 0 2
B= 0
(34)
800
0 0 0 1 0
1
0
20
1
0 0 0 0 1
0
20
with C = I 0 ∈ R3×5 . We used process and measurement noise with ΣW = 10−8 and ΣZ = 10−5 , respectively.
10
20
30
40
50
Time(s)
Time(s)
Fig. 2.
Deviation of (12) in Simulation of Autonomous Vehicle
Our simulations used the wind model: dn+1 = 0.9dn + χn ,
where χn are i.i.d. zero mean Gaussians with σχ2 = 2×10−6 ,
and the wind state d entered additively into the y dynamics.
We applied our tests using a dynamic watermark with
variance ΣE = 12 I, where K and L were chosen to stabilize
the closed-loop system without an attack. We conducted
four simulations: Un-attacked and attacked simulations were
conducted with a test computed without wind in the system model, and un-attacked and attacked simulations were
conducted with a test computed with wind in the system
model. In both attack simulations, we chose an attacker with
α = −0.6,P
ξ0 = 0, ΣO = 10−8 , and ΣS = 10−8 . Fig. VI
N −1
1
shows k N n=0 (C x̂n −yn )(C x̂n −yn )T −CΣ∆ C T −ΣZ k,
PN −1
and Fig. VI shows k N1 n=0 (C x̂n −yn )eTn−k0 −1 k. If the test
is detection consistent, then these values go to zero. The plots
show dynamic watermarking cannot detect the presence or
absence of an attack when wind affects the system dynamics
but is not included in the test, while our test sdetect the
presence or absence of an attack when a model of wind is
included in the test. Fig. VII shows the results of applying
our statistical test (28), and the same behavior is seen.
VIII. C ONCLUSION
This paper constructed a dynamic watermarking approach
for detecting malicious sensor attacks for general LTI systems, and the two main contributions were: to extend dynamic watermarking to general LTI systems under a specific
attack model that is more general than replay attacks, and to
show that modeling is important for designing watermarking
techniques by demonstrating how persistent disturbances can
negatively affect the accuracy of dynamic watermarking. Our
Statistical Test
Negative Log Liklihood
650
600
550
500
Un-Attacked
Attacked
(0.05)
450
400
0
5
10
15
Negative Log Liklihood
540
20
25
30
Time (s)
Statistical Test with Wind Model
35
40
45
50
520
500
480
Un-Attacked
Attacked
(0.05)
460
440
420
0
Fig. 3.
5
10
15
20
25
Time (s)
30
35
40
45
50
Value of (28) for Simulation of Autonomous Vehicle, with a Negative Log-Likelihood Threshold for α = 0.05 False Detection Error Rate
approach to resolve this issue was to incorporate a model of
the persistent disturbance via the internal model principle.
Future work includes generalizing the attack models that
can be detected by our approach. An additional direction
for future work is to study the problem of robust controller
design in the regime of when an attack is detected.
R EFERENCES
[1] M. Abrams and J. Weiss, “Malicious control system cyber security
attack case study–Maroochy water services, australia,” MITRE, 2008.
[2] R. Langner, “Stuxnet: Dissecting a cyberwarfare weapon,” IEEE
Security & Privacy, vol. 9, no. 3, pp. 49–51, 2011.
[3] A. A. Cárdenas, S. Amin, and S. Sastry, “Research challenges for the
security of control systems.” in HotSec, 2008.
[4] B. Parno, M. Luk, E. Gaustad, and A. Perrig, “Secure sensor network
routing: A clean-slate approach,” in ACM CoNEXT, 2006.
[5] V. Kumar, J. Srivastava, and A. Lazarevic, Managing cyber threats:
issues, approaches, and challenges. Springer, 2006, vol. 5.
[6] W. Wang and Z. Lu, “Cyber security in the smart grid: Survey and
challenges,” Computer Networks, vol. 57, no. 5, pp. 1344–1371, 2013.
[7] K.-D. Kim and P. R. Kumar, “Cyber–physical systems: A perspective
at the centennial,” Proc. of IEEE, vol. 100, pp. 1287–1308, 2012.
[8] B. Satchidanandan and P. Kumar, “Dynamic watermarking: Active
defense of networked cyber-physical systems,” Proc. of IEEE, 2016.
[9] S. Weerakkody, Y. Mo, and B. Sinopoli, “Detecting integrity attacks
on control systems using robust physical watermarking,” in Proc. of
IEEE CDC, 2014, pp. 3757–3764.
[10] W.-H. Ko, B. Satchidanandan, and P. Kumar, “Theory and implementation of dynamic watermarking for cybersecurity of advanced
transportation systems,” in Proc. of IEEE CNS, 2016, pp. 416–420.
[11] A. A. Cardenas, S. Amin, and S. Sastry, “Secure control: Towards
survivable cyber-physical systems,” in ICDCS, 2008, pp. 495–500.
[12] F. Pasqualetti, F. Dörfler, and F. Bullo, “Attack detection and identification in cyber-physical systems,” IEEE Transactions on Automatic
Control, vol. 58, no. 11, pp. 2715–2729, 2013.
[13] H. Fawzi, P. Tabuada, and S. Diggavi, “Secure estimation and control
for cyber-physical systems under adversarial attacks,” IEEE Transactions on Automatic Control, vol. 59, no. 6, pp. 1454–1467, 2014.
[14] ——, “Secure state-estimation for dynamical systems under active
adversaries,” in Allerton Conference. IEEE, 2011, pp. 337–344.
[15] C.-Z. Bai, F. Pasqualetti, and V. Gupta, “Security in stochastic control
systems: Fundamental limitations and performance bounds,” in American Control Conference (ACC), 2015. IEEE, 2015, pp. 195–200.
[16] Y. Mo and B. Sinopoli, “Secure control against replay attacks,” in
Allerton Conference. IEEE, 2009, pp. 911–918.
[17] Y. Mo, E. Garone, A. Casavola, and B. Sinopoli, “False data injection
attacks against state estimation in wireless sensor networks,” in Proc.
of IEEE CDC, 2010, pp. 5967–5972.
[18] Y. Mo, R. Chabukswar, and B. Sinopoli, “Detecting integrity attacks
on scada systems,” IEEE CST, vol. 22, no. 4, pp. 1396–1407, 2014.
[19] Y. Mo, S. Weerakkody, and B. Sinopoli, “Physical authentication
of control systems: Designing watermarked control inputs to detect
counterfeit sensor outputs,” IEEE Control Systems, vol. 35, no. 1, pp.
93–109, 2015.
[20] H. Gonzalez and E. Polak, “On the perpetual collision-free rhc of
fleets of vehicles,” Journal of optimization theory and applications,
vol. 145, no. 1, pp. 76–92, 2010.
[21] A. Aswani and C. Tomlin, “Game-theoretic routing of gps-assisted
vehicles for energy efficiency,” in ACC, 2011, pp. 3375–3380.
[22] W. Zhang, M. Kamgarpour, D. Sun, and C. J. Tomlin, “A hierarchical
flight planning framework for air traffic management,” Proceedings of
the IEEE, vol. 100, no. 1, pp. 179–194, 2012.
[23] R. Vasudevan, V. Shia, Y. Gao, R. Cervera-Navarro, R. Bajcsy, and
F. Borrelli, “Safe semi-autonomous control with enhanced driver
modeling,” in ACC, 2012, pp. 2896–2903.
[24] S. Mohan and R. Vasudevan, “Convex computation of the reachable
set for hybrid systems with parametric uncertainty,” in Proc. of ACC,
2016, pp. 5141–5147.
[25] G. Como, E. Lovisari, and K. Savla, “Convexity and robustness of
dynamic network traffic assignment for control of freeway networks,”
IFAC-PapersOnLine, vol. 49, no. 3, pp. 335–340, 2016.
[26] V. Turri, A. Carvalho, H. Tseng, K. Johansson, and F. Borrelli, “Linear
model predictive control for lane keeping and obstacle avoidance on
low curvature roads,” in Proc. of IEEE ITSC, 2013, pp. 378–383.
| 3cs.SY
|
arXiv:1208.3167v1 [math.AC] 15 Aug 2012
THE MAXIMAL DENUMERANT OF A NUMERICAL
SEMIGROUP
LANCE BRYANT AND JAMES HAMBLIN
Abstract. Given a numerical semigroup S = ha0 , a1 , a2 , . . . , at i
and n ∈ S, we consider the factorization n = c0 a0 +c
P1 a1 +· · ·+ct at
where ci ≥ 0. Such a factorization is maximal if
ci is a maximum over all such factorizations of n. We provide an algorithm for
computing the maximum number of maximal factorizations possible for an element in S, which is called the maximal denumerant
of S. We also consider various cases that have connections to the
Cohen-Macualay and Gorenstein properties of associated graded
rings for which this algorithm simplifies.
1. Introduction
Let N denote the nonnegative integers. A numerical semigroup S
is a subsemigroup of N that contains 0 and has a finite complement
in N. For two elements u and u′ in S, write u u′ if there exists
an s ∈ S such that u + s = u′. This defines a partial ordering on
S. The minimal elements in S \ {0} with respect to this ordering
form the unique minimal set of generators of S, which is denoted by
{a0 , a1 , a2 , . . . , at } where a0 < a1 < aP
2 < · · · < at .
The numerical semigroup S = { ti=0 ci ai : ci ≥ 0} is represented
using the notation S = ha0 , a1 , a2 , . . . , at i. Since the minimal generators of S are distinct modulo a0 , the set of minimal generators is
finite. Furthermore, S having finite complement in N is equivalent to
gcd (a0 , a1 , . . . , at ) = 1. The cardinality, t + 1, of the set of minimal
generators of a semigroup S is called the embedding dimension of S,
written ν = ν(S). The element a0 is called the multiplicity of S, also
sometimes written e = e(S). When S 6= N, we have 2 ≤ ν ≤ e.
By definition, if n ∈ S = ha0 , a1 , . . . , at i, then there exists
Pa (t + 1)tuple of nonnegative integers c = (c0 , c1 , . . . , ct ) such that
ci ai = n.
We call c a factorization of n in S,
or
simply
an
S-factorization
of
P
n. The length of c, written |c|, is
ci . We say that c is maximal if
no S-factorization of n has length greater than |c|, and minimal if no
S-factorization of n has length less then |c|. For n ∈ S, the order of n,
written ord(n; S) or ord(n), is the length of a maximal S-factorization
1
2
LANCE BRYANT AND JAMES HAMBLIN
of n. Similarly, the length of a minimal S-factorization of n is denoted
by min ord(n; S).
Provided S 6= N, it is a finitely generated monoid which fails to
have unique factorization with respect to its minimal generators. For a
given element of s ∈ S, a basic arithmetic constant that measures this
failure is the cardinality of the set of factorizations of s. This is called
the denumerant of s in S, denoted by d(s; S). See [24] for an exhaustive
view of related results. This is one of several numerical invariants that
have appeared in recent papers exploring the factorization properties
of numerical semigroups; for example, see [2, 3, 4, 6, 14, 15, 16].
In [13], the authors considered a variation of the denumerant.
Definition 1.1. The maximal denumerant of an element s in S is the
number of factorizations of s that have maximal length and is denoted
by dmax (s; S). The maximal denumerant of the semigroup S is defined
to be dmax (S) = maxs∈S {dmax (s; S)}.
It was shown in [13] that the maximal denumerant of a semigroup S is
always finite, and the following formulas were given for this value when
the embedding dimension of S is less than four: Let S be (perhaps nonminimally) generated by a1 , a2 , and a3 , and let g = gcd(a2 −a1 , a3 −a1 ),
m = (a2 − a1 )/g, and n = (a3 − a1 )/g. Then we can write
S = ha1 , a1 + gm, a1 + gni.
If 0 ≤ α < mn such that α ≡ −a1 mod mn, we have
l a1 m
if α ∈ hm, ni
mn ,
dmax (S) =
(1)
l
m
a1 + 1, otherwise.
mn
Moreover, if x and y are integers such that mx + ny = a1 , we have
lxm l y m
(2)
+
.
dmax (S) =
n
m
It is not clear how these formulas might extend to semigroups with
higher embedding dimension, and in this paper we are concerned with
computing the maximal denumerant of an arbitrarily given semigroup.
Notice that the formulas above make use of the values a1 , a2 − a1 , and
a3 − a1 . These integers generate the blowup semigroup of S, see [7].
This suggests that the blowup semigroup may be useful for computing the maximal denumerant in general. In Section 3, we show that
this is indeed the case. In essence, the problem of counting maximal
factorizations in a semigroup S corresponds to counting factorizations
THE MAXIMAL DENUMERANT OF A NUMERICAL SEMIGROUP
3
with bounded length in the blowup semigroup, and Theorem 3.9 is the
crucial result.
Section 4 is concerned with a special class of semigroups, namely
additive semigroups (also called good [5] and M-additive [12]). These
are the semigroups for which the associated graded ring with respect
to the maximal ideal of the the corresponding local semigroup ring is
Cohen-Macaulay, see [5, 8, 9, 11, 12, 17, 19]. Supersymmetric semigroups (also called M-symmetric [11]) are important examples of additive semigroups since the associated graded ring is Gorenstein in this
case, see [11, 12, 17, 18]. The method for computing the maximal denumerant outlined in Section 3 simplifies when the semigroup is additive,
and especially when the semigroup is supersymmetric.
In the next section, we begin by establishing the basic connection
between factorizations in a semigroup and its blowup that will be used
throughout the paper.
2. Factorizations and the blowup semigroup
Let S = he, a1 , a2 , . . . , at i be a numerical semigroup with mutiplicity
e and embedding dimension ν = t + 1, where e < a1 < a2 < · · · < at .
As noted in the introduction, the ai are distinct modulo e. They are,
in fact, the least elements of S in their respective congruence classes.
The set consisting of the least elements in S for each congruence class
modulo e is called the Apéry set of S. More generally, an Apéry set
with respect to an element u ∈ S is defined to be Ap(S; u) = {w ∈ S :
w − u 6∈ S}. Additionally, we denote the elements of S congruent to i
modulo e by Si , so that Ap(S; e) = {min(Si ) : 0 ≤ i < e}.
For our study of the maximal denumerant of S, we consider the
blowup of S, sometimes called the Lipman semigroup of S in honor of
[22].
Definition 2.1. Given a numerical semigroup S = he, a1 , a2 , . . . , at i,
the blowup of S is the semigroup B = he, a1 − e, a2 − e, . . . , at − ei.
We will write di = ai − e, so that B = he, d1 , d2 , . . . , dt i, and set
D = {e, d1 , d2 , . . . , dt }.
Note that the element e is not necessarily the multiplicity of B,
and ν = t + 1 is not necessarily the embedding dimenstion of B. For
example, if we have S = h4, 5, 6i with multiplicity 4 and embedding
dimesnion 3, then B = h4, 1, 2i = N with multiplicity 1 and embedding
dimension 1. In general, we have e(B) ≤ e(S) and ν(B) ≤ ν(S). Thus
D is always a generating set of B, but not necessarily minimal.
We wish to consider B with respect to the generating set
so we say
PD,
t
D
that (x0 , x1 , . . . , xt ) is a B -factorization of b ∈ B if x0 e+ i=1 xi di = b.
4
LANCE BRYANT AND JAMES HAMBLIN
For example, B = N is the blowup of S = h4, 5, 6i and D = {4, 1, 2}.
Thus, (0, 3, 1) is a B D -factorization of 5 since 5 = 0(4) + 3(1) + 1(2),
and d(5; B D ) = 4 is the denumerant of 5 in B with respect to D since
there are a total of four B D -factorizations of 5.
Lemma 2.2. With the definitions above, let s ∈ S and r ∈ N.
(1) If x = (x0 , x1 , . . . , xt ) is an S-factorization of s with length r,
then (0, x1 , x2 , . . . , xt ) is a B D -factorization of s−re with length
at most r.
(2) If y = (y0 , y1 , . . . , yt ) is a B D -factorization of s − re with length
at most r, then (2y0 +r −|y|, y1, y2 , . . . , yt ) is an S-factorization
of s with length r + y0 . In particular, ord(s; S) ≥ r.
Proof. For the proof of 1., let x = (x0 , x1 , . . . , xt ) be an S-factorization
of s with length r. Then
s = x0 e + x1 a1 + · · · + xt at
s − re = (x0 e − x0 e) + (x1 a1 − x1 e) + · · · + (xt at − xt e)
s − re = 0e + x1 d1 + · · · + xt dt
Hence (0, x1 , x2 , . . . , xt ) is a B D -factorization of s − re with length at
most r.
For the proof of 2., let y = (y0 , y1 , . . . , yt) be a B D -factorization of
s − re with |y| ≤ r. Hence
(3)
s − re = y0 e + y1 d1 + · · · + yt dt .
Let v = |y| − y0 and note that v ≤ r. Now add ve to both sides of (3)
to obtain
(4)
s − re + ve = y0 e + y1 a1 + · · · + yt at .
Next, adding (r − v)e to both sides of (4) gives
s = (y0 + r − v)e + y1 a1 + y2 a2 + · · · + yt at .
Since y0 +r−v = 2y0 +r−|y|, this tells us that (2y0 +r−|y|,
y1 , y2 , . . . , yt )
P
is an S-factorization of s with length y0 + r − v + ti=1 yi = y0 + r.
3. The Maximal Denumerant
Recall that if s ∈ S has a maximal S-factorization x, then |x| =
ord(s). By Lemma 2.2 we have a corresponding B D -factorization of
s − ord(s)e. This situation will occur frequently in this section, so we
have the following definition.
Definition 3.1. Given s ∈ S, the adjustment of s is adj(s) = s −
ord(s)e. Morevoer, for U ⊂ S, we set adj(U) = {adj(s) : s ∈ U} and
call this the adjustment of U.
THE MAXIMAL DENUMERANT OF A NUMERICAL SEMIGROUP
5
Using Lemma 2.2, we have a way to find dmax (s; S) via B D -factorizations.
This is the content of Proposition 3.3.
Lemma 3.2. Let s ∈ S. If (x0 , x1 , . . . , xt ) is a B D -factorization of
adj(s) with length at most ord(s), then x0 = 0.
Proof. By statement 2. of Lemma 2.2, we see that s has an S-factorization
of length x0 + ord(s). By definition, this implies x0 = 0.
Proposition 3.3. If s ∈ S, then dmax (s; S) is the number of B D factorizations of adj(s) of length at most ord(s).
P
Proof. Given a t-tuple (x1 , . . . , xt ), write x0 = ord(s) − ti=1 xi . It is
sufficient to prove that (x0 , x1 , . . . , xt ) is a maximal S-factorization of s
if and only if (0, x1 , . . . , xt ) is a B D -factorization of adj(s) with length
at most ord(s).
Let (x0 , x1 , . . . , xt ) be a maximal S-factorization of s. Then by statement 1. of Lemma 2.2 with r = ord(s), we see that (0, x1 , . . . , xt ) is a
B D -factorization of adj(s) with length at most ord(s).
Conversely, let (0, x1 , . . . , xt ) be a B D -factorization of adj(s) with
length at most ord(s). Note that by Lemma 3.2, any such factorization
has a P
zero in the first component. By statement 2. of Lemma 2.2,
t
since
i=0 xi = ord(s), we see that (x0 , x1 , . . . , xt ) is a maximal Sfactorization of s.
From Proposition 3.3 we know that dmax (S) can be computed by
considering the B D -factorizations of elements in adj(S). The rest of
this section details how to carry out this computation.
Proposition 3.4. Let f ∈ Ap(B; e), s = f + min ord(f ; B D )e, and
s′ = f + ord(f ; B D )e. Then s and s′ are in S, and
(1) adj(s + ke) = f , for k ≥ 0.
(2) dmax (s′ + ke) = d(f ; B D ), for k ≥ 0.
Proof. For 1., by Lemma 2.2, we have that ord(s+ke; S) ≥ min ord(f ; B D ).
Since (s + ke) − (min ord(f ; B D ) + k + 1)e = f − e ∈
/ B, it follows by
Lemma 2.2 that s + ke does not have an S-factorization of length
greater than min ord(f ; B D ), hence ord(s + ke; S) = min ord(f ; B D ).
Hence adj(s + ke) = (s + ke) − (min ord(f ; B D ) + k)e = f .
For 2., notice that adj(s′ ) = f by 1. Using Proposition 3.3, it suffices to show that ord(s′ + ke; S) ≥ ord(f ; B D ). From the proof of
1., we know that ord(s′ + ke; S) = min ord(f ; B D ) + k + (s′ − s)/e =
min ord(f ; B D )+k+ord(f ; B D )−min ord(f ; B D ) = ord(f ; B D )+k.
Corollary 3.5. The adjustment of S is finite, and Ap(B; e) ⊂ adj(S) ⊂
B.
6
LANCE BRYANT AND JAMES HAMBLIN
Proof. That Ap(B; e) ⊂ adj(S) follows from statement 1. of Proposition 3.4. By [7], B = {s − ke : ord(s; S) ≥ k ≥ 1}, and so
adj(S) ⊂ B.
Now we define two sets of B D -factorizations that we need to consider.
Recall that Si is the set of elements of S that are congruent to i modulo
e.
Definition 3.6. Fix 0 ≤ i < e and write adj(Si ) = {u0 , u1 , u2, . . .}
in increasing order. For any b ∈ B, write P(b) for the set of all B D factorizations of b. For uj ∈ Si , define R(uj ) recursively:
• R(u0 ) = P(u0 )
• If j > 0, then
uj − uj−1
D
R(uj ) = x ∈ P(uj ) : |x| < min ord(uj−1; B ) −
.
e
With these definitions, if u ∈ adj(S), then P(u) is the set of all B D
factorizations of u, and R(u) is a subset of P(u) that contains only
those B D -factorizations that have a certain bounded length. As we
will see, it is exactly these sets R(u) that allow us to compute dmax (S).
Theorem 3.7. If n ∈ S, then dmax (n; S) ≤ |R(adj(n))|. In particular,
each set R(adj(n)) is nonempty.
Proof. Let n ∈ S and write n mod a = i. By Proposition 3.3, we have
that dmax (n; S) is the number of B D -factorizations of adj(n) of length
at most ord(n). So it suffices to show that if x is a B D -factorization of
adj(n) of length at most ord(n), then x ∈ R(adj(n)).
As before, write adj(Si ) = {u0, u1 , u2 , . . .} in increasing order, so that
adj(n) = uj for some j. Let x ∈ P(uj ) with |x| ≤ ord(n).
If j = 0, then x ∈ P(uj ) = R(uj ) as desired.
If j > 0, write m = min ord(uj−1; B D ). Since uj−1 < uj , we have
uj−1 = n − ra for some r > ord(n). It follows from Lemma 2.2 that
uj−1 has no B D -factorizations of length at most r. Hence m > r. Also,
uj − uj−1 = (n − ord(n)e) − (n − re) = (r − ord(n))e.
Thus
uj − uj−1
= m − (r − ord(n)) > ord(n).
e
u −u
Therefore |x| ≤ ord(n) < m − j e j−1 and so x ∈ R(uj ) as desired.
m−
The set R(uj ) is defined recursively in terms of R(uj−1 ). In the
following lemma we see that we can consider the elements of R(uj ) in
terms of R(uk ) for any k < j.
THE MAXIMAL DENUMERANT OF A NUMERICAL SEMIGROUP
7
Lemma 3.8. Let 0 ≤ i < e and write adj(Si ) = {u0 , u1 , u2 , . . .} in
increasing order. Let j > 0 and 0 ≤ k < j, and let x ∈ R(uj ). Then
u −u
|x| < min ord(uk ; B D ) − j e k .
Proof. By the definition of R(uj ), we have |x| < min ord(uj−1; B D ) −
uj −uj−1
. If k = j − 1, then we are done. If not, then note that uj−1 =
e
u
−u
uk + re where r = j−1e k > 0.
Write m = min ord(uk ; B D ). Let y = (y0 , y1 , . . . , yt ) ∈ P(uk ) with
|y| = m. Now (y0 + r, y1 , . . . , yt ) is a B D -factorization of uj−1 with
length m + r, and so m + r ≥ min ord(uj−1; B D ). Now
|x| < min ord(uj−1 ; B D ) −
uj − uj−1
e
uj − uk
= m−
,
e
uj − uj−1
e
≤ m+r−
which is the desired result.
We are now ready to prove the connection between maximal denumerants and the R sets.
Theorem 3.9. For all u ∈ adj(Si ), there exists s ∈ Si with dmax (s; S) =
|R(u)|.
Proof. Let uj ∈ Si . Let r be the length of the longest factorization in
R(uj ), and let s = uj + re. Since s − re = uj has a B D -factorization
of length at most r, by Lemma 2.2 we have ord(s) ≥ r.
Suppose ord(s) > r. Then adj(s) = uk for some k < j. From
u −u
Lemma 3.8 we have r < min ord(uk ; B D ) − j e k . Now uj = s − re
and uk = s − ord(s)e, so we have r < min ord(uk ; B D ) − (ord(s) − r),
and so min ord(uk ; B D ) > ord(s). So the shortest B D -factorization of
adj(s) has length greater than ord(s). This is a contradiction, since
by Proposition 3.3 we have that adj(s) always has at least one B D factorization of length at most ord(s). This contradiction proves that
ord(s) = r, and so uj = adj(s).
Now by Proposition 3.3, we know that dmax (s; S) equals the number
of B D -factorizations of uj of length at most r. However, since r is
the length of the longest factorization in R(uj ), it follows that R(uj )
is the set of all B D -factorizations of uj of length at most r. Thus
dmax (s; S) = |R(uj )|.
The next corollaries follow immediately from Theorems 3.7 and 3.9.
8
LANCE BRYANT AND JAMES HAMBLIN
Corollary 3.10. Let S be a semigroup with multiplicity e, and let
0 ≤ i < e. Then dmax (Si ) = maxu∈adj(Si ) |R(u)|.
Corollary 3.11. Let S be a semigroup with the notation defined above.
dmax (S) =
max
0≤i<e,u∈adj(Si )
|R(u)|.
We will now show an example to illustrate how to compute dmax (S)
working with B D -factorizations.
Example 3.12. Let S = h15, 17, 36, 38, 71i.
For this example, we will let i = 11 and compute dmax (S11 ). The
computation of dmax (Si ) for the other values of i is similar.
Now D = {15, 2, 21, 23, 56} and B = h15, 2, 21, 23, 56i. The element
of Ap(B; 15) congruent to 11 mod 15 is f11 = 26. To compute adj(S11 ),
we start with the smallest element of S11 (which is 71) and work up
to f11 + min ord(f11 ; B D ), by Lemma 3.4. Since f11 only has one B D factorization (of length 13), we have the following computation:
s ∈ S11 71 86 101 116 131 146 161 176 191 206 221
ord(s) 1 2
3
4
5
6
7
8
10 11 13
adj(s) 56 56 56 56 56 56 56 56 41 41 26
We see that adj(S11 ) = {26, 41, 56}.
Now R(26) = P(26) is the set of all B D -factorizations of 26, of which
there is only one: (0, 13, 0, 0, 0). So M(26) = 13.
Next, R(41) is the set of B D -factorizations of 41 with length less than
min ord(26; B D ) − 41−26
= 12. There are two of these: (0, 9, 0, 1, 0) and
15
(0, 10, 1, 0, 0). So min ord(41; B D ) = 10.
Finally, R(56) is the set of B D -factorizations of 56 with length less
than 10− 56−41
= 9. There are three of these: (0, 0, 0, 0, 1), (0, 5, 0, 2, 0),
15
and (0, 6, 1, 1, 0).
The largest of these three sets is |R(56)| = 3, so by Corollary 3.10
we have dmax (S11 ) = 3.
4. Additive Semigroups
When computing the maximal denumerant of S using the set adj(S) ⊂
B, we are concerned with B D factorizations with bounded length. However, for elements of adj(S) contained in Ap(B; e), this restriction is
removed and we consider the denumerant of these elements (with respect to D). Thus, we have this result: If adj(S) = Ap(B; e), then
dmax (S) = max{d(f ; B D ) : f ∈ Ap(B; e)}.
THE MAXIMAL DENUMERANT OF A NUMERICAL SEMIGROUP
9
We can refine this even more. Recall from the introduction that u u′
if there exists an s ∈ S such that u + s = u′ . This defines a partial
ordering on S.
Definition 4.1. For a semigroup S and element u ∈ S, we define
maxAp(S; u) = {w ∈ Ap(S; u) : w is maximal in Ap(S; u) \ {0} with
respect to }.
Proposition 4.2. Let S be a semigroup with multiplicity e and blowup
B. If adj(S) = Ap(B; e), then
dmax (S) = max{d(f ; B D ) : f ∈ maxAp(B; e)}.
Proof. Let g ∈ Ap(B; e) \ maxAp(B; e). It suffices to show that there
exists an element f ∈ maxAp(B; e) such that d(f ; B D ) ≥ d(g; B D ).
To do this, we choose f ∈ maxAp(B; e) so that f = g + g ′ where
g ′ ∈ Ap(B; e). Let C = {ci } be the set of B D -factorizations of g and
fix a factorization d of g ′ . The ci + d is a B D -factorization of f for
each i. Furthermore, ci = cj if and only if i = j. Thus, we have
d(f ; B D ) ≥ d(g; B D ).
Proposition 4.2 can greatly reduce the amount of work needed to
compute the maximal denumerant of a semigroup S, especially in light
of the result contained in Corollary 4.5, which states when we need
only compute the denumerant of a single element of B. This follows
from Proposition 4.4 found in [11].
Definition 4.3. The Frobenius number, F (S), of a semigroup S is the
largest integer not contained in S. The semigroup S is called symmetric
if whenever x + y = F (S) for x, y ∈ Z, then exactly one of x and y
belongs to S.
It is easy to verify that N is symmetric, and it is classically known
that any semigroup with embedding dimension 2 is symmetric, see
[10, 25]. For such semigroups, there is only one maximal element in an
Apéry set with respect to the partial ordering .
Proposition 4.4. Let S be a semigroup, u ∈ S, and Ap(S; u) =
{w0 , w1, . . . , wu−1 } where w0 < w1 < · · · < wu−1. Then the following are equivalent.
(1)
(2)
(3)
(4)
S is symmetric
wi + wj = wu−1 whenever i + j = u − 1
w wu−1 for all w ∈ Ap(S; u)
maxAp(S; u) = {F (S) + u}
10
LANCE BRYANT AND JAMES HAMBLIN
Corollary 4.5. Let S be a semigroup with multiplicity e and blowup B.
If adj(S) = Ap(B; e) and B is symmetric, then dmax (S) = d(F (B) +
e; B D ).
The next definition distinguishes an important class of numerical
semigroups, which turn out to be precisely the semigroups we are considering in this section.
Definition 4.6. A numerical semigroup S with multiplicity e is additive if
(5)
ord(u + e; S) = ord(u; S) + 1
for all u ∈ S.
Additive semigroups
semigroups for which the associated graded
L∞ are
i
i+1
m
/m
of the corresponding ring (R, m) =
ring grm (R) =
i=0
e
a1
a2
at
k[[x , x , x , . . . , x ]] with respect to the maximal ideal m is CohenMacaulay.
Proposition 4.7. A semigroup S with multiplicity e and blowup B is
additive if and only if adj(S) = Ap(B; e).
Proof. First assume that S is additive and let fi ∈ Ap(B; e) where fi ≡
i mod e. By Proposition 3.3, si = fi + min ord(fi ; B D )e is an element
of S with adj(si ) = fi . For any integer k, such that si +ke ∈ S, we have
adj(si +ke) = (si +ke)−ord(si +ke; S)e = (si +ke)−ord(si ; S)e−ke =
si − ord(si ; S)e = adj(si ) = fi . It follows that adj(Si ) = fi . Since this
holds for all 0 ≤ i < e, we have adj(S) = Ap(B; e).
Now assume that adj(S) = Ap(B; e) and let u ∈ S. Choose f ∈
Ap(B; e) such that u ≡ f mod e. Then u − ord(u; S)e = f = (u +
e) − ord(u + e; S)e, and we have ord(u + e; S) = ord(u; S) + 1.
We can now restate Proposition 4.2 and Corollary 4.5 as follows.
Theorem 4.8. Let S be an additive semigroup with multiplicity e and
blowup B. Then dmax (S) = max{d(f ; B D ) : f ∈ maxAp(B; e)}.
Corollary 4.9. Let S be an additive semigroup with multiplicity e and
symmetric blowup B. Then dmax (S) = d(F (B) + e; B D ).
Before looking at some examples, there is an important class of additive semigroups in connection to ring theory for which Corollary 4.9
applies.
Definition 4.10. Let S be a semigroup with multiplicity e and Ap(S; e) =
{w0 , w1, . . . , we−1 } where w0 < w1 < · · · < we−1 . The semigroup S is
called supersymmetric if S is additive and, in addition, wi + wj = we−1
and ord(wi ; S) + ord(wj ; S) = ord(we−1 ; S) whenever i + j = e − 1.
THE MAXIMAL DENUMERANT OF A NUMERICAL SEMIGROUP
11
Supersymmetric semigroups
are semigroups for which the associL∞
i
i+1
of the corresponding ring
ated graded ring grm (R) =
i=0 m /m
e
a1
a2
at
(R, m) = k[[x , x , x , . . . , x ]] with respect to the maximal ideal m
is Gorenstein.
By definition, a supersymmetric semigroup S is additive, and it follows from Proposition 4.4 that S is symmetric. We will show that the
blowup semigroup is also symmetric.
Lemma 4.11. Let S be a supersymmetric semigroup. Then the blowup
semigroup of S is symmetric.
Proof. Since S is additive, by Proposition 4.7, adj(Ap(S; e)) ⊂ adj(S) =
Ap(B; e). These sets have the same cardinality, so, in fact, adj(Ap(S; e)) =
Ap(B; e).
Consider f ∈ Ap(B; e) where e is the multiplicity of S. From above,
we know that f = wi − ord(wi )e for some wi ∈ Ap(S; e). Since S is
supersymmetric, we have [wi − ord(wi )e] + [wj − ord(wj )e] = we−1 −
ord(we−1)e where i + j = e − 1. Furthermore wj − ord(wj )e = adj(wj )
and we−1 − ord(we−1 )e = adj(we−1 ) are both in Ap(B; e). Thus f
adj(we−1) for all f ∈ Ap(B; e). This forces adj(we−1 ) to be the largest
element of Ap(B; e), and by Proposition 4.4, B is symmetric.
Theorem 4.12. Let S be a supersymmetric semigroup with multiplicity
e and blowup B. Then dmax (S) = d(F (B) + e; B D ).
We conclude with two more results.
Proposition 4.13. If S is additive and B = N, or equivalently, the difference between the two smallest minimal generators is 1, then dmax (S)
is the denumerant of e − 1 in B with respect to D.
Proof. This follows from Corollary 4.9 since B is symmetric and F (B) =
−1.
Proposition 4.14. If S is generated by an arithmetic sequence, i.e., it
is of the form S = he, e + d, e + 2d, . . . e + tdi where gcd(e, d) = 1, then
dmax (S) is the number of integer partitions of e − 1 using the numbers
1, 2, . . . , t.
Proof. First we note that by [8, 23], S is additive. Moreover, B =
he, d, 2d, . . . , tdi = he, di is symmetric with F (B) = ed − e − d, see [25].
Thus by Corollary 4.9, dmax (S) is the denumerant of F (B)+e = ed−e−
d + e = (e − 1)d with respect to the generating set D = {d, 2d, . . . , td}.
By factoring out d, we obtain our result.
From Proposition 4.14, we see that if S has embedding dimension 3
with multiplicity e, and is generated by an arithmetic sequence, then
12
LANCE BRYANT AND JAMES HAMBLIN
dmax (S) is the number of integer partitions of e − 1 in which all parts
are either 1 or 2. This is given by the formula
l m
e
e−1
.
+1 =
2
2
Thus, for this particular case, we obtain the formula given in Equation
1 in the Introduction using different results than those found in [13].
If S has embedding dimension 4 with multiplicity e, and is generated by an arithmetic sequence, then dmax (S) is the number of integer
partitions of e − 1 in which all parts are either 1, 2, or 3. This is given
by the formula
(e + 2)2
,
12
where [x] denotes the integer nearest to x, see [20]. Hence, we obtain a
formula of the type given in [13] that is valid for embedding dimension
4, albeit, for only a special class of semigroups.
References
[1] Ramı́rez Alfonsı́n, J. L., The Diophantine Frobenius Problem, Oxford Lecture
Series in Mathematics and its Applications 30, Oxford University Press, Oxford, (2005).
[2] Amos, J.; Chapman, S. T.; Hine, N.; Paixão, J. Sets of lengths do not characterize numerical monoids. Integers 7 (2007), A50.
[3] Aguiló-Gost, F.; Garcı́a-Sánchez, P. A. Factorization and catenary degree in
3-generated numerical semigroups. European Conference on Combinatorics,
Graph Theory and Applications (EuroComb 2009), 157-161, Electron. Notes
Discrete Math., 34, Elsevier Sci. B. V., Amsterdam.
[4] —, Factoring in embedding dimension three numerical semigroups. Electron.
J. Combin. 17 (2010).
[5] Barucci, V. (2006). Numerical semigroup algebras. In: Brewer, J., Glaz, S.,
Heinzer, W., and Olberding, B. Multiplicative Ideal Theory in Commutative
Algebra. New York: Springer, pp. 3953.
[6] Bowles, C.; Chapman, S. T.; Kaplan, N.; Reiser, D. On delta sets of numerical
monoids. J. Algebra Appl. 5 (2006), 695-718.
[7] V. Barucci, D. Dobbs, M. Fontana, Maximality properties in numerical semigroups and applications to one-dimensional analytically irreducible local domains. Mem. Amer. Math. Soc. 125:vii-77.
[8] V. Barucci, R. Fröberg, Associated graded rings of one-dimensional analytically irreducible rings, Journal of Algebra 304 (2006) 349–358.
[9] —, Associated graded rings of one-dimensional analytically irreducible rings
II, Journal of Algebra 336 (2011) 279–285.
[10] A. Brauer, On a problem of partitions, Amer. J. Math. 64 (1942), 299-312.
[11] L. Bryant, Filtered numerical semigroups and applications to one-dimensional
rings, PhD thesis, Purdue Univ., 2009.
THE MAXIMAL DENUMERANT OF A NUMERICAL SEMIGROUP
13
[12] —, Goto Numbers of a Numerical Semigroup Ring and the Gorensteiness of
Associated Graded Rings, Comm. Alg. 38 n.6 (2010), 2092 - 2128.
[13] L. Bryant, J. Hamblin, L. Jones, M aximal denumerant of a numerical semigroup with embedding dimension less than four, to appear in Journal of Commutative Algebra.
[14] Chapman, S. T.; Daigle, J.; Hoyer, R.; Kaplan, N. Delta sets of numerical
monoids using nonminimal sets of generators. Comm. Algebra 38 (2010), 26222634.
[15] Chapman, S. T.; Hoyer, R.; Kaplan, N. Delta sets of numerical monoids are
eventually periodic. Aequationes Math. 77 (2009), 273-279.
[16] Chapman, S. T.; Garcı́a-Sánchez, P. A.; Llena, D.; Marshall, J. Elements in
a numerical semigroup with factorizations of the same length. Canad. Math.
Bull. 54 (2011), 39-43.
[17] M. DAnna, V. Micale, A. Sammartano, On the Associated Graded Ring of a
Semigroup Ring, Journal of Commutative Algebra 3 n.2 (2011), 147- 168.
[18] —, When the associated graded ring of a semigroup ring is Complete Intersection, in press.
[19] A. Garcia, Cohen-Macaulayness of the Associated Graded Ring of a Semigroup
Ring, Comm. Algebra 10 (1982), 393-415.
[20] G. Hardy, Some Famous Problems of the Theory of Numbers. Clarendon Press,
1920.
[21] J. C. Rosales, On symmetric numerical semigroups, Journal of Algebra 182
(1996), 422–434. MR 98h:20099
[22] J. Lipman, Stable ideals and Arf rings, Amer. J. Math. 93 (1971), 649685
[23] S. Molinelli and G. Tamone, On the Hilbert function of certain rings of monomial curves, J. Pure Appl. Algebra 101 (1995), no. 2, 191206.
[24] Ramı́rez Alfonsı́n, J. L. The Diophantine Frobenius Problem, Oxford Lecture
Series in Mathematics and its Applications 30, Oxford University Press, Oxford, (2005).
[25] J. Sylvester, Mathematical question with their solutions, Educational Times
41 (1884), 21.
| 0math.AC
|
Tight tradeoffs for approximating palindromes in streams∗
Paweł Gawrychowski1 and Przemysław Uznański2
1
arXiv:1410.6433v5 [cs.DS] 7 Apr 2016
2
Institute of Informatics, University of Warsaw, Poland
Department of Computer Science, ETH Zürich, Switzerland
Abstract
We consider computing the longest palindrome in a text of length n in the streaming model, where
the characters arrive one-by-one, and we do not have random access to the input. While computing the
answer exactly using sublinear memory is not possible in such a setting, one can still hope for a good
approximation guarantee.
We focus on the two most natural variants, where we aim for either additive or multiplicative
approximation of the length of the longest palindrome. We first show that there is no point in considering
Las Vegas algorithms in such a setting, as they cannot achieve sublinear space complexity. For Monte
n
Carlo algorithms, we provide a lower bound of Ω( E
) bits for approximating the answer with additive
log n
error E, and Ω( log(1+ε) ) bits for approximating the answer with multiplicative error (1 + ε) for the
binary alphabet. Then, we construct a generic Monte Carlo algorithm, which by choosing the parameters
appropriately achieves space complexity matching up to a logarithmic factor for both variants. This
substantially improves the previous results by Berenbrink et al. (STACS 2014) and essentially settles the
space complexity.
1
Introduction
A recent trend in algorithms on strings is to develop efficient algorithms in the streaming model, where
characters arrive one-by-one, and we do not have random access to the input. The main goal is to minimize
the space complexity, i.e., avoid storing the already seen prefix of the text explicitly. One usually allows
randomization and requires that the answer should be correct with high probability. We consider computing
the longest palindrome in this model, where a palindrome is a fragment which reads the same in both
directions. This is one of the basic questions concerning regularities in texts and it has been extensively
studied in the classical non-streaming setting, see [1, 9, 15, 17] and the references therein. The notion of
palindromes, but with a slightly different meaning, is very important in computational biology, where one
considers strings over {A, T, C, G} and a palindrome is a sequence equal to its reverse complement (a reverse
complement reverses the sequences and interchanges A with T and C with G); see [10] and the references
therein for a discussion of their algorithmic aspects. Our results generalize to biological palindromes in a
straightforward manner.
Computing the longest palindrome in the streaming model was recently considered by Berenbrink
et al. [2], who developed tradeoffs between the bound on the error and the space complexity for additive
and multiplicative variants of the problem, that is, for approximating the length of the longest palindrome
with either additive or a multiplicative error. Their algorithms were Monte Carlo, i.e., returned the correct
answer with high probability. They also proved that any Las Vegas algorithm achieving additive error E
n
must necessarily use Ω( E
log |Σ|) bits
√ of memory, which matches the space complexity of their solution up to
a logarithmic factor in the E ∈ [1, n] range, but leaves at least two questions. Firstly, does the lower √
bound
still hold for Monte Carlo algorithms? Secondly, what is the best possible space complexity when E ∈ ( n, n]
in the additive variant, and what about the multiplicative version? We answer all these questions.
∗ up
to a logarithmic factor.
1
Related work. The most basic problem in algorithms on strings is pattern matching, where we want
to detect an occurrence of a a pattern in a given text. It is somewhat surprising that one can actually
solve it using polylogarithmic space in the streaming model, as proved by Porat and Porat [18]. A simpler
solution was later given by Ergün et al. [6], and Breslauer and Galil [3]. Similar questions studied in such
setting include multiple-pattern matching [4], approximate pattern matching [5], and parametrized pattern
matching [12].
Pattern matching is also very closely related to detecting periodicities, and in fact Ergün et al. [6]
also developed an efficient algorithm for computing the smallest period, where p is a period of T [1..n] if
T [i] = T [i + p] for all i = 1, 2, . . . , n − p. Also palindromes are closely connected to periodicities. Informally,
two long palindromes occurring close to each other imply a periodicity of the underlying fragment of the text
(to the best of our knowledge, this has been first explicitly stated by Apostolico et al. [1]). Similar insights
have been used to partition the text into the smallest number of palindromes [7, 11] and recognizing the
so-called Palk language [16]. At a very high level, the idea there is to consider longer and longer prefixes
of the text and maintain a succinct description of all palindromic suffixes of the current prefix. Naturally,
our algorithm is based on the same high-level idea, but there are multiple non-trivial technical difficulties
stemming from the fact that we cannot provide random access to the already seen part of the text, so we can
only approximate such information.
Model. We work in the streaming model and consider additive and multiplicative variant of the problem.
The model works as follows: we are first given the length of the text n and the bound on the desired error
E (in the additive variant) or ε (in the multiplicative variant), then the characters T [1], T [2], . . . , T [n] ∈ Σ
arrive one-by-one. In the h-th step we receive T [h] and we are required to output a number `, such that the
length of the longest palindrome in T [1..h] is either between ` and ` + E (in the additive variant) or between
` and (1 + ε) · ` (in the multiplicative variant). We have s(n) bits of memory available, where we can store
an arbitrary data. It is important to remember that the procedure operates in steps corresponding to the
characters and we cannot retrieve an already seen character unless it has been stored in memory.
Now we are interested in the possible tradeoffs between s(n) and the bound on the error. We consider Las
Vegas and Monte Carlo algorithms. A Las Vegas algorithm always returns a correct answer, but its memory
usage s(n) is a random variable. A Monte Carlo algorithm returns a correct answer with high probability,
and its memory usage s(n) does not depend on the random choices, where high probability means 1 − n1c , for
arbitrarily large constant c.
We assume that the memory consists of words of size Ω(log max{n, |Σ|}) and basic operations take O(1)
time on such words. Bounds on the space are expressed in such words unless stated otherwise.
Previous work. The longest palindrome can be found in O(n) time and space (cf. Manacher [17]).
n
Berenbrink et al. [2] constructed a streaming algorithm achieving additive error E using O( E
) space and
√
n1.5
O( E ) total time for any E ∈ [1, n], and a streaming algorithm guaranteeing multiplicative error (1 + ε)
log n
n log n
using O( ε log(1+ε)
) space and O( ε log(1+ε)
) total time for any ε ∈ (0, 1], both Monte Carlo. They also proved
n
that any Las Vegas algorithm with additive error E must necessarily use Ω( E
log |Σ|) bits of space.
Our results. We significantly improve on the previous results as follows and essentially settle the space
complexity of the problem in both variants (see Table 1 for summary).
Firstly, we prove that any Las Vegas algorithm approximating (in either variant) the length of the longest
palindrome inside a text of length n over an alphabet Σ must necessarily use Ω(n log |Σ|) bits of memory (see
Theorem 6.3). Hence Las Vegas randomization is simply not the right model for this particular problem.
Then we move to Monte Carlo algorithms, and prove the following lower bounds on their space complexity:
n
n
• Ω( E
log min{|Σ|, E
}) bits to achieve additive error E with high probability if E ∈ [1, 0.49n] (see
∗
Theorem 6.6),
√
can be strengthened to 0.5n − ω( n), but for the sake of clarity we prefer to state a weaker bound, here and in
subsequent similar places.
∗ This
2
Las Vegas approximation
n
Ω( E
log |Σ|)
Ω(n log |Σ|)
n
O( E
) space, O( nE
Monte Carlo additive approximation
√
n
) time, E ∈ [1, n]
O( E
) space, O(n log n) time, E ∈ [1, n]
n
n
–
Ω( E log min{|Σ|}, E
}) bits, E ∈ [1, 0.49n]
log n
O( ε log(1+ε)
)
n log n
O( ε log(1+ε)
)
1.5
Monte Carlo multiplicative approximation
space,
–
time, ε ∈ (0, 1]
log(nε)
O( log(1+ε)
) space, O(n log n) time, ε ∈ [ n2 , n]
log n
log n
Ω( log(1+ε) log min{|Σ|, log(1+ε)
}) bits, ε ∈ [n−0.98 , n0.49 ]
Table 1: A comparison of previous (on the left, c.f. [2]) and our (on the right) results. Lower bounds are in
bits, and upper bounds in words consisting of Ω(log max{n, |Σ|}) bits.
log n
log n
log min{|Σ|, log(1+ε)
}) bits to achieve multiplicative error (1 + ε) with high probability if
• Ω( log(1+ε)
−0.98
0.49
ε ∈ [n
, n ] (see Theorem 6.7).†
Secondly, we construct a generic Monte Carlo approximation algorithm, which by adjusting the parameters
appropriately matches our lower bounds up to a logarithmic multiplicative factor. In more detail, our algorithm
n
uses O( E
) words of space for any E ∈ [1, n] in the additive variant (see Theorem 3.2 and Theorem 4.6) and
log(nε)
O( log(1+ε) ) words of space for any ε ∈ [ n2 , n] in the multiplicative variant (see Theorem 3.3 and Theorem 4.6).‡
This essentially settles the space complexity of the problem, as it can be seen that our lower and upper bounds
differ by at most a logarithmic factor for any E ∈ [1, 0.49n] and ε ∈ [n−0.98 , n0.49 ]. The time complexity of
our algorithm is always O(n log n) (see Theorem 5.7).
Overview of the methods. As usual in the streaming model, we apply Karp-Rabin fingerprints. We
store such fingerprints for some carefully chosen prefixes of the already seen part of the text. Informally, these
chosen prefixes become more and more sparse as we move closer to the beginning, with the details depending
on the variant. We call such fingerprints of prefixes landmarks, and formalize this notion in Section 2. Then,
in Section 3, we present a generic algorithm. The idea is that for every possible palindrome center we create a
separate process which maintains the corresponding palindromic radius (or, more precisely, its approximation).
By adjusting the parameters of the generic algorithm we are able to guarantee good bound on the error
in both variants. For E ∈ Ω( logn n ) or ε ∈ Ω(1) there are only few landmarks and such generic algorithm
is already efficient enough when implemented naively. To implement it efficiently for smaller E or ε, we
need to avoid running processes which have already found a mismatch, but maintaining such a list explicitly
might take too much space. However, multiple sufficiently long palindromes appearing close to each other
imply periodicity of the corresponding fragment of the text, which can be exploited to concisely describe the
whole situation. This insight (dating back to Apostolico et al. [1]) allows us to approximate the information
about all active processes in logarithmic space as explained in Section 4. Then in Section 5 we use it to avoid
running all active processes after reading every character. Finally, in Section 6 we apply the Yao’s minimax
principle to derive the lower bounds. For Las Vegas algorithms, this is straightforward, but requires more
work for Monte Carlo algorithms.
Comparison with previous √
work. The additive approximation algorithm proposed by
√ Berenbrink
et al. [2] uses a flat structure of n fingerprints (called checkpoints). The most recently seen n characters
are stored explicitly, and the information of all palindromes with larger radius is compressed using the
√
−0.98 can be replaced by any constant larger than −1, and n0.49 can be strengthened to o( n).
log(nε)
log n
small ε this is O( ε ), and for large ε becomes O( log(1+ε) ). Note that this does not contradict the lower bound,
† Here
‡ For
because log(nε) = Θ(log n) for ε ∈ [n−0.98 , n0.49 ].
3
periodicity lemma. For multiplicative approximation, a sparse structure of checkpoints is used. Our technical
contribution is of several flavors. Firstly, we use a single generic construction for both variants of the problem.
Secondly, in all variants we use a hierarchical structure of fingerprints, with the fingerprints becoming more
and more sparse as we move closer to the beginning. This in particular allows us to avoid storing a long
suffix explicitly in the additive version. Thirdly, also the periodicity compression is applied in a hierarchical
manner: we maintain a partition of the text into segments with lengths exponentially increasing with the
distance from the most recently seen character
work,
√ and compress each such segment separately. In previous
√
a rigid partition into segments of length n was used. Storing such rigid partition requires Ω( n) space,
which might be too much when the allowed error is large. Working with segments of exponentially increasing
lengths allows us to decrease this additional memory usage to only O(log n), but also makes the details more
involved.
2
Preliminaries
For a word w ∈ Σ∗ , we denote its length by |w|, and its i-th letter by w[i] for any i = 1, 2, . . . , |w|. Similarly,
w[i..j] denotes the fragment starting at the i-th and ending at the j-th character, and wR denotes the
reversal, i.e., w[|w|]w[|w| − 1]..w[1]. The period per(w) of w is the smallest natural number such that
w[i] = w[i + per(w)] for all i = 1, 2, . . . , |w| − per(w). The well-known periodicity lemma [8] states that if p
and q are periods of w and p + q ≤ |w|, then gcd(p, q) is also a period of w. We focus on detecting palindromes
of even length (odd palindromes can be detected with the standard trick of duplicating every letter, see [1]).
The palindromic radius at c is the largest R(c) such that T [c..(c + R(c) − 1)] = T [(c − R(c))..(c − 1)]R . c is
the center of a palindrome T [(c − R(c))..(c + R(c) − 1)].
Karp-Rabin fingerprints. We use the Karp-Rabin fingerprints [14] to quickly check equality of long
strings. We choose a large prime p ≥ max(|Σ|, poly(n)) and draw x ∈ Zp uniformly at random.Then define
fx (w[1..k]) = (w[1] + w[2]x + . . . + w[k]xk−1 ) mod p and define fingerprint Φ(w) of a word w to consist of |w|,
fx (w), fx (wR ), x|w| , x−|w| , which takes O(1) space if |w| ≤ n. The following operations take O(1) time:
concatenation
erasing a prefix
erasing a suffix
reversal
given Φ(w) and Φ(v), find Φ(wv),
given Φ(wv) and Φ(w), find Φ(v),
given Φ(wv) and Φ(v), find Φ(w),
given Φ(w), find Φ(wR ).
The fingerprints allow us to check if two strings are the same. Formally, we assume that |Σ| ≤ poly(n).
Then, to check if u = v we compare Φ(u) and Φ(v). If u = v then Φ(u) = Φ(v), and if u =
6 v while |u|, |v| ≤ n
n
then Φ(u) = Φ(v) with probability at most poly(n)
. The latter situation is called a false positive. Because
the running time of our algorithms will be always polynomial in n, and we will be operating on strings of
length at most n, by the union bound the probability of a false positive can be made n1c for any c by choosing
exponent in poly(n) large enough. When analyzing the correctness, we assume no false positives.
Landmarks. Our algorithms stores some values Φ(i) = Φ(T [1..i]). The intuition is that after reading T [h]
we calculate the fingerprint of the currently seen prefix and keep it available for some time. A landmark is a
position i such that Φ(i) is currently stored. If additionally i = 2λ · j for some j, i is a landmark on level λ,
and if j is odd then i is a landmark strictly on level λ. Yλ is the set of landmarks on level λ and Y is the
set of all landmarks. Observe that knowing Φ(t) for all t ∈ Y is enough to calculate Φ(T [t + 1 .. t0 ]) for any
t, t0 ∈ Y in O(1) time. Technically, Y depends on the current value of h, and “t is a landmark at h” means
that t ∈ Y just after reading T [h].
For each level of landmarks we fix its size bλ . After reading T [h], the λ-th level consists of the bλ most
recently seen positions of the form 2λ · j, that is, 2λ (b 2hλ c − i) for i = 0, 1, . . . , bλ − 1. The sizes bλ are chosen
differently depending on the desired approximation guarantee. In all versions, the last level has number
L ≤ log n, and b0 = ... = bL−1 , while no restriction is put on bL .
4
Φ(2c − h − 2)
Φ(h)
T [(2c − h − 1)..(c − 1)]
h−c+1
T [c..h]
h−c+1
Figure 1: Checking if R(c) ≥ h − c + 1.
For such choice of landmarks, after increasing h by one we need to add and remove at most one landmark
per level, which takes O(log h) time in total. The landmarks on each level are kept in a random access array
with cyclic addressing, thus using O(bλ + 1) space while allowing accesses and updates in O(1) time.
3
Space-efficient algorithm
We start with the basic algorithm. A proper choice of all bλ guarantees small additive or multiplicative error,
but the time and space complexity might be high. Nevertheless, the basic algorithm serves as a good starting
point for developing first the space efficient version, and then finally the time efficient solution.
The idea of the algorithm is that for every possible center c we create a process P (c), which keeps on
computing the corresponding radius R(c). We call a process alive if it has not found T [c + ∆] such that
T [c−1−∆] 6= T [c+∆] yet, and dead otherwise. The process starts with R(c) = 0 and then uses the landmarks
to update the value of R(c) (and also the final answer) whenever possible. To verify if R(c) ≥ h − c + 1
we need to check if T [(2c − h − 1)..h] is a palindrome. This requires accessing Φ(h) and Φ(2c − h − 2) to
calculate Φ(T [(2c − h − 1)..h]) and then Φ((T [(2c − h − 1)..h])R ), which are then compared to each other, see
Fig. 1. We can simply maintain the current value of Φ(h) but retrieving Φ(2c − h − 2) is only possible when
2c − h − 2 is a landmark. Therefore, the process P (c) can update its R(c) only when 2c − h − 2 is a landmark,
and doing so will be referred to as running P (c) using 2c − h − 2. If T [(2c − h − 1)..h] is a palindrome, we
say that P (c) succeeds, and otherwise fails.
We would like to guarantee that running a process is a O(1) time procedure, so we need to quickly check
if 2c − h − 2 is currently a landmark (and if so, access the stored Φ(2c − h − 2)). This can be easily done by
iterating through all possible levels, but we want a faster method. We consider the last L-th level separately
in O(1) time. For all lower levels, the values bλ are all the same. We compute the largest power of 2 dividing
2c − h − 2, call it 2λ , then 2c − h − 2 cannot be a landmark on level larger than λ. On the other hand, if
2c − h − 2 is a landmark on level λ0 < λ, then it is also a landmark on level λ. Therefore, we only need to
consider the λ-th level. This allows us to run any process in O(1) time, and furthermore the state of any
P (c) can be fully described just by specifying its center c (R(c) is not stored explicitly, unless mentioned
otherwise). Observe that even if a process is dead, there is no harm in running it again.
The basic algorithm simply runs all processes after reading the next T [h] using appropriately defined
landmarks (depending on variant and desired error guarantee). For E ∈ Ω( logn n ) or ε ∈ Ω(1), it needs O(log n)
time to process T [h].
Theorem 3.1. The basic algorithm approximates the longest palindrome with additive error E ∈ [1, n]
n
using O( E
) time and words of memory to process T [h], and the longest palindrome with multiplicative error
log(nε)
2
ε ∈ [ n , n] using O( log(1+ε)
) time and words of memory to process T [h].
Proof. The algorithm runs, after reading every T [h], the process P ( h+y
2 ) for every landmark y. For additive
approximation, the landmarks are defined as in Theorem 3.2 with L = log E. For multiplicative approximation
and n2 ≤ ε ≤ 1, the landmarks are defined as in Theorem 3.3 with D = Θ( 1ε ), but when 1 ≤ ε ≤ n we slightly
change the definition by choosing k = log(1 + ε) and keeping as a landmark, for every i = 1, 2, . . . , logk n , the
last position divisible by 2k·i .
5
An optimized version of the basic algorithm will be referred to as a scheduling scheme. A scheduling
scheme should guarantee that any alive P (c) such that 2c − h − 2 is a landmark is run, unless we can either
be sure that it would fail anyway, or there is another P (c0 ) such that c0 < c and 2c0 − h − 2 is also a landmark,
and we can be sure that it succeeds (in particular, P (c0 ) is still alive).
Theorem 3.2. Any scheduling scheme with bL = ∞ approximates the longest palindrome with additive error
2L .
Proof. Consider an arbitrary palindrome T [(c − x)..(c + x − 1)]. We will show that any scheduling scheme
with bL = ∞ returns at least x − 2L . We can assume x ≥ 2L , then there exists y ∈ (x − 2L , x] such that
2L c − y − 1. Therefore, c − y − 1 is permanently a landmark on level L. P (c) will be alive at c + y − 1, so
by the properties of a scheduling scheme we will run a process detecting a palindrome with radius at least
y > x − 2L . Thus any scheduling scheme with bL = ∞ approximates the longest palindrome with additive
error 2L .
Theorem 3.3. Any scheduling scheme with b0 = b1 = . . . , blog(n/D) = D for D ≥ 6 approximates the longest
palindrome with multiplicative error 1 + O(1/D).
Proof. Consider an arbitrary palindrome T [(c − x)..(c + x − 1)]. We will show that any such scheduling
scheme returns at least x/(1 + O(1/D)). Let λ be the smallest integer such that (D − 1) · 2λ ≥ 2 · x. We have
two cases.
λ = 0 After reading T [c + x − 1], all c + x − 1, c + x − 2, . . . , c + x − D are landmarks on level 0, and c − x − 1
is one of them because 2x < D, so T [(c − x)..(c + x − 1)] or a longer palindrome is detected.
λ > 0 In the interval [c − x − 1, c + x − 1] there are at most 22xλ + 1 ≤ D numbers divisible by 2λ , thus
there exists y ∈ (x − 2λ , x] such that c − y − 1 was a landmark on level λ after reading T [c + x − 1].
As P (c) is still alive at c + x − 1, we will detect a palindrome of radius at least y. In other words, we
will approximate x with additive error 2λ . However, since λ was chosen to be minimal, we have that
x
2 · x > (D − 1) · 2λ−1 , so we can bound the multiplicative error from above by x−2
λ , which is at most
1
1 + Ω(D) for D ≥ 6.
Therefore, any such scheduling scheme approximates the longest palindrome with multiplicative error
1 + O(1/D).
Lemma 3.4. If bλ ≥ 12 for all λ ≤ L, then for any ∆ = 2` − 1 and for any c there is at least one
h ∈ [c + 5∆, c + 6∆] such that 2c − h − 2 is a landmark at h.
Proof. Observe that there exists unique h ∈ [c+5∆, c+6∆] such that 2` 2c−h−2. Because h−(2c−h−2) ≤
2 + 12 · ∆ < 12 · 2` , after reading T [h] there are two possibilities. If ` ≤ L then 2c − h − 2 is among the 12
last seen positions divisible by 2` . If ` > L then 2c − h − 2 is definitely a landmark anyway.
4
Maintaining the alive processes
To implement a scheduling scheme, we want to know which processes are alive. Maintaining them explicitly
is too space-expensive, though. Therefore, we will store a compressed approximate representation of all alive
processes. Intuitively, we will group together nearby alive processes using the notion of a partition scheme
described below. The representation will not be exact in the sense that it might report some dead processes
as alive (but then they will have some additional properties). Such information is not providing any speedup
by itself yet, but later in Section 5 we will use it to implement any scheduling scheme efficiently. In this
section, we focus on maintaining the information.
6
Partition scheme. We maintain a partition of T [1..h] into disjoint segments stored in a linked list. The
length of every segment is a power of 2, their lengths are nonincreasing as one moves to the right, and there
is M such that we have between A and B segments of length 2` for every ` = 0, 1, . . . , M − 1, and between 1
and B segments of length 2M . A and B are constants to be specified later. After increasing h by one, a new
segment of length 20 appears, then we possibly take two adjacent segments of length 2` such that the segment
on their left (if any) is longer, and merge them into one segment of length 2`+1 . We call this a partition
scheme, as there is some flexibility as to when the merging happens.
Lemma 4.1. There is a partition scheme with A = 3 and B = 5, which guarantees that after adding a new
segment of length 20 we can merge in O(1) time at most one pair of adjacent segments of length 2` , such that
there are 3 segments of the same length 2` on their right.
Proof. This is a simple example of the recursive slow-down method of Kaplan and Tarjan [13]. Let
2a1 , 2a2 , 2a3 , . . . be the lengths of the segments in the current partition, where a1 ≤ a2 ≤ a3 ≤ . . .. We group
together all segments with the same length, and denote the number of segments of length 2` by c` . We will
show how to maintain c` ∈ {3, 4, 5} for every ` = 0, 1, 2, . . . , M − 1 and cM ∈ {1, 2, 3, 4, 5}, where 2M is the
maximum length of a segment in the current partition. To this end, we will keep the following invariant: if
ci = 5 then there exists j ∈ {0, 1, . . . , i − 1} such that cj = 3 and cj+1 = . . . = ci−2 = ci−1 = 4. We call such
partition valid.
We must show that, given a valid partition of T [1..h], we can construct in O(1) time a valid partition of
T [1..h + 1]. We start with creating a new segment of length 20 and adding it to the previous partition, which
increases c0 by one. Now there are two cases.
c0 = 5 We merge two (leftmost) segments of length 20 into a segment of length 21 , or in other words we
decrease c0 by two (c0 is now equal to 3) and increase c1 by one. Because the initial value of c0 was 4,
the only way the invariant could have been broken is that c1 was 3, c2 = . . . = ci−1 = 4 and ci = 5 for
some i ≥ 3. But then the new c1 becomes 4, and all c2 , c3 , . . . , ci−1 are now 4, so the invariant holds.
c0 = 4 If there is i such that c1 = . . . = ci−2 = ci−1 = 4 and ci = 5, then we merge the two (leftmost)
segments of length 2i into a segment of length 2i+1 , which decreases ci by two and increases ci+1
by one. As in the previous case, the only way the invariant could have been broken is that ci+1
was 3, ci+2 = ci+3 = . . . = cj−1 = 4 and cj = 5 for some j ≥ i + 2. Then ci becomes 3, and all
ci+1 , ci+2 , . . . , cj−1 are now 4, so the invariant holds.
To implement the update, we group together all consecutive i’s with the same value of ci . In other words,
we store a list of lists of segments. This allows us to find i from the second case in O(1) time.
Instead of storing every alive P (c) explicitly, for every segment we group together all alive processes such
that c lies inside. We need the following result, which follows from a definition of a palindrome, see [1].
Lemma 4.2. If c < c0 , c0 − c ≤ 2` and R(c), R(c0 ) ≥ 2` , then 2(c0 − c) is a period of T [(c − 2` )..(c0 + 2` − 1)].
The intuition is that in a segment of length 2` either there are at most 4 alive processes which can be kept
explicitly, or there are at least 5 of them and the whole segment is periodic with period at most 2`−1 . Hence
for every segment we store either a sparse or a dense description, depending (roughly) on the periodicity of
the corresponding fragment.
Sparse description. We explicitly store a list of all processes inside the segment, which can be potentially
still alive. We guarantee that there are at most 4 processes on that list, and that if a process is not on
the list, it is surely dead. We do not guarantee that all processes on the list are still alive, but whenever
we run one of them and the check fails, we declare it dead and remove from the list. The processes
currently on the list are called relevant.
7
P (c)
P (c + 4|w|)
s
w
wR
w
wR
wR
w
w
wR
w
wR
Figure 2: Alive processes inside s with a dense description are of the form P (c + α|w|).
Dense description. We guarantee that there exists a word w such that |w| ≤ 14 2` for which the whole
segment of length 2` is a subword of (wwR )∞ , see Fig. 2. Denoting the segment by s, this implies that
per(s) ≤ 12 |s| and s has a palindromic subword of length per(s). In such a case we store a multiple of
the period, denoted by p = k per(s) ≤ 12 2` , such that the only alive processes inside the segment are of
the form P (c + α p2 ) for α ≥ 0, where T [c..(c + p − 1)] is an even palindrome fully within the segment.
(We do not require that all such processes are still alive.) We store c and p, which is enough to run any
relevant process inside s in O(1) time, where relevant means of the form P (c + α p2 ). No other process
inside s can be alive.
We use Lemma 4.1 to maintain a partition of T [1..h] into segments. The description of every segment
requires just O(1) space, making the total additional space complexity O(log n). After reading T [h] we create
a sparse description of the new segment of length 20 and then need to merge at most one pair of adjacent
segments. After having updated the partition, we can simply run all relevant processes. Therefore, now we
need to show how to merge a pair of adjacent segments s and s0 of length 2` as to obtain a new segment
ss0 . If their descriptions are sparse, we merge the lists of s and s0 and either get at most 4 processes, which
constitute a valid sparse description, or at least 5 processes P (c1 ), P (c2 ), . . . , P (c5 ). The following observation
follows from the Lemma 4.1.
Observation 4.3. When a segment of length 2` is being created, the number of already seen characters on
its right is at most 3 · 2`−1 + 5(2`−1 − 1) = 2`+2 − 5. When it is being destroyed, there are at least 3(2`+1 − 1)
of them.
Hence any P (ci ) could have been run everywhere in the interval [ci + 2` + 2`+2 − 5, ci + 3(2`+1 − 1)],
and by Lemma 3.4 had at least one landmark available in that interval, so its radius must be at least
2` + 2`+2 − 4 ≥ 2`+1 . Therefore, we have a list of 5 processes inside a segment of length 2`+1 , all of which
have radii at least 2`+1 (` = 0 must be considered separately). This suffices to construct a dense description
by the following lemma.
Lemma 4.4. Given a list of m ≥ 5 processes P (c1 ), P (c2 ), . . . , P (cm ) inside a segment of length 2` , such
that their radii are all at least 2` and no other process inside is alive, we can construct in O(m + log n) time
a dense description.
Proof. We rearrange the processes so that c1 < c2 < . . . < cm and define ∆i = ci+1 − ci . Every 2∆i is a
period of the segment by Lemma 4.2. We claim that by the periodicity lemma also gcd(2∆1 , 2∆2 , . . . , 2∆m−1 )
is a period of the segment. This can be seen by the following reasoning: if the radii at c < c0 < c00 are
all at least 2` , c00 − c ≤ 2` , 2d 2(c0 − c) is a period of T [(c − 2` )..(c0 + 2` − 1)] and 2d0 2(c00 − c0 ) is
a period of T [(c0 − 2` )..(c00 + 2` − 1)], then by the periodicity lemma 2 gcd(d, d0 ) is a period of the whole
T [(c−2` )..(c00 +2` −1)]. Then by induction p = 2 gcd(∆1 , ∆2 , . . . , ∆m−1 ) is a period of T [(c1 −2` )..(ck +2` −1)],
which contains the whole segment inside. Because ∆1 + ∆2 + . . . + ∆m ≤ 2` and m ≥ 5, ∆i ≤ 14 2` for at
least one i, so p ≤ 12 2` and consequently p must be a multiple of per(s).
Now we can construct a dense description. We compute p in O(m + log n) with m − 1 applications of the
Euclidean algorithm, and set c = c1 . Because p 2∆i for every i, all ci are of the form c + α p2 . Furthermore,
because p ≤ min(∆1 , ∆2 ), c + p ≤ c2 , so T [c..(c + p − 1)] is fully within the segment. Finally, we must argue
that T [c..(c + p − 1)] is an even palindrome. First observe that T [(c3 − p)..(c3 + p)] lies fully within the
segment, and consider two cases.
8
• If c3 = c + αp, then T [(c3 − p)..(c3 + p)] = T [c..(p − 1)]2 . Because the palindromic radius at c3 is at
least 2` ≥ p, T [c..(c + p − 1)] is a palindrome.
• If c3 = c + p2 + αp, then T [(c3 − p2 )..(c3 + p2 )] = T [c..(c + p − 1)]. Because the palindromic radius at c3
is at least 2` ≥ p2 , T [c..(c + p − 1)] is a palindrome.
This settles the situation when both descriptions are sparse. Before we move to the remaining case, we
need an additional tool. If a description of a segment is dense, we maintain some additional information about
the processes inside. Informally, we would like to know which of them are still alive, but of course we cannot
afford to explicitly maintain such information. We can only afford to store a short buffer, where we keep
information about a few most recently run processes. Formally, the buffer is a list of processes P (c) together
with their corresponding values of R(c). We do not require that P (c) is still alive, so it might have happened
that it has been run again after reading T [h0 ] with h < h0 , but the more recent run was unsuccessful. The
buffer is updated whenever we successfully run a process P (c) inside the segment. There either P (c) was in
the buffer, so we move it to the front and update the corresponding R(c), and otherwise we prepend it to the
buffer together with the current R(c), and if the length of the buffer is now 6 we remove the last element
from there. Hence the buffer is of length at most 5. A less trivial consequence is as follows.
Lemma 4.5. If a segment with dense description of length 2` is being destroyed while at most 4 processes in
its buffer have radii at least 2`+1 , then no other process inside the segment can be still alive.
Proof. By Observation 4.3 and how we process segments with dense descriptions, any P (c) which might be
still alive could have been run everywhere in the interval [c + 2` + 2`+2 − 5, c + 3(2`+1 − 1)], and by Lemma 3.4
it had at least one landmark available in that interval. Also, whenever we run any P (c) inside the segment in
the interval [c + 2` + 2`+2 − 5, ∞), and it succeeds, R(c) is set to at least 2`+1 (except when ` = 0, but then
there is just one process inside the segment, so the buffer surely contains it). Therefore, if the buffer contains
at most 4 processes with radii at least 2`+1 , any P (c) such that R(c) ≥ 2`+1 is stored in the buffer, and no
other process can be still alive.
If at least one description is dense, by applying Lemma 4.5 to s (if its description is dense) or s0 (if its
description is dense), we either get that one of these segments contains at least 5 processes with radii at
least 2`+1 in its buffer, or we get a list of at most 4 potentially still alive processes inside each segment. In
the latter case we concatenate the lists to get a list of at most 8 processes inside ss0 such that all other
processes inside are dead. If the list contains at most 4 processes, we construct a sparse description of ss0 ,
and otherwise we apply Lemma 4.4 to construct a dense description of ss0 in O(log n) time. In the former
case we get a list of between 5 and 10 alive processes P (c1 ), P (c2 ), . . . , P (cm ) inside ss0 . It might be the case
that there are also some other processes inside the segment which are still alive, but they are not stored in the
buffer of the corresponding segment. Nevertheless, proceeding as in the proof of Lemma 4.4 we can compute
in O(log n) time p and c such that T [c..(c + p − 1)] is an even palindrome fully within ss0 , all ci are of the
form c + α p2 , and p ≤ 12 |ss0 | is a period of ss0 . This is not a valid dense description yet, as s or s0 (or both)
might have dense descriptions, and we cannot guarantee that all alive processes there are of the form c + α p2 .
Consider the case when s has a dense description, meaning that we have p0 and c0 such T [c0 ..(c0 + p0 − 1)]
0
is an even palindrome fully within s, all alive processes there are of the form c0 + α p2 , and p0 ≤ 12 |s| is a
period of s. If p0 p there is nothing to do. Otherwise, because the list P (c1 ), P (c2 ), . . . , P (cm ) contains at
least 5 processes inside s we have p ≤ 21 |s| and by the periodicity lemma gcd(p, p0 ) is a period of s. Then
gcd(p, p0 ) must be actually a period of the whole ss0 . Now we claim that p can be, in fact, replaced by
gcd(p, p0 ). This is because if a power of a word is a palindrome, the word itself must be a palindrome, so
T [c0 ..(c0 + gcd(p, p0 ) − 1)] is an even palindrome.
The case when s0 has a dense description, or both s and s0 have dense descriptions, can be dealt with
similarly.
Theorem 4.6. For any scheduling scheme with bλ ≥ 12 for all λ ≤ L, descriptions of all segments in the
current partition of T [1..h] can be maintained in O(log n) space and O(log n) time plus the time to run all
relevant processes.
9
5
Time-efficient algorithm
The simulation from the previous section was space-efficient, but not time-efficient yet, because there might be
segments with dense descriptions and small periods, which in turn requires running many relevant processes.
This is the only reason the time to process T [h] might exceed O(log n), as merging at most one pair of
segments and running the processes in all segments with sparse descriptions takes just O(log n) time. In this
section we show how to simulate running all relevant processes in a segment with dense description in O(1)
time.
Consider a dense description of a segment s. Recall that it consists of c and p, such that T [c..(c + p − 1)]
is an even palindrome and p is a period of the whole segment, and we want to run all processes P (c0 ) inside s
of the form c0 = c + α p2 , where 2c − h − 2 is a landmark. We can construct and run all relevant processes in
O(1) time each, but there might be many of them. However, there are only two consequences of running
such a P (c0 ): we might update the final answer, and we might also store it in the buffer (or move it to the
front there). Therefore, if we can guarantee that a particular P (c0 ) will fail anyway, we can avoid running it
altogether. Similarly, if we can guarantee that many processes P (c0 ) will succeed, it is enough to run just the
5 leftmost of them. We will build on these observations to simulate running all processes of such form in a
single segment with a dense description in O(1) total time. This is the most technical part, so we start with
an overview.
Overview. We start with observing in Lemma 5.1 that, when considering such a segment, just a constant
number of associated landmark levels needs to be considered. Then we analyze which relevant processes
inside a segment should be run because of a landmark on level λ. After some basic arithmetical manipulation,
we get a succinct description of all such values of c0 . To avoid considering all of them, which might be too
costly, we apply two lemmas characterizing the structure of palindromes in a sufficiently periodic fragment of
the text, described in Lemma 5.5 and Lemma 5.6 (these observations go back to [1], but we need a slightly
different formulation). To apply them, we need to compute how far the periodicity of a segment with a dense
description continues to the left and to the right. To this end, we relax the notion of landmarks, introducing
the so-called ghost landmarks, which allow us to operate on a longer suffix of the already seen T [1..h]. Then,
using the ghost landmarks, we binary search to compute how far the periodicity extends, and apply the
structural results to isolate at most 5 relevant processes, which should be run as to guarantee the correctness.
To achieve the final complexity of O(log n) to process T [h], we precompute how far the periodicity continues
when creating the segment, and then maintain this information in O(1) time.
Associated landmark levels. Consider a segment s. If, for some c inside s, 2c − h − 2 is a landmark
strictly strictly on level λ at h, we say that λ is a landmark level associated to s.
Lemma 5.1. There are at most 4 landmark levels associated to a single segment, and they can be all
determined O(log n) time.
Proof. Consider a segment s of length 2` and any c inside. By Observation 4.3, when the segment is being
created by merging two segments of length 2`−1 we have h − c ≥ 3(2` − 1). Similarly, when the segment is
being destroyed by merging with an adjacent segment of length 2` to form a segment of length 2`+1 we have
h − c < 2`+1 + 2`+3 − 5 = 5(2`+1 − 1). Consequently, we can bound 2(h − c + 1), which is the number of
already seen characters on the right of 2c − h − 2, as follows:
2(h − c + 1)
2(h − c + 1)
< 10 · 2`+1 − 8
≥ 6 · 2` − 4
If 2c − h − 2 is a landmark strictly on level λ < L, then the number of already seen characters on its right
belongs to [b0 · 2λ−1 , b0 · 2λ ). Bounding the number of different landmark levels associated to s requires
counting λ < L such that [b0 · 2λ−1 , b0 · 2λ ) ∩ [6 · 2` − 4, 10 · 2`+1 − 8) 6= ∅. The condition translates into:
b0 · 2λ−1
b0 · 2 λ − 1
≤ 10 · 2`+1 − 8 − 1
≥ 6 · 2` − 4
10
which is equivalent to 2λ · b0 ∈ [6 · 2` − 3, 40 · 2` − 18]. If λ = L, the number of already seen characters on
the right is at least b0 · 2λ−1 , so the condition becomes 2λ · b0 ≤ 40 · 2` − 18. All in all, there are at most 4
different possible values of λ.
Generating the landmark levels associated with a given segment can be done in O(log n) time by performing
the above calculation.
Due to the above Lemma 5.1, to achieve the claimed O(log n) time complexity for processing T [h], we only
need to show how to run all relevant processes inside a segment with a dense description using landmarks on
a particular level λ associated to that segment in O(1) time.
Relevant processes. We need to consider all relevant processes P (c0 ), such that 2c0 − h − 2 is a landmark
on level λ, implying that 2λ 2c0 − h − 2. The condition is equivalent to:
α · p = h + 2 − 2c (mod 2λ )
(1)
which, denoting 2` = gcd(p, 2λ ), is in turn equivalent to:
α·
h + 2 − 2c
p
=
2`
2`
(mod 2λ−` )
(unless 2` does not divide h + 2 − 2c, when no c0 needs to be considered), so by computing the multiplicative
inverse we finally get a base solution to (1):
h + 2 − 2c p −1
α0 =
· `
(mod 2λ−` )
2`
2
and the general solution is:
α = α0 + t · 2λ−`
for t ∈ {. . . , −1, 0, 1, . . .}
Thus we also get the solution to the original equation:
p
1
c0 = c00 + t · 2λ−` = c00 + t · lcm(2` , p)
2
2
p
where c00 = c + α0 .
2
(2)
Therefore, with a simple calculation we get a succinct description of all values of c0 which should be taken
into the account. Before we proceed further, let us comment on the complexity of the calculation. Since λ is
fixed, both values of
p −1
2` = gcd(p, 2λ ) and
mod 2L−`
2`
can be computed in O(log n) time when we create the segment and stored there.
The situation now is that we have a dense description of a segment, and want to run all processes P (c0 ) of
the form (2) inside the segment. Additionally, because we do not necessarily have all possible landmarks on
level λ, just a few most recent, we are interested only in sufficiently large c0 . Observe, that we can analyze
separately processes of the following two forms:
P (c00 + t · lcm(2λ , p))
P (c000 + t · lcm(2λ , p))
(3)
where c000 = c00 +
1
2
lcm(2λ , p)
(4)
From now on we will only consider the former, as the whole reasoning still holds after replacing c00 by c000 .
We will also assume that only t ≥ 0 need to be considered, which can be ensured by decreasing c00 by an
appropriate multiple of lcm(2λ , p).
Because p is a period of the whole segment, lcm(2λ , p) is its period as well, and furthermore we can assume
that lcm(2λ , p) ≤ 12 |s|, as otherwise there are just at most two relevant processes to run. Intuitively, knowing
how far the period extends to the left and to the right allows us to restrict the number of processes to run by
an argument based on the combinatorial properties of palindromes. While computing how far the period
extends exactly is not possible in our setting, it can be approximated quite well using the landmarks. First,
we need to introduce the notion of ghost landmarks.
11
T [1..h]
2λ
∆
w
w
w
w
Figure 3: A number of repetitions of w such that 2λ |w| implies that |w| is a period of a certain full fragment
between two ghost landmarks.
Ghost landmarks. For every level of landmarks λ, we store fλ = 4 · bλ most recently seen landmarks
on level λ. We call them ghost landmarks on level λ. All ghost landmarks can be maintained in the same
manner as the regular landmarks, so storing them does not change the complexity of our algorithm.
Lemma 5.2. If λ is a landmark level associated to a segment s, then for any c inside s there exists at least
one ghost landmark on level λ in T [1..(2c − h − 2)].
Proof. Consider a segment s of length 2` . By Observation 4.3, the number of already seen characters on the
right of s when it is being created is at least 3(2` −1). Let c0 be any position inside s causing λ to be associated
to s, i.e., 2λ 2c0 −h0 −2, and denote x0 = 2c0 −h0 −2. Notice that h0 might be either smaller or larger than the
current h. Because c0 is a landmark on level λ at h0 , we have that 2λ ·bλ ≥ h0 −x0 = 2(h0 −c0 )+2 ≥ 2+6(2` −1).
Now consider any c inside s and denote x = 2c − h − 2. Since c and c0 both belong to the same segment of
length 2` , c0 −c ≤ 2` −1. Applying Observation 4.3 again, we also get that h−h0 ≤ 5(2`+1 −1)−1−3(2` −1) =
7 · 2` − 3.
Thus the number of already seen characters on the right of x can be bounded as follows:
h − x = 2h − 2c + 2 ≤ (2h0 − 2c0 + 2) + 2(2` − 1) + 14 · 2` − 6.
8
0
0
Because h0 − x0 = 2(h0 − c0 ) + 2 ≥ 2 + 6(2` − 1), we have 16 · 2` − 32
3 ≤ 3 · (h − x ), so the above bound can
be rewritten as:
11
8
11
8
h−x≤
· (h0 − x0 ) + ≤
bλ · 2λ + ≤ (4bλ − 1) · 2λ
3
3
3
3
where the last inequality holds because bλ ≥ 12. Since the leftmost ghost landmark on level λ has at least
(4bλ −1)·2λ already seen characters on its right, by the above calculation it must be on the left of x = 2c−h−2
as claimed.
Now going back to approximating how far the period extends to the left and to the right, we proceed as
follows. We choose w of length lcm(2λ , p) starting at T [c00 ]. Because we have adjusted c00 so that only t ≥ 0
need to be considered and |w| ≤ 12 |s|, we can assume that w is fully within the segment. We know that the
whole segment can be covered by repeating w to the left and to the right (where, possibly, the last repetition
is a suffix or a prefix of w, respectively), and would like to figure out how far we can continue that until we
hit either a boundary of the already seen T [1..h], or a subword of length |w| which is different than w. This
can be approximated quite well using the ghost landmarks, if w repeats at least twice.
Lemma 5.3. For any w such that T [i..(i + 2|w| − 1)] = w2 , 2λ |w|, and T [1..(i − 1)] contains at least one
ghost landmark on level λ, we can compute in O(log h) time r ≥ 2 such that T [i..(i + r|w| − 1)] = wr and
either i + (r + 2)|w| > h or T [i..(i + (r + 2)|w| − 1)] 6= wr+2 .
Proof. By the assumption about ghost landmark on level λ, we can access any Φ(2λ · j) with j ≥ i−1
in
2λ
λ
O(1) time. Hence if we are lucky and i = 2 · j + 1, we can compute Φ(T [(i + α|w|)..(i + β|w| − 1)] for any
0 ≤ α ≤ β in O(1) time, which allows us to binary search for r in O(log h) time. In more detail, to check if
T [i..(i + r|w| − 1)] = wr we check if |w| is a period of T [i..(i + r|w| − 1), which can be done by comparing
Φ(T [(i + |w|)..(i + r|w| − 1)]) and Φ(T [i..(i + (r − 1)|w| − 1)]).
12
In the general case, let i = 2λ · j + 1 + ∆, where ∆ ∈ [0, 2λ ). If T [i..(i + r|w| − 1)] = wr and r ≥ 2, then |w|
is a period of T [(2λ · j + 2λ + 1)..(2λ · j + 2λ + α|w|)], see Fig. 3, where α = r − 1. In the other direction, if |w| is
a period of T [(2λ ·j +2λ +1)..(2λ ·j +2λ +α|w|)] and r ≥ 2, then r ≥ α. (The assumption that r ≥ 2 is crucial.)
Hence we can determine the largest α such that |w| is a period of T [(2λ · j + 2λ + 1)..(2λ · j + 2λ + α|w|)] in
O(log h) time using ghost landmarks on level λ, and then simply return α, which guarantees r ∈ {α, α+1}.
Lemma 5.4. For any w such that T [i..(i + 2|w| − 1)] = w2 , 2λ |w|, and T [1..(2i − h − 2)] contains at least
one ghost landmark on level λ, we can compute in O(log h) time ` ≥ 0 such that T [(i − `|w|)..(i − 1)] = w`
and either i − (` + 2)|w| < 2i − h − 2 or T [(i − (` + 2)|w|)..(i − 1)] 6= w`+2 .
Proof. The proof will be very similar to the proof of Lemma 5.3, except that we have to take into the account
the fact that while there might be many more repetitions of w to the left, we might not have enough ghost
landmarks on level λ to detect them.
Let i = 2λ · j + 1 + ∆, where ∆ ∈ [0, 2λ ). If T [(i − `|w|)..(i − 1)] = w` , then |w| is a period of
T [(2λ · j − α|w| + 2λ + 1)..(2λ · j + |w| + 2λ )], where α = `. In the other direction, if |w| is a period of
T [(2λ · j − α|w| + 2λ + 1)..(2λ · j + |w| + 2λ )], then ` ≥ α − 1. So we only need to binary search for the largest
α such that |w| is a period of T [(2λ · j − α|w| + 2λ + 1)..(2λ · j + |w| + 2λ )] and return max(0, α − 1). The
remaining difficulty is that 2λ · j − α|w| + 2λ might lie too far on the left to be a ghost landmark on level λ,
so the binary search needs to be slightly modified. We first choose the largest α0 such that 2λ · j − α0 |w| + 2λ
is a ghost landmark on level λ. There are two possibilities.
1. |w| is a period of T [(2λ · j − α0 |w| + 2λ + 1)..(2λ · j + |w| + 2λ )], then the largest α might exceed α0 .
But we can return ` = max(0, α0 − 1), because then i − (` + 1)|w| ≤ i − α0 |w|, and the choice of α0
and the assumption, by Lemma 5.2, implies i − (α0 + 1)|w| < 2i − h − 2, so i − (` + 2)|w| < 2i − h − 2.
2. Otherwise, we binary search over all α ≤ α0 , and return max(0, α − 1).
We can binary search for α0 in O(log h) time, so the total time is O(log h).
We apply Lemma 5.3 and Lemma 5.4 to approximate how many times w repeats on its right and on its
left in T [(2c00 − h − 2)..(c00 − 1)] with accuracy 1, assuming that w2 occurs at T [c00 ]. Notice that there might
be many more repetitions to the left in the whole T [1..(c00 − 1)], but Lemma 5.4 does not allow us to detect
all of them. Now the crucial insight is that even though we do now know the exact number of repetitions, we
can iterate through the at most 4 possible combinations of the number of of repetitions to the left and to the
right right, and the additionally consider the possibility that there is only a single occurrence of w in the
segment. Hence we need to iterate through 5 possibilities in total. For each such combination, we will restrict
the number of processes which should be run, therefore by running the processes determined for each of these
combinations we will not lose the correctness. Hence from now on we assume that we know the exact number
of repetitions of w to the left and to the right.
We need the following two simple structural results, which allow us to bound the palindromic radius in a
sufficiently periodic subword of the text. A similar (in spirit) argument appeared already in [1], but we need
a slightly different formulation. We say that a palindrome centered at c reaches h if R(c) ≥ h − c + 1.
Lemma 5.5. Consider uwk v starting at position i in T [1..h], where |u| = |w| = |v|, w is a palindrome, and
u, v 6= w. For any α ∈ {1, 2, . . . , k}, if the palindrome centered at i + α|u| reaches h then α = k2 + 1.
Proof. Take any α ∈ {1, 2, . . . , k}. For a palindrome centered at i + α|u| to reach c, R(i + α|u|) must be at
least min(α − 1, k + 1 − α)|u|. But then either u = wR or v = wR , which is a contradiction.
Lemma 5.6. Consider uwk starting at position i in T [1..h], where |u| = |w|, w is a palindrome, u 6= w, and
h − i − (k + 1)|u| + 1 < |w|. For any α ∈ {1, 2, . . . , d k2 e}, the palindrome centered at i + α|u| cannot reach h.
Additionally, either all palindromes centered at i + α|u| with α ∈ {d k2 e + 1, . . . , k} reach h, or none of them
do.
13
T [1..h]
< |u|
u
w
w
R(i + 4|u|)
w
w
w
v
R(i + 4|u|)
Figure 4: All palindromes centered at positions i + α|u| with α ∈ {d k2 e + 1, . . . , k} reach h if v is a prefix of w.
Proof. Take any α ∈ {1, 2, . . . , k}. If α ≤ d k2 e, then because u 6= w the radius at i + α|u| is too small for
the palindrome centered at i + α|u| to reach h. Otherwise, let T [i..h] = uwk v, where |v| < |u| because
h − i − (k + 1)|u| + 1, see Fig. 4. Now either v is not a prefix of w, and we actually get the situation from
Lemma 5.5, so only α = k2 + 1 can possibly correspond to a palindrome reaching h, or v is a prefix of w, and
for all α ≥ d k2 + 1e the palindrome centered at i + α|u| reaches h.
Recall that we want to run all P (c00 + t|w|) inside the segment with t ≥ 0, and our w starts at T [c00 ]. We
know that w repeats ` times to the left in T [(2c00 − h − 2)..(c00 − 1)] and r times to the right till the end of the
already seen T [1..h]. The actual number of repetitions of w to the left in the whole T 1..(c00 − 1)], denoted `0 ,
might be larger than `. By Lemma 5.5 and Lemma 5.6, either all processes of the form P (c00 + (−`0 + α)|w|)
0
`0 +r
with α ≥ d ` +r
2 e + 1 will succeed, or just the one with α =
2 + 1 will succeed. Because the size of the
buffer is 5, we only need to ensure that the 5 leftmost processes which will succeed are run. To guarantee
this, we run all processes of the form P (c00 + (max(−` + d `+r
2 e + 1, 0) + x)|w|) for x = 0, 1, 2, 3, 4 which are
still inside the segment. This is correct, as following two cases show.
0
1. The process P (c00 + (−`0 + α)|w|) with α = ` +r
2 + 1 is on the left of the segment, so either all or none
processes of such form in the segment are alive.
0
0
2. The process P (c00 + (−`0 + α)|w|) with α = ` +r
2 + 1 is inside the segment, so ` cannot be too large.
0
0
More precisely, ` ≤ r, and consequently ` = ` .
We run a constant number of processes, each of them in O(1) time, but to ensure that every segment is
processed in such complexity, we also need to remove the binary search used to approximate how many times
w can be repeated to the left and to the right.
Recall that |w| = lcm(2λ , p), w starts at T [c00 ] and lies fully within a segment s, and furthermore |w| is a
period of the whole s. As mentioned before, we can also assume that |w| ≤ 12 |s|, as otherwise there are at
most two processes which might need to be run. We can compute how many times w can be repeated to its
left (or rather approximate this value as described in Lemma 5.4) when the segment is created, as the result
does not depend on the current value of h. Similarly, we can compute how many times it can be repeated to
the right when we create the segment, but here the important difference is that we might continue till the
very end of the current T [1..h], i.e., the next copy of w might extend beyond the current prefix T [1..h]. It
can be seen that in such a case the next time we need to deal with the same segment, at most one additional
copy of w fits inside T [1..h]. This happens because the segment is relevant when 2` h + 2 − 2c, and |w| ≥ 2` .
Therefore, the number of times w repeats to the right can be maintained in O(1) time.
Theorem 5.7. Any scheduling scheme with bλ ≥ 12 for all λ ≤ L can be simulated using O(log n) additional
space on the top of the space taken by the landmarks and O(log n) time to process T [h].
6
Lower bounds
In this section we use Yao’s minimax principle [19] to prove lower bounds on the space complexity of computing
the largest radius of a palindrome in a word of length n over an alphabet Σ in the streaming model. We
14
denote this problem by PALINΣ [n].
Theorem 6.1 (Yao’s minimax principle for randomized algorithms). Let X be the set of inputs for a problem
and A be the set of all deterministic algorithms solving it. Then, for any x ∈ X and A ∈ A, the cost of
running A on x is denoted by c(a, x) ≥ 0.
Let p be the probability distribution over A, and let A be an algorithm chosen at random according to p.
Let q be the probability distribution over X , and let X be an input chosen at random according to q. Then the
worst-case expected cost of the randomized algorithm is at least as large as the cost of the best deterministic
algorithm against the chosen distribution on the inputs:
max E[c(A, x)] ≥ min E[c(a, X)].
x∈X
a∈A
We use the above theorem for both Las Vegas and Monte Carlo algorithms. For Las Vegas algorithms, we
consider only correct algorithms, and c(x, a) is the memory usage. For Monte Carlo algorithms, we consider
all algorithms (not necessarily correct) with memory usage not exceeding a certain threshold, and c(x, a) is
the correctness indicator function, i.e., c(x, a) = 0 if the algorithm is correct and c(x, a) = 1 otherwise.
Our proofs will be based on appropriately chosen padding. The padding requires a constant number of
fresh characters. If Σ is twice as large as the number of required fresh characters, we can still use half of it to
construct a difficult input instance, which does not affect the asymptotics. Otherwise, we construct a difficult
input instance over Σ, then add enough new fresh characters to facilitate the padding, and finally reduce the
resulting larger alphabet to binary at the expense of increasing the size of the input by a constant factor.
Lemma 6.2. For any alphabet Σ = {1, 2, . . . , σ} there exists a morphism h : Σ∗ → {0, 1}∗ such that, for any
c ∈ Σ, |h(c)| = 2σ + 6 and, for any word w, w contains a palindrome of length ` if and only if h(w) contains
a palindrome of length (2σ + 6) · `.
Proof. We set:
h(c) = 11s 01s−c 10011s−c 01c 1.
Clearly |h(c)| = 2σ + 6 and, because every h(c) is a palindrome, if w contains a palindrome of length ` then
h(w) contains a palindrome of length (2σ + 6) · `. Now assume that h(w) contains a palindrome of length
(2σ + 6) · `, where ` ≥ 1. If ` = 1 then we obtain that w should contain a palindrome of length 1, which
always holds. Otherwise, the palindrome contains 00 inside and we consider two cases.
1. The palindrome is centered inside 00. Then it corresponds to an odd palindrome of length ` in w.
2. The palindrome maps some 00 to another 00. Then it corresponds to an even palindrome of length ` in
w.
In either case, the claim holds.
For the padding we will often use an infinite word ν = 01 11 02 12 03 13 . . ., or more precisely its prefixes of
length d, denoted ν(d). Here 0 and 1 should be understood as two characters not belonging to the original
alphabet,
which is then reduced using the above lemma. The longest palindrome inside ν(d) has radius
√
O( d).
We first show that any Las Vegas approximation algorithm must necessarily use Ω(n log |Σ|) bits of
memory in expectation in both variants, so Las Vegas randomization is essentially useless here. By Yao’s
minimax principle, it is enough to construct a distribution over the inputs, which is hard for any deterministic
algorithm using less memory. We restrict the inputs to a family of strings of the form ν(E)x$$xR ν(E)R ,
where $ is a special character not belonging to Σ and ν(E) is a padding word of length E chosen so that
there are no long palindromes inside. Then the longest palindrome must be centered in the middle of the
whole word. By a counting argument, the state of the algorithm after having seen ν(E)x$ must be distinct
for different words x, so the required number of bits is Ω(n log |Σ|) in expectation. A bound on multiplicative
approximation follows because multiplicative approximation implies additive approximation.
15
Theorem 6.3 (Las Vegas approximation). Let A be a Las Vegas streaming algorithms solving PALINΣ [n]
with additive error E ≤ 0.49n or multiplicative error (1 + ε) ≤ 50 using s(n) bits of memory. Then
E[s(n)] = Ω(n log |Σ|).
Proof. By Theorem 6.1, it is enough to construct a probability distribution P over Σn such that for any
deterministic algorithm D, its expected memory usage on a word chosen according to P is Ω(n log |Σ|) in bits.
Consider solving PALINΣ [n] with additive error E. We define P as the uniform distribution over
0
ν(E)x$$yν(E)R , where x, y ∈ Σn , n0 = n2 − E − 1, and $ are special characters not belonging to Σ. Let us
look at the memory usage of D after having read ν(E)x. We say that x is "good" when the memory usage is
0
0
at most n2 log |Σ| and "bad" otherwise. Assume that 12 |Σ|n of all x’s are good, then there are two strings
x 6= x0 such that the state of D after having read both ν(E)x and ν(E)x0 is exactly the same. Hence the
behavior of D on ν(E)x$$xR ν(E)R and ν(E)x0 $$xR ν(E)R is exactly the same. The former is a palindrome
of radius n2 = n0 + E + 1, so D must answer at least n0 + 1, and consequently the latter also must contain a
palindrome of radius at least n0 + 1. A palindrome inside ν(E)x0 $$xR ν(E)R is either fully contained
within
√
ν(E), x0 , xR or it is a middle palindrome. But the longest palindrome inside ν(E) is of length O( E) < n0 + 1
(for n large enough) and the longest palindrome inside x or xR is of length n0 < n0 + 1, so ν(E)x0 $$xR ν(E)R
contains a middle palindrome of radius n0 + 1. This implies that x = x0 , which is a contradiction. Therefore,
0
0
at least 12 |Σ|n of all x’s are bad. But then the expected memory usage of D is at least n4 log |Σ|, which for
E ≤ 0.49n is Ω(n log |Σ|) as claimed.
Now consider solving PALINΣ [n] with multiplicative error (1 + ε). An algorithm with multiplicative error
ε
(1 + ε) can also be considered as having additive error E = n2 · 1+ε
, so if the expected memory usage of such an
algorithm is o(n log |Σ|) and (1 + ε) ≤ 50 then we obtain an algorithm with additive error E ≤ n2 49
50 = 0.49n
and expected memory usage o(n log |Σ|), which we already know to be impossible.
Now we move to Monte Carlo algorithms. We first consider exact algorithms solving PALINΣ [n]; lower
bounds on approximation algorithms will be then obtained by padding the input appropriately. We introduce
an auxiliary problem MID-PALINΣ [n], which is to compute radius of the middle palindrome in a word of
length n over an alphabet Σ. We want to show that solving MID-PALINΣ [n] exactly with error probability
1
smaller than n|Σ|
requires b n2 log |Σ|c bits of space. By Yao’s minimax principle, it is enough to construct a
distribution over the inputs, such that any deterministic algorithm using less memory is not able to distinguish
between inputs with different answers reasonably often. This can be done by considering uniform distribution
on inputs of the form x[1] . . . x[ n2 ]x[ n2 ] . . . x[k + 1]cx[k − 1] . . . x[1]. Then amplification (running multiple
instances of an algorithm in parallel) gives us a lower bound on the space complexity of any algorithm solving
MID-PALINΣ [n] exactly. The lower bound can be translated to PALINΣ [n] by padding the input in the
middle, so that the longest palindrome must be centered in the middle.
Lemma 6.4. There exists a constant γ such that any randomized Monte Carlo streaming algorithm A solving
MID-PALINΣ [n] or PALINΣ [n] exactly with probability 1 − n1 uses at least γ · n log min{|Σ|, n} bits of
memory.
Proof. First we prove that if A is a Monte Carlo streaming algorithm solving MID-PALINΣ [n] exactly
1
.
using less than b n2 log |Σ|c bits of memory, then its error probability is at least n|Σ|
By Theorem 6.1, it is enough to construct probability distribution P over Σn such that for any deterministic
algorithm D using less than b n2 log |Σ|c bits of memory, the expected probability of error on a word chosen
1
according to P is at least n|Σ|
.
Let n0 =
n
2.
0
For any x ∈ Σn , k ∈ {1, 2, . . . , n0 } and c ∈ Σ we define:
w(x, k, c) = x[1]x[2]x[3] . . . x[n0 ]x[n0 ]x[n0 − 1]x[n0 − 2] . . . x[k + 1]cx[k − 1] . . . x[2]x[1].
Now P is the uniform distribution over all such w(x, k, c).
0
0
n
Since there are |Σ|n = 2n log |Σ| ≥ 2 · 2b 2 log |Σ|c−1 possible strings of length n0 and we assume that D uses
at most b n2 log |Σ|c bits, we can partition at least half of these strings into pairs (x, x0 ), such that D is in the
same state after reading either x or x0 . (If we choose an arbitrary maximal matching of strings into pairs, at
16
most half of possible strings will be left unpaired, that is one per each possible state of D.) Let s be longest
common suffix of x and x0 , so x = vcs and x0 = v 0 c0 s, where c =
6 c0 are single characters. Then D returns
0
0
0
the same answer on w(x, n − |s|, c) and w(x , n − |s|, c), even though the radius of the middle palindrome is
exactly |s| in one of them, and at least |s| + 1 in the other one. Therefore, D errs on at least one of these two
inputs. Similarly, it errs on either w(x, n0 − |s|, c0 ) or w(x, n0 − |s|, c0 ). Thus the error probability is at least
1
1
2n0 |Σ| = n|Σ| .
Now we can prove the lemma for MID-PALINΣ [n] with a standard amplification trick. Say that we
have a Monte Carlo streaming algorithm, which solves MID-PALINΣ [n] exactly with error probability ε
using s(n) bits of memory. Then we can run its k instances simultaneously and return the most frequently
reported answer. The new algorithm needs O(k · s(n)) bits of memory and its error probability εk satisfies:
X k
εk ≤
(1 − ε)i εk−i ≤ 2k · εk/2 = (4ε)k/2 .
i
2i<k
log(4/n)
1−o(1)
log n
1
Let us choose κ = 16 log(1/(n|Σ|))
= 16 1+log
|Σ|/ log n = Θ( log n+log |Σ| ) = γ · log |Σ| log min{|Σ|, n}, for some
constant γ. Now we can prove the
1 theorem.
3 1 Assume that A uses less1 than κ · n log |Σ| = γ · n log min{|Σ|, n}
bits of memory. Then running 2κ
≥ 4 2κ (which holds since κ < 6 ) instances of A in parallel requires less
than b n2 log |Σ|c bits of memory. But then the error probability of the new algorithm is bounded from above
by:
3
16κ
18
16
4
1
1
=
≤
n
n|Σ|
n|Σ|
which we have already shown to be impossible.
The lower bound for MID-PALINΣ [n] can be translated into a lower bound for solving PALINΣ [n]
exactly by padding the input so that the longest palindrome is centered in the middle. Let n0 = n2 and
x = x[1]x[2] . . . x[n] be the input for MID-PALINΣ [n]. We define:
w(x) = x[1]x[2]x[3] . . . x[n0 ]1 000 . . . 0 1x[n0 + 1] . . . x[n].
n
Now if the radius of the middle palindrome in x is k, then w(x) contains a palindrome of radius at least
n0 + k + 1. In the other direction, any palindrome inside w(x) of radius larger than n0 must be centered
somewhere in the middle block consisting of only zeroes and both ones are mapped to each other, so it must
be the middle palindrome. Thus, the radius of the longest palindrome inside w(x) is exactly n0 + k + 1, so we
have reduced solving MID-PALINΣ [n] to solving PALINΣ [2n + 2]. We already know that solving MIDPALINΣ [n] with probability 1 − n1 requires γ · n log min{|Σ|, n} bits of memory, so solving PALINΣ [2n + 2]
1
with probability 1 − 2n+2
≥ 1 − n1 requires γ · n log{|Σ|, n} ≥ γ 0 · (2n + 2) log min{|Σ|, 2n + 2} bits of memory.
Notice that the reduction needs O(log n) additional bits of memory to count up to n, but for large n this is
much smaller than the lower bound if we choose γ 0 < γ4 .
To obtain a lower bound for Monte Carlo additive approximation, we observe that any algorithm solving
PALINΣ [n] with additive error E can be used to solve PALINΣ [ n−E
E+1 ] exactly by inserting E zeroes between
every two characters, in the very beginning, and in the very end. However, this reduction requires log E ≤ log n
additional bits of memory for counting up to E and cannot be used when the desired lower bound on the
n
n
required number of bits Ω( E
log min(|Σ|, E
) is significantly smaller than log n. Therefore, we need a separate
technical lemma which implies that either additive or multiplicative approximation with error probability n1
requires Ω(log n) bits of space.
Lemma 6.5. Let A be any randomized Monte Carlo streaming algorithm solving PALINΣ [n] with additive
error at most 0.49n or multiplicative error at most n0.49 and error probability n1 . Then A uses Ω(log n) bits
of memory.
17
Proof. By Theorem 6.1, it is enough to construct a probability distribution P over Σn , such that for any
deterministic algorithm D using at most s(n) = O(log n) bits of memory, the expected probability of error on
1
.
a word chosen according to P is 2s(n)+2
0
0
Let n = s(n) + 1. For any x, y ∈ Σn , let w(x, y) = ν( n2 − n0 )R xy R ν( n2 − n0 )R . Observe that if x = y
then w(x, y) contains
a√
palindrome of radius n2 , and otherwise the longest palindrome there has radius at
√
0
most 2n + O( n) = O( n), thus any algorithm with additive error of at most 0.49n or with a multiplicative
error at most n0.49 must be able to distinguish between these two cases (for n large enough).
0
Let S ⊆ Σn be an arbitrary family of words of length n0 such that |S| = 2 · 2s(n) , and let P be the uniform
distribution on all words of the form w(x, y), where x and y are chosen uniformly and independently from S.
0
By a counting argument, we can create at least |S|
4 pairs (x, x ) of elements from S such that the state of D
n
n
0 R
0 R 0
is the same after having read ν( 2 − n ) x and ν( 2 − n ) x . (If we create the pairs greedily, at most one
such x per state of memory can be left unpaired, so at least |S| − 2s(n) = |S|
2 elements are paired.) Thus, D
cannot distinguish between w(x, x0 ) and w(x, x), and between w(x0 , x0 ) and w(x0 , x), so its error probability
1
1
must be at least |S|/2
|S|2 = 4·2s(n) . Thus if s(n) = o(log n), the error rate is at least n for n large enough, a
contradicion.
Combining the reduction with the technical lemma and taking into account that we are reducing to a
n
problem with word length of Θ( E
), we obtain the following.
Theorem 6.6 (Monte Carlo additive approximation). Let A be any randomized Monte Carlo streaming
algorithm solving PALINΣ [n] with additive error E with probability 1 − n1 . If E ≤ 0.49n then A uses
n
n
Ω( E
log min{|Σ|, E
}) bits of memory.
n
}.
Proof. Define σ = min{|Σ|, 12 E
1
n
log min{|Σ|, E
} (which holds due to E ≤ 0.49n), it is enough to
Because of Lemma 6.5 and log σ ≥ 35
n
prove that Ω( E log σ) is a lower bound when
E≤
γ
n
·
log σ.
4 log n
(5)
Assume that there is a Monte Carlo streaming algorithm A solving PALINΣ [n] with additive error E using
n
1 n
o( E
log σ) bits of memory and probability 1 − n1 . Let n0 = n−E
E+1 ≥ 2 E (the last inequality holds because
E ≤ 0.49n and because we can assume that E > 1). Given a word x[1]x[2] . . . x[n0 ], we can simulate running
E
E
E
0 E
A on 0E x[1]0
j x[2]0
k x[3] . . . 0 x[n ]0 to calculate R (using log E ≤ log n additional bits of memory), and
R
then return E+1
. We call this new Monte Carlo streaming algorithm A0 . Recall that A reports the radius
of the longest palindrome with additive error E. Therefore, if the original word contains a palindrome
of radius r, the new word contains a palindrome of radius E2 + r(E + 1), so R ≥ r(E + 1) and A0 will
return at least r. In the other direction, if A0 returns r, then the new word contains a palindrome of radius
r(E + 1). If such palindrome is centered so that x[i] is matched with x[i + 1] for some i, then it clearly
corresponds to a palindrome of radius r in the original word. But otherwise every x[i] within the palindrome
is matched with 0, so in fact the whole palindrome corresponds to a streak of consecutive zeroes in the new
word and can be extended to the left and to the right to start and end with 0E , so again it corresponds to
a palindrome of radius r in the original word. Therefore, A0 solves PALINΣ [n0 ] exactly with probability
0
1
1 − (n0 (E+1)+E)
≥ 1 − n10 and uses o( n (E+1)+E
log σ) + log n = o(n0 log σ) + log n bits of memory. Observe
E
that by Lemma 6.4 we get a lower bound
γ · n0 log min{|Σ|, n0 } ≥
γ 0
γ n
γ
· n log σ + · log σ ≥ · n0 log σ + log n
2
4 E
2
(where the last inequality holds because of Eq.(5)). Then, for large n we obtain contradiction as follows
o(n0 log σ) + log n <
18
γ 0
· n log σ + log n.
2
Finally, we consider multiplicative approximation. Here we observe that any algorithm solving PALINΣ [n]
log n
with multiplicative error (1 + ε) can be used to solve MID-PALINΣ [2n0 ] exactly, where n0 = Θ( log(1+2ε)
),
by separating the characters appropriately. Intuitively, the padding is chosen so that the middle palindrome
has the largest radius and the larger the distance from the center the longer the separator inserted between
two consecutive characters of the original input. Again, we need log n bits for a counter and hence need
to invoke a separate technical lemma when (1 + ε) is very large. After some calculations, and taking into
log n
) we obtain the following.
account that we are reducing to a problem with word length of Θ( log(1+ε)
Theorem 6.7 (Monte Carlo multiplicative approximation). Let A be any randomized Monte Carlo streaming
algorithm solving PALINΣ [n] with multiplicative error (1 + ε) with probability 1 − n1 . If n−0.98 ≤ ε ≤ n0.49
log n
log n
log min{|Σ|, log(1+ε)
}) bits of memory.
then A uses Ω( log(1+ε)
Proof. For ε ≥ n0.001 then the claimed lower bound reduces to Ω(1) bits, which obviously holds. Thus we
can assume that ε < n0.001 . Define
σ = min{|Σ|,
1
log n
− 2}.
50 log(1 + 2ε)
log n
First we argue that it is enough to prove that A uses Ω( log(1+ε)
log σ) bits of memory. Since log(1 + 2ε) ≤
0.001 log n + o(log n), we have that:
1
log n
− 2 ≥ 18 − o(1)
50 log(1 + 2ε)
(6)
1
log n
log n
− 2 = Θ(
).
50 log(1 + 2ε)
log(1 + 2ε)
(7)
log(1 + 2ε) = Θ(log(1 + ε))
(8)
and consequently:
Finally, observe that:
because log 2(1 + ε) = Θ(log(1 + ε)) for ε ≥ 1, and log(1 + ε) = Θ(ε) for ε < 1. From (7) and (8) we conclude
that:
log n
log σ = Θ(log min{|Σ|,
}).
(9)
log(1 + ε)
log n
Because of Lemma 6.5 and equations (8) and (9), it is enough to prove that Ω( log(1+ε)
log σ) is a lower
bound when
log σ
log(1 + 2ε) ≤ γ ·
,
(10)
100
log n
log n
as otherwise Ω( log(1+ε)
log σ) = Ω( log(1+2ε)
log σ) = Ω(log n).
Assume that there is a Monte Carlo streaming algorithm A solving PALINΣ [n] with multiplicative error
log n
(1+ε) with probability 1− n1 using o( log(1+ε)
log σ) bits of memory. Let x = x[1]x[2] . . . x[n0 ]x[n0 +1] . . . x[2n0 ] be
0
an input for MID-PALINΣ [2n0 ]. We choose n0 so that n = (1+2ε)n +1 ·n0.99 . Then n0 = log(1+2ε) (n0.01 )−1 =
log n
1
d+1
0
· n0.99 e for any 0 ≤ d ≤ n0 .
100 log(1+2ε) − 1. We choose i0 , i1 , i2 , i3 , . . . , in so that i0 + . . . + id = d(1 + 2ε)
−0.98
0.99
0.01
(Observe that for ε = Ω(n
) we have i0 > n
and i1 , . . . , id > 2n
− 1.) Finally we define:
w(x) = ν(in0 )R x[1]ν(in0 −1 )R . . . x[n0 ]ν(i0 )R ν(i0 )x[n0 + 1]ν(i1 ) . . . ν(in0 −1 )x[2n0 ]ν(in0 ).
If x contains a middle palindrome of radius exactly k, then w(x) contains a middle palindrome of radius
(1 + 2ε)k+1 · n0.99 . √Also, based on the properties of ν, any non-middle centered palindrome in w(x) has
radius at most O( n), which is less than n0.99 for n large enough. Since d(1 + 2ε)k · n0.99 e · (1 + ε) <
((1 + 2ε)k · n0.99 + 1) · (1 + ε) < (1 + 2ε)k+1 · n0.99 , value of k can be extracted from the answer of A. Thus, if
A approximates the middle palindrome in w(x) with multiplicative error (1 + ε) with probability 1 − n1 using
19
log n
o( log(1+ε)
log σ) bits of memory, we can construct a new algorithm A0 solving MID-PALINΣ [2n0 ] exactly
1
with probability 1 − n1 > 1 − 2n
0 using
o(
log n
log σ) + log n
log(1 + ε)
(11)
bits of memory. By Lemma 6.4 we get a lower bound
γ · 2n0 log min{|Σ|, 2n0 }
=
≥
γ
log n
·
log σ − 2γ log σ
50 log(1 + 2ε)
γ
log n
·
log σ + log n − 2γ log σ
100 log(1 + 2ε)
(where the last inequality holds because of (10)). On the other hand, for large n
log n
γ
log n
1
·
log
σ
−
2γ
log
σ
+
log
n
=
−
2
γ log σ + log n
100 log(1+2ε)
100 log(1+2ε)
log n
= Θ log(1+ε)
log σ + log n
so (12) exceeds (11), a contradiction.
20
(12)
Acknowledgments
The first author is currently holding a post-doctoral position at Warsaw Center of Mathematics and Computer
Science. However, most of this work has been done when the first author was at Max-Planck-Institut
für Informatik and the second author at LIF, CNRS – Aix Marseille University (supported by the Labex
Archimède and by the ANR project MACARON (ANR-13-JS02-0002)).
The authors would like to thank Tomasz Syposz for a suggestion which allowed them to simplify the
algorithm.
References
[1] Alberto Apostolico, Dany Breslauer, and Zvi Galil. Parallel detection of all palindromes in a string.
Theor. Comput. Sci., 141(1&2):163–173, 1995.
[2] Petra Berenbrink, Funda Ergün, Frederik Mallmann-Trenn, and Erfan Sadeqi Azer. Palindrome
Recognition In The Streaming Model. In STACS 2014, volume 25 of LIPIcs, pages 149–161, Dagstuhl,
Germany, 2014. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik.
[3] Dany Breslauer and Zvi Galil. Real-time streaming string-matching. ACM Transactions on Algorithms,
10(4):22, 2014.
[4] Raphaël Clifford, Allyx Fontaine, Ely Porat, Benjamin Sach, and Tatiana A. Starikovskaya. Dictionary
matching in a stream. In Nikhil Bansal and Irene Finocchi, editors, Algorithms - ESA 2015 - 23rd Annual
European Symposium, Patras, Greece, September 14-16, 2015, Proceedings, volume 9294 of Lecture Notes
in Computer Science, pages 361–372. Springer, 2015.
[5] Raphaël Clifford, Allyx Fontaine, Ely Porat, Benjamin Sach, and Tatiana A. Starikovskaya. The k mismatch problem revisited. In Robert Krauthgamer, editor, Proceedings of the Twenty-Seventh Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA 2016, Arlington, VA, USA, January 10-12, 2016,
pages 2039–2052. SIAM, 2016.
[6] Funda Ergün, Hossein Jowhari, and Mert Saglam. Periodicity in streams. In Maria J. Serna, Ronen
Shaltiel, Klaus Jansen, and José D. P. Rolim, editors, Approximation, Randomization, and Combinatorial
Optimization. Algorithms and Techniques, 13th International Workshop, APPROX 2010, and 14th
International Workshop, RANDOM 2010, Barcelona, Spain, September 1-3, 2010. Proceedings, volume
6302 of Lecture Notes in Computer Science, pages 545–559. Springer, 2010.
[7] Gabriele Fici, Travis Gagie, Juha Kärkkäinen, and Dominik Kempa. A subquadratic algorithm for
minimum palindromic factorization. J. Discrete Algorithms, 28:41–48, 2014.
[8] N. J. Fine and H. S. Wilf. Uniqueness theorems for periodic functions. Proceedings of the AMS,
16:109–114, 1965.
[9] Zvi Galil and Joel Seiferas. A linear-time on-line recognition algorithm for “palstar”. J. ACM, 25(1):102–
111, January 1978.
[10] Paweł Gawrychowski, Florin Manea, and Dirk Nowotka. Testing Generalised Freeness of Words. In STACS
2014, volume 25 of LIPIcs, pages 337–349, Dagstuhl, Germany, 2014. Schloss Dagstuhl–Leibniz-Zentrum
fuer Informatik.
[11] Tomohiro I, Shiho Sugimoto, Shunsuke Inenaga, Hideo Bannai, and Masayuki Takeda. Computing
palindromic factorizations and palindromic covers on-line. In CPM 2014, volume 8486 of Lecture Notes
in Computer Science, pages 150–161. Springer, 2014.
21
[12] Markus Jalsenius, Benny Porat, and Benjamin Sach. Parameterized matching in the streaming model.
In 30th International Symposium on Theoretical Aspects of Computer Science, STACS 2013, February
27 - March 2, 2013, Kiel, Germany, pages 400–411, 2013.
[13] Haim Kaplan and Robert E. Tarjan. Persistent lists with catenation via recursive slow-down. In
Proceedings of the Twenty-seventh Annual ACM Symposium on Theory of Computing, STOC ’95, pages
93–102, New York, NY, USA, 1995. ACM.
[14] Richard M. Karp and Michael O. Rabin. Efficient randomized pattern-matching algorithms. IBM Journal
of Research and Development, 31(2):249–260, 1987.
[15] Donald E. Knuth, Jr. James H. Morris, and Vaughan R. Pratt. Fast pattern matching in strings. SIAM
Journal on Computing, 6(2):323–350, 1977.
[16] Dmitry Kosolobov, Mikhail Rubinchik, and Arseny M. Shur. Palk is linear recognizable online. In
SOFSEM 2015, volume 8939 of Lecture Notes in Computer Science, pages 289–301. Springer, 2015.
[17] Glenn K. Manacher. A new linear-time “on-line” algorithm for finding the smallest initial palindrome of
a string. J. ACM, 22(3):346–351, 1975.
[18] Benny Porat and Ely Porat. Exact and approximate pattern matching in the streaming model. In
50th Annual IEEE Symposium on Foundations of Computer Science, FOCS 2009, October 25-27, 2009,
Atlanta, Georgia, USA, pages 315–323. IEEE Computer Society, 2009.
[19] Andrew Chi-Chih Yao. Probabilistic computations: Toward a unified measure of complexity (extended
abstract). In FOCS, pages 222–227. IEEE Computer Society, 1977.
22
| 8cs.DS
|
Capacity limit for faster-than-Nyquist
non-orthogonal frequency-division multiplexing
signaling
Ji Zhou1, 2 , Yaojun Qiao1,* , Zhanyu Yang3 , Qixiang Cheng2 , Qi Wang4 , Mengqi Guo1 , and
Xizi Tang1
arXiv:1611.10309v3 [cs.IT] 27 Jun 2017
1 State
Key Laboratory of Information Photonics and Optical Communications, School of Information and
Communication Engineering, Beijing University of Posts and Telecommunications (BUPT), Beijing 100876, China
2 Department of Electrical Engineering, Columbia University, New York 10026, USA
3 Department of Electrical and Computer Engineering, University of Virginia, Virginia 22904, USA
4 School of Electronics and Computer Science, University of Southampton, Southampton SO17 1BJ, U.K.
* [email protected]
ABSTRACT
Faster-than-Nyquist (FTN) signal achieves higher spectral efficiency and capacity compared to Nyquist signal due to its smaller
pulse interval or narrower subcarrier spacing. Shannon limit typically defines the upper-limit capacity of Nyquist signal. To the
best of our knowledge, the mathematical expression for the capacity limit of FTN non-orthogonal frequency-division multiplexing
(NOFDM) signal is first demonstrated in this paper. The mathematical expression shows that FTN NOFDM signal has the
potential to achieve a higher capacity limit compared to Nyquist signal. In this paper, we demonstrate the principle of FTN
NOFDM by taking fractional cosine transform-based NOFDM (FrCT-NOFDM) for instance. FrCT-NOFDM is first proposed and
implemented by both simulation and experiment. When the bandwidth compression factor α is set to 0.8 in FrCT-NOFDM,
the subcarrier spacing is equal to 40% of the symbol rate per subcarrier, thus the transmission rate is about 25% faster than
Nyquist rate. FTN NOFDM with higher capacity would be promising in the future communication systems, especially in the
bandwidth-limited applications.
Introduction
In the 1940s, an upper limit of communication capacity was proposed for Nyquist signal in an additive white Gaussian noise
(AWGN) channel, now known as the Shannon limit1, 2 . Initially, the communication capacity is enough for the traditional data
and voice services. However, with the exponential growth of data traffic due to bandwidth-intensive applications such as high
definition TV and mobile video, the communication capacity gradually approaches Shannon limit nowadays3–5 . Increasing
the spectral efficiency is a key challenge to meet the increasing demand for higher capacity over communication channels.
Faster-than-Nyquist (FTN) signal was first proposed by Mazo in 1970s to improve the spectral efficiency6 . As the name implies,
FTN signal can achieve a symbol rate faster than Nyquist rate. Therefore, it has been widely investigated in high-capacity
wireless and optical communications7–13 .
The reported FTN scheme can be generally categorized into two types, one is compressing the duration between the adjacent
pulses in time domain6 , and the other one is compressing the baseband bandwidth in frequency domain12–14 . In time domain,
FTN binary sinc-pulse signal was proposed in 1975 by accelerating the pulse with time acceleration factor τ. The accelerated
pulses are no longer orthogonal, thus FTN binary sinc-pulse signal is a kind of FTN non-orthogonal time-division multiplexing
(NOTDM) signal. When τ is set to 0.8, known as the Mazo limit, 25% more bits can be carried in the same bandwidth and
there is not obvious deterioration in bit error rate (BER) performance by using the trellis decoding to effectively compensate
the inter-symbol interference (ISI)7 . In frequency domain, the duobinary-pulse shaping filter is employed to compress the
baseband bandwidth of single-carrier signal12, 13 or the subcarrier spacing of multi-carrier signal is compressed to obtain
narrower baseband bandwidth14 . In this paper, we focus on the multi-carrier FTN signal. Non-orthogonal frequency-division
multiplexing (NOFDM) signal has been realized by further compressing the subcarrier spacing compared to OFDM signal15–17 .
We found that only when the subcarrier spacing is less than half of the symbol rate per subcarrier, NOFDM signal can achieve a
symbol rate faster than Nyquist rate as we demonstrated in our recent work14 .
FTN signal achieves higher spectral efficiency and capacity compared to Nyquist signal due to its smaller pulse interval
or narrower subcarrier spacing. Shannon limit typically defines the upper-limit capacity of Nyquist signal. In this paper, we
Figure 1. Block diagram of FrCT-NOFDM for optical transmission system.
first give the mathematical expression for the capacity limit of FTN NOFDM signal. The mathematical expression shows that
FTN signal has the potential to achieve a higher capacity limit compared to Nyquist signal. In this paper, we demonstrate the
principle of FTN NOFDM by taking fractional cosine transform-based NOFDM (FrCT-NOFDM) for instance. FrCT-NOFDM
is first proposed and implemented by both simulation and experiment. When the bandwidth compression factor α is set to 0.8,
the subcarrier spacing is equal to 40% of the symbol rate per subcarrier, thus transmission rate is about 25% faster than Nyquist
rate.
Results
In this paper, we demonstrate the principle of FTN NOFDM by taking FrCT-NOFDM for instance. A block diagram of the
FrCT-NOFDM system is depicted in Figure 1. Different from traditional OFDM system, the inverse FrCT (IFrCT)/FrCT
algorithm is employed to implement the multiplexing/demultiplexing processing. The N-order IFrCT and FrCT are defined as
r
2 N−1
πα(2n + 1)k
xn =
W
X
cos
,
(1)
∑ k k
N k=0
2N
r
Xk =
πα(2n + 1)k
2 N−1
Wk ∑ xn cos
N n=0
2N
where 0 ≤ n ≤ N − 1, 0 ≤ k ≤ N − 1,
(
√1 , k = 0
2
Wk =
1,
k = 1, 2, · · · , N − 1
(2)
(3)
and α is the bandwidth compression factor. The α less than 1 determines the level of the bandwidth compression. When α is
equal to 1, Equation 2 is the Type-II discrete cosine transform (DCT) in which the matrix is orthogonal. The Type-II DCT is
probably the most commonly used form, and is often simply referred to as “the DCT”18 . DCT-based OFDM (DCT-OFDM)
signal can be generated by Type-II DCT19, 20 . In DCT-OFDM, the subcarrier spacing is equal to half of the symbol rate per
subcarrier, which is half of the subcarrier spacing in discrete Fourier transform-based OFDM (DFT-OFDM). Therefore, the
subcarrier spacing of DCT-OFDM is equal to 1/2T where T denotes the time duration of one DCT-OFDM symbol21, 22 . Due to
the compression of subcarrier spacing, the subcarrier spacing of FrCT-NOFDM should be smaller than 1/2T .
Spectral efficiency and capacity limit
Figure 2 shows the sketched spectra of DCT-OFDM (i.e., α = 1) and FrCT-NOFDM when the subcarrier number is set to 4.
The subcarrier spacing of FrCT-NOFDM is equal to α/2T where T denotes the time duration of one FrCT-NOFDM symbol.
All the subcarriers locate in the positive frequency domain. Meanwhile, their images fall into the negative frequency domain.
When the subcarrier number is large enough, the baseband bandwidth of FrCT-NOFDM can be calculated by
Baseband bandwidth =
(N − 1) × α 1
N ×α
+ ≈
2×T
T
2×T
(4)
where N denotes the subcarrier number.
2/11
Figure 2. Sketched spectra of DCT-OFDM (i.e., α = 1) and FrCT-NOFDM.
In general, Nyquist frequency is equal to half of the sample rate, which is equal to N/2T . When N is large enough, the
baseband bandwidth of DCT-OFDM (i.e., α = 1) is almost the same with Nyquist frequency and the baseband bandwidth of
FrCT-NOFDM (i.e., α < 1) is smaller than Nyquist frequency. It is worth noting that Nyquist frequency should not be confused
with Nyquist rate. Nyquist rate is defined as the twice of the baseband bandwidth,
RN =
N ×α
.
T
(5)
As we know, the symbol rate of FrCT-NOFDM is equal to the sample rate when all the subcarriers are valid,
RS =
N
,
T
(6)
thus the symbol rate of FrCT-NOFDM is faster than Nyquist rate. FrCT-NOFDM can be considered as a kind of FTN signal.
When α is set to 0.8, the subcarrier spacing is equal to 40% of symbol rate per subcarrier. Twenty percent baseband bandwidth
saving can be obtained and the transmission rate is about 25% faster than the Nyquist rate.
Afterwards, we discuss the capacity limit for FrCT-NOFDM signal. The channel is in the presence of AWGN. The
well-known Shannon limit for Nyquist signal can be calculated by
PS
C ≤ W log2 1 +
(7)
PN
where W is the signal bandwidth, PS is the signal power and PN is the noise power. This equation depicts the upper-limit
capacity of Nyquist signal2 .
To obtain the capacity limit of FrCT-NOFDM signal, we will apply the analysis method for single-carrier signal in Shannon’s
papers1, 2 to the multi-carrier signal. However, there is great difference between single-carrier and multi-carrier signal. The
capacity limit of the single-carrier signal can be carried out in the time domain, but the capacity limit of multi-carrier signal
should be derived out in the frequency domain. The derivation of capacity limit needs to employ the geometric methods23 . In
the following, we will briefly give the derivation of the capacity limit for FrCT-NOFDM. This derivation is similar to that in
Shannon’s papers.
In the frequency domain, there are W /(subcarrier spacing) independent amplitudes where W is the baseband bandwidth
and the subcarrier spacing is equal to α/2T . Therefore, the number of independent amplitudes is equal to 2TW /α, which
determines the dimensions of signals.
p For large W , the perturbation caused by noise can be considered as some points near the
surface of a sphere with radius 2TW PN /α centered at the signal point. The power of received signal is PS + PN . Similar
to the perturbation,
the received signal can be considered as some points whose positions are on the surface of a sphere with
p
radius
2TW
(P
+
PN )/α. The number of the distinguishable signals
S
p
p is no more than the volume of the sphere with radius
2TW (PS + PN )/α divided by the volume of the sphere with radius 2TW PN /α. The volume of an n-dimensional sphere of
3/11
Figure 3. The probability density function (PDF) of 2-PAM with inter-carrier interference (ICI) in FrCT-NOFDM with α of
0.8.
radius r can be calculated by23
V=
π n/2
rn .
Γ n/2 + 1
(8)
R
where Γ n/2 + 1 = 0∞ e−t × t n/2 dt. Therefore, the upper limit for the number of distinguishable signals is given by
r
M≤
PS + PN
PN
2W T /α
.
(9)
Consequently, the capacity limit of FrCT-NOFDM signal can be bounded by
1
PS
log2 M
≤ ×W log2 1 +
C=
T
α
PN
(10)
It is worth noting that, in FrCT-NOFDM signal, inter-carrier interference (ICI) is a serious problem that should not be
neglected because the subcarriers are no longer orthogonal. Figure 3 shows the probability density function (PDF) of ICI in
FrCT-NOFDM with α of 0.8. The simulated curve denotes the PDF of the 2-PAM with ICI in FrCT-NOFDM with α of 0.8.
The theoretical curve depicts the PDF of the 2-PAM with Gaussian noise, which can be defined as
1
PDFT heoretical (x) = √
2 2πσ
(x+1)2
(x−1)2
−
−
2
2
2σ
2σ
e
+e
(11)
where σ 2 is the variance of the Gaussian noise. The simulated curve agrees well with the theoretical curve. Therefore, the
distribution of ICI is approximately Gaussian. Moreover, the ICI is independent of the Gaussian noise in AWGN channel.
Considering both ICI and Gaussian noise in AWGN channel, the capacity limit of FrCT-NOFDM signal can be rewritten as
1
PS
C ≤ ×W log2 1 +
(12)
α
PN + PICI
where PICI is the power of ICI, which degrades the BER performance and thus decreases the capacity of FrCT-NOFDM signal.
Therefore, the capacity of FrCT-NOFDM signal is likely to approach the capacity limit in Equation 10 only when ICI has
been effectively eliminated. As verified in literatures6, 7 , when the α is set to the value between 1 and 0.8, there is not obvious
deterioration in BER performance by employing the optimal detection to eliminate the interference. The capacity limit of
FrCT-NOFDM signal is potentially higher than that of the Nyquist signal when the α is set to the value between 1 and 0.8.
How to eliminate the ICI is crucial for FrCT-NOFDM system. Afterwards, we will demonstrate the simulation and experiment
to investigate the performance of FrCT-NOFDM system and verify the feasibility of the ICI cancellation algorithm.
4/11
Figure 4. |Cl,128 | versus l for DCT-OFDM (i.e., α = 1) and FrCT-NOFDM.
Simulation and theoretical analysis
In this section, we will introduce an ICI cancellation algorithm based on iterative detection (ID) for FrCT-NOFDM. The
simulations are demonstrated to verify the feasibility of FrCT-NOFDM. In the simulation system, the AWGN channel is
employed. The number of subcarriers is set to 256 and the number of FrCT-NOFDM symbols is set to 4096. Eb /N0 denotes the
ratio between the energy per bit and the single-sided noise power spectral density.
An N × N correlation matrix C is defined to study the ICI of FrCT-NOFDM, in which the interference between subcarriers
l and m is represented by Cl, m , the cross-correlation value. When l equals to m, Cl, m represents the auto-correlation value for
each of the subcarriers. Cl, m can be calculated by
Cl, m =
απl(2n + 1)
απ(2n + 1)m
2 N−1
W
cos
·W
cos
.
m
∑ l
N n=0
2N
2N
(13)
Figure 4 presents the |Cl,128 | as a function of l for DCT-OFDM (i.e., α = 1) and FrCT-NOFDM. When α is set to 1,
the auto-correlation value for the 128th subcarrier is 1. In the same time, the cross-correlation values are almost equal to
0. Therefore, when α is set to 1, C is an identity matrix and the output of DCT is the OFDM signal. To achieve smaller
subcarrier spacing, α can be set to be smaller than 1. However, when α is less than 1, the cross-correlation values are no
longer equal to 0 and the orthogonality is destroyed. As shown in Figure 4, the interference caused by the adjacent subcarriers
is larger than that caused by farside subcarriers. The ICI increases with the decrease of α. Thus, BER performance of
FrCT-NOFDM degrades with the decrease of α. When the subcarrier spacing is equal to the half of symbol rate per subcarrier,
the subcarriers are orthogonal in DCT-OFDM, but the counterparts are no longer orthogonal in fractional Hartley transform
NOFDM (FrHT-NOFDM)14 or fractional Fourier transform-based NOFDM (FrFT-NOFDM)15 . Therefore, under the same
subcarrier spacing, the ICI in FrCT-NOFDM should be smaller than that in FrHT-NOFDM or FrFT-NOFDM.
Recently, ID algorithm has been researched to eliminate the ICI for FrFT-NOFDM15–17 . FrFT-NOFDM is modulated using
two-dimensional constellation (e.g., M-QAM), while, as Figure 1 depicts, FrCT-NOFDM is modulated using one-dimensional
constellation (e.g., M-PAM). The ID algorithm for FrCT-NOFDM with one-dimensional constellation can been given by
Si = R − (C − e)Si−1
(14)
where R is an N-dimensional vector of the received symbol demodulated by FrCT, Si is an N-dimensional vector of the
recovered symbol after ith iteration, and e is an N × N identity matrix.
The mapping strategy of 2-PAM constellation is outlined by Fig. 5. Only points that fall in the blue area can be mapped to
the corresponding constellation points. The points that fall in the uncertainty interval, which is from −d to d (i.e., the pink
area), are unchanged. The d is equal to 1 − i/I where i denotes the ith iteration and I is the total iterative number. In each
iteration, the operation of Equation (13) is implemented to eliminate the ICI. As a result, ICI is gradually reduced after each
iteration and thus the pink area can be decreased. The algorithm terminates when d is equal to zero. Since the decision for
5/11
Figure 5. Mapping strategy for 2-PAM constellation.
Figure 6. BER against Eb /N0 for DCT-OFDM (i.e., α = 1) and FrCT-NOFDM.
one-dimensional constellation is more simple than that for two-dimensional constellation, ID algorithm of FrCT-NOFDM has
lower computational complexity than that of FrFT-NOFDM.
Figure 6 depicts the BER against Eb /N0 for DCT-OFDM (i.e., α = 1) and FrCT-NOFDM. The modulated constellation
employs 2-PAM. I denotes the iterative number of ID algorithm. When α is less than 1, the BER performance of FrCT-NOFDM
without ID algorithm (i.e., I is set to 0) is seriously influenced by ICI. When I is set to 20, FrCT-NOFDM with α of 0.9 has the
same BER performance compared to the DCT-OFDM. This is because that ID algorithm can effectively eliminate the ICI in
FrCT-NOFDM with α of 0.9. When α is set to 0.8 or 0.7, the BER performance is still influenced by the residual ICI although
the ID algorithm is employed. At the 7% forward error correction (FEC) limit, the required Eb /N0 for FrCT-NOFDM with α
of 0.8 is about 2 dB higher than that for DCT-OFDM. If the FEC coding technique is employed, FrCT-NOFDM with α of 0.8
would have almost the same performance with DCT-OFDM. When α is set to 0.7, the BER of FrCT-NOFDM only achieves the
7% FEC limit. The BER performance degrades with the decrease of α due to the increase of residual ICI.
Figure 7 reveals BER versus Eb /N0 for FrCT-NOFDM and FrHT-NOFDM with different subcarrier spacing. The iterative
number of ID algorithm is set to 20. The subcarrier spacing is equal to 45%, 40% and 35% of the symbol rate per subcarrier
when α in FrCT-NOFDM is set to 0.9, 0.8 and 0.7, respectively and α in FrHT-NOFDM is set to 0.45, 0.4 and 0.35, respectively.
As discussed above, the ICI in FrCT-NOFDM should be smaller than that in FrHT-NOFDM. After ID algorithm, the residual ICI
in FrCT-NOFDM is also smaller than that in FrHT-NOFDM. Therefore, the BER performance of FrCT-NOFDM is much better
than that of FrHT-NOFDM. When the subcarrier spacing is set to 45% of symbol rate per subcarrier, the BER of FrCT-NOFDM
descends with the increase of Eb /N0 , but the BER of FrHT-NOFDM no longer decreases because the residual ICI begins to be
the major distortion when the Eb /N0 is larger than 10 dB. For generating FTN NOFDM signal, FrCT is a much better choice
than FrHT.
Figure 8 shows the BER against iterative number of ID algorithm for FrCT-NOFDM when Eb /N0 is set to 20 dB. When α
is set to 0.8, BER performance can be improved by the increase of iterative number. Therefore, it has the potential to further
eliminate the ICI in FrCT-NOFDM with α of 0.8. However, when α is set to 0.7, the increase of iterative number cannot
significantly improve the BER performance. This is because the ICI is too large to be completely eliminated by ID algorithm.
Therefore, it needs a more effective algorithm to compensate the interference. The ID-fixed sphere decoder (ID-FSD) algorithm
has a better performance than ID algorithm, which combines the ID and FSD algorithms15 . Furthermore, the channel coding is
6/11
Figure 7. BER versus Eb /N0 for FrCT-NOFDM and FrHT-NOFDM with different subcarrier spacing.
Figure 8. BER against iterative number of the ID algorithm.
effective for resisting the ICI and has been investigated for improving the BER performance of NOFDM7, 24 .
The simulation results verify the feasibility of the FrCT-NOFDM. When α is larger than a certain value (this value may
be no more than 0.8), ICI can be effectively eliminated and FrCT-NOFDM can achieve almost the same BER performance
compared to DCT-OFDM. This certain value is similar to the Mazo limit for FTN-NOTDM systems6 . Therefore, when the α is
greater than this certain value, the capacity of FrCT-NOFDM has the possibility to approach to the capacity limit.
Proof-of-concept experiment
To verify the feasibility of the FrCT-NOFDM, we set up an experiment as shown in Figure 1. At the transmitter end, the size
of FrCT was set to 256 and 2-PAM was modulated. Sixteen cyclic prefix samples were employed. One frame included 128
FrCT-NOFDM symbols, 10 training symbols and 1 synchronization symbol. The digital signal was uploaded into an arbitrary
waveform generator (Tektronix AWG7122C) operating at 10 GS/s to realize digital-to-analog conversion. The resolution of
arbitrary waveform generator was set to 8 bits. The overall link rate was 10 Gbit/s and the net bit rate was approximately 8.7
Gbit/s (1 bit/sample × 10 GS/s × 256/(256 + 16) × 128/(128 + 10 + 1) ≈ 8.7 Gbit/s). An external cavity laser (ECL) with a
linewidth of 100 kHz was used to generate the optical carrier. A Mach-Zehnder modulator (MZM) was used to modulate the
optical carrier with the generated electrical signal. The Vπ of the MZM is about 1.5 V and the bias voltage is set to about 1.5 V.
The launch optical power was set to 3 dBm. The length of standard signal mode fiber (SSMF) was 50 km. Its total loss was
approximately 10 dB. A variable optical attenuator was employed to change the received optical power.
7/11
Figure 9. Electrical spectra of DCT-OFDM (i.e., α = 1) and FrCT-NOFDM signal.
Figure 10. BER curves for DCT-OFDM (i.e., α = 1) and FrCT-NOFDM after back-to-back (BTB) and 50 km SSMF
transmission.
At the receiver end, the received optical signal can be converted into an electrical signal by the photodiode (Discovery
DSC-R401HG). The electrical signal was then filtered by a low-pass filter with a 3-dB bandwidth of 10 GHz. The filtered
electrical signal was captured by a real-time digital phosphor oscilloscope (Tektronix DPO72004C) operating at 50 GS/s
to implement analog-to-digital conversion. The generated digital signal was decoded by off-line processing in MATLAB.
The training symbols were used to estimate the channel characteristics for intra-symbol frequency-domain averaging (ISFA)
algorithm and frequency-domain equalization25 . After equalization, ID algorithm was employed to reduce the ICI.
Figure 9 shows the electrical spectra of DCT-OFDM (i.e., α = 1) and FrCT-NOFDM signal. The baseband bandwidth of
DCT-OFDM is equal to 5 GHz which is the Nyquist frequency of the signal with a 10-GS/s sample rate. In FrCT-NOFDM, the
baseband bandwidth is compressed to 4.5, 4 and 3.5 GHz and the corresponding Nyquist rate is 9, 8 and 7 Gbit/s when α is
set to 0.9, 0.8 and 0.7, respectively. In the experiment, the link rate of FrCT-NOFDM is 10 Gbit/s, which is faster than the
corresponding Nyquist rate. Therefore, the FrCT-NOFDM signal is a kind of FTN signal. When α is set to 0.8, 20% baseband
bandwidth saving can be obtained and the transmission rate is about 25% faster than the Nyquist rate. These results verify the
above theory analysis.
Figure 10 depicts the BER curves for DCT-OFDM (i.e., α = 1) and FrCT-NOFDM after back-to-back (BTB) and 50-km
SSMF transmission. The iterative number of the ID algorithm is set to 20. FrCT-NOFDM with α of 0.9 has the same BER
8/11
Figure 11. BER versus iterative number of the ID algorithm after 50-km SSMF transmission.
performance with DCT-OFDM. After 50-km SSMF transmission, the required received power at the 7% FEC limit was
measured to be approximately −12 dBm for FrCT-NOFDM with α of 0.9. Compared to BTB transmission, the power penalty
for 50-km SSMF transmission is about 3 dB. When α is set to 0.8, the ICI degrades the BER performance. After 50-km SSMF
transmission, the required received power at the 7% FEC limit was measured to be approximately −10 dBm for FrCT-NOFDM
with α of 0.8. Compared to DCT-OFDM, the power penalty for FrCT-NOFDM with α of 0.8 is about 2 dB at the 7% FEC
limit. When α is set to 0.7, the ICI severely degrades the BER performance, the BER can only achieve the 20% FEC limit.
Figure 11 depicts the BER against iterative number of the ID algorithm after 50-km SSMF transmission. The received
power is set to −8 dBm. The α is set to 0.8 and 0.7, respectively. In theory, ID algorithm with more iterative number can
eliminate more ICI. Therefore, the BER decreases with the increase of iterative number. However, when iterative number
is larger than 20, the increase of iterative number cannot significantly improve the BER performance. This may be because
the other distortions such as chromatic dispersion (CD) mainly influence the BER performance when the ICI is small. The
experiment results agree well with the simulation results. The feasibility of FrCT-NOFDM has been verified by both the
experiment and simulation.
Discussion
The history of FTN signal began with the paper of James Mazo in 19756 , who investigated the time-domain binary sinc-pulse
scheme. The FTN binary sinc-pulse scheme compresses the time period to obtain the transmission rate faster than Nyquist rate.
The pulses are accelerated with the time acceleration factor τ and become no longer non-orthogonal, thus the FTN sinc-pulse
signal can be considered as a kind of FTN NOTDM signal. Twenty-five percent more bits can be carried in the same bandwidth
while τ is set to 0.8, now known as the Mazo limit. In this paper, we demonstrate the principle of FTN NOFDM by taking
FrCT-NOFDM for instance. The subcarrier spacing is less than 50% of the symbol rate per subcarrier. When the bandwidth
compression factor α is set to 0.8, the subcarrier spacing is equal to 40% of the symbol rate per subcarrier, thus 20% baseband
bandwidth saving can be obtained and the transmission rate is about 25% faster than the Nyquist rate.
In 1940s, Shannon put forward the up limit of communication capacity for the Nyquist signal1, 2 . When the ISI is effectively
compensated, the FTN NOTDM signal can achieve a higher capacity limit than Nyquist signal7, 26 . To the best of our knowledge,
we first give the mathematical expression for capacity limit of FTN NOFDM signal,
1
PS
C ≤ ×W log2 1 +
.
(15)
α
PN + PICI
The capacity limit of FTN NOFDM signal is potentially higher than that of Nyquist signal when ICI is effectively compensated.
In other word, the capacity of FTN-NOFDM signal is likely to approach the limit only when ICI has been effectively eliminated.
Therefore, how to eliminate the ICI is crucial for FTN-NOFDM.
As verified in literatures6, 7 , when the α is set to the value between 1 and 0.8, there is not obvious deterioration in BER
performance by employing the optimal detection to compensate the interference. In this paper, we demonstrate the simulations
and experiments of FrCT-NOFDM system, which verify that ICI can be effectively eliminated when the α is set to the value
9/11
Algorithm 1 ID algorithm for 2-PAM constellation.
Input: Received symbol : R; Compression factor : α; Iterative number : I
Output: Recovered symbol : S
1: for l = 0; l < N; l + + do
2:
for m = 0; m < N; m + + do
απl(2n+1)
απ(2n+1)m
W
cos
·W
cos
.
3:
Cl, m = N2 ∑N−1
m
n=0 l
2N
2N
4:
end for
5: end for
6: Initialization : S0 = 0, d = 1
7: for i = 1; i ≤ I; i + + do
8:
Si = R − (C − e)Si−1
9:
if Si > d then
10:
Si = 1
11:
else if Si < −d then
12:
Si = −1
13:
else
14:
Si = Si
15:
end if
16:
d = 1 − i/I
17: end for
18: S = SI
19: Return S
. Calculating correlation matrix C
. Iterative operation
. Constellation mapping
. Updating d
between 1 and 0.8. As a result, the capacity limit of FrCT-NOFDM signal is potentially higher than that of the Nyquist signal
when the α is set to the value between 1 and 0.8. In the further work, more effective algorithm can be employed to eliminate
the ICI and higher-order M-PAM constellation can be investigated to obtain higher spectral efficiency. FrCT-NOFDM can be
potentially used in next-generation high-capacity wireless and optical communications.
Methods
The simulations and off-line processing in the experiment were both implemented by MATLAB. The encoding and decoding
were shown in Figure 1. In the simulation, the AWGN channel was employed. The channel equalization is not required for the
AWGN channel. In the experiment, the training symbols were used to estimate the channel characteristics. Meanwhile, the
ISFA algorithm can improve the performance of channel estimation25 . The frequency-domain equalization can compensate the
channel distortion by using the estimated channel characteristics.
In NOFDM system, ID algorithm is employed to eliminate the ICI, which is critical to improve the BER performance. The
ID algorithm for 2-PAM constellation is shown in Algorithm 1. Compared to two-dimensional constellation (i.e., M-QAM), the
constellation mapping in ID algorithm for one-dimensional constellation (i.e., M-PAM) is more simple and accurate.
References
1. Shannon C. E. A Mathematical Theory of Communication. The Bell System Technical Journal 27, 379–423 (1948).
2. Shannon C. E. Communication in the Presence of Noise. Proceedings of the IRE 37, 10–21 (1949).
3. Essiambre R. J. et al. Capacity Limits of Optical Fiber Networks. J. Lightwave Technol. 28, 662–701 (2010).
4. Essiambre R. J. and Robert W. T. Capacity Trends and Limits of Optical Communication Networks. Proceedings of the
IEEE 100, 1035–1055 (2012).
5. Liu X. et al. Phase-conjugated twin waves for communication beyond the Kerr nonlinearity limit. Nature Photonics 7,
560–568 (2013).
6. Mazo J. E. Faster-Than-Nyquist Signaling. The Bell System Technical Journal 54, 1451–1462 (1975).
7. Anderson, J. B., Rusek, F., and Viktor O. Faster-Than-Nyquist Signaling. Proceedings of the IEEE 101, 1817–1830 (2013).
8. Andrews, J. G. et al. What Will 5G Be?. IEEE Journal on Selected Areas in Communications 32, 1065–1082 (2014).
10/11
9. Zhang, J., Yu, J. & Chi, N. Transmission and full-band coherent detection of polarization multiplexed all-optical Nyquist
signals generated by Sinc-shaped Nyquist pulses. Scientific Reports 5, 13649 (2015).
10. Kaneda, N., Pfau, T. & Lee, J. Frequency Diversity MIMO Detection in Polarization Multiplexed Coherent Optical
Transmission.OptoElectronics and Communication Conference and Australian Conference on Optical Fibre Technology
(OECC/ACOFT) (Melbourne, Australia), paper TH12E-3 (2014, July. 6th).
11. Sato, M. et al. Frequency diversity MIMO detection for dual carrier DP-16QAM transmission.European Conference on
Optical Communication (ECOC) (Cannes, France), paper Th.2.5.1 (2014, Sept. 21th).
12. Igarashi, K., Tsuritani, T. & Morita,I. Bit-error rate performance of super-Nyquist-WDM DP-QPSK signals with duobinarypulse shaping. Optical Fiber Communication Conference (OFC) (San Francisco, California, USA), paper Th2A.18 (2014,
Mar. 9th).
13. Zhang, J. et al. Transmission of single-carrier 400G signals (515.2-Gb/s) based on 128.8-GBaud PDM QPSK over 10,130and 6,078 km terrestrial fiber links. Opt. Express 23, 16540–16545 (2015).
14. Zhou, J. et al. Faster-than-Nyquist non-orthogonal frequency-division multiplexing based on fractional Hartley transform.
Optics Letters 41, 4488–4491 (2016).
15. Xu, T. et al. An Improved Fixed Sphere Decoder Employing Soft Decision for the Detection of Non-orthogonal Signals.
IEEE Commun. Lett. 17, 1964–1967 (2013).
16. Darwazeh, I. et al. Optical SEFDM System; Bandwidth Saving Using Non-Orthogonal Sub-Carriers. IEEE Photonic. Tech.
Lett. 26, 352–355 (2014).
17. Nopchinda, D. et al. Dual Polarization Coherent Optical Spectrally Efficient Frequency Division Multiplexing. IEEE
Photonic. Tech. Lett. 28, 83–86 (2016).
18. Ahmed, N., Natarajan. T. & Rao, K. R. Discrete cosine transform. IEEE Tran. computers C-23, 90–93 (1974).
19. Zhou, J., Qiao, Y. J., Cai, Z. & Ji, Y. F. Asymmetrically Clipped Optical Fast OFDM Based on Discrete Cosine Transform
for IM/DD Systems. J. Lightwave Technol. 33, 1920–1927 (2015).
20. Zhou, J. et al. FOFDM Based on Discrete Cosine Transform for Intensity-Modulated and Direct-Detected Systems. J.
Lightwave Technol. 34, 3717–3725 (2016).
21. Tan, P. and Beaulieu, N. C. A Comparison of DCT-Based OFDM and DFT-Based OFDM in Frequency Offset and Fading
Channels. IEEE Trans. Commun. 54, 2113–2125 (2006).
22. Zhao, J. and Ellis, A. D. A Novel Optical Fast OFDM with Reduced Channel Spacing Equal to Half of the Symbol Rate Per
Carrier. Optical Fiber Communication Conference (OFC) (San Diego, California, USA), paper OMR1 (2010, Mar. 21th).
23. Sommerville, D. M. Y. An Introducation to the Geometry of N Dimensions. (METHUEN & CO. LTD, 1929).
24. Xu, T. et al. A Soft Detector for Spectrally Efficient Systems with Non-Orthogonal Overlapped Sub-Carriers. IEEE
Commun. Lett. 18, 1847–1850 (2014).
25. Liu, X. and Buchali, F. Intra-symbol frequency-domain averaging based channel estimation for coherent optical OFDM.
Opt. Express 16, 21944–21957 (2008).
26. Rusek, F. and Anderson, J. B. Constrained Capacities for Faster-Than-Nyquist Signaling. IEEE Trans. Inf. Theory 55,
764–775 (2009).
Acknowledgements (not compulsory)
This work was supported in part by National Natural Science Foundation of China (61427813, 61331010, 61271192); BUPT
Excellent Ph.D. Students Foundation; China Scholarship Council Foundation.
Author contributions statement
J. Z. and Y. Q. derived the theoretical results, J. Z. and Y. Q. conceived the simulation and experiment, J. Z., M. G., and X.
T. conducted the simulation and experiment, J. Z., Z. Y., Q. C., and Q. W. analyzed the results. All authors reviewed the
manuscript.
Additional information
Competing financial interests: The authors declare no competing financial interests.
11/11
| 7cs.IT
|
arXiv:1706.01230v2 [math.CO] 20 Mar 2018
On the Heapability of finite partial orders
János Balogh∗
Cosmin Bonchiş, Diana Diniş, Gabriel Istrate†
Ioan Todinca‡
March 21, 2018
Abstract
We investigate the partitioning of partial orders into a minimal number
of heapable subsets. We prove a characterization result reminiscent of the
proof of Dilworth’s theorem, which yields as a byproduct a flow-based
algorithm for computing such a minimal decomposition. On the other
hand, in the particular case of sets and sequences of intervals we prove
that this minimal decomposition can be computed by a simple greedytype algorithm. The paper ends with a couple of open problems related
to the analog of the Ulam-Hammersley problem for decompositions of sets
and sequences of random intervals into heapable sets.
Keywords: partial order; heapable sequence; Dilworth’s theorem; greedy algorithm; random intervals.
1
Introduction
The longest increasing subsequence is a classical problem in combinatorics and
algorithmics, with deep connections to statistical physics and random matrix
theory (for a very readable introduction see [23]).
Byers et al. [5] introduced an interesting variant of the concept of increasing sequence: sequence of integers A = a1 , a2 , . . . , an is heapable if numbers
a1 , a2 , . . . , an can be successively inserted as actual leaves into a heap-ordered
binary tree (not necessarily complete).
Heapability was further investigated in [15] (and, independently, in [22]).
In particular, a subgroup of the authors of the present papers showed that for
permutations one can compute in polynomial time a minimal decomposition
∗ Dept. of Applied Informatics, Gyula Juhász Faculty of Education, University of Szeged,
Hungary, [email protected]
† Dept. of Computer Science, West University of Timişoara and the e-Austria Research
Institute, Bd. V. Pârvan 4, cam. 045B, Timişoara, RO-300223, Romania. Corresponding
author’s email: [email protected]
‡ Université d’Orleans, INSA Centre Val de Loire, LIFO EA 4022, Orléans, France,
[email protected].
1
into heapable subsequences, and connected this problem to a variation on Hammersley’s interacting particle process (see also further work in [16, 2, 3], that
extended/confirmed some of the conjectures of [15]).
This paper started as a conversation on the heapability of sequences of intervals during a joint Timişoara-Szeged seminar on theoretical computer science
in November 2015. Its main purpose is to offer a different perspective on the
concept of heapability, by connecting the problem of computing a minimal decomposition of partial ordered sets into ”heapable subsequences” (that will be
parametrized in this paper by an integer k ≥ 2, and will be called k-chains)
to the classical theorems of Dilworth and Kőnig-Egerváry: we show that the
number of classes in a minimal decomposition can be obtained as a vertex cover
in a bipartite graph. As a byproduct we obtain an efficient algorithm based
on network flows for computing such a minimal decomposition. This results
and the ones from [15] (where the same parameter was computed, in the case
of integer sequences, via a direct greedy algorithm) raise the question whether
such greedy algorithms exist for other posets. We answer this question in the
affirmative way by adapting the result of [15] to sets and sequences of intervals,
ordered by the natural partial order, the setting that motivated the discussion
leading to the paper.
The structure of the paper is as follows: in Section 2 we review the main
concepts and technical results we will be concerned with. Then, in Section
3 we prove a numerical characterization inspired by (the proof of) Dilworth’s
theorem for the size of minimal partition of a partial order into k-chains. This
result provides, as a byproduct, a flow-based algorithm for the computation of
such a minimal partition. We then investigate, in Section 4 a special case of
this problem for sequences of intervals, showing that a greedy-type algorithm
computes such a minimal partition in this special case. In Section 5 we show that
the result above extends to (unordered) sets (rather than sequences) of intervals.
We extend this result in Section 6 from interval to trapezoid partial orders. In
Section 7 we investigate the maximum heapablity of interval partial orders,
proving that it has polynomial time algorithms: the status of this problem for
permutation posets is open. We conclude in Section 8 with some open questions
raised by our results.
2
Preliminaries
We will assume knowledge of standard graph-theoretic notions. In particular,
given integer k ≥ 1, rooted tree T is k-ary if every node has at most k descendants. Given a directed graph G = (V, E), we will denote by vc(G) the size of
the minimum vertex cover of G.
Let k ≥ 1. A sequence of integers A = (a1 , a2 , . . . , an ) is called k-heapable
if there exists a k-ary tree T with n nodes, such that its nodes are labeled by
a1 , a2 , . . . , an , and a bijection between vertices of T and elements of A such that
for any two nodes labeled ai , aj , if aj is a descendant of ai of T then i < j holds
for their indices and ai < aj .
2
We will be concerned with finite partially ordered sets (posets) only. Sequences of elements from a poset naturally embed into this framework by associating, to every sequence A = (a1 , . . . , an ) the poset QA = {(i, ai ) : 1 ≤ i ≤ n}
with partial order (i, ai ) ≤ (j, aj ) if and only if i ≤ j and ai ≤ aj . Given poset
Q, its subset B is a chain if the partial order of Q is a total order on A. Subset
C is an antichain if no two elements a, b of C are comparable with respect to
the patial order relation of Q. Dilworth’s theorem [7] states that the minimum
number of classes in a chain decomposition of a partial order Q is equal to the
size of the largest antichain of Q.
One particular type of partial orders we consider is that of permutation
orders: partial order is a permutation order if there exists a permutation
π of the elements of U such that, for all a, b ∈ U , rankπ (a) < rankπ (b) and
π[a] < π[b].
Another particular case we will be concerned with is that of interval orders.
Without loss of generality all our intervals will be closed subsets of (0, 1). We
define a partial order of them as follows: Given intervals I1 = [a1 , b1 ] and
I2 = [a2 , b2 ] with a1 < b1 and a2 < b2 , we say that I1 ≤ I2 if and only if the
entire interval I1 lies to the left of I2 on the real numbers axis, that is b1 ≤ a2 .
For technical reasons we will also require a total ordering of intervals, denoted
by ⊑ and defined as follows: I1 ⊑ I2 if either b1 < b2 or b1 = b2 and a1 < a2 .
On the other hand, by the phrase ”random intervals” we will mean, similarly
to the model in [17], random subintervals of (0, 1). A random sample I can
be constructed iteratively at each step by choosing two random real numbers
a, b ∈ (0, 1) and taking I = [min(a, b), max(a, b)].
We need to briefly review one classical proof of Dilworth’s theorem, due
to Fulkerson [11]. This proof employs the Kőnig-Egerváry theorem [18, 8],
stating that the size of a maximum matching in a bipartite graph is equal
to the minimum vertex cover in the same graph. This result is applied as
follows: first, we prove that each matching in GQ uniquely corresponds to a chain
decomposition of Q. Next we apply Kőnig’s theorem to the so-called split graph
associated to poset Q = (U, ≤) [10], i.e. the bipartite graph GQ = (V1 , V2 , E),
where V1 = {x− : x ∈ U } and V2 = {y + : y ∈ U } are independent copies of
U , and given x ≤ y ∈ U we add to E edge x− y + . The conclusion is that the
cardinality of a minimum chain decomposition of Q is equal to n, the number
of elements in U , minus the size of the smallest vertex cover in GQ defined
above. Finally, this latter quantity is proved to coincide to the size of the
largest antichain of Q.
We extend the definition of k-heapable sequences to general posets as follows:
given integer k ≥ 1 and poset Q = (U, ≤) a subset A of the base set U is called
k-heapable (or, equivalently, a k-chain of Q) if there exists a k-ary rooted tree T
and a bijection between A and the vertices of T such that for every i, j ∈ A, if
j is a descendant of i in T then i < j in Q. The k-width of poset Q, denoted by
k-wd(Q), was defined in [16] as the smallest number of classes in a partition of Q
into k-chains. For k = 1, by Dilworth’s theorem, we recover the usual definition
of poset width [24]. On the other hand, when π ∈ Sn is a permutation of n
elements, k-wd(Qπ ) particularizes to the parameter whose scaling properties
3
were investigated in [15].
Two important (unpublished) minimax theorems attributed in [13] to Tibor
Gallai deal with partitioning sets of intervals. We will only be concerned with
the first of them, that states that, given a set of intervals J on the real numbers
line the following equality holds:
Proposition 1. The minimum number of partition classes of J into pairwise
disjoint intervals is equal to the maximum number of pairwise intersecting intervals in J.
Of course, the first quantity in Proposition 1 is nothing but the 1-width of
the partial order ≤ on intervals. As for random permutations (the well-known
Ulam-Hamersley problem [19, 25]), the scaling of (the expected value of) this
parameter for sets R of n random intervals has been settled and has the form
[17]
2 √
E[1-wd(R)] ∼ √
n.
π
2.1
A graph-theoretic restatement
Problems that we are concerned with have, it turns out, an algorithmic interpretation that is strongly related to problems of computing the maximum
independent sets and the chromatic number of various classes of graphs. The
idea first appeared (somewhat implicitly) in [22]. One can make it fully explicit
as follows: Given directed graph P = (U, ≤), call a set of elements W ⊆ U a
k-treelike independent set of P if there is a k-ary tree T , not necessarily complete, and a bijection V (T ) → W such that, for all vertices v, w of T , if v is an
ancestor of w in T then f (v) and f (w) are not connected in P .
For k = 1 the previous concept coincides with that of an independent set.
Paralleling the case k = 1 (for which a polynomial time algorithm is known, in
the form of patience sorting [20]), we can restate the open problem from [5] as
the problem of computing a maximal 2-treelike independent set:
Proposition 2. Computing the longest k-heapable subsequence of an arbitary
permutation is equivalent to computing a maximum k-treelike independent sets
in the permutation digraph induced by the Hasse diagram of P .
When k is implied, we will informally use the name maximum heapable subset
for the problem of computing a maximum k-chain of an arbitrary permutation.
Similarly, the problem of partitioning the set into k-chains can be regarded as
a generalization of graph coloring: define a A k-treelike coloring of a partial
order P to be a partition of the universe U into classes that induce k-treelike
independent sets. The k-treelike chromatic number of a partial order P is the
minimum number of colors in a k-treelike coloring of P .
The main result from [15] is equivalent to the following statement:
Proposition 3. For k ≥ 2 the greedy algorithm of Figure 1 computes an optimal
k-treelike coloring of permutation partial orders.
4
Input: Permutation order P specified by permutation π ∈ Sn .
Output: A k-treelike coloring of P .
let F = ∅
for i := 1 to n do:
if π(i) cannot become the child of any node in F .
then
Start a new tree with root π(i).
else
make π(i) a child of the largest value π(j) < π(i),
j < i that has fewer than two children.
Figure 1: The greedy best-fit algorithm for tree-like colorings of permutation
posets.
Again, let us remark that the corresponding statement for k = 1 is well
known, as it is equivalent to the greedy coloring of permutation graphs, accomplished by Patience Sorting.
3
Main result
Our main result gives a characterization of the k-width of a finite poset that is
strongly related to the proof of Dilworth’s theorem sketched above:
Definition 1. Given poset Q = (U, ≤) an integer k ≥ 1, the k-split graph
associated to poset Q is the bipartite graph GQ,k = (V1 , V2 , E), where
−
• V1 = {x−
1 , . . . , xk : x ∈ U }
• V2 = {y + : y ∈ U }
+
• given x, y ∈ U, x ≤ y and 1 ≤ i ≤ k, add to E edge x−
i y .
Theorem 1. Let Q = (U, ≤) be a finite poset with n elements and a fixed integer
k ≥ 1. Then
k-wd(Q) = n − vc(GQ,k ).
(1)
Proof. Define a left outgoing k-matching of graph GQ to be any set of edges
A ⊆ E such that for every x, y ∈ U , degA (x− ) ≤ k and degA (y + ) ≤ 1.
5
Claim 1. Partitions of Q into k-chains bijectively correspond to left k-matchings of GQ . The number of classes of a partition is equal to n minus the number
of edges in the associated left k-matching.
Proof. Consider a left k-matching A in GQ . Define the partition PA as follows:
roots of the k-chains consist of those x ∈ U for which degA (x+ ) = 0. There
must be some element x ∈ U satisfying this condition, as the minimal elements
of P with respect to ≤ satisfy this condition.
Now we recursively add elements of U to the partition P (in parallel) as
follows:
1. All elements y ∈ U not yet added to any k-chain, and such that y + is
connected to some x− by an edge in A are added to the k-chain containing
x, as direct descendants of x. Note that element x (if there exists at
least one y with this property) is unique (since degA (y + ) = 1), so the
specification of the k-chain to add y to is well defined. On the other hand,
each operation adds at most k successors of any x to its k-chain, since
degA (x− ) ≤ k.
2. If all direct predecessors of an element x ∈ U have been added to some
k-chain and are no longer leaves of that k-chain then x becomes the root
of a new k-chain.
Conversely, given any partition P of U into k-chains, define set A consisting of
edges x− , y + such that x is the father of y in a k-chain. It is immediate that A
is a left k-matching.
Corollary 1. There is a bijective mapping between partitions of Q into k-chains
and matchings of GQ,k such that the number of k-chains in a partition is n minus
the number of edges in the matching.
Proof. We will actually show how to associate left k-matchings of GQ to matchings of GQ,k . The idea is simple: given a node x− of GQ with l ≤ k neighbors in
−
V2 , construct a matching in GQ,k by giving each of x−
1 , . . . , xl a single neighbor
−
from the neighbors of x . In the other direction, if e = xi y is an edge in the
matching of GQ,k then consider the edge in appropriate x, y edge in GQ . It is
easy to check that it gives a k−chain in GQ , and the number of k−chains is the
number of edges in the left k−matching in GQ which is the same as n minus
vc(GQ ).
We complete the proof of Theorem 1 by applying in a straightforward way
(based on Claim 1 and Corollary 1) the Kőnig’s theorem to graph GQ,k .
Corollary 2. One can compute parameter k-wd(Q) by creating a flow network
ZQ and computing the value of the maximum flow of ZQ consisting of :
• vertices and edges of GQ , with capacity 1.
• a source s, connected to nodes in V1 by directed edges of capacity k,
6
• a sink t, that all nodes in V2 connect to via oriented edges of capacity 1.
computing the maximum s-t network flow value f in network ZQ and outputting
k-wd(Q)=n-f.
Proof. Straightforward, this is simply the maximal flow algorithm for computing
the maximal size left k-matching in GQ , similar to the construction for maximum
matchings in bipartite graphs in the literature.
4
Heapability of sequences of intervals: a greedy
approach
The result in the previous section naturally raises the following question: for
random permutations a greedy approach, extending the well-known patience
sorting algorithm [20] works, as it was shown in [15]. On the other hand network
flow algorithms are not greedy. One is naturally led to inquire whether such
an algorithm can be defined/works for any other partial order. In the sequel we
provide an affirmative answer to this question, in the case of interval sequences:
Theorem 2. For every fixed k ≥ 1 there exists a (polynomial time) greedy algorithm that, given the sequence of intervals S = (I1 , I2 , ..., In ) as input, computes
a minimal partition of S into k-chains.
Before proceeding with the proof of theorem 2, let us remark a potential
application of a variant of this result to parallel computing: many algorithms
in this area (e.g. algorithms using a parallel prefix-sum design methodology [4])
require the computation of all prefixes of an associative operation A1 ∗A2 ∗. . . An .
Operations being performed (each corresponding to one computation of a *product) are arranged on a binary tree. In the (completely equivalent) maxheap variant of theorem 2, children intervals are required to be less or equal
to the parent interval with respect to ordering ≤. This is quite natural from
the standpoint of parallel computing: consider the setting of a parallel-prefix
problem where each intermediate *-computation is a rather costly operation;
intervals now represent times when these operations can be scheduled. The
requirement that the parent interval be larger than child intervals with respect
to ≤ is completely natural, as child computations need to complete in order
to feed their results to the parent computation. Thus our result answers the
question whether all the time intervals can be scheduled on a single heap-ordered
binary tree, and gives such a scheduling, if the answer is affirmative.
4.1
Proof of theorem 2
Without loss of generality, we will present the proof of the theorem for k = 2,
but the proof is similar for all k’s. The proof employs the concept of slots,
adapted for interval sequences from similar concepts for permutations [5, 15]:
7
Definition 2. When a new interval is added to a k-chain it opens k new positions to possibly insert other intervals as direct successors into this node. Each
position has an associated integer value that will be called its slot. The value of
all empty slots created by inserting I1 = [a1 , b1 ] into a k-chain will be b1 , the
right endpoint of I1 .
Definition 3. An interval I is compatible with a slot with value x if all of I
lies in [x, ∞).
Intuitively, x is the smallest value of the left endpoint of an interval that
can be inserted in the k-chain as a child of an interval I1 with right endpoint x
while respecting the heap property. Indeed, as k-chains are (min-)heap ordered,
an insertion of an interval I into a k-chain as a child of I1 is legal if the interval
I is greater than I1 with respect to ≤ relation. This readily translates to the
stated condition, that the start point of I must be greater or equal than the slot
value of its parent.
Input: A sequence of intervals I = (I1 , I2 , . . . , In ).
Output: A partition H of I into k-chains.
for i := 1 to n do:
if Ii = [li , ri ] can be inserted into some empty slot
then insert Ii in the highest-valued compatible slot (a child
of the node with this slot).
else create new k-chain rooted at Ii
Figure 2: The greedy best-fit algorithm for k-chain partition of sequences of
intervals.
The proposed greedy best-fit algorithm for computing a minimum partition
into 2-chains a sequence of n intervals is described in Figure 2. As an example, consider the sequence of intervals S1 below, with k = 2. The resulting
configuration is shown in Figure 3.
S1 = {[1, 7], [1, 11], [11, 12], [15, 16], [7, 9], [8, 16], [1, 2], [3, 19], [13, 17], [5, 7]}
By choosing the highest valued slot available for insertion for Ik = [ak , bk ],
we make sure that the difference between the slot value s and ak is minimal.
This is desirable because there may be some interval further down the sequence
with starting point value in between the s and ak that fits slot s but cannot be
inserted there, as the slot is no longer available.
We define the concepts of signature of a multuset of slots and domination
between such multisets in a similar way to the corresponding concepts for permutations in [5]:
8
H3
H1
[1, 2]
[1, 7]
7
7
9
16
[7, 9]
9
2
2
19
7
[3, 19]
[8, 16]
19
16
[5, 7]
7
H2
[1, 11]
11
11
[11, 12]
12
12
16
17
[15, 16]
16
[13, 17]
17
Figure 3: The 2-chain configuration corresponding to S1 .
Definition 4. Given multiset of slots H, we call the vector of slots, sorted in
increasing order, the signature of H. We shall denote this by sig(H). By slightly
abusing notation, we will employ the previous definition in the obvious way when
H is a union of 2-chains as well.
Definition 5. Multiset P dominates multiset Q (denoted P Q) if
|sig(P )| ≤ |sig(Q)| and for all 1 ≤ i ≤ |sig(P )| we have sig(P )[i] ≤ sig(Q)[i].
For example, for the 2-chains in the Figure 3 their corresponding signatures
are, respectively:
sig(H1 ) = [9, 9, 16, 16];
sig(H2 ) = [11, 16, 16, 17, 17];
sig(H3 ) = [7, 7, 19, 19].
Therefore, in our example H1 H2 , (H1 dominates H2 ), but no other
domination relations between H1 , H2 , H3 are true.
Lemma 1. Assume that A, B are multisets of slots and A dominates B. Then
the multisets A′ and B ′ obtained after inserting a new interval into the largest
compatible slot of A, and into an arbitrary compatible slot of B, propagates the
domination property, i.e. A′ dominates B ′ .
Proof. Let sig(A) = [a1 , a2 , . . . , a|sig(A)| ] and sig(B) = [b1 , b2 , . . . , b|sig(B)| ].
Also, by convention, define a0 = b0 = −∞ and a|sig(A)|+1 = b|sig(B)|+1 = +∞.
Proving that A′ B ′ is equivalent to proving that |sig(A′ )| ≤ |sig(B ′ )| and
for all indices 1 ≤ l ≤ |sig(A′ )|:
sig(A′ )[l] ≤ sig(B ′ )[l].
9
(2)
j′
k
i′
ai
sig(A)
bk
sig(B)
bj
bj ′
aj
sig(A′ )
sig(B ′ )
bk+1
➀
bj bj+1
ai′
ai+1
bj ′ y
y bj ′ +1
➁
ai′ y
y ai′ +1
bi′ bi′ +1
➂
➃
Figure 4: The various cases of inserting a new interval [x, y] into B and A.
The cardinality condition can be easily verifies: indeed, |sig(A′ )| is either |sig(A)|+
1 (if the process adds two copies of y but also kills a lifeline) or |sig(A)| + 2
(no slot exists with a value less or equal to x), and similarly for |sig(B ′ )|. It
follows easily from the domination property that when |sig(A′ )| is larger by
2, than |sig(A)| then so is |sig(B ′ )| with respect to |sig(B)|. Indeed, since
|sig(A′ )| = |sig(A)| + 2, no slot of A has value lowe than x (or it would lose a
lifeline). Thus a1 > x and since A B, b1 ≥ a1 > x. So no slot of B is smaller
than x either.
As for the second condition, consider the slots from A and B which interfere
with the newly arrived interval as follows:
• Let i and i′ be the (unique) indices such that ai ≤ x < ai+1 and ai′ ≤ y <
ai′ +1 hold in A, with i, i′ ∈ 0, · · · , |sig(A)|.
• Similarly, let j and j ′ be the unique indices such that bj ≤ x < bj+1 and
bj ′ ≤ y < bj ′ +1 hold in B, with j, j ′ ∈ 0, · · · , |sig(B)|.
Since A B, it follows that i ≥ j and i′ ≥ j ′ . Suppose that we insert the
interval [x, y] in B in an arbitrary slot bk ≤ bj (thus removing one life of slot bk
and inserting two copies of slot y). The rest of the proof is by a case analysis.
The four cases which can be distinguished (Fig. 4) are:
Case 1. l < k:
In this case, none of the signatures were affected at position l by the
insertion of [x, y], hence:
sig(A′ )[l] = sig(A)[l] ≤ sig(B)[l] = sig(B ′ )[l].
Case 2. l ∈ [k, j ′ ):
In this range all slots from B ′ have been shifted by one position to the left
compared to B due to the removal of bk . Consequently:
sig(A′ )[l] = sig(A)[l] ≤ sig(B)[l + 1] = sig(B ′ )[l].
10
Case 3. l ∈ [j ′ , i′ ):
Knowing that j ′ is the position in A where we inserted two new slots with
value y, i′ is the position in B where we inserted these same slots and
i′ ≥ j ′ , the following is true:
sig(A′ )[l] ≤ sig(B)[i′ ] = y = sig(B ′ )[j ′ ] ≤ sig(B ′ )[l].
Case 4. l ≥ i′ :
a) For l = i′ and l = i′ + 1 : sig(Hq′ )[l] = sig(Hp′ )[j ′ ] = y. Since
j ′ < i′ < i′ + 1, then:
sig(A′ )[i′ ] = y = sig(B ′ )[j ′ ] ≤ sig(B ′ )[i′ ].
sig(A′ )[i′ + 1] = y = sig(B ′ )[j ′ ] ≤ sig(B ′ )[i′ + 1].
b) For l > i′ + 1 the two signatures do not have any shifted components
compared to the original signatures of A and B, so:
sig(A′ )[i′ + k] = sig(A)[i′ + k − 1] ≤ sig(B)[i′ + k − 1] = sig(B ′ )[i′ + k]
In conclusion, sig(A′ )[l] ≤ sig(B ′ )[l] for any l, and relation A′ B ′ follows.
Lemma 2. Given sequence I of intervals, consider the optimal way to partition
I into k-chains. Let r be a stage where the greedy best-fit algorithm creates a
new k-chain. Then the optimal way also creates a new k-chain.
Proof. We use the fact that, by Lemma 1, before every step r of the algorithm
the multiset Γr−1 of slots created by our greedy algorithm dominates the multiset Ωr−1 created by the optimal insertion.
Suppose that at stage r the newly inserted interval Ir element causes a new
k-chain to be created. That means that the left endpoint lr of Ir is lower than
any of the elements of multiset Γr−1 . By domination, the minimum slot of Ωr−1
is higher than the minimum slot of Γr−1 . Therefore lr is also lower than any
slot of Ωr−1 , which means that the optimal algorithm also creates a new k-chain
when inserting Ir .
Lemma 2 proves that the Greedy algorithm for insertion of new intervals is
optimal.
4.2
The interval Hammersley (tree) process
One of the most fruitful avenues for the investigation of the scaling properties
of the LIS (Longest increasing subsequence) of a random permutation is made
via the study of the (so-called hydrodynamic) limit behavior of an interacting
particle system known as Hammersley’s process [1]. This is a stochastic process
that, for the purposes of this paper can be defined (in a somewhat simplified
form) as follows: random numbers X0 , X1 , . . . , Xn , . . . ∈ (0, 1) arrive at integer
moments. Each value Xj eliminates (”kills”) the smallest Xi > Xj that is still
11
alive at moment j. Intuitively, ”live” particles represent the top of the stacks
in the so-called patience sorting algorithm [20] that computes parameter LIS.
The problem of partitioning a random permutation into a minimal set of kheapable subsequences is similarly connected to a variant of the above process,
introduced in [15] and further studied in [2, 3], where it was baptized Hammersley’s trees process. Now particles come with k lives, and each particle Xj merely
takes one life of the smallest Xi > Xj , if any (instead of outright killing it).
The proof of theorem 2 shows that a similar connection holds for sequences of
intervals. The Hammersley interval tree process is defined as follows: ”particles”
are still numbers in (0, 1), that may have up to k lives. However now the
sequence I0 , I1 , . . . , In , . . . is comprised of random intervals in (0, 1). When
interval In = [an , bn ] arrives, it is an that takes a life from the largest live
particle y < an . However, it is bn that is inserted as a new particle, initially
with k lives.
Corollary 3. Live particles in the above Hammersley interval process correspond to slots in our greedy insertion algorithm above. The newly created kchains correspond to local minima (particle insertions that have a value lower
than the value of any particle that is live at that particular moment).
5
From sequences to sets of intervals
Theorem 2 dealt with sequences of intervals. On the other hand a set of intervals
does not come with any particular listing order on the constituent intervals.
Nevertheless, the problem can be easily reduced to the sequence case by the
following:
Theorem 3. Let k ≥ 1 and Q be a set of intervals. Then the k-width of Q
is equal to the k-width of GrQ , the sequence of intervals obtained by listing the
intervals in the increasing order of their right endpoints (with earlier starting
intervals being preferred in the case of ties).
Proof. Clearly k-wd(Q)≤ k-wd(GQ ), since a partition of GrQ into k-chains is
also a partition of Q.
To prove the opposite direction we need the following:
Lemma 3. Let S be a multiset of slots. Let I1 = [a1 , b1 ], I2 = [a2 , b2 ] be two
intervals such that b1 < b2 or b1 = b2 and a1 < a2 . Let S1 and S2 be the
multisets of slots obtained by inserting the two intervals in the order (I1 , I2 )
and (I2 , I1 ), respectively. Then S1 dominates S2 .
Proof. The only nontrivial cases are those for which S1 6= S2 . This condition
can only happen when the insertions of I1 , I2 “interact”.
Indeed, let x, y be the largest slots less or equal to a1 , a2 , respectively. If
x 6= y and b1 does not become (after insertion of I1 ) the slot occupied by b2
instead of y (nor does the symmetric situation for the insertion order I2 , I1 hold)
12
then S1 , S2 are obtained by adding k copies of b1 , b2 each, and deleting x, y, so
S1 = S2 .
If instead x = y then the two slots to be deleted (for both insertion orders)
are x and the largest slot smaller or equal than x. Again S1 = S2 .
The only remaining case is when the slot removed after insertion of I2 is
b1 . In this case S cannot contain any element in the range (b1 , a2 ). Insertion
(I1 , I2 ) removes p, adds k − 1 copies of b1 and k copies of b2 . On the other hand,
insertion (I2 , I1 ) may remove some element x′ . It is certainly x ≤ x′ ≤ b1 . It
then adds k copies of b2 . Then it removes some x′′ ≤ x and adds k copies of b1 .
Thus both sets S1 , S2 can be described as adding k copies of b1 and k copies of
b2 to S, and then deleting one or two elements,
Two elements
- x and b1 in the case of order (I1 , I2 )
- x′ and x′′ in the case of order (I2 , I1 ).
in the case when some element of S is strictly less than a1 ,
And one element
- b1 in the case of order (I1 , I2 )
- x′′ in the case of order (I2 , I1 ).
In the second case the result follows by taking into account the fact that x′′ ≤ b1
and the following lemma:
Lemma 4. Let S be a multiset of slots. Let s2 ≤ s1 ∈ S, and let S1 , S2 be
the sets obtained by deleting from S elements s1 (or s2 , respectively). Then S1
dominates S2 .
Proof. It follows easily from the simple fact that for every r ≥ 1 and multiset
W the function that maps an integer x ∈ W to the r’th smallest element of
W \ x is non-increasing (more precisely a function that jump from the r + 1’th
down to the r’th smallest element of W )
The first case is only slightly more involved. Since x′ ≤ x, by applying
Lemma 4 twice we infer:
S1 =
S \ {x, b1 } = (S \ {b1 }) \ {x} (S \ {b1 }) \ {x′ } = (S \ {x′ }) \ {b1 }
(S \ {x′ }) \ {x′′ } = S2 .
which is what we wanted to prove.
From Lemma 3 we infer the following result
13
Input: A set of intervals I.
Output: A partition H of I into k-chains.
Sort the intervals w.r.t. ⊑: I = (I1 , . . . , In ).
For i := 1 to n do:
If Ii = [ai , bi ] can be inserted into some empty slot
then insert Ii in the highest-valued compatible slot.
else create new k-chain rooted at Ii
Figure 5: The greedy algorithm for sets of intervals.
Lemma 5. Let 1 ≤ r ≤ n and let X, Y be two permutations of intervals I1 ,
I2 , . . . , In ,
X = (I1 , . . . , Ir−1 , Ir , Ir+1 , . . . In ),
Y = (I1 , . . . , Ir−1 , Ir+1 , Ir , . . . In ).
respectively (i.e. X, Y differ by a transposition). If Ir ≤ Ir+1 then multisets of
slots SX , SY obtained by inserting intervals according to the listing specified by
X and Y, respectively, satisfy
SX SY .
Proof. WIthout loss of generality one may assume that r = n − 1 (as the result
for a general n follows from this special case by repeatedly applying Lemma
1). Let S be the multiset of slots obtained by inserting (in this order) intervals
I1 , . . . , Ir−1 . Applying Lemma 3 to intervals Ir , Ir+1 we complete the proof of
Lemma 5.
Now the opposite direction in the proof of theorem 3 (and the theorem)
follows: the multiset of labels obtained by inserting the intervals according to
GrQ dominates any multiset of labels arising from a different permutation, since
one can ”bubble down” smaller intervals (as in bubble sort), until we obtain GrQ .
As we do so, at each step, the new multiset of labels dominates the old one.
Hence SGr(Q) dominates all multisets arising from permutations of I1 , . . . , In , so
sequence GrQ minimizes the parameter k-wd among all permutations of Q.
Corollary 4. The greedy algorithm in Figure 5 computes the k-width of an
arbitrary set of intervals.
14
Corollary 5. Modify the Hammersley interval process to work on sets of intervals by considering them in non-decreasing order according to relation ⊑.
Then live particles in the modified process correspond to slots obtained using
our greedy insertion algorithm for sets of intervals. New k-chains correspond to
local minima (such particle insertions that have a value lower than the value of
any particle that is live at that particular moment).
6
Extension to sequences of elements from a
trapezoid partial order
The theorem in the previous section is strongly reminiscent of the fact that for
interval partial orders a greedy best-fit algorithm computes the chromatic number [21]. This result has an extension to an even more general class of graphs,
that of trapezoid graphs [6]. Trapezoid graphs are an extension of both interval
and permutation graphs that unify many natural algorithms (for problems such
as maximum independent set, coloring) for the two classes of graphs.
As noted in [9], trapezoid graphs can be equivalently defined as the cocomparability graphs of two-dimensional boxes, with sides parallel to the coordinate
axes:
Definition 6. A box is the set of points in R2 defined as
B = {(x1 , x2 ) ∈ R2 | liB ≤ xi ≤ riB , i = 1, 2}
for some numbers liB ≤ riB ∈ R, where lB = (l1B , l2B ) is the lower corner of the
box B and rB = (r1B , r2B ) is the upper corner.
The interval I(B) associated to box B is the projection onto the y axis of
box B. Clearly I(B) = [l2B , r2B ].
The dominance partial order on intervals naturally extends to boxes:
Definition 7. The dominance partial order among boxes is defined as follows:
box B2 dominates box B1 if point LB2 dominates point rB1 , i.e.
riB1 ≤ liB2 for i = 1, 2.
A trapezoid partial order is a poset P induced by the dominance partial order on
a finite set V of boxes. l and u refer to the lower and upper corner coordinate
functions defining the boxes. That is, for every v ∈ V , l(v) is the lower corner
of box v and u(v) is its upper corner.
The box representation allowed the authors of [9] to give a “sweep-line algorithm” for coloring trapezoid graphs that improved the coloring algorithm
in [6]. Such an algorithm can be described as moving a (vertical) sweep line
L from left to right through the x positions of the lower/upper corners of the
boxes from B.
As noted in Section 2.1, partition into k-heapable sequences is a natural generalization of coloring permutation graphs. In the sequel we give an algorithm
15
that generalizes the sweep-line coloring algorithm for trapezoid graphs from [6]
to the partition into k-chains of a particular class of sequences of boxes.
Theorem 4. Let B = (B1 , B2 , . . . , Bn ) be a set of two-dimensional axis-parallel
boxes, totally ordered by the x-coordinates of their right endpoints. Then the
greedy sweep-line algorithm in Figure 6 computes an optimal partition of sequence B into k-chains.
Input: A sequence of boxes B = (B1 , B2 , . . . , Bn ).
Output: A partition H of B into k-chains.
initialize S = {d}, where d is a real number smaller
than all the y coordinates of corners in l(B) ∪ u(B),
marked available
foreach p = (p1 , p2 ) ∈ l(B) ∪ u(B) from left to right do:
q ← first available slot below p2 in S
if p = l(v) for some v ∈ B then
if q = u(w)2 for some w ∈ B then
insert v in the k-chain of w as a child of this node
remove q from S
add k copies of p2 to S, marking them unavailable
else // (q == d)
start a new k-chain rooted at v
add k copies of p2 to S, marking them unavailable
if p = u(v) for some v ∈ B then
mark all slots with value p2 in S as available
return the set of k-chains constructed by the algorithm.
Figure 6: The greedy sweep-line algorithm for x-sorted trapezoid sequences.
Proof. The proof of Theorem 4 leverages and extends the method used for intervals in the proof of Theorem 2. The fact that sequence B is sorted in increasing
order of the x-coordonate of the endpoints allows us to see the problem as one
ofn intervals: instead of dealing with boxes B1 , B2 , . . . , Bn we will instead deal
with the associated intervals I(B1 ), I(B2 ), . . . , I(Bn ).
16
We will assume that the x and the y coordinates of all boxes in B are all
distinct. As usual with such algorithms, the truth of our statement does not
rely on this assumption: if the statement is not true then points can be slightly
perturbed to make the assumption true. The algorithm we give then extends
to the degenerate cases as well.
The sweep line maintains a multiset S of slots, that correspond to right endpoints of the intervals associated with the boxes processed so far. The difference
with respect to the case of Theorem 2 is that the slots are now of one of two
types:
- unavailable: a slot of this type is present in multiset S but cannot be
used to process intevals.
- available: a slot of this type can be fully employed when processing a
new interval.
The action of a sweep line comprises two types of actions:
- insertion: This happens when the sweep line reaches the left corner of
some box Bj . We process the interval I(Bj ) similarly to the process in
Theorem 2. Namely, we remove one lifeline from the largest available
slot with value at most lj2 , and insert into S k slots with value rj2 . These
slots are marked unavailable.
- state change: This happens when the sweep line reaches the right corner of
some box Bj . All the slots with value rj2 change marking, from unavailable
to available.
The intuitive explanation for this modification is clear: a box D can become
the parent of another box E only when the right corner of D is to the left (on
the x axis) of the left corner of E. Thus, if the sweep line has not reached the
right endpoint of a box D then this box is not eligible to become the parent of
any currently processed box.
Let σt , be the multiset of all slots created by the sweep-line algorithm up
to at a given moment t. Let OP Tt be the multiset of slots corresponding to
′
the optimal partition into k-chains. Let σt be the corresponding (multi)set of
′
all slots marked available at moment t. Let OP Tt be the set of slots in OP Tt
marked available at moment t.
The basis for the proof of Theorem 4 is the following adaptation of Lemma 1
to the setting of slots with availability statuses:
Lemma 6. For every t ≥ 0, sig(σt′ ) sig(OP Tt′ ).
Proof. By induction on t. The statement is clear for t = 0, since both σ0 and
′
OP T0 are empty. Assume the claim is true for all t′ < t. Let sig(σt−1
) =
′
′
′
)
=
[b
,
b
,
.
.
.
,
b
].
Also,
by
[a1 , a2 , . . . , a|sig(σt−1
]
and
sig(OP
T
1 2
)|
|sig(OP Tt−1 )|
t−1
′
′
convention, define a0 = b0 = −∞ and a|sig(σt−1
=
b
=
+∞.
)|+1
|sig(OP Tt−1 )|+1
Finally, let [ls2 , rs2 ] is the interval to be processed at stage t.
17
′
′
Proving that σt−1
OP Tt−1
entails proving that |sig(σt′ )| ≤ |sig(OP Tt′ )|
′
and for all indices 1 ≤ l ≤ |sig(σt )|:
sig(σt′ )[l] ≤ sig(OP Tt′ )[l].
(3)
• Case 1: t is an insertion step:
′
σt′ is obtained from σt−1
by removing the largest available slot less or
2
′
′
equal to ls . OP Tt is obtained from OP Tt−1
by perhaps removing some
2
slot with value at most ls .
′
The inequality |σt′ | ≤ |OP Tt−1
| follows easilty from the corresponding
′
inequality at stage t − 1. Indeed, σt−1
may lose an element (in which case
′
OP Tt−1 may also lose at most one element) or stay the same (in which
′
case OP Tt−1
also stays the same (because there is no available slot to lose
a lifeline.)
′
As for inequality (3): if OP Tt−1
does not lose an element the inequality
follows easily from the induction hypothesis for stage t − 1 and the fact
′
that elements of σt−1
stay in place or shift to the left.
′
′
If both σt−1
and OP Tt−1
lose one element to yield σt′ , OP Tt′ , there are
four types of positions l:
– Those below both deleted positions. They are unchanged as we move
′
′
from σt−1
and OP Tt−1
to σt′ , OP Tt′ . That is:
′
′
σt′ [l] = σt−1
[l] and OP Tt′ [l] = OP Tt−1
[l].
Hence inequality (3) is true by the induction hypothesis.
– Those above both deleted positions. They get shifted to the left by
one in both σt′ and OP Tt′ . That is:
′
′
σt′ [l] = σt−1
[l + 1] and OP Tt′ [l] = OP Tt−1
[l + 1].
Hence again inequality (3) is true by the induction hypothesis.
– Those below one but above the other deleted position. By the fact
′
′
that the deleted slot in OP Tt−1
is less or equal than the one in σt−1
2
(since this is the largest available slot less or equal to ls ), it follows
that
′
′
σt′ [l] = σt−1
[l] and OP Tt′ [l] = OP Tt−1
[l + 1].
Hence
′
′
′
σt′ [l] = σt−1
[l] ≤ OP Tt−1
[l] ≤ OP Tt−1
[l + 1] = OP Tt′ [l].
so relation (3) holds in all cases.
• Case 2: t is a change step: The effect of such a step is that, both in
σt−1 and OP Tt−1 the k slots with value rs2 (that were previously inserted,
but marked unavailable) become available.
18
′
′
Since |σt′ | = |σt−1
| + k and, similarly, |OP Tt′ | = |OP Tt−1
| + k, statement
′
′
|σt | ≤ |OP Tt | follows from the analogous statement for t − 1.
Positions l before the insertion points of the k copies of rs2 are not modified
in σt′ , OP Tt′ so equation (3) follows for such l’s from the corresponding
inequalities for stage t − 1. Similarly, positions larger than both insertion
points get shifted by exactly k, so inequality (3) also follows for such l’s
by the corresponding inequality for stage t − 1.
These two cases cover all positions l for which both values are different
from the newly inserted values equal to rs2 . If both positions are equal to rs2
the inequality also follows. The only remaining cases are those l for which
one of σt′ [l], OP Tt′ [l] is equal to the newly inserted value rs2 but the other
is not. On inspection, though, inequality (3) is true in these cases as well:
′
because of dominance, the insertion point into OP Tt−1
is to the left (or
′
′
equal) to the insertion point into σt−1 . So if σt [l] 6= rs2 but OP Tt′ [l] = rs2
then σt′ [l] < rs2 = OP Tt′ [l]. The reasoning in the case σt′ [l] = rs2 but
OP Tt′ [l] 6= rs2 is analogous, the conclusion being that σt′ [l] = rs2 < OP Tt′ [l].
Using Lemma 6 we infer that at every step t when the greedy best-fit algorithm creates a new k-chain (because there is no available slot to lose a lifeline),
so does the optimal algorithm (for the very same reason). Hence the greedy
best-fit algorithm is optimal.
7
Maximal heapable subsets of interval orders
Finally, note that we are unable to solve the open problem in [5] on the complexity of computing a maximum treelike 2-independent set in permutation posets.
Instead, we settle this problem on a different class of partial orders, that of
interval orders:
Theorem 5. Given a set of intervals I, the best-fit algorithm in Figure 7 computes a longest heapable subset of intervals of the set I.
Proof. We prove by induction the following result, which implies the statement
of the theorem:
Lemma 7. For every 1 ≤ r ≤ n there exists a longest heapable subsequence Γr
of I and a tree Tr witnessing the optimality of Γr such that the tree TG built
by the greedy algorithm of Figure 7 agrees with Tr with respect to the intervals
I1 , I2 . . . , Ir (with respect to the presence or absence of these intervals).
Proof. By induction.
The result is simple for r = 1: if I1 is part of an optimal solution then
there is nothing to prove, since I1 can only participate in a heap-ordered tree
as its root. If, on the other hand, I1 is not part of an optimal solution, then
consider such a solution Γ. Create a new solution Γ1 with the same tree shape
19
Input: A set of intervals I.
Output: A k-chain J ⊆ I of maximum cardinality.
Sort the intervals w.r.t. ⊑: I = (I1 , . . . , In ).
For i := 1 to n do:
If Ii = [li , ri ] can be inserted into some empty slot
then
J = J ∪ {i}.
insert Ii in the highest-valued compatible slot.
else
reject Ii from J.
Figure 7: The greedy best-fit algorithm for sets of intervals.
by replacing the root interval I of Γ by I1 (thus obtaining tree T1 ). This is legal,
since I1 ends earlier than I.
Assume we have proved the result for 1 ≤ i ≤ r − 1. Consider now the
optimal solution Γr−1 (with tree Tr−1 ) agreeing with the greedy solution TG on
the presence aof the first r − 1 intervals.
• If Ir ∈ Γr−1 then there is nothing to prove.
• If Ir cannot be inserted into Γr−1 after I1 , I2 , . . . , Ir−1 it means that no
slot is available for Ir . By domination (Lemma 1) this must true for both
the Greedy and the optimal solution Γr−1 . Therefore these two solutions
coincide on the set I1 , I2 , . . . , Ir as well.
• The remaining case is when Ir can be inserted in a slot (thus is present
in the greedy solution) but is not present in the optimal solution Γr−1 .
Let J be the interval that sits in the same position in Γr−1 as Ir in the
greedy solution (Figure 8). J must exist, otherwise one could extend Γr−1
by inserting Ir at that position.
By the order we considered, the intervals Ir ends no later than J does.
So by inserting Ir instead of J we obtain an optimal solution Γr = Γr−1 \
{J} ∪ {Ir } that satisfies the induction property.
The conclusion of the induction argument is that
there exists an optimal tree containing exactly the same intervals as TG . Thus
the greedy best-fit algorithm produces an optimal solution.
20
Γr−1
Γr
TG
y
y
J
y
⇒
Ir
Ir
Figure 8: Creating optimal solution Γr = Γr−1 \ {J} ∪ {Ir }.
8
Open questions and future work
Our Theorem 1 is very similar to (the proof of) Dilworth’s theorem. It is not
yet a proper generalization of this result to the case k ≥ 1 because of the lack
of a suitable extension of the notion of antichain:
Open problem 1. Is there a suitable definition of the concept of k-antichain,
that coincides with this concept for k = 1 and leads (via our theorem 1) to an
extension of Dilworth’s theorem ?
On the other hand, results in Section 4 naturally raise the following:
Open problem 2. For which partial orders Q can one compute the parameter k-wd(Q) (and an associate optimal k-chain decomposition) via a greedy
algorithm ?
Several open problem concerns the limit behavior of the expected value of
the k-width of a set of random intervals, for k ≥ 1. As discussed in Section 2,
for k = 1 the scaling behavior of this parameter is known [17]. However, in the
case of random permutations, the most illuminating description of this scaling
behavior is by analyzing the hydrodynamic limit of the Hammersley process
[1, 12]. As shown in Section 5, the difference between sequences and sets of
intervals is not substantial.
We ask, therefore, whether the success in analyzing this process for random
permutations can be replicated in the case of sequences/sets of random intervals:
Open problem 3. Analyze the hydrodynamic limit of the Hammersley process
for sequences/sets of random intervals.
When k ≥ 2 even the scaling behavior is not known, for both sequences and
sets of random intervals. The connection with the interval Hammersley process
given by Corollaries 3 and 5 provides a convenient, lean way to simulate the
dynamics, leading to experimental observations on the scaling constants. A
C++ program used to perform these experiments is publicly available at [14].
Based on these experiments we would like to raise the following:
21
Conjecture 1. For every k ≥ 2 there exists a positive constant ck > 0 such
that, if Rn is a sequence of n random intervals then
lim
n→∞
Moreover ck =
E[k-wd(Rn )]
= ck
n
(4)
1
k+1 .
According to this conjecture, just as in the case of random permutations,
the scaling behavior of the k-width changes when going from
√ k = 1 to k = 2.
Note, though, that the
direction
of
change
is
different
(Θ(
n) to Θ(log n) for
√
integer sequences, Θ( n) to Θ(n) for sequences of intervals). On the other
hand, somewhat surprisingly, the scaling behavior of sets of random intervals
seems to be similar to that for sequences:
Conjecture 2. For every k ≥ 2 there exists a positive constant dk > 0 such
that, if Wn is a set of n random intervals then
lim
n→∞
E[k-wd(Wn )]
= dk
n
Experiments suggest that d2 = c2 = 13 , and similarly for k = 3, 4 dk = ck =
Therefore we conjecture that
dk = ck =
9
1
for all k ≥ 2.
k+1
(5)
1
k+1 .
(6)
Acknowledgments
Research of J.B. was supported by the European Union, co-financed by the
European Social Fund (EFOP-3.6.3-VEKOP-16-2017-00002).
Research of C.B. was supported in part by a grant of the Romanian National
Authority for Scientific Research and Innovation, CNCS-UEFISCDI, project no.
PN-II-RU-TE-2014-4-0270 - FraDys: ”Theoretical and Numerical Analysis of
Fractional-Order Dynamical Systems and Applications”.
G. I. and D.D. were supported in part by a grant of Ministry of Research
and Innovation, CNCS - UEFISCDI, project number PN-III-P4-ID-PCE-20160842, within PNCDI III, ATCO: ”Advanced techniques in optimization and
computational complexity”.
I.T. was partially supported by the ANR (Agence Nationale pour la Recherche)
project GraphEn, ANR-15-CE40-0009.
References
[1] David Aldous and Persi Diaconis. Hammersley’s interacting particle process
and longest increasing subsequences. Probability theory and related fields,
103(2):199–213, 1995.
22
[2] A-L Basdevant, Lucas Gerin, J-B Gouéré, and Arvind Singh. From Hammersleys lines to Hammersleys trees. Probability Theory and Related Fields,
pages 1–51, 2016.
[3] Anne-Laure Basdevant and Arvind Singh.
Almost-sure asymptotic
for the number of heaps inside a random sequence. arXiv preprint
arXiv:1702.06444, 2017.
[4] Guy E. Blelloch. Prefix sums and their applications. Technical Report
CMU-CS-90-190, School of Computer Science, Carnegie Mellon University,
November 1990.
[5] John Byers, Brent Heeringa, Michael Mitzenmacher, and Georgios Zervas.
Heapable sequences and subseqeuences. In Proceedings of the Eighth Workshop on Analytic Algorithmics and Combinatorics (ANALCO’2011), pages
33–44. SIAM Press, 2011.
[6] Ido Dagan, Martin Charles Golumbic, and Ron Yair Pinter. Trapezoid
graphs and their coloring. Discrete Applied Mathematics, 21(1):35–46,
1988.
[7] Robert P. Dilworth. A decomposition theorem for partially ordered sets.
Annals of Mathematics, pages 161–166, 1950.
[8] Jeno Egerváry. Matrixok kombinatorius tulajdonságairól. Matematikai és
Fizikai Lapok, 38(1931):16–28, 1931.
[9] Stefan Felsner, Rudolf Müller, and Lorenz Wernisch. Trapezoid graphs and
generalizations, geometry and algorithms. Discrete Applied Mathematics,
74(1):13–32, 1997.
[10] Stefan Felsner, Vijay Raghavan, and Jeremy Spinrad. Recognition algorithms for orders of small width and graphs of small Dilworth number.
Order, 20(4):351–364, 2003.
[11] D.R. Fulkerson. Note on Dilworth’s decomposition theorem for partially
ordered sets. Proceedings of the American Mathematical Society, 7(4):701–
702, 1956.
[12] Piet Groeneboom. Hydrodynamical methods for analyzing longest increasing subsequences. Journal of Computational and Applied Mathematics,
142(1):83–105, 2002.
[13] András Gyárfás and Jenö Lehel. Covering and coloring problems for relatives of intervals. Discrete Mathematics, 55(2):167–180, 1985.
[14] Gabriel
Istrate.
C++
program
interval
for
experiments
on
the
heapability
of
intervals.
https://bitbucket.org/gabriel-istrate/heapable-intervals,
2017. Accessed: 2017-05-06.
23
[15] Gabriel Istrate and Cosmin Bonchiş. Partition into heapable sequences,
heap tableaux and a multiset extension of Hammersley’s process. In Proceedings of the 26th Annual Symposium on Combinatorial Pattern Matching (CPM’15), Ischia, Italy, volume 9133 of Lecture Notes in Computer
Science, pages 261–271. Springer, 2015.
[16] Gabriel Istrate and Cosmin Bonchiş. Heapability, interactive particle systems, partial orders: Results and open problems. In Proceedings of the
18th International Conference on Descriptional Complexity of Formal Systems (DCFS’2016), Bucharest, Romania, volume 9777 of Lecture Notes in
Computer Science, pages 18–28. Springer, 2016.
[17] Joyce Justicz, Edward R. Scheinerman, and Peter M. Winkler. Random
intervals. Amer. Math. Monthly, 97(10):881–889, 1990.
[18] Denés Konig. Gráfok és mátrixok. Matematikai és Fizikai Lapok, 38:116–
119, 1931.
[19] Benjamin F. Logan and Larry A. Shepp. A variational problem for random
Young tableaux. Advances in mathematics, 26(2):206–222, 1977.
[20] Colin L Mallows. Patience sorting. SIAM Review, 5(4):375–376, 1963.
[21] Stephan Olariu. An optimal greedy heuristic to color interval graphs. Information Processing Letters, 37(1):21–25, 1991.
[22] Jaclyn Porfilio.
A combinatorial characterization of heapability.
Master’s thesis, Williams College, May 2015. available from
https://unbound.williams.edu/theses/islandora/object/studenttheses%3A907.
Accessed: March 2018.
[23] Dan Romik. The surprising mathematics of longest increasing subsequences. Cambridge University Press, 2015.
[24] William T. Trotter. Partially ordered sets. Handbook of combinatorics,
1:433–480, 1995.
[25] Anatoly M. Vershik and Sergei V. Kerov. Asymptotics of Plancherel
measure of symmetrical group and limit form of Young tables. Doklady
Akademii Nauk SSSR, 233(6):1024–1027, 1977.
24
| 8cs.DS
|
Functional Choice and Non-significance Regions in Regression
Laurie Davies
Fakulät Mathematik
Universität Duisburg-Essen
arXiv:1605.01936v1 [math.ST] 6 May 2016
[email protected]
Abstract
Given data y and k covariates x the problem is to decide which covariates to include when approximating y by a linear function of the covariates. The decision is based on replacing subsets of the covariates by
i.i.d. normal random variables and comparing the error with that obtained
by retaining the subsets. If the two errors are not significantly different
for a particular subset it is concluded that the covariates in this subset
are no better than random noise and they are not included in the linear
approximation to y.
1
1.1
Introduction
Notation
Consider n measurements y n = (y1 , . . . , yn )t of a variable y and for each yi
concomitant measurements of k covariables xj , j = 1, . . . , k, given by xi· =
(xi1 , . . . , xik )t forming an n × k matrix xn with jth column x·j . A subset of
the covariates will be denoted by a row vector e = (e1 , . . . , ek ) with ej ∈ {0, 1}
whereby ej = 1 means that the jth covariate is included. A model e will be
Pk
encoded as j=1 ej 2j−1 . The subset consisting of all covariates will be denoted
Pk
by ef . Given an e with
j=1 ej = k(e) the n × k(e) matrix with columns
corresponding to those covariates with ej = 1 will be denoted by xn (e) with
xi· (e), x·j (e) and x(e) having the corresponding interpretations. The empirical
measure of the data will be denoted by Pn
n
Pn = Pn ((y n , xn )) =
1X
δy , x
n i=1 i i·
1
(1)
with the corresponding definition of Pn,e for any subset e. The L1 and L2 norms
will be denoted by k k1 and k k2 respectively.
1.2
The problem
The problem to decide which if any of the covariates xj influence the value of
y. There are many proposals for doing this. Some such as AIC (Akaike (1973,
1974, 1981)) or BIC (Schwarz (1978)) require an explicit model such as
Y = xt β + ε
(2)
where xt = (x1 , . . . , xk ), β = (β1 , . . . , βk )t and the errors ε are random variables
with an explicit distribution. Others such as Lasso (Tibshirani (1996))
k
n
X
X
|βj |
(yi − xi· β)2 + λ
argminβ
i=1
(3)
j=1
may or may not require an explicit model to determine the choice of the smoothing parameter λ.
The following is based on a simple idea. Let s2 denotes the least sum of
squares based on all covariates and for a given subset let Se2 denote the least
sum of squares when all the covariates with ej = 0 are replaced by i.i.d. N (0, 1)
random variables. If s2 is not significantly small than Se2 the conclusion is that
the omitted covariates are no better than random noise. To define ‘significantly’
the process is repeated a large number of times. For a given α, α = 0.95 for
example, s2 is significantly smaller the Se2 if in at least 100α% the simulations
s2 ≤ Se2 . The P-value pe is the proportion of simulations for which Se2 < s2 ,
so that the excluded covariates are significantly better than random noise if
pe ≤ 1 − α. A small values of pe indicates that at least some of the omitted
covariates are relevant. A large value of pe indicates that in toto the omitted
covariates are no better than random noise.
The method is not restricted to least squares regression. It can be equally
well applied to L1 regression or more generally to any measure of discrepancy
d(y n , xn ).
As an example consider the stack loss data of Brownlee (1960). It is one of
the data sets provided by R Core Team (2013) and is used in Koenker (2010).
There are 21 observations with one dependent variable ‘Stack.Loss’ and the
2
three covariates ‘Air.Flow’, ‘Water.Temp’ and ‘Acid.Conc’ labelled from one to
three. In the following the intercept will always be included. There are eight
possible models. The results of an L1 regression for the stack loss data are given
in Table 1. The total computing time was 102 seconds using Koenker (2010). The
only subset with a large P -value is the subset encoded as 3 which corresponds
to e = (1, 1, 0).
subset
0
1
2
3
4
5
6
7
P -value
0.000
0.015
0.000
0.231
0.000
0.007
0.000
1.000
Table 1: Encoded subsets and P -values for the stack loss data based on 5000
simulations.
1.3
Non-significance regions
Given a subset e of covariates the best linear approximation to the variable y n
in the L1 norm is
xn (e)β 1,n (e)
(4)
where
β1,n (e) = argminβ (e) ky n − xn (e)β(e)k1 .
(5)
A single value is not sufficient to answer many questions of interest which require
a range of plausible values. In frequentist statistics such a range is provided by
a confidence region. This option is not available in the present context as a confidence region assumes that there is a ‘true’ value to be covered. The confidence
region will be replaced by a non-significance region whose construction will be
illustrated for the median.
Given data y n the median minimizes s1 (y n ) =
other value m 6= med(y n )
n
X
i=1
|yi − med(y n )| <
n
X
i=1
Pn
i=1
|yi − med(y n )|. For any
|yi − m|
A value m will be considered as not being significantly different from the median
3
med(y n ) if the difference
n
X
i=1
|yi − m| −
n
X
i=1
|yi − med(y n )|
is of the order attainable by a random perturbation of the y-values. More precisely if
P
inf
b
n
X
i=1
!
|yi + bZi − m| < s1 (y n )
≥ 1 − α.
(6)
The set of values m which satisfy (6) can be determined by simulations. For
the y n of the stack loss data the 0.95-non-significance region is [11.86, 18.71]
which can be compared with the 0.95-confidence region [11, 18] based on the
order statistics. For any m the P-value p(m) is defined as
!
n
X
|yi + bZi − m| < s1 (y n ) .
p(m) = P inf
b
2
(7)
i=1
Choice of functional
The procedure expounded in the previous section makes no use of a model of
the form (2). It solely based on the approximation of y n by a linear combination
of the covariates as measured in the L1 and L2 norms. There is no mention of
an error term. It therefore makes little sense to describe the procedure as one of
model sense. It makes more sense to interpret it as one of functional choice. There
does not seem to be any immediate connection with ’wrong model’ approaches
as in Berk et al. (2013) and Lindsay and Liu (2009).
For a given subset e the L1 function T1,e is defined by
Z
T1,e (Pn ) = argminβ(e) |y − x(e)t β(e)| dPn (y, x(e))
=
argminβ(e) kyn − xn (e)β(e)k1
with the corresponding definition of the L2 functional
Z
T2,e (Pn ) = argminβ(e) (y − x(e)t β(e))2 dPn (y, x(e))
=
argminβ(e) kyn − xn (e)β(e)k2 .
More generally an M -functional Tρ,e can be defined as
Z
y − x(e)t β(e)
Tρ,e (Pn ) = argminβ (e) ρ
dP (y, x(e))
σn
n
yi − xi· (e)t β(e)
1X
ρ
.
= argminβ (e)
n
σn
i=1
4
(8)
(9)
(10)
The function ρ is taken to be convex with a bounded first derivative. This is the
case for the default choice in this paper namely the Huber ρ-function defined
by
ρc (u) =
1 2
2u
c|u| − 1 c2
2
: |u| ≤ c
(11)
: |u| > c
where c is a tuning constant. The functional can be calculated using the iterative
scheme described in Chapter 7.8.2 of Huber and Ronchetti (2009).
For reasons of equivariance (10) contains a scale parameter σn which may be
external or part of the definition of Tρ (see Chapter 7.8 of Huber and Ronchetti
(2009)). The default choice in this paper is the Median Absolute Deviation of
the residuals from an L1 fit:
σn = mad(y n − xn β1,n (ef )).
(12)
One use of M -functionals is to protect against outlying y-values. The choice
(12) preserves this property.
2.1
L1 regression
The best linear fit based on all covariates is determined by
n
T1,ef (Pn ) = β 1,n (ef ) = argminβ
1X
|yi −xti· β| = argminβ ky n −xn βk1 (13)
n i=1
with mean sum of absolute deviations
n
s1,n (ef ) =
1X
|yi − xti· β 1,n (ef )| = kyn − xn β 1,n (ef )k1 .
n i=1
(14)
Let Z n be a n × k matrix with elements Zij which are i.i.d. N (0, 1). Given
e replace the covariates with ej = 0 by the Zij , that is, put Wi,j (e) = xi,j if
ej = 1 and Wij (e) = Zij if ej = 0. Denote the relevant matrices by W n (e) and
Z n (ec ) and the empirical measure by P̃n,e . The best linear fit based on these
covariates is determined
T1,ef (P̃n,e ) = β̃ 1,n (e) = argminβ ky n − W n (e)βk1
(15)
with mean sum of absolute deviations
S1,n (e) = ky n − W n (e)β̃ 1,n (e)k1
(16)
5
The quantity S1,n (e) is a random variable. The P -value pn (e) is defined by
pn (e) = P (S1,n (e) ≤ s1,n (ef )) .
(17)
There is no explicit expression for the P -values in the case of L1 regression.
They must be calculated using simulations as in Table 1. This results in four
of the P -values being zero and so no comparison between them. A comparison
can be obtained as follows. Simulate the distribution of
s1,n (e) − S1,n (e)
(18)
where
n
s1,n (e) =
1X
|yi − xti· β1,n (e)| = ky n − xn β 1,n (e)k1
n i=1
(19)
and then approximate it by a Γ-distribution with the shape and scale parameters
ˆ
sh(e) and sc(e) estimated from the simulations as sh(e)
and sc(e)
ˆ
respectively.
The resulting estimated P -values are given by
ˆ
p̂n (e) = 1 − pgamma(s1,n (e) − s1,n (ef ), sh(e),
sc(e))
ˆ
.
(20)
The results for the stack loss data are given in Table 2 and may be compared
with the P -values of Table 1.
functional
0
1
2
3
4
5
6
7
P -value
1.93e-7
1.41e-2
4.90e-4
2.32e-1
5.02e-9
7.43e-3
2.57e-4
1.00
Table 2: Encoded L1 -functionals and P -values for the stack loss data based on
1000 simulations using the Γ-approximation (20),
Small P -values indicate that covariables have been omitted which have a
significant effect on the dependent variable. This excludes the functionals encoded as 0, 1, 2, 4, 5, 6 although the functional encoded as 1 could possible be
retained. The functional 3 with P -value 0.232 omits the covariable Acid.Conc.
As the functional 7 differs from 3 only through the inclusion of Acid.Conc the
conclusion is that it contains a covariate which is little better than random
noise. Thus an analysis of the P -values leads to the choice of the functional 3.
6
The interpretation of P -values and the choice of functional will be considered
in greater detail in Sections 2.5 and 2.6 respectively.
The second running example is the low birth weight data of Hosmer and Lemeshow
(1989) with n = 189 and k = 9. The dependent variable is the weight of the
child at birth. The nine covariates range from the weight and age of the mother
to hypertension and indicators of race. There are in all 512 different functionals.
In the context of model choice it is considered in Claeskens and Hjort (2003).
For this data set the computing time using 1000 simulations is about 50
minutes. This can be reduced by a factor of about ten by approximating the
modulus function |x| by the Huber ρ-function (11) with a small value of the
tuning constant c, for example c = 0.01 (see Section 2.2). Care must be taken
in interpreting the decrease in computing time as the L1 -functional was calculated using package Koenker (2010) whereas the program for the M -functional
was written entirely in Fortran using the algorithm given in Chapter 7.8 of
Huber and Ronchetti (2009) (see also Dutter (1977b) and Dutter (1977a)) and
the pseudo-random number generator ran2 (see Press et al. (2003)). A pure Fortran program for the L1 -functional may be much faster (see Koenker and Portnoy
(1997)).
2.2
M-regression functionals
The M -functionals can be treated in the same manner as the L1 functional but
with the added advantage that for large values of the tuning constant c in (11)
7
there exist asymptotic approximations for the P -values. On writing
n
yi − xti· β
1X
ρ
Tρ,ef (Pn ) = β ρ (ef ) = argminβ
n i=1
σn
n
yi − xti· β ρ (ef )
1X
ρ
sρ,n (ef ) =
n i=1
σn
n
1X
yi − xi· (e)t β
Tρ,e (Pn ) = β ρ (e) = argminβ
ρ
n i=1
σn
n
t
yi − xi· (e) βρ (e)
1X
sρ (e) =
ρ
n i=1
σn
n
1X
yi − W i· (e)t β(e)
Tρ,e (P̃n ) = β̃ ρ (e) = argminβ (e)
ρ
n i=1
σn
!
n
yi − W i· (e)t β̃ρ (e)
1X
ρ
Sρ,n (e) =
n i=1
σn
(21)
(22)
(23)
(24)
(25)
(26)
a second order Taylor expansion gives
2
1 Pn
(1) ri (e)
χ2k−k(e)
ρ
σn
1 n i=1
Sρ,n (e) ≈ sρ,n (e) −
Pn
1
2
ρ(2) ri (e)
i=1
n
(27)
σn
where ρ(1) and ρ(2) are first and second derivatives of ρ respectively, ri (e) = yi −
xi· (e)t β(e) and χ2k−k(e) is a chi-squared random variable with k − k(e) degrees
of freedom. The inequality Sρ,n (e) ≤ sρ,n (ef ) is asymptotically equivalent to
P
n
2
(2) ri (e)
(sρ,n (e) − sρ,n (ef ))
ρ
i=1
n
σn
(28)
≤ χ2k−k(e)
2
P
n
r
(e)
1
i
(1)
ρ
i=1
n
σn
with asymptotic P -value
pn (e) ≈ 1 − pchisq
2
n
Pn
(2)
i=1 ρ
1
n
ri (e)
σn
Pn
i=1
ρ(1)
(sρ,n (e) − sρ,n (ef ))
, k − k(e) .
2
ri (e)
σn
(29)
(1)
As the tuning constant c tends to zero the terms ρc
ri (e)
σn
2
(2)
and ρc
become one and zero respectively and the approximation breaks down.
ri (e)
σn
The results for the stack loss data with c = 1.5 are given in Table 3 and may
be compared with those given in Table 1 for the L1 -functional.
8
functional
0
1
2
3
4
5
6
7
P -value
0.000
0.012
0.000
0.293
0.000
0.006
0.000
1.000
P -value
3.19e-7
1.23e-2
4.40e-4
3.03e-1
3.79e-8
5.74e-3
3.59e-5
1.00
P -value
1.89e-6
9.96e-3
1.81e-3
2.33e-1
4.63e-11
2.67e-3
6.66e-5
1.00
Table 3: Encoded M -functionals (c=1.5) and P -values for the stack loss data
based on 5000 simulations: first row the raw values, second row the values based
on the Γ-approximation (20), third row the values based on the asymptotic
approximation (29).
2.3
Least squares regression
The L2 -regression functionals are a special case of the M -functionals for a sufficiently large tuning constant c. The P -values can either be estimated directly
using simulations or using the Γ-approximation (20) or using the asymptotic
approximation (29) which takes the form
pn (e) ≈ 1 − pchisq
!
n(ky n − xn (e)β 2,n (e)k22 − kyn − xn β 2,n k22 )
, k − k(e) .
ky n − xn (e)β2,n (e)k22
(30)
If a lower bound α is given for pn (e) then (30) is asymptotically equivalent to
the F -test in the linear regression model for testing the null hypothesis that the
coefficients of the covariates not included are zero.
The P -values for the stack loss data are given in Table 4.
functional
0
1
2
3
4
5
6
7
P -value
0.000
0.011
0.000
0.427
0.000
0.005
0.000
1.000
P -value
1.89e-4
1.07e-2
1.28e-4
4.35e-1
2.80e-5
4.42e-3
4.24e-4
1.00
P -value
2.53e-4
1.51e-2
1.04e-4
3.11e-1
8.14e-5
4.69e-3
2.20e-4
1.00
Table 4: Encoded L2 -functionals and P -values for the stack loss data based on
5000 simulations: first row the raw values, second row the values based on the Γapproximation (20), third row the values based on the asymptotic approximation
(30).
9
2.4
Non-linear regression
The ideas can be applied mutatis mutandis to non-linear regression
Tnl,1,e (Pn ) = argminβ(e) kyn − g(xn (e), β(e))k1
(31)
with corresponding definitions for Tnl,ρ,e and Tnl,2,e . The computational cost is
much higher so that only small values of k are possible.
2.5
Lower bounds for P -values
In contrast to AIC and BIC the P -values do not order the different functionals.
One possibility it to choose a cut-off value p0 (n, k) for pn (e) and consider only
those functionals Te with pn (e) ≥ p0 (n, k). A possible value for p0 (n, k) can be
obtained by considering the size of the P -values when all covariates are noise,
xn = Z̃ n . For each such xn = Z̃ n the minimum value of pn (e) over all e can be
calculated and then simulated for different Z̃ n . The α quantile with for example
α = 0.05 can then be taken as the value of p0 (n, k) = p0 (n, k, α).
The minimum of the p(e)-values can only be determined by simulation and
then further simulations are required in order to determine the quantiles of
the minimum values. For L1 -functionals for the stack loss data with n = 21 and
k = 3 the time required with 1000 simulations for each p(e) and 2000 simulations
for the quantile was 10 minutes using the approximation to |x| based on the
Huber ρ-function (11) with tuning constant c = 0.01. The results using the
Γ-approximation (20) are given in the first line of Table 5.
The computational time for the low birth weight data with k = 9 was considerably higher. The time required for 1000 simulations for the minimum values
of the pn (e) each of which was also based on 500 simulations was 34 hours. The
results using the Γ-approximation (20) are given in the second line of Table 5.
The corresponding p0 values for the M -functional and the L2 -functional
are based on 500 simulations with 250 simulations for each pn (e) value. The
computing time for the low birth weight data was 2 1/2 hours in each case. The
results are given in the lines 3-6 of Table 5.
For the M -functionals with a not too smaller a value of c in (11) and the
L2 -functionals use can be made of the approximations (29) and (30) respectively
10
L1 -functional
p0 (21, 3, 0.01) = 0.00368
p0 (21, 3, 0.05) = 0.0155
p0 (21, 3, 0.10) = 0.0340
p0 (189, 9, 0.01) = 0.00044
p0 (189, 9, 0.05) = 0.0031
p0 (189, 9, 0.10) = 0.0068
p0 (21, 3, 0.01) = 0.0055
p0 (21, 3, 0.05) = 0.0193
p0 (21, 3, 0.10) = 0.036
p0 (189, 9, 0.01) = 0.00011
p0 (189, 9, 0.05) = 0.0020
p0 (189, 9, 0.10) = 0.0056
L2 -functional
Table 5: The P -values p0 (n, k, α) for the L1 - and L2 -functionals.
which allow simulations for larger values of k. The computational load can be
further reduced as follows. Without loss of generality suppose kyn k2 = 1. As all
the xn = Z̃ n are standard Gaussian white noise random variables it follows for
the L2 -functionals
X
kZ n (ec )t (y n − xn (e)β 2,n (e))k22
kZ n (ecj )t y n k22 (32)
≈ kZ n (ec )t y n k22 ≈
2
ky n − xn (e)β 2,n (e)k2
j
P
where ec = j ecj and each ecj has only one element not equal to zero. Furthermore the Z n (ecj )t y n are independent N (0, 1) random variables. This yields the
asymptotic approximation
X
p̃(e) = 1 − pchisq
χ21 (j), |S|
(33)
j∈S
where the χ21 (j), j = 1, . . . , k are independent χ2 random variables with one
degree of freedom. Taking the minimum over e, simulating sets of χ2 random
variables and then taking the α quantile gives a value p̃0 (k, α). It is only necessary to perform the simulations one for each value of k. The p̃0 (k, α) can be
approximated by
p̃0 (k, α) ≈ exp(c1 (k) + c2 (k) log(α) + c3 (k) log(α)2 )
(34)
for α < 0.5 (see Chapter 2.9 of Davies (2014)) and can be used in place of the
p0 (n, k, α). For k = 17 the time required on a standard laptop was 7 hours 42
minutes. As an example
p̃0 (9, 0.01) = 0.00028, p̃0 (9, 0.05) = 0.0025, p̃0 (9, 0.10) = 0.0059 .
The results compare well with those based on simulations as given in Table 5
and their computing costs are essentially zero. This suggests that they can be
used as guidelines when simulations are too expensive.
11
2.6
Choosing functionals
In view of the interpretation of the P -value of a functional the first step is to
decide on a cut-off value p0 and then restrict consideration to those functionals
Te with pn (e) > p0 . A possible choice of p0 is p0 = p0 (n, k, α) and this will be
done below. The choice may be further restricted by requiring that for each such
e and for all e′ with e′ < e pointwise all the omitted covariates are relevant with
respect to e. More precisely pn (e′ , e) < p0 (n, k(e), α) for all e′ < e where pn (e′ , e)
is the P -value of e′ calculated with respect to the covariates x(e). A final choice
may be made by choosing that functional Te with the highest pn (e)-value.
The results of applying the above strategy to the stack loss data with the
L1 -norm are as follows. Taking the cut-off value to be p0 (21, 3, 0.01) = 0.00368
(based on Table 5) the first step results in the functionals based on e1 = (1, 0, 0),
e2 = (1, 1, 0) and e3 = (1, 0, 1). The second step eliminates the functionals based
on e2 and e3 . The choices α = 0.05 and α = 0.1 both lead to e = e2 with
pn (e2 ) = 0.232.
The results for the low birth weight data are as follows. The first step with
α = 0.01 leads to 379 functionals. The second step results in the single functional
encoded as 260 with just two variables and a P -value of 4.25e-3. Putting α =
0.05 results in 221 functionals after the first step. The second step yields the
five functionals encoded as 292, 166, 260, 36 and 60. The functional 292 has
the highest P -value with pn (e) = 0.074. Finally the choice α = 0.1 results in
149 functionals after the first step. The second step reduces this to the seven
functionals 308, 292, 166, 262, 38, 52 and 260 of which the functional encoded
as 308 has the highest P -value equal to 0.321.
The strategy described above ‘guarantees’ that all included covariates are
relevant. If it is more important not to exclude covariates which may have an
influence at the possible cost of including some irrelevant covariates the this
may be done by increasing the value of α in p0 (n, k, α) or by simply specifying
some cut-off level p0 judged to be appropriate.
Although AIC and BIC list the models in order of preference they give no
indication as to whether any of the models under consideration is an adequate
approximation to the data or not. Presumably this is the responsibility of the
user before applying the criterion. The first ten models for the birth weight data
12
based on BIC are encoded as
308, 310, 436, 438, 294, 292, 182, 316, 422, 309 .
(35)
The functionals obtained using the P -values strategy are encoded as 260, 292
and 308. Their positions in the BIC list are 66, 6 and 1 respectively. The second
model on the BIC list is encoded as 310 and includes the additional covariate ‘weight of mother’ compared to the 308 model. If one uses R Core Team
(2013) to do an L1 regression based on the covariates corresponding to 310
the 95% confidence interval for ‘weight of mother’ includes zero. This may be
interpreted as a non-significant effect given the other covariates. This interpretation is consistent with the P -value strategy with α = 0.1 where the encoded
value 310 is not included in the list (35). The reason is that the P -value for the
functional excluding ‘weight of mother’ has a P -value of 0.112 which exceeds
p0 (189, 5, 0.1) = 0.017.
This shows that models may be high in the BIC list although they contain
variables which are not significantly better than random noise. This can be
made more explicit by replacing the all covariates by random noise and using
simulations to determine how often a model containing a random noise covariate
is first on the BIC list. This was simulated 500 times with the weight of the child
as the dependent variable. This happened in 43% of the cases. With α = 0.1
the P -value strategy is calibrated to do this in 100α% = 10% of the cases. The
simulations resulted in 9%.
3
3.1
Non-Significance regions
The median and M-functionals
The 0.95-non-significance region for the median of the stack loss data was defined
and calculated in Section 1.3 with the result [11.86, 18.71]. In general the α-nonsignificance region is defined by y n is
N S(y n , med, α)
)
(
n
n
X
X
|yi − med(y n )| ≤ ql1(α, m, y n )
|yi − m| −
=
m:
i=1
i=1
13
(36)
where ql1(α, m, y n ) is the α-quantile of
n
X
i=1
|yi − m| − inf
b
n
X
i=1
|yi − m − bZi |
(37)
and the Zi are standard Gaussian white noise.
The non-significance region (36) can be calculated as follows. Put
f (m, α, y n ) = ql1(α, m, y n ) −
n
X
i=1
|yi − m| +
n
X
i=1
|yi − med(y n )|
(38)
and note that f (med(y n ), α, y n ) ≥ 0. Now determine an order statistic y(nl) with
nl = qbinom((1 − β)/2, n, 0.5)) for a suitably large β such that f (y(nl) , α, y n ) <
0. Interval bisection combined with simulations can now be used to find an
approximate solution mlb of f (m, α, y n ) = 0. This gives a lower bound and the
same process can be used to get an upper bound mub to give N S(y n , med, α) =
[mlb , mub ].
Non-significance regions for M -functionals Tρ are defined analogously by
replacing (36) by
N S(y n , Tρ , α)
(39)
)
(
n
n
X
X
yi − Tρ (Pn )
yi − m
ρ
ρ
−
≤ qrho(α, m, y n )
=
m:
σn
σn
i=1
i=1
where qrho(α, m, y n ) is the α-quantile of
n
n
X
X
yi − m
yi − m − bZi
ρ
ρ
− inf
,
b
σn
σn
i=1
i=1
(40)
the Zi are standard Gaussian white noise and σn is a scale functional whose
default value in this situation is (12).
For smooth functions ρ an asymptotic expression for the non-significance
region is available. Let ρ(1) and ρ(2) denote the first and second derivative of ρ.
A Taylor series expansion results in
N S(y n , Tρ , α) ≈
q
m : |Tρ (Pn ) − m| ≤ qnorm((1 + α)/2)σn v(Tρ , Pn )/n
(41)
where
1
n
Pn
(1)
i=1 ρ
v(Tρ , Pn ) = P
n
1
n
i=1
ρ(2)
yi −Tρ (Pn )
σn
yi −Tρ (Pn )
σn
14
2
2 .
(42)
This latter expression is well known in robust statistics and corresponds to the
asymptotic variance of an M -location functional: the non-significance region
(41) is the corresponding α-confidence region for the ‘unknown’ Tρ (P ). In the
special case ρ(u) = u2 /2 (41) is the asymptotic α-confidence region for the mean
based on Gaussian errors.
3.2
L1 regression
The idea carries over to the L1 regression functional. For any β put
Γ(y n , xn , β, Z n ) = ky n − xn βk1 − inf ky n − xn β − Z n bk1
b
(43)
and denote the α-quantile of Γ(y n , x, β, Z n ) by q1(α, β, y n , xn ). An α-nonsignificance region is then defined as
N S(y n , xn , α, T1 ) = {β : ky n − xn βk1 − kyn − xn β 1,n k1 ≤ q1(α, β, yn , xn )}
(44)
where β 1,n = T1 (Pn ).
As it stands the non-significance region is difficult to calculate as it requires
a grid of values for the possible values of β and the values of q1(α, β, y n , xn )
have to be estimated using simulations. If the quantiles are largely independent
of the β-values then q1(α, β 1 , y n , xn ) can be used with a large reduction in
computation. Section 3.5 contains some asymptotics which suggest that the
independence may hold for large sample sizes n. The defining inequality in (44)
will still have to be checked over a grid of values.
Most software packages provide only confidence regions for the individual
components of β. Corresponding component wise non-significance regions can
be defined with a large reduction in the computational overload. For the first
component β1 of T1 (Pn ) the α-non-significance region is given by
n
N S(y n , xn , α, T1,1 ) = β1 :
inf
β2 ,...,βk
y n − x·1 β1 −
− y n − xn β1,n
15
1
k
X
j=2
x·j βj
1
o
≤ q1(α, β1 , y n , xn )
(45)
where q1(α, β1 , y n , xn ) is the α-quantile of
Γ1 (y n , xn , β1 , Z·1 ) =
inf
β2 ,...,βk
y n − x·1 β1 −
k
X
j=2
y n − x·1 β1 −
inf
b1 ,β2 ,...,βk
x·j βj
k
X
j=2
1
−
(46)
x·j βj − Z ·1 b1
1
The non-significance intervals of the stack loss data and for comparison the
0.95-confidence intervals are given in Table 6.
Air.Flow
Water.Temp
Acid.Conc
Non-sig. intervals (45)
(0.552,1.082)
(0.225,1.603)
(-0.345,0.102)
L1 confidence intervals
(0.509,1.168)
(0.272,3.037)
(-0.278,0.015)
Table 6: First line: 0.95-non-significance intervals for the stack loss data. Second
line: 0.95-confidence intervals produced by Koenker (2010) for the default choice
‘se=rank’.
3.3
M-regression functionals
Non-significance regions for M -regression functionals are defined in the same
manner as for L1 regression. Just as in Section 2.2 the computational burden
can be reduced for large n by using the asymptotic expressions. These result in
(
X
n
n
X
yi − xti· βρ
yi − xti· β
ρ
ρ
N S(y n , xn , α, Tρ ) = β :
−
≤
(47)
σn
σn
i=1
i=1
t
Pn
(1)2 yi −xi· β )
σn
qchisq(α, k) i=1 ρ
Pn
y −x it β
2
ρ(2) i ·
i=1
σn
where β ρ = Tρ (Pn ). This can be further simplified to
(
N S(y n , xn , α, Tρ ) =
β : (β − β ρ )t xtn xn (β − β ρ ) ≤
i=1 ρ
.
qchisq(α, k)
Pn
yi −x· it β ρ
(2)
i=1 ρ
σ
Pn
(1)2
(48)
yi −xti· β ρ
σn
n
16
3.4
Least squares regression
The method goes through for the least squares functional with the advantage
that explicit expressions are available. The result corresponding to (47) is
(
N S(y n , xn , α, T2 ) =
β : ky n − xn βk22 − ky n − xn β 2,n k22 ≤
(49)
)
ky n − xn βk22
qchisq(α, k)
n
which is the same as
(
N S(y n , xn , α, T2 ) =
β : (β − β 2,n )t xtn xn (β − β2,n ) ≤
ky n − xn β 2,n k22 qchisq(α, k)
n − qchisq(α, k)
)
(50)
.
where β 2,n = T2 (Pn ). The region is asymptotically equivalent to a standard
α-confidence region for the ‘true’ parameter value.
3.5
Covering properties
The concept of a non-significance region makes no mention of a model or true
values. Nevertheless there are situations where a model and its parameters are
well founded and relate to well-defined properties of the real world. In such
cases there is an interest in specifying a region which includes the real world
value with the required frequency in repeated measurements. It has to be kept
in mind however that covering true parameter values in simulations is not the
same as covering the corresponding real values for real data (see Chapter 5.5 of
Davies (2014), Stigler (1977), Chapter 8.1 of Hampel et al. (1986), Kunsch et al.
(1993)).
Given this there is an interest in the covering properties of non-significance
regions. Table 7 gives the frequencies with which the non-significance intervals
(36) and the confidence intervals based on the rank statistics cover the population median and also the average lengths of the intervals. The results are for
the normal, Cauchy, χ21 and the Poisson Po(4) distributions and four different
sample sizes n = 10, 20, 50, 100 and are based on 1000 simulations. The discreteness of Poisson distribution was taken into account in the calculations of the
non-significance region as follows. If an non-significance interval [ℓ, u] contains
17
N (0, 1)
C(0, 1)
χ21
Pois(4)
n
10
20
50
100
(36)
0.940 1.512
0.954 1.040
0.948 0.648
0.942 0.464
rank
0.968 2.046
0.968 1.198
0.970 0.767
0.964 0.530
(36)
0.960 3.318
0.956 1.670
0.960 0.958
0.952 0.629
rank
0.978 5.791
0.950 1.850
0.968 1.069
0.964 0.700
(36)
0.944 1.368
0.936 0.877
0.932 0.550
0.942 0.396
rank
0.982 2.064
0.958 1.086
0.970 0.675
0.968 0.452
(36)
0.934 1.918
0.925 0.993
0.926 0.288
0.938 0.071
rank
0.996 3.948
0.964 2.342
0.997 1.573
1.000 1.085
Table 7: Covering frequencies and average interval lengths based on 1000 simulations for the median for the 0.95-non-significance intervals as defined by (36)
and (37) with Z = N (0, 1) and the 0.95-confidence intervals based on the ranks.
For each sample size the first column gives the covering frequency and the second
the average interval length.
an integer it is by [⌈ℓ⌉, ⌊u⌋]. If it does not contain an integer it is replaced by
[⌊ℓ⌋, ⌈u⌉]. The covering frequencies and lengths refer to this modified interval.
In this well defined situation Table 7 indicates that the 0.95-non-significance
intervals also have covering probabilities of about 0.95. The finite sample behaviour seems to be better than that of the ranks procedure. Both methods
have approximately the correct covering frequencies but the lengths of the nonsignificance intervals are uniformly smaller than the lengths of the confidence
intervals.
There is some theoretical explanation as to why the non-significance regions
have covering frequencies given by α, at least asymptotically. Consider firstly
i.i.d. integer valued random variables Yj with a unique median ν. Then for a
large sample size n
n
X
j=1
|Yj − ν − bZj |
is, with large probability, minimized by putting b = 0. In other words the 0.95non-significance interval is simply [ν, ν] with a covering probability tending to
one. This is illustrated by the Poisson distribution in Table 7.
18
Suppose that the Yj are continuous random variables with median 0 and a
density f which is continuous at 0 with f (0) > 0. Then the approximation
n
X
i=1
n
X
bZi
|Yi | − bN (0, 1) + f (0)b2
Yi − √ ≈
n
i=1
(51)
holds (see the Appendix for a heuristic proof) and minimizing over b gives
inf
b
n
X
i=1
n
X
bZi
χ21
Yi − √ ≈
|Yi | −
.
4f (0)
n
i=1
(52)
Moreover the same proof gives
inf
b
n
X
i=1
n
X
θ
χ2
bZi
Yi − med(Y n ) − √ − √ ≈
|Yi −med(Y n )|+f (0)θ2 − 1 (53)
4f (0)
n
n
i=1
from which the asymptotic α-non-significance interval
s
s
#
"
qchisq(α, 1)
qchisq(α, 1)
med(Y n ) −
, med(Y n ) +
4f (0)2 n
4f (0)2 n
(54)
as defined in (36) and (37) follows. This latter interval is the same as the
asymptotic confidence interval based on the median. Just as for the inverse
rank method it does not require an estimate of f (0).
L1 linear regression can be treated in the same manner. Corresponding to
(51) one has
n
X
i=1
n
X
Zt b
|Yi | − N (0, I k )t b + f (0)kbk22 .
Yi − √i· ≈
n
i=1
(55)
Applying this to the L1 regression functional gives
inf
b
n
X
i=1
n
Yi −
where Qn =
xti· β 1,n
X
xt θ Z t b
χ2k
− √i· − √i· ≈
|Yi −xti· β1,n |+f (0)θt Qn θ−
n
n
4f (0)2
i=1
(56)
1 t
n xn xn .
From this the asymptotic α-non-significance region
(β − β 1,n )t Qn (β − β1,n ) ≤
qchisq(α, k)
4f (0)2 n
(57)
follows. It is the same as the α-confidence region based on the L1 regression
estimate β1 , see for example Zhou and Portnoy (1996).
Table 8 gives the covering frequencies and average interval lengths for data
generated according to
Y = −39.69+0.832·Air.F low+0.574·W ater.T emp−0.061·Acid.Conc+ε (58)
19
using the L1 coefficients for the stack loss data. The sample size is n = 21. The
following four distributions for the error term ε are used: ε = N (0, 1) ∗ Res,
ε = σN (0, 1) for the normal distribution, ε = σL∗ for the Laplace distribution
and ε = σC ∗ for the Cauchy distribution where L∗ and C ∗ are respectively the
Laplace and Cauchy distributions closest to the N (0, 1) distribution, Res are
the residuals and σ the mean absolute deviation of the residuals of the stack
loss data.
residuals
Normal
Laplace
Cauchy
β2
β3
β4
(45)
0.944 0.265
0.982 0.682
0.998 0.248
rank
0.976 0.390
0.970 1.205
0.970 0.273
(45)
0.954 0.381
0.946 1.042
0.964 0.442
rank
0.974 0.435
0.956 1.208
0.962 0.542
(45)
0.953 0.501
0.959 1.375
0.952 0.580
rank
0.966 0.594
0.959 1.697
0.960 0.761
(45)
0.928 1.467
0.942 4.052
0.936 1.731
rank
0.936 1.948
0.946 5.676
0.942 2.984
Table 8: Covering frequencies and average interval lengths for data generated
according to (58) with different distributions for the error term: α = 0.95.
Finally, in the case of non-linear L1 regression the asymptotic α-non-significance
is, under suitable regularity conditions, given by
(β − β nlr1,n )t Qn (β − β nlr1,n ) ≤
qchisq(α, k)
4f (0)2 n
(59)
where
n
Qn =
1X
∇i ∇ti
n i=1
∇i =
and
4
∂m(xi· , θ)
∂m(xi· , θ)
,...,
∂θ1
∂θk
t
.
Choice of noise
It is possible to use random variables other than Gaussian. As an example
the 0.95-non-significance intervals for the median of the stack loss data us20
ing N (0, 1), ±1, U (−1, 1), ±beta(5, 5) and the standard Cauchy distribution
are (11.88, 18.63), (11.86, 18.25) (11.83, 18.16), (11.93, 18.21), (11.83, 18.16) and
(11.00, 18.56) respectively. It is clear that the results depend on the choice of
noise to some extent but that at least in this example the dependence is weak.
Given the advantages of Gaussian noise are the easily available asymptotic expressions such (29) it would seem to be the default choice of noise.
Other possibilities are to make the noise dependent on the size of the covariates as in Wij = xij Zij or to randomly permute the covariates (see Anderson and Robinson
(2001), Klingbiel (2009)).
5
Appendix
Consider
n
X
i=1
Ut b
εi − √i·
n
(60)
where the εi are symmetric, i.i.d. random variables with a continuously differentiable density at u = 0 with f (0) > 0. The U i· are k dimensional random
variables U i· = (Ui,1 , . . . , Ui,k )t where the Uij are symmetric i.i.d. random variable with unit variance. The sum (60) may be decomposed as
n
X
i=1
Ut b
εi − √i·
n
=
εi ≤−
+
X
εi ≥
|εi |≤
n
X
i=1
X
X
|εi |≤
n
√
n
U ti· b/
|εi | +
+
√
U ti· b/
+
=
√
n
U ti· b/
X
n
X
√
−εi + U ti· b/ n
√
εi − U ti· b/ n
Ut b
εi + √i·
n
Ut b
± √i· −
n
i=1
U√ti· b
X
√
U ti· b/ n
U ti· b
− |εi | .
εi − √
n
n
The random variables
U ti· b
U ti· b
Vi = √
|εi | ≤ √
n
n
21
|εi |≤
U ti· b
√
n
are i.i.d with mean zero and variance
t
U i· b
1
√
−F
E U (U ti· b)2 F
n
n
U i· tb
− √
n
!!!
=o
kbk22
n
.
This together with the central limit theorem implies
n
n
X
X
X
U ti· b
U ti· b
t
εi − √
|εi | + Z b +
εi − √
=
− |εi | + o kbk22
n
n
√
t
i=1
i=1
|εi |≤ U i· b/ n
D
where Z = N (0, Ik ). Denote the distribution function of U ti· b by H. Then
Z ∞
U ti· b
2
U ti· b
w
|εi | ≤ √
= √
dH(w)
E U εi − √
w |εi | ≤ √
n
n
n 0
n
E
and taking the expected value with respect to εi gives
U ti· b
U ti· b
|εi | ≤ √
εi − √
n
n
Z ∞
2
w
w
= √
− F −√
dH(w)
w F √
n 0
n
n
Z
4f (0) ∞ 2
2f (0)kbk22
≈
w dH(w) =
n
n
0
as the Uij are symmetric random variables with variance 1. A similar calculation
gives
f (0)kbk22
U ti· b
≈
.
E |εi | |εi | ≤ √
n
n
Putting this together leads to
n
X
i=1
n
X
f (0)kbk22
Ut b
|εi | + Z t b +
εi − √i· ≈
n
n
i=1
and minimizing over b results in
inf
b
n
X
i=1
n
X
χ2k
Ut b
|εi | −
εi − √i· ≈
n
4f (0)
i=1
where χ2k is a chi-squared random variable with k degrees of freedom.
References
Akaike, H. (1973), Information theory and an extension of the maximum likelihood principle, in B. Petrov and F. Csaki, eds, ‘Second international symposium on information theory’, Acadamiai Kiado, Budapest, pp. 267–281.
22
Akaike, H. (1974), ‘A new look at the statistical model identification’, IEEE
Transactions on Automatic Control 19, 716–723.
Akaike, H. (1981), ‘Likelihood of a model and information criteria’, Journal of
Econometrics 16, 3–14.
Anderson, M. and Robinson, J. (2001), ‘Permutation tests for linear models’,
Australian and New Zealand Journal of Statistics 43(1), 75–88.
Berk, R., Brown, L., Buja, A., Zhang, K. and Zhao, L. (2013), ‘Valid postselection inference’, Annals of Statistics 41(2), 802–837.
Brownlee, K. A. (1960), Statistical Theory and Methodology in Science and Engineering, 2nd edn, Wiley, New York.
Claeskens, G. and Hjort, N. L. (2003), ‘Focused information criterion’, Journal
of the American Statistical Association 98, 900–916.
Davies, L. (2014), Data Analysis and Approximate Models, Monographs on
Statistics and Applied Probability 133, CRC Press.
Dutter, R. (1977a), ‘Algorithms for the Huber estimator in multiple regression’,
Computing 18, 167–176.
Dutter, R. (1977b), ‘Numerical solution of robust regression problems: Computational aspects, a comparison’, Journal of Statistical Computing and Simulation 5, 207–238.
Hampel, F. R., Ronchetti, E. M., Rousseeuw, P. J. and Stahel, W. A. (1986),
Robust Statistics: The Approach Based on Influence Functions, Wiley, New
York.
Hosmer, D. W. and Lemeshow, S. (1989), Applied Logistic Regression, Wiley,
New York.
Huber, P. J. and Ronchetti, E. M. (2009), Robust Statistics, second edn, Wiley,
New Jersey.
Klingbiel, D. (2009), Permutation Tests Applied to Multivariate Regression,
PhD thesis, Philosophisch-naturwissenschaftlice Fakultät der Universität
Bern, Bern, Switzerland.
23
Koenker,
R.
(2010),
‘quantreg:
Quantile
regression’,
http://CRAN.R-project.org/package=quantreg. R package version 4.53.
Koenker, R. and Portnoy, S. (1997), ‘The Gaussian hare and the Laplacian tortoise: computability of squared-error versus absolute-error estimators’, Statist.
Sci. 12(4), 279–300.
Kunsch, H., Beran, J. and Hampel, F. (1993), ‘Contrasts under long-range correlations’, Annals of Statistics 21, 943–964.
Lindsay, B. and Liu, J. (2009), ‘Model assessment tools for a model false world’,
Statistical Science 24(3), 303–318.
Press, W. H., Teukolsky, S. A., Vetterling, W. T. and Flannery, B. P. (2003),
Numerical Recipes in Fortran 77: The Art of Scientific Computing, Vol. 1,
second edn, Cambridge University Press.
R Core Team (2013), R: A Language and Environment for Statistical Computing, R Foundation for Statistical Computing, Vienna, Austria.
URL: http://www.R-project.org/
Schwarz, G. E. (1978), ‘Estimating the dimension of a model’, Annals of Statistics 6(2), 461–464.
Stigler, S. M. (1977), ‘Do robust estimators work with real data? (with discussion)’, Annals of Statistics 5(6), 1055–1098.
Tibshirani, R. (1996), ‘Regression shrinkage and selection via the lasso’, J.
Royal. Statist. Soc B. 58(1), 267–288.
Zhou, K. Q. and Portnoy, S. L. (1996), ‘Direct use of regression quantiles to
construct confidence sets in linear models’, Annals of Statistics 24(1), 287–
306.
24
| 10math.ST
|
WRPN: Training and Inference using Wide Reduced-Precision Networks
Asit Mishra
Jeffrey J Cook
Eriko Nurvitadhi
Debbie Marr
arXiv:1704.03079v1 [cs.LG] 10 Apr 2017
Accelerator Architecture Lab, Intel Labs
Abstract
rameters) and increase the number of filter maps in a layer.
We call networks using this scheme wide reduced-precision
networks (WRPN) and find that this scheme compensates
or surpasses the accuracy of the baseline full-precision network. Although the number of raw compute operations increase as we increase the number of filter maps in a layer,
the compute bits required per operation is now a fraction of
what is required when using full-precision operations.
We present results using our scheme on AlexNet and
ResNet on ILSVRC-12 dataset. Our results show that the
proposed scheme offer better accuracies on ILSVRC-12
dataset, while being computationally less expensive compared to previously reported reduced-precision networks.
Further, our reduced-precision quantization scheme is hardware friendly allowing for efficient hardware implementations of such networks for servers, deeply-embedded and
real-time deployments.
For computer vision applications, prior works have
shown the efficacy of reducing numeric precision of model
parameters (network weights) in deep neural networks but
also that reducing the precision of activations hurts model
accuracy much more than reducing the precision of model
parameters. We study schemes to train networks from
scratch using reduced-precision activations without hurting
the model accuracy. We reduce the precision of activation
maps (along with model parameters) using a novel quantization scheme and increase the number of filter maps in a
layer, and find that this scheme compensates or surpasses
the accuracy of the baseline full-precision network. As a
result, one can significantly reduce the dynamic memory
footprint, memory bandwidth, computational energy and
speed up the training and inference process with appropriate hardware support. We call our scheme WRPN - wide
reduced-precision networks. We report results using our
proposed schemes and show that our results are better than
previously reported accuracies on ILSVRC-12 dataset while
being computationally less expensive compared to previously reported reduced-precision networks.
2. WRPN Scheme
While most prior works proposing reduced-precision
networks work with low precision weights (e.g. [1, 5, 4, 3]),
we find that activation maps occupy a larger memory footprint when using mini-batches of inputs. Using minibatches of inputs is typical in training of DNNs and for
multi-modal inference [2]. Figure 1 shows the memory
footprint of activation maps and filter maps as batch size
changes for 4 different networks during the training and inference steps. As batch-size increases, because of the filter
reuse aspect across batches of inputs, the activation maps
occupy significantly larger fraction of memory compared to
the filter weights.
Based on this observation, we reduce the precision of activation maps for DNNs to speed up training and inference
steps as well as cut down on memory requirements. However, a straightforward reduction in precision of activation
maps leads to significant reduction in model accuracy. This
has been reported in prior work as well [4]. We conduct
a sensitivity study where we reduce the precision of activation maps and model weights for AlexNet network using
ILSVRC-12 dataset. Table 1 reports our findings. We find
that reducing the precision of activation maps hurts model
accuracy much more than reducing the precision of the filter
parameters.
To re-gain the model accuracy while working with
reduced-precision operands, we increase the number of filter maps in a layer. Increasing the number of filter maps
increases the layer compute complexity linearly. Another
knob we tried was to increase the size of the filter maps.
However increasing the filter map spatial dimension in-
1. Introduction
Deep learning for robotics vision demands highly efficient real-time solutions. A promising approach to deliver
one such extremely efficient solution is through the use of
low numeric precision deep learning algorithms. Operating in lower precision mode reduces computation as well
as data movement and storage requirements. Due to such
efficiency benefits, there are many existing works that have
proposed low-precision deep neural networks (DNNs), even
down to 2-bit ternary mode [5] and 1-bit mode [4, 1]. However, the majority of existing works in low-precision DNNs
sacrifice accuracy over the baseline full-precision networks.
Further, most prior works target reducing the precision of
the model parameters (network weights). This primarily
benefits the inference step only when batch sizes are small.
We observe that activation maps (neuron outputs) occupy
more memory compared to the model parameters for batch
sizes typical during training. This observation holds even
during inference when batch size is 8 or more for modern
networks. Based on this observation, we study schemes for
training and inference using low-precision DNNs where we
reduce the precision of activation maps as well as the model
parameters without sacrificing the network accuracy. We reduce the precision of activation maps (along with model pa1
2x the number of filters, reduced-precision AlexNet is just
14% of the total compute cost1 of the full-precision baseline.
We also study how our scheme applies to bigger networks. For this we study widening the number of filters in
ResNet-34. When doubling the number of filters in ResNet34, we find that 4b weights and 8b activations surpass
the baseline accuracy (baseline full-precision ResNet-34 is
73.2% top-1 accuracy and our reduced-precision ResNet-34
is 73.8% top-1 accuracy). The reduced-precision ResNet is
15.1% the compute cost of full-precision ResNet.
!"#$%$%&
%89/9:4;8 <::2=4312
%8>?
)'#!%
'$#+%
&#'%
+
$)
!)#*%
&'#"%
$)
34 5)
!#!%
&#(%
(#&%
+
')#$%
'$#(%
',#,%
&(#&%
-./01 /2
)!#)%
$+#"%
$,#,%
!"#$%
%8-@2?
+
$)
4 /61 /2 7 ("
+
$)
4 /61 /2 7 +"+
!"#$%$"&$
%89/9:4;8 <::2=4312
%8>?
2.1. Hardware friendly quantization scheme
%8-@2?
For quantizing tensor values, we first constrain the
weight values (W) to lie within the range {-1,+1} and activation values (A) within the range {0,1}. This is followed
by a k-bit quantization of the values within the respective interval. Since a k-bit number can represent 2k numbers, we
k−1
−1)∗W )
represent the quantized weight values as round((2
2k−1 −1
&#$%
$#'%
+)#*%
+,#$%
$&#(%
((#,%
,'#"%
'+#!%
'(#$%
'*#"%
!'#(%
!"#$%
*(#*%
&&#+%
")#'%
!#&%
"
&+
-./01 /2
"
&+
34 5+
"
&+
4 /61 /2 7 $)
"
&+
4 /61 /2 7 ")"
k
Figure 1. Memory footprint of activations and weights during
training and inference for mini-batch sizes 1 and 32.
Top-1
Acc.
32b W
8b W
4b W
2b W
1b W
32b A
1x/2x
57.2/60.5
-/-/57.5/56.8/-
8b A
1x/2x
54.3/58.9
54.5/59.0
54.2/58.8
50.2/-/-
4b A
1x/2x
54.4/58.6
53.2/58.8
54.4/58.6
50.5/57.2
-/-
2b A
1x/2x
52.7/57.5
51.5/57.1
52.4/57.3
51.3/55.8
-
1b A
1x/2x
-/52.0
-/50.8
-/-/44.2/-
Table 1. AlexNet on ILSVRC-12 Top-1 accuracy % as precision
of activations (A) and weights (W) changes. - is a data-point we
did not experiment for. Bold numbers are top-1 accuracy with
2x filters. All results are with end-to-end training of the network
from scratch. Noteworthy data-point is the comparison of 32b W
and 32b A with 1x filter (baseline; 57.2%) with 4b W and 4b A
with 2x filter (58.6%).
creases the compute complexity quadratically. Although the
number of raw compute operations increase linearly as we
increase the number of filter maps in a layer, the compute
bits required per operation is now a fraction of what is required when using full-precision operations. As a result,
with appropriate support, one can significantly reduce the
dynamic memory requirements, memory bandwidth, computational energy and speed up the training and inference
process.
Table 1 also reports the accuracy of AlexNet when we
double the number of filter maps in a layer. We find that
with doubling of filter maps, AlexNet with 4b weights and
2b activations exhibits accuracy at-par with full-precision
networks. Operating with 4b weights and 4b activations,
surpasses the baseline accuracy by 1.44%. When doubling
the number of filter maps, AlexNet’s raw compute operations grow by 3.9x compared to the baseline full-precision
network, however by using reduced-precision operands the
overall compute complexity is a fraction of the baseline. For
example, with 4b operands for weights and activations and
−1)∗A)
. One
and the quantized activation values as round((2
2k−1
bit is reserved for sign-bit in case of weight values hence
the use of 2k−1 for these quantized values. With appropriate affine transformations, the convolution operations (the
bulk of the compute operations in the network) can be done
using quantized values followed by scaling with floatingpoint constants. This makes our scheme hardware friendly
and can be used in embedded systems. This is the quantization scheme for all the results reported above for AlexNet
and ResNet.
3. Conclusions
Vision, speech and NLP based applications have seen
tremendous success with DNNs. However, DNNs are compute intensive workloads. In this paper, we present WRPN
scheme to reduce the compute requirements of DNNs.
While most prior works look at reducing the precision of
weights, we find that activations contribute significantly
more to the memory footprint than weight parameters when
using mini-batches and thus aggressively reduce the precision of activation values. Further, we increase the number
of filter maps in a layer and show that widening of filters and
reducing precision of network parameters does not sacrifice
model accuracy. Our scheme is hardware friendly making
it viable for deeply embedded system deployments which is
useful for robotics applications.
References
[1] M. Courbariaux and Y. Bengio. Binarynet: Training deep neural networks with
weights and activations constrained to +1 or -1. CoRR, abs/1602.02830, 2016. 1
[2] I. V. Serban, A. G. O. II, J. Pineau, and A. C. Courville. Multi-modal variational
encoder-decoders. CoRR, abs/1612.00377, 2016. 1
[3] G. Venkatesh, E. Nurvitadhi, and D. Marr. Accelerating deep convolutional networks using low-precision and sparsity. CoRR, abs/1610.00324, 2016. 1
[4] S. Zhou, Z. Ni, X. Zhou, H. Wen, Y. Wu, and Y. Zou. Dorefa-net: Training
low bitwidth convolutional neural networks with low bitwidth gradients. CoRR,
abs/1606.06160, 2016. 1
[5] C. Zhu, S. Han, H. Mao, and W. J. Dally. Trained ternary quantization. CoRR,
abs/1612.01064, 2016. 1
1 Compute
cost is the product of the number of FMA operations and the
width of the activation and weight operands.
| 9cs.NE
|
Minimax Regret Bounds for Reinforcement Learning
Mohammad Gheshlaghi Azar, Ian Osband, and Remi Munos
DeepMind, London, UK
arXiv:1703.05449v2 [stat.ML] 1 Jul 2017
July 4, 2017
Abstract
We consider the problem of provably optimal exploration in reinforcement learning
√ for finite horizon MDPs.√We
e HSAT + H 2 S 2 A + H T )
show that an optimistic modification to value iteration achieves a regret bound of O(
where H is the time horizon, S the number of states, A the√
number of actions and T the number of time-steps. This
e
AT ) achieved by the UCRL2 algorithm of Jaksch et al.
result improves over the best previous known bound O(HS
(2010).
The
key
significance
of
our
new
results
is
that
when
T√ ≥ H 3 S 3 A and SA ≥ H, it leads to a regret of
√
e
O( HSAT ) that matches the established lower bound of Ω( HSAT ) up to a logarithmic factor. Our analysis
contains two key insights. We use careful application of concentration inequalities to the optimal value function
as a whole, rather than to the transitions probabilities (to improve scaling in S), and we define Bernstein-based
"exploration bonuses" that use the empirical variance of the estimated values at the next states (to improve scaling in
H).
1 Introduction
We consider the reinforcement learning (RL) problem of an agent interacting with an environment in order to maximize
its cumulative rewards through time (Burnetas & Katehakis, 1997; Sutton & Barto, 1998). We model the environment
as a Markov decision process (MDP) whose transition dynamics are unknown from the agent. As the agent interacts
with the environment it observes the states, actions and rewards generated by the system dynamics. This leads to a
fundamental trade off: should the agent explore poorly-understood states and actions to gain information and improve
future performance, or exploit its knowledge to optimize short-run rewards.
The most common approach to this learning problem is to separate the process of estimation and optimization. In
this paradigm, point estimates of the unknown quantities are used in place of the unknown parameters and a plan is
made with respect to these estimates. Naive optimization with respect to these point estimates can lead to premature
exploitation and so may never learn the optimal policy. Dithering approaches to exploration (e.g., ǫ-greedy) address
this failing through random action selection. However, as this exploration is not directed the resultant algorithms may
take exponentially long to learn (Kearns & Singh, 2002). In order to learn efficiently it is necessary that the agent
prioritizes potentially informative states and actions. To do this, it is important that the agent maintains some notion
of its own uncertainty. In some sense, given any prior belief, the optimal solution to this exploration/exploitation
dilemma is given by the dynamic programming in the extended Bayesian belief state (Bertsekas, 2007). However, the
computational demands of this method become intractable for even small problems (Guez et al., 2013) while finite
approximations can be arbitrarily poor (Munos, 2014).
To combat these failings, the majority of provably efficient learning algorithms employ a heuristic principle known
as optimism in the face of uncertainty (OFU). In these algorithms, each state and action is afforded some “optimism”
such that its imagined value is as high as statistically plausible. The agent then chooses a policy under this optimistic
view of the world. This allows for efficient exploration since poorly-understood states and actions are afforded higher
optimistic bonus. As the agent resolves its uncertainty, the effects of optimism will reduce and the agent’s policy
will approach optimality. Almost all reinforcement learning algorithms with polynomial bounds on sample complexity employ optimism to guide exploration (Kearns & Singh, 2002; Brafman & Tennenholtz, 2002; Strehl et al., 2006;
Dann et al., 2017).
An alternative principle motivated by the Thompson sampling (Thompson, 1933) has emerged as a practical competitor to optimism. The algorithm posterior sampling reinforcement learning (PSRL) maintains a posterior distribution for MDPs and, at each episode of interaction, follows a policy which is optimal for a single random sample
(Strens, 2000).
1
Previous works have argue for the potential benefits of such PSRL methods over existing optimistic approaches
(Osband et al., 2013; Osband & Van Roy, 2016b) but they come with guarantees on the Bayesian regret only.
However a very recent work Agrawal & Jia (2017) have shown that an√
optimistic version of posterior sampling
e
(using a max over several samples) achieves a frequentist regret bound O(H
SAT ) (for large T ) in the more general
setting of weakly communicating MDPs.
In this paper we present a conceptually simple and computationally efficient approach to optimistic reinforcement
learning in finite-horizon MDPs and report results for the frequentist regret. Our algorithm, upper confidence bound
value iteration (UCBVI) is similar to model-based interval estimation (MBIE-EB) (Strehl & Littman, 2005) with a
delicate alteration to the form of the “exploration bonus”. In particular UCBVI replaces the universal scalar of the
bonus in MBIE-EB with the empirical variance of the next-state
√ value function of each state-action pair. This alteration
e
e H).
is essential to improve the regret bound from O(H)
to O(
√
√
e HSAT + H 2 S 2 A + H T ) where S
Our key contribution is to establish a high probability regret bound O(
is the number of states, A is the number of actions, H is the episode length and T is the total number of√time-steps
e ignores logarithmic factors). Importantly, for T > H 3 S 3 A and SA ≥ H this bound is O(
e HSAT ),
(and where O
which matches the established lower bound for this problem, up to logarithmic factors (Osband & Van Roy, 2016a).1
This positive result is the first of its kind and helps to address an ongoing question about where the fundamental
lower bounds lie for reinforcement learning in finite horizon MDPs (Bartlett & Tewari, 2009; Dann & Brunskill, 2015;
Osband & Van Roy, 2016a). Our refined analysis contains two key ingredients:
• We use careful application of Bernstein and Freedman inequalities (Bernstein, 1927; Freedman, 1975) to the
concentration of the optimal value function directly, rather than building confidence sets for the transitions
probabilities and rewards, like in UCRL2 (Jaksch et al., 2010) and UCFH (Dann & Brunskill, 2015).
• We use empirical-variance exploration bonuses based on Bernstein’s inequality, which together with a recursive
Bellman-type Law of Total Variance (LTV) provide tight bounds on the expected sum of the variances of the
value estimates, in a similar spirit to the analysis from Azar et al. (2013); Lattimore & Hutter (2012).
At a high level, this work addresses the noted shortcomings of existing RL algorithms (Bartlett & Tewari, 2009;
Jaksch et al., 2010; Osband & Van Roy, 2016b), in terms of dependency on S and H. We demonstrates that it is
possible to design a simple and computationally efficient optimistic algorithm
√ that simultaneously address both the
loose scaling in S and H to obtain the first regret bounds that match the Ω( HSAT ) lower bounds as T becomes
large.
We should be careful to mention the current limitations of our work, each of which may provide fruitful ground for
future research. First, we study the setting of episodic, finite horizon MDPs and not the more general setting of weakly
communicating systems (Bartlett & Tewari, 2009; Jaksch et al., 2010). Also we assume
length H is
√ that the horizon
e
AT ) for T > H 3 S 3 A.
known to the learner. Further, our bounds only improve over previous scaling O(HS
We hope that this work will serve to elucidate several of the existing shortcomings of exploration in the tabular
setting and help further the direction of research towards provably optimal exploration in reinforcement learning.
2 Problem formulation
In this section, we briefly review some notation, as well as some standard concepts and definitions from the theory of
Markov decision processes (MDPs).
Markov Decision Problems We consider the problem of undiscounted episodic reinforcement learning (RL) (Bertsekas & Tsitsiklis,
1996), where an RL agent interacts with a stochastic environment and this interaction is modeled as a discrete-time
MDP. An MDP is a quintuple hS, A, P, R, Hi, where S and A are the set of states and actions, P is the state transition
distribution, The function R : S × A → ℜ is a real-valued function on the state-action space and the horizon H
is the length of episode. We denote by P (·|x, a) and R(x, a) the probability distribution over the next state and the
immediate reward of taking action a at state x, respectively. The agent interacts with the environment in a sequence of
episodes. The interaction between the agent and environment at every episode2 k ∈ [K] is as follows: starting from
xk,1 ∈ S which is chosen by the environment, the agent interacts with the environment for H steps by following a
sequence of actions chosen in A and observes a sequence of next-states and rewards until the end of episode. The
initial state xk,1 may change arbitrarily from one episode to the next. We also use the notation k · k1 for the ℓ1 norm
throughout this paper.
1 In fact the lower bound of (Jaksch et al., 2010) is for the more general setting of the weakly communicating MDPs and it doesn’t directly
apply to our setting. But a similar approach can be used to prove a lower bound of same order for the finite-horizon MDPs, as it is already used in
(Osband & Van Roy, 2016a).
2 We write [n] for {i ∈ N | 1 ≤ i ≤ n}.
2
Assumption 1 (MDP Regularity). We assume S and A are finite sets with cardinalities S, A, respectively. We also
assume that the immediate reward R(x, a) is deterministic and belongs to the interval [0, 1].3
In this paper we focus on the setting where the reward function R is known, but extending our algorithm to
unknown stochastic rewards poses no real difficulty.
The policy during an episode is expressed as a mapping π : S × [H] → A. The value Vhπ : S → R denotes
the value function at every step h = 1, 2, . . . , H and state x ∈ S such that Vhπ (x) corresponds to the expected sum
of H − h rewards received under policy π, starting from xh = x ∈ S. Under Assumption 1 there exists always a
def
policy π ∗ which attains the best possible values, and we define the optimal value function Vh∗ (x) = supπ Vhπ (x) for
π
all x ∈ S and h ≥ 1. The policy π at every step h defines the state transition kernel Ph and the reward function
def
def
rhπ as Phπ (y|x) = P (y|x, π(x, h)) and rhπ (x) = R(x, π(x, h)) for all x ∈ S. For every V : S → R the rightdef P
linear operators P · and Phπ · are also defined as (P V )(x, a) =
y∈S P (y|x, a)V (y) for all (x, a) ∈ S × A and
def P
π
π
(Ph V )(x) =
y∈S Ph (y|s)V (y) for all x ∈ S, respectively. The Bellman operator for the policy π, at every step
def
h > 0 and x ∈ S, is defined as (Thπ V )(x) = rhπ (x) + (Phπ V )(x). We also define the state-action Bellman operator
def
for all (x, a) ∈ S × A as (T V )(x, a) = R(x, a) + (P V )(x, a) and the optimality Bellman operator for all x ∈ S as
def
(T ∗ V )(x) = maxa∈A (T V )(x, a). For ease of exposition, we remove the dependence on x and (x, a), e.g., writing
P V for (P V )(x, a) and V for V (x), when there is no possible confusion.
We measure the performance of the learner over T = KH steps4 by the regret Regret(K), defined as
def
Regret(K) =
K
X
k=1
V1∗ (xk,1 ) − V1πk (xk,1 ),
where πk is the control policy followed by the learner at episode k. Thus the regret measures the expected loss
of following the policy produced by the learner instead of the optimal policy. So the goal of learner is to follow a
sequence of policies π1 , π2 , . . . , πK such that Regret(K) is as small as possible.
3 Upper confidence bound value iteration
In this section we introduce two variants of the algorithm that we investigate in this paper. We call the algorithm
upper confidence bound value iteration (UCBVI). UCBVI is an extension of value iteration which guarantees that the
resultant value function is a (high-probability) upper confidence bound (UCB) on the optimal value function. This
algorithm is related to the model based interval estimation (MBIE-EB) algorithm (Strehl & Littman, 2008). Our key
contribution is the precise design of the upper confidence sets, and the analysis which lead to tight regret bounds.
UCBVI, described in Algorithm 1, calls UCB-Q-values (Algorithm 2) which returns UCBs on the Q-values
computed by value iteration using an empirical Bellman operator to which is added a confidence bonus bonus. We
consider two variants of UCBVI depending on the structure of bonus, which we present in Algorithms 3 and 4.
Algorithm 1 UCBVI
Initialize data H = ∅
for episode k = 1, 2, . . . , K do
Qk,h = UCB − Q − values(H)
for step h = 1, . . . , H do
Take action ak,h = arg maxa Qk,h (xk,h , a)
Update H = H ∪ (xk,h , ak,h , xk,h+1 )
end for
end for
3 For
rewards in [Rmin , Rmax ] simply rescale these bounds.
this paper we will often substitute T =KH to highlight various dependencies relative to the existing literature. This equivalence should be
kept in mind by the reader.
4 In
3
Algorithm 2 UCB-Q-values
Require: Bonus algorithm bonus, Data H
Compute, forP
all (x,a,y) ∈ S × A × S,
Nk (x,a,y) = (x′ ,a′ ,y′ )∈H I(x′ = x,a′ = a,y ′ = y)
P
Nk (x,a) = P
y∈S Nk (x,a,y)
′
Nk,h
(x,a) = (xi,h ,ai,h ,xi,h+1 )∈H I(xi,h = x,ai,h = a)
Let K = {(x,a) ∈ S × A, Nk (x,a) > 0}
Estimate Pbk (y|x,a) = NNkk(x,a,y)
(x,a) for all (x,a) ∈ K
Initialize Vk,H+1 (x) = 0 for all (x,a) ∈ S × A
for h = H,H − 1,...,1 do
for (x,a) ∈ S × A do
if (x,a) ∈ K then
′
bk,h (x,a) = bonus(Pbk ,Vk,h+1 ,Nk ,Nk,h
)
Qk,h (x,a) = min Qk−1,h (x,a),H,
R(x,a) + (Pbk Vk,h+1 )(x,a) + bk,h (x,a)
else
Qk,h (x,a) = H
end if
Vk,h (x) = maxa∈A Qk,h (x,a)
end for
end for
return Q-values Qk,h
The first of these UCBVI-CH is based upon Chernoff-Hoeffding’s concentration inequality, considers UCBVI with
bonus = bonus_1. bonus_1 is a very simple bound which only assumes that values are bounded
in [0, H]. We will
√
e
see in Theorem 1 that this very simple algorithm can already
achieve
a
regret
bound
of
O(H
SAT
), thus improving
√
the best previously known regret bounds from a S to a S dependence. The intuition for this improved S-dependence
is that our algorithm (as well as our analysis) does not consider confidence sets on the transition dynamics P (y|x, a)
like UCRL2 and UCFH do, but instead directly maintains confidence intervals on the optimal value function. This
is crucial as, for any given (x, a), the transition dynamics are S-dimensional whereas the Q-value function is onedimensional.
Algorithm 3 bonus_1
Require: Pbk (x, a),
qNk (x, a)
1
where L = ln(5SAT /δ),
b(x, a) = 7HL Nk (x,a)
return b
However, the loose form of UCB given by UCBVI-CH does not look at the value function of the next state, and
just consider it as being bounded in [0, H]. However, much better bounds can be obtained by looking at the variance
of the next state values. Our main result relies upon UCBVI with bonus = bonus_2, which we refer to as UCBVI-BF
as it relies on Bernstein-Freedman’s concentration inequalities to build the confidence set. UCBVI-BF builds upon
the intuition for UCBVI-CH but also incorporates a variance-dependent
exploration bonus. This leads to tighter
√
e HSAT ).
exploration bonuses and an improved regret bound of O(
Compared to UCBVI-BF here we use a bonus built from the empirical variance of the estimated next values. The
idea is that if we had knowledge of the optimal value V ∗ , we could build tight confidence bounds using the variance
of the optimal value function at the next state in place of the loose bound of H. Since however V ∗ is unknown, here
we use as a surrogate the empirical variance of the estimated values. As more data is gathered, this variance estimate
will converge to the variance of V ∗ . Now we need to make sure our estimates Vk,h are optimistic (i.e., that they
upper bound Vh∗ ) at all times. This is achieved by adding an additional bonus (last term in b(x, a)), which guarantees
that we upper bound the variance of V ∗ . Now, using an iterative -Bellman-type- Law of Total Variance, we have
(see proof) that the sum of the next-state variances of V ∗ (over H time steps) (which is related to the sum of the
exploration bonuses over H steps) is bounded by the variance of the H-steps return. Thus the size of the bonuses built
4
Algorithm 4 bonus_2
Require: Pbk (x, a), Vk,h+1 , Nk , N ′
k,h
b(x,a)=
s
v
i
h
u P
u8
bk (y|x,a) min 1002′H 3 S 2 AL2 ,H 2
P
8LVarY ∼Pbk (·|x,a)(Vk,h+1 (Y ))
y
t
Nk,h+1 (y)
14HL
+
+
Nk (x,a)
3Nk (x,a)
Nk (x,a)
where L=ln(5SAT /δ)
return b
by UCBVI-BF
are constrained over the H steps. And we prove that the sum√of those bonuses do not grow linearly in
√
H but in H only. This is the key for our improved dependence from H to H.
4 Main results
In this section we present the main results of the paper, which are upper bounds on the regret of UCBVI-CH and
UCBVI-BF algorithms. We assume Assumption 1 holds.
Theorem 1 (Regret bound for UCBVI-CH ).
Consider a parameter δ > 0. Then the regret of UCBVI-CH is bounded w.p. at least 1 − δ, by
√
Regret(K) ≤ 20H 3/2 L SAK + 250H 2 S 2 AL2 ,
where L = ln(5HSAT /δ).
√
e
For T ≥ HS 3 A and SA ≥ H this bound translates to a regret bound of O(H
SAT ), where T = KH is the total
number of time-steps at the end of episode K.
√
Theorem 1 is significant in that, for large T , it improves the regret dependence from S to S, compared to the
best known bound of Jaksch et al. (2010). The main intuition for this improved S-dependence is that we bound the
estimation error of the next-state value function directly, instead of the transition probabilities.
More precisely, instead of bounding the estimation error (Pbkπk − P πk )Vk,h+1 by kPbkπk − P πk k1 kVk,h+1 k∞ (as is
∗
done in Jaksch et al. (2010) for example), we bound (Pbkπk −P πk )Vh+1
instead (for which a bound with no dependence
∗
∗
on S can be achieved since V is deterministic) and handle carefully the correction term (Pbkπk −P πk )(Vk,h+1 −Vh+1
).
Our second result, Theorem 2, demonstrates that we can improve upon the H-dependence by using a more refined,
Bernstein-Friedman-type, exploration bonus.
Theorem 2 (Regret bound for UCBVI-BF ).
Consider a parameter δ > 0. Then the regret of UCBVI-BF is bounded w.p. 1 − δ, by
√
√
Regret(K) ≤ 30HL SAK + 2500H 2 S 2 AL2 + 4H 3/2 KL,
where L = ln(5HSAT /δ).
√
e HSAT ). This result
We note that for T ≥ H 3 S 3 A and SA ≥ H this bound translates to a regret bound√of O(
3 3
e
is particularly significant since,
√ for T large enough (i.e., T ≥ H S A), our bound is O( HSAT ) which matches the
established lower bound Ω( HSAT ) of (Jaksch et al., 2010; Osband & Van Roy, 2016a) up to logarithmic factors.
The key insight is to apply concentration inequalities to bound the estimation errors and the exploration bonuses
in terms of the variance of V ∗ at the next state. We then use the fact that the sum of these variances is bounded by the
variance of the return (see e.g., Munos√
& Moore, 1999; Azar et al., 2013; Lattimore & Hutter, 2012), which shows
that the estimation errors accumulate as H instead of linearly in H, thus implying the improved H-dependence.
Computational efficiency
Theorems 1 and 2 guarantee the statistical efficiency of UCBVI. In addition, both UCBVI-CH and UCBVI-BF are
computationally tractable. Each episode both algorithms perform an optimistic value iteration with computational
cost of the same order as solving a known MDP. In fact, the computational cost of these algorithms can be further
reduced by only selectively recomputing UCBVI after sufficiently many observations. This technique is common
5
to the literature Jaksch et al. (2010); Dann & Brunskill (2015) and would not affect the Õ statistical efficiency. The
e
computational cost of this variant of UCBVI then amounts to O(SA
min(SA, T ) min(T, S)) as it only needs to update
e
the model O(SA) times (Jaksch et al., 2010).
Weakly communicating MDPs
In this short paper we focus on the setting of finite horizon MDPs. By comparison, previous optimistic approaches to
exploration, such as UCRL2, provide bounds for the more general setting of weakly communicating MDPs (Jaksch et al.,
2010; Bartlett & Tewari, 2009).
However, we believe that much of the insight from the UCBVI algorithm (and its analysis) will carry over to this
more general setting using existing techniques such as ‘the doubling trick‘ (Jaksch et al., 2010).
5 Proof sketch
Here we provide the sketch proof of our results. The full proof is deferred to the appendix.
5.1 Sketch Proof of Theorem 1
Let Ω = {Vk,h ≥ Vh∗ , ∀k, h} be the event under which all computed Vk,h values are upper bounds on the optimal
value function. Using backward induction on h (and standard concentration inequalities) one can prove that Ω holds
with high probability (see Lem. 18 in the appendix). To simplify notations in this sketch of proof we will not make the
numerical constants explicit, and instead we will denote by a numerical constant which can vary from line to line.
The exact values of these constants are provided in the full proof. We will also make use of simplified notations, such
as using L to represent the logarithmic term L = ln(HSAT /δ).
def
def P
πk
∗
^
The cumulative regret at episode K is Regret(K) =
1≤k≤K V1 (xk,1 ) − V1 (xk,1 ). Define Regret(K) =
P
πk
^
^
1≤k≤K Vk,1 (xk,1 ) − V1 (xk,1 ). Under Ω we have Regret(K) ≤ Regret(K), so we now bound Regret(K). Define
def
def
e k,h = Vk,h − V πk . Thus
∆k,h = Vh∗ − Vhπk and ∆
h
e k,h+1 + bk,h .
e k,h = Pb πk Vk,h+1 + bk,h − P πk V πk = (Pbπk − P πk )Vk,h+1 + P πk ∆
∆k,h ≤ ∆
k
h+1
k
The difficulty in bounding (Pbkπk − P πk )Vk,h+1 is that both Vk,h+1 and Pbkπk are random variables and are not independent (the value function Vk,h+1 computed at h + 1 may depend on the samples collected from state xh,k ), thus a
straightforward application of Chernoff-Hoeffding (CH) inequality does not work here. In √
Jaksch et al. (2010), this
issue is addressed by bounding it by kPbkπk − P πk k1 kVk,h+1 k∞ at the price of an additional S.
√
√
e
SAT ) bound (which removes a S factor compared to the previous bound
The main contribution of our O(H
of Jaksch et al. (2010)) is to handle this term more properly. Instead of directly bounding (Pbkπk − P πk )Vk,h+1 , we
√
∗
∗
is
bound (Pbkπk − P πk )Vh+1
, using straightforward application of CH (which removes the S factor since Vh+1
πk
πk
∗
b
deterministic), and deal with the correction term (Pk − P )(Vk,h+1 − Vh+1 ). We have
∗
e k,h+1 + bk,h + ek,h ,
e k,h = (Pbπk − P πk )(Vk,h+1 − Vh+1
∆
) + P πk ∆
k
def
∗
where ek,h = (Pbkπk − P πk )Vh+1
(xk,h ) is the estimation error of the optimal value function at the next state. Defining
def
e k,h (xk,h ), we have
δek,h = ∆
def
δek,h ≤ (Pbkπk − P πk )∆k,h+1 (xk,h ) + δek,h+1 + ǫk,h + bk,h + ek,h ,
where ǫk,h = P πk ∆k,h+1 (xk,h ) − ∆k,h+1 (xk,h+1 ).
Step 1: bound on the correction term (Pbkπk − P πk )∆k,h+1 (xk,h ). Using Bernstein’s inequality (B), this term is
bounded by
s
X
SHL
L
∆k,h+1 (y) +
,
P πk (y|xk,h )
πk (y|x
P
)n
nk,h
k,h
k,h
y
6
def
where nk,h = Nk (xk,h , πk (xk,h )). Now considering only the y such that P πk (y|xk,h )nk,h ≥ H 2 L, and since
e k,h+1 , then (Pbπk − P πk )∆k,h+1 (xk,h ) is bounded by
0 ≤ ∆k,h+1 ≤ ∆
k
s
1
SHL
SHL
L
≤ ǭk,h + δek,h+1 +
,
δek,h+1 +
ǭk,h +
P πk (xk,h+1 |xk,h )nk,h
nk,h
H
nk,h
def
where ǭk,h =
q
L
nk,h
P
e k,h+1 (y)
∆
πk
(y|xk,h ) √
yP
−√
P πk (y|xk,h )
ek,h+1
δ
P πk (xk,h+1 |xk,h )
.
The sum over the neglected y such that P πk (y|xk,h )nk,h < H 2 L contributes to an additional term
s
X P πk (y|xk,h )nk,h L
SH 2 L
.
∆k,h+1 (y) ≤
2
nk,h
nk,h
y
Neglecting this term (and the smaller order term SHL/nk,h) for now (by the pigeon-hole principle we can prove
that these terms contribute to the final regret by a constant at most S 2 AH 2 L2 ), we have
H−1
1 H X
1 e
δk,h+1 + ǫk,h + ǭk,h + bk,h + ek,h ≤ 1 +
ǫk,i + ǭk,i + bk,i + ek,i .
δek,h ≤ 1 +
H
|
{zH } i=h
≤e
The regret is thus bounded by
^
Regret(K)
≤
X
(ǫk,h + ǭk,h + bk,h + ek,h ).
(1)
k,h
P
P
and k,h ǭk,h are sums of martingale differences,
√
e
which are bounded using Azuma’s inequality, and lead to a regret of O(H
T ) without dependence on the
Psize of state
and action space. TheP
leading terms in the regret bound comes from the sum of the exploration bonuses k,h bk,h and
the estimation errors k,h ek,h .
We now bound those 4 terms. It is easy to check that
Step 2: Bounding the martingales
X
k,h
P
k,h ǫk,h
k,h
P
k,h ǭk,h .
X
(Az)
√
ǫk,h ≤ H T L,
Step 3: Bounding the exploration bonuses
X
and
k,h ǫk,h
bk,h = HL
X
k,h
s
k,h
P
k,h bk,h :
1
nk,h
Using Azuma’s inequality we deduce
(Az)
ǭk,h ≤
√
T L.
(2)
Using the pigeon-hole principle, we have
= HL
(x,a)
X NKX
x,a
n=1
r
√
1
≤ HL SAT .
n
(3)
(CH)
P
∗
≤
Step 4: Bounding on the estimation errors k,h ek,h . Using CH, w.h.p. we have ek,h = (Pbkπk − P πk )Vh+1
q
L
H nk,h . Thus this bound on the estimation errors are of the same order as the exploration bonuses (which is the
reason we choose those bonuses...).
Putting everything together: Plugging Eq. 2 and Eq. 3 into Eq. 1 (and adding the smaller order term) we deduce
√
3
^
Regret(K) ≤ Regret(K)
≤ H 2 L SAK + H 2 S 2 AL2 .
5.2 Sketch Proof of Theorem 2
The proof of Theorem 1 relied on proving by a straightforward induction over h that Ω = {Vk,h ≥ Vh∗ , ∀k, h} hold
with high probability. In the case of exploration bonuses defined by:
7
v
u
u min H 3 S 2 AL2 P Pbk′ (y|x,a) , H 2
LVY ∼Pbπk (·|x,a)
y Nk,h+1 (y)
t
HL
k
+
,
bk,h (x, a) =
+
Nk (x, a)
Nk (x, a)
Nk (x, a)
|
{z
} |
{z
}
empirical Bernstein
additional bonus
s
Vk,h+1 (Y )
(4)
∗
the backward induction over h is not straightforward. Indeed, if the Vk,h+1 are upper bounds on Vh+1
, it is not
∗
necessarily the case that the empirical variance of Vk,h+1 are upper bound on the empirical variance of Vh+1 . However
∗
we can prove by (backward) induction over h that Vk,h+1 is sufficiently close to Vh+1
to guarantee that the variance
of those terms are sufficiently close to each other so that the additional bonus (additional bonus in Eq. 4) will make
sure that Vk,h is still an upper-bound on Vh∗ . More precisely, define the set of indices:
[k, h]hist
def
= {(i, j), s.t.(1 ≤ i ≤ k ∧ 1 ≤ j ≤ H) ∨ (i = k ∧ h < j ≤ H)},
def
and the event Ωk,h = {Vi,j ≥ Vh∗ , (i, j) ∈ [k, h]hist }. Our induction is the following:
q
∗
• Assume that Ωk,h holds. Then we prove that (Vk,h+1 − Vh+1
)(y) ≤ H N ′ SAL(y) .
k,h+1
P
b
∗
• We deduce that VY ∼Pbk (·|x,a) Vk,h+1 (Y ) + H 3 S 2 AL2 y NP′k (y|x,a)
≥ VY ∼Pbk (·|x,a) Vh+1
(Y ) , so the additional
(y)
k,h+1
∗
bonus compensates for the possible variance difference. Thus Vk,h ≥ Vh and Ωk,h−1 holds.
So in order to prove that all values computed by the algorithm
are upper bounding V ∗ , we just need to prove that under
q
∗
Ωk,h , we have (Vk,h+1 − Vh+1
)(y) ≤ min(H 1.5 SL N ′ A (y) , H), which is obtained by deriving the following
k,h+1
regret bound on
ek,h (y) def
R
=
X
i≤k
q
πi
′
(y).
(Vi,h+1 − Vh+1
)(xi,h+1 )I{xi,h+1 = y} ≤ HL SANk,h+1
(5)
Indeed, since {Vi,h }i is a decreasing sequence in i, we have
q
∗
′
ek,h+1 (y)/N ′
SA/Nk,h+1
(y).
(Vk,h+1 − Vh+1
)(y) ≤ R
(y)
≤
HL
k,h+1
Once we have proven that w.h.p., all computed values are upper bounds on V ∗ (i.e. event Ω), then we prove that under
Ω, the following regret bound holds:
√
^
Regret(K) ≤ Regret(K)
≤ (HL SAK + H 2 S 2 AL2 ).
(6)
The proof of Eq. 5 relies on the same derivations as those used for proving Eq. 6. The only two differences being
′
that (i) HK is replaced by Nk,h+1
(y), the number of times a state y was reached at time h + 1, up to episode k, and
√
′
(y) can only tick once, whereas
(ii) the additional H factor which comes from the fact that at any episode, Nk,h+1
the total number of transitions from y during any episode can be as large as H. The full proof of Eq. 5 will be given
in details in the appendix. We now give a proof sketch of Eq. 6 under Ω.
Similar steps used for proving Theorem 1 apply. The main difference compared to Theorem 1 is the bound on the
sum of the exploration
√ bonuses and the estimation errors (which we consider in Steps 3’ and 4’ below). This is where
we can remove the H factor. The use of the Bernstein inequality makes it possible to bound both of those terms in
terms of the expected sum of variances (under the current policy πk at any episode k) of the next-state values (for that
policy), and then using recursively the Law of Total Variance to conclude that this quantity is nothing but the variance
of the returns. This step is detailed now. For simplicity of the exposition of this sketch we neglect second order terms.
Step 3’: Bounding the sum of exploration bonuses bk,h . We have
v
s
u
u min H 3 S 2 AL2 P Pbk′ (y|x,a) , H 2
X
X L
√ X VY ∼Pbπk (·|xk,h ) Vk,h+1 (Y )
y Nk,h+1 (y)
t
h
+
.
+
bk,h = L
nk,h
Nk (x, a)
Nk (x, a)
k,h
k,h
k,h
{z
} |
{z
}
|
main term
second order term
8
P
1/2
P
1
b
By Cauchy-Schwarz, the main term is bounded by
, where
k,h nk,h
k,h Vk,h+1
P
def
1
b k,h+1 = V bπk
V
k,h nk,h ≤ SA ln(T ) by the pigeon-hole principle, we now focus
Y ∼P (·|x
) Vk,h+1 (Y ) . Since
P hb k,h
on the term k,h Vk,h+1 .
def
πk
b k,h+1 is close to Vπk
π
We now prove that V
k,h+1 = VY ∼P k (·|xk,h ) Vh+1 (Y ) by bounding the following quantity:
h
b k,h+1 − Vπk
V
k,h+1
πk 2
πk 2
2
Pb πk Vk,h+1
− (Pb πk Vk,h+1 )2 − P πk (Vh+1
) + (P πk Vh+1
)
=
(i)
πk 2
2
∗
Pb πk Vk,h+1
− P πk (Vh+1
) + 2H(P πk − Pb πk )Vh+1
≤
(ii)
2
2
+ P πk (Vk,h+1
−
(Pbπk − P πk )Vk,h+1
{z
} |
{z
|
≤
(ak,h )
πk 2
(Vh+1
) ) +H 2
}
(a′k,h )
s
L
,
nk,h
(7)
where (i) holds since under Ωk,h , Vk,h ≥ Vh∗ ≥ Vhπk and (ii) holds due to Chernoff Hoeffding.
Step 3’-a: bounding
that
P
k,h
b k,h+1 − Vπk .
V
k,h+1
ak,h
def
Using similar argument as those used in Jaksch et al. (2010), we have
q
≤ H 2 kPb πk − P πk k1 ≤ H 2 SL/nk,h ,
(where nk,h = Nk (xk,h , πk (xk,h ))). Thus from the pigeon-hole principle,
Now a′k,h is bounded as
P
k,h
√
ak,h ≤ H 2 S AT L.
b k,h .
a′k,h ≤ 2HP πk (Vk,h − Vhπk ) = 2HP πk ∆
Thus using Azuma’s inequality,
X
k,h
(Az)
a′k,h ≤ 2H
X
k,h
√
√
δbk,h+1 + H 2 T L ≤ 2H 2 U + H 2 T L,
def
where U is defined as an upper-bound on the pseudo regret: U =
the r.h.s. of Eq. 1).
P
k,h (bk,h
√
+ ek,h ) + H T (an upper bound on
P
πk
. (Dominant term)
Step 3’-b: bounding k,h
PVk,h+1
k
|Hk ] is the expected sum of variances of the value function Vkπ (y) at the next
For any episode k, E[ h Vπk,h+1
πk
state y ∼ P (·|xk,h ) under the true transition model for the current policy. A recursive application of the law of
total variance (see e.g., Munos & Moore, 1999; Azar et al., 2013; Lattimore & Hutter, P
2012) shows that this quantity
is nothing else than the variance of the return (sum of H rewards) under policy
π
:
V
k
h r(xk,h , πk (xk,h )) , which
P
k
by its expectation (see the
is thus bounded by H 2 . Finally, using Freedman’s (Fr) inequality to bound k,h Vπk,h+1
exact derivation in the appendix), we deduce
X
k,h
(F r)
k
Vπk,h+1
≤
i
X hX π
√
√
k
E
Vk,h+1
|Hk + H 2 T L ≤ T H + H 2 T L.
k
Thus, using Eq. 8, Eq. 7 and the bounds on
X
k,h
P
k,h
ek,h =
X
k,h
(Pbkπk
ak,h and
P
a′k,h , we deduce that
p
bk,h ≤ L (T H + H 2 U )SA.
Step 4’: Bounding the sum of estimation errors
tion errors
X
(8)
h
−P
πk
P
k,h ek,h .
∗
)Vh+1
(xk,h )
9
We now use Bernstein inequality to bound the estima-
≤
X
k,h
s
V∗k,h+1
nk,h
+
HL
,
nk,h
def
∗
where V∗k,h+1 = VY ∼P πk (·|xk,h ) Vh+1
(Y ) . Now,
similar way as in Step 3’ above, we relate V∗k,h+1 to
P in a very
πk
πk
Vk,h+1 and use the Law of total variance to bound k,h Vk,h+1 by HT and deduce that
X
k,h
p
ek,h ≤ L (T H + H 2 U )SA.
√
p
From Eq. 1 we see that U ≤ L (T H + H√2 U )SA thus U ≤ (L HSAT + H 2 SAL2 ). This implies Eq. 6.
So the reason we are able to remove the H factor from the regret bound comes from the fact that the sum, over
H steps, of the variances of the next state values (which define the amplitude of the confidence intervals) is at most
bounded by the variance of the return.
√ Intuitively this means that the size of the confidence intervals do not add up
linearly over H steps but grows as H only. Although the sequence of estimation errors are not independent over
time, we are able to demonstrate a concentration of measure phenomenon that shows that those estimation errors
concentrate as if they were independent.
6 Conclusion
In this paper we refine the familiar concept of optimism in the face of uncertainty. Our key contribution is the design
and analysis of the algorithm UCBVI-BF , which addresses two key shortcomings in existing algorithms for optimistic exploration in finite MDPs. First we apply
a concentration to the value as a whole, rather than the transition
√
estimates, this leads to a reduction from S to S. Next we apply a recursive law of total variance
√ to couple estimates
across an episode, rather than at each time step individually, this leads to a reduction from H to H.
Theorem 2 provides the first regret bounds which, for sufficiently large T , match the lower bounds for the problem
√
e HSAT ) up to logarithmic factors. It remains an open problem whether we can match the lower bound using this
O(
2
e 2 S 2 A) to O(HS
e
approach for small T . We believe that the higher order term can be improved from O(H
A) by a
more careful
analysis,
i.e.,
a
more
extensive
use
of
Freedman-Bernstein
inequalities.
The
same
applies
to
the
term of
√
√
order H T which can be improved to HT .
These results are particularly
√ significant because they help to estabilish the information-theoretic lower bound
of reinforcement learning at Ω( HSAT )√Osband & Van Roy (2016a), whereas it was suggested in some previous
work that lower-bound should be of Ω(H SAT ). Moving from this big-picture insight to an analytically rigorous
bound is non-trivial. Although we push many of the technical details to the appendix, our paper also makes several
contributions in terms of analytical tools that may be useful in subsequent work. In particular we believe that the way
we construct the exploration bonus and confidence intervals in UCBVI-CH is novel to the literature of RL. Also the
constructive approach in the proof of UCBVI-CH, which bootstraps the regret bounds to prove that Vk,h s are ucbs, is
another analytical contribution of this paper.
Acknowledgements
The authors would like to thank Marc Bellemare and all the other wonderful colleagues at DeepMind for many hours
of discussion and insight leading to this research. We are also grateful for the anonymous reviewers for their helpful
comments and for fixing several mistakes in an earlier version of this paper.
10
References
Agrawal, Shipra and Jia, Randy. Posterior sampling for reinforcement learning: worst-case regret bounds. arXiv
preprint arXiv:1705.07041, 2017.
Azar, Mohammad Gheshlaghi, Munos, Rémi, and Kappen, Hilbert J. Minimax pac bounds on the sample complexity
of reinforcement learning with a generative model. Machine learning, 91(3):325–349, 2013.
Bartlett, Peter L. and Tewari, Ambuj. REGAL: A regularization based algorithm for reinforcement learning in weakly
communicating MDPs. In Proceedings of the 25th Conference on Uncertainty in Artificial Intelligence (UAI2009),
pp. 35–42, June 2009.
Bernstein, S. Theory of probability, 1927.
Bertsekas, D. P. Dynamic Programming and Optimal Control, volume I. Athena Scientific, Belmount, Massachusetts,
third edition, 2007.
Bertsekas, D. P. and Tsitsiklis, J. N. Neuro-Dynamic Programming. Athena Scientific, Belmont, Massachusetts, 1996.
Brafman, Ronen I. and Tennenholtz, Moshe. R-max - a general polynomial time algorithm for near-optimal reinforcement learning. Journal of Machine Learning Research, 3:213–231, 2002.
Bubeck, Sébastien and Cesa-Bianchi, Nicolò. Regret analysis of stochastic and nonstochastic multi-armed bandit
problems. CoRR, abs/1204.5721, 2012. URL http://arxiv.org/abs/1204.5721.
Bubeck, Sébastien, Munos, Rémi, Stoltz, Gilles, and Szepesvári, Csaba. X-armed bandits. Journal of Machine
Learning Research, 12:1587–1627, 2011.
Burnetas, Apostolos N and Katehakis, Michael N. Optimal adaptive policies for markov decision processes. Mathematics of Operations Research, 22(1):222–255, 1997.
Cesa-Bianchi, N. and Lugosi, G. Prediction, Learning, and Games. Cambridge University Press, New York, NY,
USA, 2006.
Dann, Christoph and Brunskill, Emma. Sample complexity of episodic fixed-horizon reinforcement learning. In
Advances in Neural Information Processing Systems, 2015.
Dann, Christoph, Lattimore, Tor, and Brunskill, Emma. Ubev-a more practical algorithm for episodic rl with nearoptimal pac and regret guarantees. arXiv preprint arXiv:1703.07710, 2017.
Freedman, David A. On tail probabilities for martingales. the Annals of Probability, pp. 100–118, 1975.
Guez, Arthur, Silver, David, and Dayan, Peter. Scalable and efficient bayes-adaptive reinforcement learning based on
monte-carlo tree search. Journal of Artificial Intelligence Research, pp. 841–883, 2013.
Jaksch, T., Ortner, R., and Auer, P. Near-optimal regret bounds for reinforcement learning. Journal of Machine
Learning Research, 11:1563–1600, 2010.
Kearns, Michael J. and Singh, Satinder P. Near-optimal reinforcement learning in polynomial time. Machine Learning,
49(2-3):209–232, 2002.
Lattimore, Tor and Hutter, Marcus. PAC bounds for discounted MDPs. CoRR, abs/1202.3890, 2012.
Maurer, Andreas and Pontil, Massimiliano. Empirical bernstein bounds and sample variance penalization. stat, 1050:
21, 2009.
Munos, R. and Moore, A. Influence and variance of a Markov chain : Application to adaptive discretizations in optimal
control. In Proceedings of the 38th IEEE Conference on Decision and Control, 1999.
Munos, Rémi. From bandits to Monte-Carlo Tree Search: The optimistic principle applied to optimization and planning. Foundations and Trends R in Machine Learning, 7(1):1–129, 2014.
Osband, Ian and Van Roy, Benjamin. On lower bounds for regret in reinforcement learning. stat, 1050:9, 2016a.
11
Osband, Ian and Van Roy, Benjamin. Why is posterior sampling better than optimism for reinforcement learning.
arXiv preprint arXiv:1607.00215, 2016b.
Osband, Ian, Russo, Dan, and Van Roy, Benjamin. (more) efficient reinforcement learning via posterior sampling. In
Advances in Neural Information Processing Systems, pp. 3003–3011, 2013.
Strehl, Alexander L and Littman, Michael L. A theoretical analysis of model-based interval estimation. In Proceedings
of the 22nd international conference on Machine learning, pp. 856–863. ACM, 2005.
Strehl, Alexander L and Littman, Michael L. An analysis of model-based interval estimation for markov decision
processes. Journal of Computer and System Sciences, 74(8):1309–1331, 2008.
Strehl, Alexander L., Li, Lihong, Wiewiora, Eric, Langford, John, and Littman, Michael L. PAC model-free reinforcement learning. In ICML, pp. 881–888, 2006.
Strens, Malcolm J. A. A Bayesian framework for reinforcement learning. In ICML, pp. 943–950, 2000.
Sutton, Richard and Barto, Andrew. Reinforcement Learning: An Introduction. MIT Press, March 1998.
Thompson, W.R. On the likelihood that one unknown probability exceeds another in view of the evidence of two
samples. Biometrika, 25(3/4):285–294, 1933.
Weissman, Tsachy, Ordentlich, Erik, Seroussi, Gadiel, Verdu, Sergio, and Weinberger, Marcelo J. Inequalities for the
l1 deviation of the empirical distribution. Hewlett-Packard Labs, Tech. Rep, 2003.
12
Appendices
We begin by introducing some notation in Sect. B and Sect. A. We then provide the full analysis of UCBVI in Sect. C.
A
Table of Notation
Symbol
S
A
πk
P
R
S
A
H
T and Tk
K
Nk (x, a)
Vh∗
T
Vk,h
Qk,h
b
L
Nk (x, a, y)
′
Nk,h
(x, a)
′
Nk,h
(x)
b
Pk (y|x, a)
b k,h (x, a)
V
V∗h (x, a)
b ∗ (x, a)
V
k,h
Vπh (x, a)
b′i,j (x)
[(x, a)]k
[k]typ
[y]k,x,a
Regret(k)
^
Regret(k)
Regret(k, x, h)
^
Regret(k,
x, h)
∆k,h
e k,h
∆
e
∆typ,k,h
Mt
εk,h and ε̄k,h
c1 (v, n), c2 (p, n) and c3 (n)
Ck,h
Bk,h
E
Ω
Ht
Explanation
The state space
The action space
The policy at episode k
The transition distribution
The reward function
Size of state space
Size of action space
The horizon length
The total number of steps and number of steps up to episode k
The total number of episodes
Number of visits to state-action pair (x, a) up to episode k
Optimal value function V ∗
Bellman operator
The estimate of value function at step h of episode k
The estimate of action value function at step h of episode k
The exploration bonus
ln(5SAT /δ)
Number of transitions from x to y upon taking action a up to episode k
Number of visits to state-action pair (x, a) at step h up to episode k
Number of visits to state x at step h up to episode k
The empirical transition distribution from x to y upon taking action a up to episode k
The empirical next-state variance of Vk,h for every (x, a)
The next-state variance of V ∗ for every state-action pair (x, a)
The empirical next-state variance of Vh∗ for every state-action pair (x, a) at episode k
π
The next-state
variance of
Vh for every state-action pair (x, a)
min
1002 S 2 H 2 AL2
, H2
′ (x)
Ni,j
Set of typical state-action pairs
Set of typical episodes
Set of typical next states at every episode k for every (x, a)
The regret after k episodes
The upper-bound regret after k episodes
The regret upon encountering state x at step h after k episodes
The regret upon encountering state x at step h after k episodes
One step regret at step h of episode k
One step upper-bound regret at step h of episode k
One step upper-bound regret at step h of episode k for typical episodes
The martingale operator
Martingale difference terms
The confidence intervals for the value function and transition distribution
Sum of confidence intervals c1 up to step h of episode k
Sum of exploration bonuses b up to step h of episode k
The high probability event under which the concentration inequality holds
The high probability event under which the estimates Vk,h are ucbs
The history of all random events up to time step t
13
B Notation
Let denote the total number of times that we visit state
Px while′ taking action a at step h of all episodes up to episode k
′
′
(x, a) for the total number of visits to state x at time
by Nk,h
(x, a). We also use the notation Nk,h
(x) = a∈A Nk,h
b k,h (x, a), the next-state variance of optimal value
step h up to episode k. Also define the empirical next-state variance V
∗
b ∗ (x, a) and the next-state variance of
function Vh (x, a) the next-state empirical variance of optimal value function V
k,h
π
Vh as
b k,h (x, a)
V
def
=
Vary∼Pbk (·|x,a) (Vk,h+1 (y)),
V∗h (x, a)
def
=
Vary∼P (·|x,a) (Vh∗ (y)),
b ∗ (x, a)
V
k,h
def
=
Vary∼Pbk (·|x,a) (Vh∗ (y)),
def
Vary∼P (·|x,a) (Vhπ (y)).
Vπh (x, a)
=
for every (x, a) ∈ S × A and k ∈ [K] and h ∈ [H]. We further introduce some short-hand notation: we use the lower
case to denote the functions evaluated at the current state-action pair, e.g., we write nk,h for Nk (xk,h , πk (xk,h , h)) and
k
k
vk,h for Vk,h (xk,h ). Let also denote V∗k,h = V∗k,h (xk,h , πk (xk,h , h)) and Vπk,h
= Vπk,h
(xk,h , πk (xk,h , h)) for every
2 2 2 2
100 S H AL
2
′
for every x ∈ S.
k ∈ [K] and h ∈ [H]. Also define bi,j (x) = min
N′
(x) , H
i,j+1
B.1
“Typical” state-actions and steps
In our analysis we split the episodes into 2 sets: the set of “typical” episodes in which the number of visits to the
encountered state-actions are large and the rest of the episodes. We then prove a tight regret bound for the typical
episodes. As the total count of other episodes is bounded this technique provides us with the desired result. The set of
typical state-actions pairs for every episode k is defined as follows
[(x, a)]k
def
=
′
{(x, a) : (x, a) ∈ S × A, Nh (x, a) ≥ H, Nk,h
(x) ≥ H}.
Based on the definition of [(x, a)]typ we define the set of typical episodes and the set of typical state-dependent
episodes as follow
[k]typ
def
[k]typ,x
def
=
=
{i : i ∈ [k], ∀h ∈ [H], (xi,h , πi (xi,h , h)) ∈ [(x, a)]k , i ≥ 250HS 2AL},
′
{i : i ∈ [k], ∀h ∈ [H], (xi,h , πi (xi,h , h)) ∈ [(x, a)]k , Nk,h
(x) ≥ 250HS 2AL}.
Also for every (x, a) ∈ S × A the set of typical next states at every episode k is defined as follows
[y]k,x,a
def
=
{y : y ∈ S, Nk (x, a)P (y|x, a) ≥ 2H 2 L}.
Finally let denote [y]k,h = [y]k,xk,h ,πk (xxk,h ) for every k ∈ [K] and h ∈ [H].
B.2
Surrogate regrets
Our ultimate goal is to prove bound on the regret Regret(k). However in our analysis we mostly focus on bounding
e k,h (x) def
the surrogate regrets. Let ∆
= Vk,h (x)−V πk (x) for every x ∈ S, h ∈ [H] and k ∈ [K]. Then the upper-bound
^ defined as follows
regret Regret
h
^
Regret(k)
def
=
k
X
i=1
δei,1 .
^
Regret(k)
is useful in our analysis since it provides an upperbound on the true regret Regret(k). So one can
^
bound Regret(k) as a surrogate for Regret(k).
14
We also define the corresponding per state-step regret and upper-bound regret for every state x ∈ X and step
h ∈ [H], respectively, as follows
Regret(k, x, h)
k
X
def
=
I(xi,h = x)δi,h ,
i=1
^
Regret(k,
x, h)
k
X
def
=
i=1
B.3
I(xi,h = x)δei,h .
Martingale difference sequences
In our analysis we rely heavily on the theory of martingale sequences to prove bound on the regret incurred due to
encountering a random sequence of states. We now provide some definitions and notation in that regard.
We define the following martingale operator for every k ∈ [K], h ∈ [H] and F : S → ℜ. Also let t = (k−1)H +h
denote the time stamp at step h of episode k then
Mt F
def
=
Phπk F − F (xk,h+1 ).
Let Ht be the history of all random events up to (and including) step h of episode k then we have that E(Mt F |Ht ) =
0. Thus Mt F is a martingale difference w.r.t. Ht . Also let G be a real-value function depends on Ht+s for some
integer s > 0. Then we generalize our definition of operator Mt as
Mt G
def
=
E ( G(Ht+s )| Ht ) − G(Ht+s ),
where E is over the randomization of the sequence of states generated by the sequence of policies πk , πk+1 , . . . . Here
also Mt G is a martingale difference w.r.t. Ht .
Let define ∆typ,k,h : S → ℜ as follows for every k ∈ [K] and h ∈ [H] and y ∈ S
e typ,k,h+1 (y)
∆
def
=
s
Ik,h (y) e
∆k,h+1 (y),
nk,h pk,h (y)
where the function pk,h : S → [0, 1] is defined as pk,h (y) = Phπk (y|xk,h ) and Ik,h (y) writes for Ik,h (y) = I(y ∈
[y]k,h ) for every y ∈ X . We also define the following martingale differences which we use frequently
B.4
εk,h
def
ε̄k,h
def
e k,h+1 ,
Mt ∆
=
e typ,k,h+1 .
Mt ∆
=
High probability events
We now introduce the high probability events E and Ωk,h under which the regret is small.
def
Let use the shorthand notation L = ln 5SAT
. Also for every v > 0, p ∈ [0, 1] and n > 0 let define the
δ
confidence intervals c1 , c2 and c3 , respectively, as follow
c1 (v, n)
def
c2 (p, n)
def
c3 (n)
def
=
=
=
2
r
r
vL 14HL
+
,
n
3n
p(1 − p)L 2L
+
,
n
3n
r
SL
2
.
n
2
15
Let P be the set of all probability distributions on S. Define the following confidence set for every k = 1, . . . , K,
n > 0 and (x, a) ∈ S × A
P(k, h, n, x, a, y)
def
=
n
b ∗ (x, a), n
Pe (·|x, a) ∈ P : |(Pe − P )Vh∗ (x, a)| ≤ min c1 (V∗h (x, a), n) , c1 V
k,h
|Pe (y|x, a) − P (y|x, a)| ≤ c2 (P (y|x, a), n) ,
o
kPe(·|x, a) − P (·|x, a)k1 ≤ c3 (n) .
We now define the random event EPb as follows
EPb
def
=
n
o
Pbk (y|x, a) ∈ P(k, h, Nk (x, a), x, a, y), ∀k ∈ [K], ∀h ∈ [H], ∀(y, x, a) ∈ S × S × A .
Let t be a positive integer. Let F = {fs }s∈[t] be a set of real-value functions on Ht+s , for some integer s > 0. We
now define the following random events for every w̄ > 0 and ū > 0 and c̄ > 0:
Eaz (F , ū, c̄)
Efr (F , w̄, ū, c̄)
def
=
def
=
(
(
t
X
s=1
t
X
)
√
Ms fs ≤ 2 tū2 c̄ ,
√
14ūc̄
Ms fs ≤ 4 w̄c +
3
s=1
)
.
We also use the short-hand notation Eaz (F , ū) and Efr (F , w̄, ū) for Eaz (F , ū, L) and Efr (F , w̄, ū, L), respectively.
Now let define the following sets of random variables for every k ∈ [K] and h ∈ [H]:
F∆,k,h
e
def
′
F∆,k,h
e
def
F∆,k,h,x
e
def
′
F∆,k,h,x
e
def
GV,k,h
def
GV,k,h,x
def
n
o
e i,j : i ∈ [k], h < j ∈ [H − 1] ,
∆
n
o
e typ,i,j : i ∈ [k], h < j ∈ [H] ,
∆
o
n
e i,j I(xi,h = x) : i ∈ [k], h < j ∈ [H] ,
∆
n
o
e typ,i,j I(xi,h = x) : i ∈ [k], h < j ∈ [H] ,
∆
H
X
Vπj i : i ∈ [k], h < j ∈ [H] ,
j=h+1
H
X
Vπj i I(xi,h = x) : i ∈ [k], h < j ∈ [H] ,
=
=
=
=
=
=
j=h+1
Fb′ ,k,h
def
Fb′ ,k,h,x
=
=
def
b′i,j : i ∈ [k], h < j ∈ [H − 1] ,
b′i,j I(xi,h = x) : i ∈ [k], h < j ∈ [H] .
We now define the high probability event E as follows
E
def
=
EPb
\
\ \
k∈[K]
h∈[H]
x∈S
Eaz (F∆,k,h
, H)
e
4
3
Efr (GV,k,h , H T, H )
\
\
\
√
√ \
′
′
, H) Eaz (F∆,k,x,h
, 1/ L)
Eaz (F∆,k,h
, 1/ L) Eaz (F∆,k,h,x
e
e
e
Eaz (GV,k,h,x , H
5
′
Nk,h
(x), H 3 )
\
Eaz (Fb′ ,k,h , H )
The following lemma shows that the event E holds with high probability:
Lemma 1. Let δ > 0 be a real scalar. Then the event E holds w.p. at least 1 − δ.
16
2
\
Eaz (Fb′ ,k,h,x , H ) .
2
Proof. To prove this result we need to show that a set of concentration inequalities with regard to the empirical model
Pbk holds simultaneously. For every h ∈ [H] the Bernstein inequality combined with a union bound argument, to
take into account that Nk (x, a) ∈ [T ] is a random number, leads to the following inequality w.p. 1 − δ (see, e.g.,
Cesa-Bianchi & Lugosi, 2006; Bubeck & Cesa-Bianchi, 2012, for the statement of the Bernstein inequality and the
application of the union bound in similar cases, respectively.)
i
h
(P − Pbk )Vh∗ (x, a)
s
≤
2V∗h (x, a) ln
Nk (x, a)
2T
δ
2H ln 2T
δ
+
,
3Nk (x, a)
(9)
where we rely on the fact that Vh∗ is uniformly bounded by H. Using the same argument but this time with the
Empirical Bernstein inequality (see, e.g., Maurer & Pontil, 2009), for Nk (x, a) > 1, leads to
i
h
(P − Pbk )Vh∗ (x, a)
≤
s
b ∗ (x, a) ln
2V
k,h
2T
δ
Nk (x, a)
7H ln 2T
δ
+
.
3Nk (x, a)
(10)
The Bernstein inequality combined with a union bound argument on Nk (x, a) also implies the following bound
w.p. 1 − δ
|Nk (y, x, a) − Nk (x, a)P (y|x, a)| ≤
s
2Nk (x, a)Varz∼P (·|x,a) (1(z = y)) ln
2T
δ
+
2T
δ
2 ln
3
,
which implies the following bound w.p. 1 − δ:
Pbk (y|x, a) − P (y|x, a)
s
≤
P (y|x, a)(1 − P (y|x, a)) ln
Nk (x, a)
2T
δ
2 ln 2T
δ
+
.
3Nk (x, a)
(11)
A similar result holds on ℓ1 -normed estimation error of the transition distribution. The result of (Weissman et al.,
2003) combined with a union bound on Nk (x, a) ∈ [T ] implies w.p. 1 − δ
Pbk (·|x, a) − P (·|x, a)
1
≤
s
2S ln 2T
δ
.
Nk (x, a)
(12)
We now focus on bounding the sequence of martingales. Let n > 0 be an integer and u, δ > 0 be some real
scalars. Let the sequence of random variables {X1 , X2 , . . . , Xn } be a sequence of martingale differences w.r.t. to
some filtration Fn . Let this sequence be uniformly bounded from above and below by u. Then the Azuma’s inequality
(see, e.g., Cesa-Bianchi & Lugosi, 2006) implies that w.p. 1 − δ
n
X
Xi
i=1
When the sum of the variances
Freedman (1975) holds w.p. 1 − δ
Pn
i=1
n
X
i=1
≤
s
2nu ln
1
.
δ
(13)
Var(Xi |Fi ) ≤ w for some w > 0 then the following sharper bound due to
Xi
≤
s
2u ln
1
+
2w ln
δ
3
1
δ
.
(14)
Let k ∈ [K], h ∈ [H] and x ∈ X . Then the inequality of Eq. 13 immediately implies that the following events
holds w.p. 1 − δ:
, H, ln (1/δ) ,
Eaz F∆,k,h
e
√
′
,
1/
L,
ln
(1/δ)
,
Eaz F∆,k,h
e
Eaz Fb′ ,k,h , H 2 , ln (1/δ) .
17
(15)
(16)
(17)
′
Also Eq. 13 combined with a union bound argument over all Nk,h
(x) ∈ [T ] (see, e.g., Bubeck et al., 2011, for
the full description of the application of union bound argument in the case of martigale process with random stopping
time) implies that the following events hold w.p. 1 − δ
Eaz F∆,k,h,x
, H, ln (T /δ) ,
e
√
′
,
1/
Eaz F∆,k,h,x
L,
ln
(T
/δ)
,
e
Eaz Fb′ ,k,h,x , H 2 , ln (T /δ) .
(18)
(19)
(20)
Similarly the inequality of Eq. 14 leads to the following events hold w.p. 1 − δ
Efr GV,k,h , w̄k,h , , H 3 , ln (T /δ) ,
Efr
(21)
GV,k,h,x , w̄k,h,x , , H , ln (1/δ) ,
3
(22)
where w̄k,h and w̄k,h,x are upper bounds on Wk,h and Wk,h,x , respectively, defined as
def
=
Wk,h
k
X
i=1
def
=
Wk,h,x
k
X
i=1
Var
H−1
X
j=h
Vπi,j+1 Hi,1 ,
I(xi,h = x)E
H−1
X
j=h
(23)
Vπi,j+1 Hi,1 .
(24)
So to establish a value for w̄k,h and w̄k,h,x we need to prove bound on Wk,h and Wk,h,x . Here we only prove this
bound for Wk,h as the proof techniques to bound Wk,h,x is identical to the way we bound Wk,h .
Wk,h ≤
k
X
i=1
E
H−1
X
j=h
2
k
Hk ≤ H 3
Vπi,j+1
k
X
i=1
E
H−1
X
j=h
k
Hk .
Vπi,j+1
(25)
Now let the sequence {x1 , x2 , . . . , xH } be the sequence of states encountered by following some policy π throughout an episode k. Then the recursive application of LTV leads to (see e.g., Munos & Moore, 1999; Lattimore & Hutter,
2012, for the proof.)
E
H−1
X
j=h
Vπ (xj , π(xj , j))
=
By combining Eq. 26 into Eq. 25 we deduce
Wk,h
≤ H3
k
X
i=1
Var
H−1
X
j=h
Similarly the following bound holds on Wk,h,x
Wk,h,x
≤
Var
H−1
X
j=h
rπ (xj ) .
(26)
rk,h Hk ≤ H 5 k = H 4 Tk .
(27)
H 5 Nk,h (x).
(28)
Plugging the bounds of Eq. 27 and Eq. 28 in to the bounds of Eq. 21 and Eq. 22 and a union bound over all
Nk,h (x) ∈ [T ] leads to the following events hold w.p. 1 − δ:
Efr GV,k,h , H 4 T, H 3 , ln (1/δ) ,
Efr
GV,k,h,x , H Nk,h (x), H , ln (T /δ) .
5
3
(29)
(30)
Combining the results of Eq. 9, Eq. 10, Eq. 11, Eq. 12, Eq. 15, Eq. 16 Eq. 17, Eq. 18, Eq. 19, Eq. 20, Eq. 29 and
Eq. 30 and taking a union bound over these random events as well as all possible k ∈ [K], h ∈ [H] and (s, a) ∈ S × A
proves the result.
18
B.4.1 UCB Events
Let k ∈ [K] and h ∈ [H]. Denote the set of steps for which the value functions are obtained before Vk,h as
[k, h]hist = {(i, j) : i ∈ [K], j ∈ [H], i < k ∨ (i = k ∧ j > h)}.
Let Ωk,h = {Vi,j ≥ Vh∗ , ∀(i, j) ∈ [k, h]hist } be the event under which Vi,j prior to Vk,h computation are upper
bounds on the optimal value functions. Using backward induction on h (and standard concentration inequalities) we
will prove that Ωk,h holds under the event E (see Lem. 19).
B.5
Other useful notation
Here we define some other notation that we use throughout the proof. We denote the total count of steps up to episode
def
k ∈ [K] by Tk = H(k − 1). We first define c4,k,h , for every h ∈ [H] and k ∈ [K], as follow
c4,k,h =
4H 2 SAL
.
nk,h
for every k ∈ [K] , h ∈ [H] and x ∈ [x] we also introduce the following notation which we use later when we sum
up the regret:
Ck,h
def
=
k
X
i=1
Bk,h
def
=
k
X
i=1
Ck,h,x
def
=
k
X
i=1
Bk,h,x
def
=
k
X
i=1
I(i ∈ [k]typ )
I(i ∈ [k]typ )
H−1
X
c1,i,j ,
j=h
H−1
X
bi,j ,
j=h
I(i ∈ [k]typ,x , xk,h = x)
I(i ∈ [k]typ,x , xk,h = x)
H−1
X
c1,i,j ,
j=h
H−1
X
bi,j ,
j=h
∗
where c1,k,h is the shorthand-notation for c1 (vk,h
, nk,h ). We also define the upper bound Uk,h and Uk,h,x for every
k ∈ [K] , h ∈ [H] and x ∈ S as follows, respectively
Uk,h
def
=
e
k H−1
X
X
i=1 j=h
Uk,h,x
def
=
e
k H−1
X
X
p
[bi,j + c1,i,j + c4,i,j ] + (H + 1) Tk L,
[bi,j + c1,i,j + c4,i,j ] + (H + 1)3/2
i=1 j=h
C
q
′ (x)L,
Nk,h
Proof of the Regret Bounds
Before we start the main analysis we state the following useful lemma that will be used frequently in the analysis:
Lemma 2. let X ∈ R and Y ∈ R be two random variables. Then following bound holds for their variances
Var(X) ≤ 2 [Var(Y ) + Var(X − Y )] .
Proof. The following sequence of inequalities hold
Var(X) = E(X − Y − E(X − Y ) + Y − E(Y ))2 ≤ 2E(X − Y − E(X − Y ))2 + 2E(Y − E(Y ))2 .
The result follows from the definition of variance.
19
We proceed by proving the following key lemma which shows that proves bound on ∆k,h under the assumption
that Vk,h is UCB w.r.t. Vh∗ .
Lemma 3. Let k ∈ [K] and h ∈ [H]. Let the events E and Ωk,h hold. Then the following bound holds on δk,h and
δek,h :
δk,h ≤ δek,h ≤ e
H−1
Xh
i=h
i
√
εk,i + 2 Lε̄k,i + c1,k,i + bk,i + c4,k,i .
(31)
Proof. For the ease of exposition we abuse the notation and drop the dependencies on k, e.g., we write x1 , π and V1
for xk,1 , πk and Vk,1 , respectively. We proceed by bounding δeh under the event E at every step 0 < h < H:
δeh
π
= Th Vh+1 (xh ) − Thπ Vh+1
(xh )
π
π
b
= [Ph Vh+1 ](xh ) + bh − [Phπ Vh+1
](xh )
∗
∗
π
= bh + [(Pbhπ − Phπ )Vh+1
](xh ) + [(Pbhπ − Phπ )(Vh+1 − Vh+1
)](xh ) + [Phπ (Vh+1 − Vh+1
)](xh )
π
π
∗
≤ δeh+1 + εh + bh + c1,h + [(Pbh − Ph )(Vh+1 − Vh+1 )](xh ),
|
{z
}
(32)
(a)
∗
](xh ) ≤ c1,h . We
where the last inequality follows from the fact that under the event E we have that [(Pbhπ − Phπ )Vh+1
now bound (a):
(a)
X
=
y∈S
∗
(Pbhπ (y|xh ) − Phπ (y|xh ))(Vh+1 (y) − Vh+1
(y))
4L
p
(y)(1
−
p
(y))L
h
h
∆h+1 (y)
2
+
nh
3nh
y∈S
s
√ X ph (y)
e h+1 (y) + 4SHL ,
2 L
∆
nh
3nh
y∈S
{z
}
|
X
(I)
≤
≤
s
(b)
where (I) holds under the event E. We proceed by bounding (b):
(b)
=
X
y∈[y]h
s
|
(c)
=
≤ ε̄h +
s
Phπ (y|xh )
y∈[y]h
r
s
(c)
The term (c) can be bounded as follows
X
X
ph (y) e
∆h+1 (y) +
nh
y ∈[y]
/ h
{z
} |
1
e h+1 (y) = ε̄h +
∆
nh ph (y)
1 e
δh+1 ,
4LH 2
s
ph (y) e
∆h+1 (y) .
nh
{z
}
(33)
(d)
1
I(xh+1 ∈ [y]h )δeh+1
nh ph (xh+1 )
(34)
where in the last line we rely on the definition of [y]h . We now bound (d):
s
√
X
ph (y)nh e
SH 4LH 2
(d) =
∆
(y)
≤
.
h+1
n2h
nh
y ∈[y]
/ h
20
(35)
By combining Eq. 34 and Eq. 35 into Eq. 33 we deduce
r
√
SH 4LH 2
1 e
(b) ≤
+
δh+1 + ε̄h .
nh
4LH 2
(36)
By combining Eq. 36 and Eq. 33 into Eq. C we deduce
√
1 e
e
δh+1 .
δh ≤ εh + 2 Lε̄h + bh + c1,h + c4,h + 1 +
H
Let denote γh = (1 + 1/H)h . The previous bound combined with an induction argument implies that
δeh
≤
PH−1
i=h
h
i
√
γi−h εi + 2 Lε̄i + c1,i + c4,i + bi .
The inequality ln(1 + x) ≤ x for every x > −1 leads to γh ≤ γH ≤ e for every h ∈ [H]. This combined with the
assumption that vh ≥ vh∗ under the event Ωh completes the proof.
Lemma 4. Let k ∈ [k] and h ∈ [H]. Let the events E and Ωk,h hold. Then
k−1
X
i=1
δi,h ≤
k−1
X
δei,h ≤ e
i=1
k−1 H−1
i
X
Xh
√
εi,j + 2 Lε̄i,j + bi,j + c1,i,j + c4,i,j .
i=1 j=h
Proof. The proof follows by summing up the bounds of Lem. 3 and taking into acoount the fact if Ωk,h holds then
Ωi,j for all (i, j) ∈ [k, h]hist hold.
To simplify the bound of Lem. 4 we prove bound on sum of the martingales εk,h and ε̄k,h
Lemma 5. Let k ∈ [k] and h ∈ [H]. Let the events E and Ωk,h hold. Then the following bound holds
k H−1
X
X
≤
H
ε̄i,j
≤
p
p
(H − h)k ≤ Tk .
i=1 j=h
k H−1
X
X
p
p
(H − h)kL ≤ H Tk L,
εi,j
i=1 j=h
(37)
(38)
Also the following bounds holds for every x ∈ X and h ∈ H:
k
X
I(xi,h = x)
i=1
k
X
i=1
H−1
X
≤
H
ε̄i,j
≤
q
′ (x).
(H − h)Nk,h
j=h
I(xi,h = x)
H−1
X
j=h
q
′ (x)L,
(H − h)Nk,h
εi,j
(39)
(40)
′
Proof. The fact that the event E holds implies that the events Eaz (F∆,k,h
, H), Eaz (F∆,k,h
, √1L ) , Eaz (F∆,k,h,x
, H) and
e
e
e
′
Eaz (F∆,x,k,h
, √1L ) hold. Under these events the inequalities of the statement hold. This combined with the fact that
e
(H − h)k ≤ Tk completes the proof.
We now bound the sum of δs in terms of the upper-bound U :
Lemma 6. Let k ∈ [K] and h ∈ [H]. Let the events E and Ωk,h holds. Then the following bounds hold for every
h ∈ [H] x ∈ S
21
k
X
i=1
k
X
k
X
≤
δi,h
I(xi,h = x)δi,h
i=1
k
X
≤
i=1
i=1
δei,h ≤ Uk,h ≤ Uk,1 ,
I(xi,h = x)δei,h ≤ Uk,h,x . ≤ Uk,1,x .
Proof. The proof follows by incorporating the result of Lem. 5 into Lem. 4 and taking into account that for every
h ∈ [H] the term Uk,h (Uk,h,x ) is a summation of non-negative terms which are also contained in Uk,1 (U1,h,x).
Lemma 7. Let k ∈ [K] and h ∈ [H]. Let the events E and Ωk,h holds. Then the following bounds hold for every
x∈S
k X
H
X
≤
δi,j
i=1 j=h
k
X
I(xi,h = x)
i=1
H
X
≤
δi,j
j=h
k X
H
X
i=1 j=h
k
X
δei,j ≤ HUk,1 ,
I(xi,h = x)
i=1
H
X
j=h
Proof. The proof follows by summing up the bounds of Lem. 6.
δei,j ≤ HUk,1,x .
We now focus on bounding the terms Ck,h (Ck,h,x ) and Bk,h (Bk,h,x ) in Lem. 11 and Lem. 12, respectively.
Before we proceed with the proof of Lem. 11 and Lem. 12. we prove the following key result which bounds sum of
π
the variances of Vk,h
using an LTV argument:
Lemma 8. Let k ∈ [K] and h ∈ [H]. Then under the events E and Ωk,h the following hold for every x ∈ S
k H−1
X
X
p
4H 3 L
≤ Tk H + 2 H 4 Tk L +
,
3
Vπi,j+1
i=1 j=h
k
X
I(xi,h = x)
i=1
H−1
X
q
3
′
′ (x)L + 4H L .
≤ Nk,h
(x)H 2 + 2 H 5 Nk,h
3
Vπi,j+1
j=h
Proof. Under E the events Efr (GV,k,h , H 4 Tk , H 3 ) and Efr (GV,k,h,x , H 5 Nk,h (x), H 3 ) hold which then imply:
k H−1
X
X
Vπi,j+1
i=1 j=h
k
X
I(xi,h = x)
i=1
H−1
X
Vπi,j+1
j=h
≤
k
X
≤
k
X
i=1
i=1
The LTV argument of Eq. 26 then leads to
k
X
i=1
k
X
i=1
E
I(xi,h = x)E
H−1
X
j=h
H−1
X
j=h
E
H−1
X
j=h
Vπi,j+1 Hk,h + 2
I(xi,h = x)E
Vπi,j+1 Hi,h
=
Vπi,j+1 Hi,h
=
k
X
i=1
k
X
i=1
H−1
X
Vπi,j+1
j=h
Var
H
X
j=h+1
p
4H 3 L
,
H 4 Tk L +
3
q
3
′ L + 4H L .
Hk,h + 2 H 5 Nk,h
3
π
≤ KH 2 = T H,
ri,j
I(xi,h = x)Var
H
X
j=h+1
π
′
ri,j
≤ Nk,h
(x)H 2 .
Eq. 41 and Eq. 42 combined with Eq. 43 and Eq. 44, respectively, complete the proof.
22
(41)
(42)
(43)
(44)
Lemma 9. Let k ∈ [K] and h ∈ [H]. Then under the events E and Ωk,h the following hold for every x ∈ S
k H−1
X
X
i=1 j=h
k
X
I(xi,h = x)
i=1
H−1
X
j=h
V∗i,j+1 − Vπi,j+1
V∗i,j+1 − Vπi,j+1
p
Tk L,
≤
2H 2 Uk,h + 4H 2
≤
2H 2 Uk,h,x + 4H 2
(45)
q
′ (x, a)L.
HNk,h
(46)
Proof. We begin by the following sequence of inequalities:
k H−1
X
X
i=1 j=h
V∗i,j+1 − Vπi,j+1
(I)
≤
=
k H−1
X
X
i=1 j=h
k H−1
X
X
i=1 j=h
≤
Ey∼pi,j
2H
h
2 i
2
πi
∗
(y)
Vi,j+1
(y) − Vi,j+1
∗
πi
πi
∗
Ey∼pi,j (Vj+1
(y))
(y)(Vj+1
(y) + Vj+1
(y) − Vj+1
k H−1
X
X
i=1 j=h
|
πi
∗
Ey∼pi,j Vj+1
(y) ,
(y) − Vj+1
{z
(47)
}
(a)
∗
π
where (I) is obtained from the definition of the variance as well as the fact that Vi,j
≥ Vk,h
. The last line also follows
πk
∗
from the fact that V ≤ Vh ≤ H.
Using an identical argument we can also prove the following bound for state-dependent difference:
k
X
I(xi,h = x)
i=1
H−1
X
j=h
V∗i,j+1 − Vπi,j+1
≤
2H
k
X
I(xi,h = x)
i=1
H−1
X
j=h
|
πi
∗
Ey∼pi,j Vj+1
(y) ,
(y) − Vj+1
{z
(b)
(48)
}
To bound (a) we use the fact that under the event E the event Eaz (F∆,k,h
, H) also holds. This combined with the
e
e
fact that under the event Ωk,h the inequality δk,h ≤ δk,h holds implies that
(a)
≤
k H−1
X
X
i=1 j=h
δei,j+1 + 2H
≤ HU1,h + 2H
p
Tk L,
p
Tk L
(49)
where in the last line we rely on the result of Lem. 7. Similarly we can prove the following bound for (b) under the
events Ωk,h and Eaz (F∆,k,h,x
, H):
e
(b) ≤
k
X
I(xi,h = x)
i=1
H−1
X
j=h
≤ HUk,h,x + 2H 1.5
e i,j+1 + 2H 1.5
∆
q
′ (x)L.
Nk,h
q
′ (x)L
Nk,h
(50)
The result then follows by incorporating the results of Eq. 49 and Eq. 50 into Eq. 47 and Eq. 48, respectively.
Lemma 10. Let k ∈ [K] and h ∈ [H]. Then under the events E and Ωk,h the following hold for every x ∈ S
k H−1
X
X
i=1 j=h
k
X
i=1
I(xi,h = x)
H−1
X
j=h
b i,j+1 − Vπ
V
i,j+1
b i,j+1 − Vπ
V
i,j+1
≤ 2H 2 Uk,1 + 15H 2 S
p
ATk L,
≤ 2H 2 Uk,h,x + 15H 2 S
23
q
′ (x)L.
HANk,h
(51)
(52)
Proof. Here we only prove the bound on Eq. 51. The proof for the bound of Eq. 52 can be done in a very similar
′
manner, as it is shown in the previous lemmas (the only difference is that HNk,h
(x) and Uk,h,x replace Tk and Uk,1 ,
respectively). The following sequence of inequalities hold:
k H−1
X
X
i=1 j=h
b i,j+1 − Vπ
V
i,j+1
k H−1
X
X
(I)
≤
i=1 j=h
+
2
2
πi
Ey∼b
pi,j (Vi,j+1 (y)) − Ey∼pi,j Vj+1 (y)
k H−1
X
X
i=1 j=h
k H−1
X
X
(II)
≤
i=1 j=h
|
2
2
∗
∗
(y) − Ey∼b
Ey∼pi,j Vj+1
pi,j Vj+1 (y)
2
Ey∼b
pi,j (Vi,j+1 (y)) −
{z
H−1
X
Ey∼pi,j (Vi,j+1 (y))
j=1
(a)
+
k H−1
X
X
i=1 j=h
+
|
k H−1
X
X
i=1 j=h
|
h
2 i
2
πi
(y)
Ey∼pi,j (Vi,j+1 (y)) − Vj+1
4H 2
{z
s
(c)
{z
(b)
L
,
nk,h
}
2
}
}
(53)
where (I) holds due to the fact that under Ωk,h , Vi,j ≥ Vj∗ ≥ Vjπi and (II) holds under the event E.
We now bound (a):
(a)
(I)
≤
(II)
≤
k H−1
X
X
2H
i=1 j=h
3H 2 S
2
s
SL
nk,h
p
ATk L,
where (I) holds under the event E and (II) holds due to the pigeon-hole argument (see, e.g., Jaksch et al., 2010, for
the proof).
Using an identical analysis to the one in Lem. 10 and taking into account that Vi,j ≥ Vj∗ under the event Ωk,h and
E we can bound (b)
(I)
(b) ≤ 2H
k H−1
X
X
i=1 j=h
p
δei,j+1 + 2H Tk L)
p
≤ 2H 2 Uk,1 + 2 Tk L ,
where (I) holds since under the event E the event Eaz (F∆,k,h
, H) holds. Another application of pigeon-hole principle
e
√
leads to a bound of 6H 2 SATk L on (c). We then combine this with the bounds on (a) and (b) to bound Eq. 53,
which proves the result.
We now bound Ck,h and Ck,h,x :
Lemma 11. Let k ∈ [K] and h ∈ [H]. Then under the events E and Ωk,h the following hold for every x ∈ S
Ck,h
≤
Ck,h,x
≤
q
p
4 HSATk + 4 H 2 Uk,1 SAL2 ,
q
q
4 H 2 SANk,h (x) + 4 H 2 Uk,h,x SAL2 .
24
(54)
(55)
Proof. Here we only prove the bound on Eq. 54. The proof for the bound of Eq. 55 can be done in a very similar
′
manner, as it is shown in the previous lemmas (the only difference is that HNk,h
(x) and Uk,h,x replace Tk and Uk,1 ,
respectively). The Cauchy–Schwarz inequality leads to the following sequence of inequalities:
Ck,h
=
k
X
i=1
≤
H−1
X
I(j ∈ [k]typ )
2
j=h
s
V∗i,j+1 L 4HL
+
ni,j
3ni,j
v
v
u k
u k H−1
H−1
H−1
k
X 1
X
X 4HL
X
uX
√ uX
∗
u
u
I(i ∈ [k]typ )
Vi,j+1 u
2 Lu
I(i ∈ [k]typ )
+
ni,j
3ni,j
u i=1
u i=1 j=h
i=1
j=h
h=j
t|
{z
}t|
{z
}
(a)
(56)
(b)
We now prove bounds on (a) and (b) respectively
(a) =
k H−1
X
X
Vπi,j+1 +
i=1 j=h
|
{z
(c)
}
k H−1
X
X
i=1 j=h
V∗i,j+1 − Vπi,j+1 .
|
{z
(d)
(57)
}
(c) and (d) can be bounded under the events E and Ωk,h using the results of Lem. 8 and Lem.9. We then deduce
(a)
≤ HTk + 2H 2 Uk,1 + 6H 2
≤ 2HTk + 2H 2 Uk,1 ,
p
4H 2 L
Tk L +
3
where the last line follows by the fact that for the typical episodes Tk ≥ 250H 2 S 2 AL2 . Thus if Tk ≤ H 2 L the term
Ck,h trivially equals to 0 otherwise the higher order terms are bounded by O(HTk ).
We now bound (b) using a pigeon-hole argument
(b) ≤ 2
X
(x,a)∈S×A
Nk (x,a)
X
n=1
T
X
1
1
≤ 2SA
≤ 2SA ln(3T ).
n
n
n=1
Plugging the bound on (a) and (b) into Eq. 56 and taking in to account that for the typical episodes [k]typ we have
that T ≥ H 2 L completes the proof.
We now bound Bk,h :
Lemma 12. Let k ∈ [K] and h ∈ [H]. Let the bonus is defined according to Algo. 4. Then under the events E and
Ωk,h the following hold for every x ∈ S,
Bk,h
≤
Bk,h,x
≤
q
p
11L Tk HSA + 12 H 2 SAL2 Uk,1 + 570H 2 S 2 AL2 ,
q
q
2 2
2
′ (x)HSA + 12 H 2 SAL2 U
11L Nk,h
k,h,x + 570H S AL ,
(58)
(59)
Proof. Here we only prove the bound on Eq. 58. The proof for the bound of Eq. 59 can be done in a very similar
′
manner, as it is shown in the previous lemmas (the only difference is that HNk,h
(x) and Uk,h,x replace Tk and Uk,1 ,
respectively). We first notice that the following holds:
Bk,h ≤
k
X
i=1
|
I(i ∈ [k]typ )
H−1
X
j=h
{z
(a)
s
v
u
H−1
k
X u 8 X
X
b i,j+1 L
8V
t
I(i ∈ [k]typ )
pbi,j (y) min
+L
ni,j
ni,j
i=1
y∈S
j=h
}
|
{z
(b)
25
!
1002 S 2 H 2 AL2
, H2 .
′
(y)
Ni,j+1
}
We first note that the bound on Bk,h is similar to the bound on Ck,h . The main difference (beside the difference in
b i,j+1 . So in our proof we first focus on dealing with this difference.
H.O.Ts) is that here V∗h+1 is replaced by V
The Cauchy–Schwarz inequality leads to:
v
v
u k
u k H−1
H−1
uX
X 1
X
√ uX
u
b
,
8Lu
I(k
∈
[k]
)
Vi,j+1 (x, a)u
typ
u
ni,j
u i=1
u i=1 j=h
j=h
t|
{z
}t|
{z
}
(a) ≤
(c)
(d)
The bound on (d) is identical to the corresponding bound in Lem. 11. So we only focus on bounding (c):
(c) =
k H−1
X
X
Vπi,j+1
i=1 j=h
|
{z
(e)
+
}
k H−1
X
X
i=1 j=h
|
b i,j+1 − Vπ
V
i,j+1 .
{z
(60)
}
(f )
(e) and (f ) can be bounded in high probability using the results of Lem. 8 and Lem.10. This implies
(c)
≤ HTk + 3H 2 Uk,1 + 15H 2 S
≤ 2HTk + 3H 2 Uk,1 ,
p
4H 2 L
ATk L +
3
where the last line follows by the fact that for the typical episodes Tk ≥ 250H 2 S 2 AL. Thus if Tk ≤ 250H 2S 2 L then
Bk,h trivially equals to 0 otherwise the higher order terms are bounded by O(HT ). Combining the bound on (b) and
(c) leads to the following bound on (a):
(a) ≤
p
p
8L HSATk + 12HL SAUk,1 .
To bound (b) we make use of Cauchy-Schwarz inequality again.
(b) ≤
v
u k
H−1
k H−1
XX
u X
X
X I(i ∈ [k]typ )
u8
I(i ∈ [k]typ )
pbi,j (y)b′i,j+1 (y)
.
u
ni,j
u i=1
i=1 j=h
j=h y∈S
t |
{z
}|
{z
}
(g)
(h)
The term (h) bounded by 2SAL using a pigeon-hole argument (see Lem. 11). We proceed by bounding (g):
(g) ≤
k H−1
X
X
i=1 j=h
|
(b
pi,j − pi,j )b′i,j+1 +
{z
(i)
}
k H−1
X
X
i=1 j=h
|
(pi,j b′V − b′i,j+1 (xi,j+1 )) +
{z
(j)
}
k
X
i=1
|
I(i ∈ [k]typ )
H−1
X
j=h
{z
k
b′i,j+1 (xi,j+1 ) .
}
√
√
Given that the event E holds the term (i) bounded by 2 2H 2 S ALTk by using the pigeon-hole√argument. Under
the event E the event Eaz (Fb′ ,k,h , H 2 ) holds. This implies that the term (j) is also bounded by 2H 2 Tk L as it is sum
of the martingale differences. The term (k) is also bounded by 20000H 3S 3 A3 L3 using the pigeon-hole argument.
Combining all these bounds together leads to the following bound on (b)
(b) ≤
q √
p
p
32 2H 2 S 2 Tk AL3 + 32H 2 SA Tk L3 + 320000S 4H 4 A2 L3 .
Combining this with the bound on (a) and taking into account the fact that we only bound the Bk,h for the typical
episodes, in which Tk ≥ 250H 2 S 2 AL2 , completes the proof.
26
Lemma 13. Let the bonus is defined according to Algo. 4. Then under the events E and ΩK,1 the following hold
√
√
p
^
Regret(K) ≤ Regret(K)
≤ UK,1 ≤ 15L HSAT + 16HL SAUk,1 + 820H 2S 2 A2 L + 2H T L.
(61)
Proof. We first notice that Regret(K) and Regret(K) are bounded by Uk,1 due to Lem.6. To bound Uk,1 we sum up
PK PH
the regret due to Bk,h and Ck,h from Lem. 11 and Lem. 12. We also bound the sum k=1 h=1 c4,k,h by 2HSAL
using a pigeon hole argument. We also note that Bk,h and Ck,h only account for the regret of typical episodes in
which T ≥ H 2 S 2 A2 L. The regret of those episodes which do not belong to the typical set [k]typ , can be bounded by
O(H 2 S 2 A2 L2 ), trivially.
The following lemma establishes an explicit bound on the regret:
Lemma 14. Let the bonus is defined according to Algo. 4. Then under the events E and ΩK,1 the following hold
√
√
^
Regret(K) ≤ Regret(K)
≤ UK,1 ≤ 30L HSAT + 2500H 2S 2 AL2 + 4H T L.
(62)
Proof. The proof follows by solving the bound of Lem. 13 in terms of Uk,1 . which only contributes to the additional
regret of O(H 2 L2 SA).
Lemma 15. Let the bonus is defined according to Algo. 3. Then under the events E and ΩK,1 the following holds
√
^
Regret(K) ≤ Regret(K)
≤ UK,1 ≤ 20HL SAT + 250H 2 S 2 AL2 .
(63)
Proof. The proof up to Lem. 11 is identicalq
to the proof of Lem. 14. The main difference is to prove bound on Ck,h
and Bk,h here we use a loose bound of O(H SAL
nk,h ) for both exploration bonus bk,h and the confidence interval c1,k,h
and then sum these
√ terms using a pigeon-hole argument (The proof is provided in Jaksch et al., 2010) which leads to a
bound of O(H SAT L) on both BK,1 and CK,1 . Plugging these results into the bound of Lem. 7 combined with the
regret of non-typical episodes complete the proof
Lemma 16. Let the bonus is defined according to Algo. 4. Let k ∈ [K] and h ∈ [H]. Then under the events E and
Ωk,h the following hold for every x ∈ S,
Regret(k, x, h) ≤
≤
q
q
′ (x) + 2500H 2 S 2 AL2 + 4H 1.5 N ′ (x)L
^
Regret(k,
x, h) ≤ 30HL SANk,h
k,h
q
′ (s).
100H 1.5SL ANk,h
Proof. The proof is similar to the proof of total regret. Here also we use Lem. 12, Lem. 11 and a pigeon-hole argument
to bound the regrets due to Bk,h , Ck,h and c4,k,h . We then incorporate these terms into Lem.6 to bound the regret in
terms of Uk,h,x . The result follows by solving the bound w.r.t. the upper bound Uk,h,x .
Lemma 17. Let the bonus b is defined according to Algo. 4. Let k ∈ [K] and h ∈ [H]. Then under the events E and
Ωk,h the following hold for every x ∈ S
Vk,h (x) −
Vh∗ (x)
≤ 100
s
H 3 S 2 AL2
′ (s) .
Nk,h
Proof. From Lem. 16 we have that
q
′ (s)
100H 1.5 SL ANk,h
≥
k
X
i=1
I(xi,h = x)(Vi,h (x) − Vhπi (x))
≥ (Vk,h (x) − Vh∗ (x))
k
X
i=1
′
I(xi,h = x) = Nk,h
(x)(Vk,h (x) − Vh∗ (x)),
27
where the last inequality holds due to the fact that Vk,h by definition is monotonically non-increasing in k. The proof
then follows by collecting terms.
Lemma 18. Let the bonus b is defined according to Algo. 3. Then under the event E the set of events {Ωk,h }k∈[K],h∈H
hold.
Proof. We prove this result by induction. First we notice that for h = H by definition Vk,h = Vh∗ thus the inequality
Vk,h ≥ Vh∗ trivially holds. Thus to prove this result for h < H we only need to show that if the inequality Vk,h ≥ Vh∗
holds for h it also holds for h − 1 for every h < H:
Vk,h (x) − Vh∗ (x)
∗
π∗
= Tk Vk,h−1 (x) − T V ∗ (x) ≥ bk (x, πh∗ (x)) + Pbk,h
Vk,h+1 (x) − Phπ Vk,h+1 (x)
π
∗
π
∗
= bk (x, πh∗ (x)) + Pbk,h
(Vk,h+1 − Vh+1
)(x) + (Pbk,h
− Phπ )Vh+1
(x)
∗
∗
∗
∗
π∗
∗
≥ bk (x, πh∗ (x)) + (Pbk,h
− Phπ )Vh+1
(x),
∗
where the last line follows by the induction condition that Vk,h+1 ≥ Vh+1
. The fact that the event E hols implies that
∗
∗
π
π
∗
∗
∗
b
(Ph − Pk,h )Vh+1 (x) ≤ c1 (Nk (x, πh (x))) ≤ bk (x, πh (x)), which completes the proof.
Lemma 19. Let the bonus b is defined according to Algo. 4. Then under the event E the set of events {Ωk,h }k∈[K],h∈H
hold.
Proof. We prove this result by induction. We first notice that in the case of the first episode V1,h = H ≥ Vh∗ .
To prove this result by induction in the case of 1 < k ∈ [K] we need to show that in the case of h ∈ [H − 1] if
Ωk,h+1 holds then Ωk,h also holds.
If Ωk,h−1 holds then Vi,j ≥ Vj∗ for every (i, j) ∈ [k, h]hist . We can then invoke the result of Lem. 17 which implies
Vk,h+1 (x) −
∗
Vh+1
(x)
≤
√
100H 1.5 SL A
q
.
′
(x)
Nk,h+1
∗
Using this result which guarantees that Vk,h+1 is close to Vh+1
we prove that Vk,h − Vh∗ ≥ 0, that is the event Ωk,h
holds.
Vk,h − Vh∗ = min(Vk−1,h , Tk,h Vi,j+1 , H) − Vh∗
If Vk−1,h ≤ Tk,h Vi,j+1 the result Vk,h − Vh∗ = Vk−1,h − Vh∗ ≥ 0 holds trivially. Also if Vk−1,h ≥ H the result
trivially holds. So we only need to consider the case that Tk,h Vi,j+1 ≤ Vk−1,h ≤ H in that case we have w
Vk,h (x) − Vh∗ (x)
≥
∗
Tk,h Vi,j+1 (x) − T Vh+1
(x)
≥
∗
bk,h (x, π ∗ (x, h)) + Pbhπ Vi,j+1 (x) − Phπ∗ Vh+1
(x)
∗
∗
∗
∗
(x) + Pbhπ∗ (Vi,j+1 − Vh+1
)(x)
bk,h (x, π ∗ (x, h)) + (Pbhπ − Phπ )Vh+1
(I)
=
(II)
≥
∗
∗
bk,h (x, π ∗ (x, h)) + (Pbhπ − Phπ )Vh+1
(x),
∗
∗
where in (I) we rely on the fact that πk,h is the greedy policy w.r.t. Vk,h . Thus
bk,h (x, π ∗ (x, h)) + Pbhπ Vi,j+1 (x)
∗
≤ bk,h (x, πk (x, h)) + Pbhπk Vi,j+1 (x).
Also (II) follows from the induction assumption. Under the event E we have
28
Vk,h − Vh∗
b ∗ , Nk )
≥ bk,h − c1 (V
h
s
s
b ∗ L 14L
b
V
8Vk,h L
h
−2
≥
−
Nk
Nk
3Nk
|
{z
}
(a)
We now prove a lower bound on (a):
v h
i
u
u Pbk 8 min 1002 H′3 S 2 A2 L2 , H 2
t
Nk,h+1
14L
+
.
+
Nk
3Nk
s
b ∗ − 8V
b k,h
4V
k,h
−
(a) ≥
Nk
0
b k,h ≤ V∗ ,
V
otherwise.
b ∗ in terms of V
b k,h from above:
We proceed by bounding V
k,h
b∗
V
k,h
(I)
≤
b k,h + 2Var b (Vk,h+1 (y) − V ∗ (y)) ≤ 2V
b k,h + 2 Pbk (Vk,h+1 − V ∗ )2 ,
2V
h+1
h+1
y∼Pk
|
{z
}
(b)
where (I) is an application of Lem. 2. We now bound (b). Combining this result with the result of Eq. 64 leads to the
following bound on (a)
v
i
h
u
u 8Pbk min 1002 H′ 3 S 2 AL2 , H 2
t
Nk,h+1
(a) ≥ −
N
k
0
b k,h ≤ V
b ∗,
V
otherwise,
where the last inequality holds under the event E. The proof is completed by plugging (a) and (b) into Eq. 64 which
proves that Vk,h ≥ Vh∗ thus the event Ωk,h holds.
C.1 Proof of Thm. 1
The result is a direct consequence of Lem. 18 and Lem. 15 and the fact that the high probability event E holds w.p.
1 − δ.
C.2 Proof of Thm. 2
The result is a direct consequence of Lem. 19 and Lem. 14 and the fact that the high probability event E holds w.p.
1 − δ.
29
| 2cs.AI
|
INDUCTIVE FREENESS OF ZIEGLER’S CANONICAL
MULTIDERIVATIONS FOR REFLECTION ARRANGEMENTS
arXiv:1705.02767v1 [math.CO] 8 May 2017
TORSTEN HOGE AND GERHARD RÖHRLE
Abstract. Let A be a free hyperplane arrangement. In 1989, Ziegler showed that the
restriction A ′′ of A to any hyperplane endowed with the natural multiplicity is then a free
multiarrangement. We initiate a study of the stronger freeness property of inductive freeness
for these canonical free multiarrangements and investigate them for the underlying class of
reflection arrangements.
More precisely, let A = A (W ) be the reflection arrangement of a complex reflection
group W . By work of Terao, each such reflection arrangement is free. Thus so is Ziegler’s
canonical multiplicity on the restriction A ′′ of A to a hyperplane. We show that the latter
is inductively free as a multiarrangement if and only if A ′′ itself is inductively free.
Contents
1
4
11
14
22
1. Introduction
2. Recollections and Preliminaries
3. Filtrations of Free Multiplicities
4. Proofs of Theorems 1.4 and 1.9
References
1. Introduction
The class of free arrangements, respectively free multiarrangments, plays a fundamental role
in the theory of hyperplane arrangements, respectively multiarrangements. In his seminal
work [Z89], Ziegler introduced the notion of multiarrangements and initiated the study of
their freeness. We begin by recalling Ziegler’s fundamental construction from loc. cit.
Definition 1.1. Let A be an arrangement. Fix H0 ∈ A and consider the restriction A ′′
with respect to H0 . Define the canonical multiplicity κ on A ′′ as follows. For Y ∈ A ′′ set
κ(Y ) := |AY | − 1,
i.e., κ(Y ) is the number of hyperplanes in A \ {H0 } lying above Y . Ziegler showed that
freeness of A implies freeness of (A ′′ , κ) as follows.
2010 Mathematics Subject Classification. Primary 20F55, 51F15, 52C35, 14N20, 32S22; Secondary 51D20.
Key words and phrases. Multiarrangement, Ziegler multiplicity, free arrangement, inductively free arrangement, reflection arrangement.
1
Theorem 1.2 ([Z89, Thm. 11]). Let A be a free arrangement with exponents exp A =
{1, e2 , . . . , eℓ }. Let H0 ∈ A and consider the restriction A ′′ with respect to H0 . Then the
multiarrangement (A ′′ , κ) is free with exponents exp(A ′′ , κ) = {e2 , . . . , eℓ }.
Note that the converse of Theorem 1.2 is false. For let A be a non-free 3-arrangement,
cf. [OT92, Ex. 4.34]. Since A ′′ is of rank 2, (A ′′ , κ) is free, [Z89, Cor. 7]. Nevertheless,
Ziegler’s construction and in particular the question of a converse of Theorem 1.2 under suitable additional hypotheses play an important role in the study of free simple arrangements,
[Y04, Thms. 2.1, 2.2], [Y05], [AY13, Cor. 4.2], [S12, Thm. 2] and [Y14, Cor. 1.35].
Because of the relevance of Ziegler’s multiplicity in the theory of free arrangements, it is
natural to investigate stronger freeness properties for (A ′′ , κ) and specifically to ask for an
analogue of Theorem 1.2 for inductive freeness.
Question 1.3. Is it the case that (A ′′ , κ) is inductively free whenever A is?
In this paper we initialize the study of this question and examine the multiarrangements
(A ′′ , κ) where the underlying class are reflection arrangements.
It can be quite cumbersome to decide whether a given (multi)arrangement is inductively free,
see for instance [BC12, §5.2], [AHR14, Lem. 4.2], [HR15, Lem. 3.5], and [CR16, Thm. 1.1].
This might entail an extensive perusal of a large number of chains of free subarrangements.
If A = A (W ) is a reflection arrangement of the complex reflection group W , then A is
free, thanks to work of Terao [T80b]. Thus (A ′′ , κ) is also free, by Theorem 1.2.
In our main result, we classify all instances when Ziegler’s canonical multiplicity (A ′′ , κ)
is inductively free in case the underlying arrangement A is a reflection arrangement. In
particular, we answer Question 1.3 in the affirmative for this class of arrangements.
Because of the compatibility of the product constructions for inductive freeness for simple arrangements, [HR15, Prop. 2.10], as well as for multiarrangements, [HRS17, Thm. 4]
(cf. Theorem 2.23), the question of inductive freeness of (A ′′ , κ) readily reduces to the case
when A is irreducible. Thus for a reflection arrangement A (W ), we may assume that W is
irreducible.
Theorem 1.4. Let A = A (W ) be the reflection arrangement of the irreducible complex
reflection group W . Let A ′′ be the restriction of A to a hyperplane in A . Then (A ′′ , κ) is
inductively free if and only if one of the following holds:
(i) A is inductively free; or
(ii) A is non-inductively free of rank at most 4.
We record several consequences of Theorem 1.4. Our first observation follows from the theorem along with the classification of all inductively free restrictions of reflection arrangements
from [AHR14, Thm. 1.2], see Theorem 2.4 below.
Corollary 1.5. Let A = A (W ) be the reflection arrangement of the complex reflection
group W . Then (A ′′ , κ) is inductively free if and only if A ′′ itself is inductively free.
2
We mention that Corollary 1.5 is false in general. For the failure of the forward implication,
see [HR15, Ex. 2.16]. There we present an example of an inductively free arrangement
A where a particular restriction to a hyperplane A ′′ fails to be free. In particular, A ′′
is not inductively free. Nevertheless, since A is free, so is (A ′′ , κ), thanks to Theorem
1.2. One can check that in this instance (A ′′ , κ) is actually inductively free. Nevertheless,
this example is consistent with the assertion in Question 1.3. Example 2.24 shows that
the reverse implication in Corollary 1.5 also fails in general. In view of these elementary
examples, the equivalence of Corollary 1.5 is rather striking and underlines the very special
nature of reflection arrangements.
In addition to the canonical multiplicity from Definition 1.1, we study multiplicities which are
concentrated at a single hyperplane. These were introduced by Abe, Terao and Wakefield,
[ATW08, §5]. While in general, (A , µ) need not be free for a free hyperplane arrangement
A and an arbitrary multiplicity µ, e.g. see [Z89, Ex. 14], for these concentrated multiplicities
freeness is also inherited from the simple arrangement. It turns out that they are closely
related to Ziegler’s canonical multiplicity, see Propositions 2.14 and 2.15.
Definition 1.6. Let A be a simple arrangement. Fix H0 ∈ A and m0 ∈ Z≥1 and define
the multiplicity δ concentrated at H0 by
(
m0 if H = H0 ,
δ(H) := δH0 ,m0 (H) :=
1
else.
It turns out that in this instance, both A and (A , δ) inherit freeness from each other, as
opposed to the general case indicated above. We record this in our next result.
Theorem 1.7. Let A be an arrangement. Fix H0 ∈ A , m0 ∈ Z≥1 and let δ = δH0 ,m0 be the
multiplicity concentrated at H0 , as in Definition 1.6. Then A is free with exponents exp A =
{1, e2 , . . . , eℓ } if and only if (A , δ) is free with exponents exp(A , δ) = {m0 , e2 , . . . , eℓ }.
The forward implication of Theorem 1.7 is [ATW08, Prop. 5.2] and the reverse implication
is Proposition 2.15(i).
As in the case of Ziegler’s canonical multiplicity, one might also speculate whether the
analogue for inductive freeness holds for the forward implication of Theorem 1.7.
Question 1.8. Is it the case that (A , δ) is inductively free whenever A is?
It turns out that Questions 1.3 and 1.8 are closely related: the assertion of Question 1.8
follows from that in Question 1.3, see Corollary 2.20.
Let A = A (W ) be the reflection arrangement of the complex reflection group W . Since A
is free, it follows from Proposition 2.14(i) that the multiarrangement (A , δ) is also free. In
[HR15, Thm. 1.1], we classified all inductively free reflection arrangements. In our second
main result we extend this classification to multiarrangements of the form (A (W ), δ). This
in particular gives an affirmative answer to Question 1.8 for the class of reflection arrangements.
Theorem 1.9. Let A = A (W ) be the reflection arrangement of the complex reflection group
W . Then (A , δ) is inductively free if and only if A is inductively free.
3
The paper is organized as follows. In Section 2.2 we recall the fundamental results for free arrangements, in particular Terao’s Addition Deletion Theorem 2.1 and the subsequent notion
of an inductively free arrangement. The classification of the inductively free reflection arrangements from [HR15, Thm. 1.1], and the classification of the inductively free restrictions
of reflection arrangements from [AHR14, Thm. 1.2] are restated in Section 2.3. Section 2.5
is devoted to multiarrangements and their freeness. Here we present the Addition Deletion
theorem due to Abe, Terao, and Wakefield, [ATW08, Thm. 0.8]. In Section 2.6 we treat
Ziegler’s canonical and concentrated multiplicities. This is followed by a discussion of inductive freeness for multiarrangements in Section 2.7. Here we recall results from [HRS17] which
show the compatibility of this notion with products and localization for multiarrangements
that are used in the sequel.
Section 3 is central to our general arguments. Here we prove some general results for free
arrangements which show that under weak conditions on the exponents of a free arrangement A and a given free restriction A H , every multiplicity between the simple and the
Ziegler multiplicity on A H must already be free, see Corollary 3.9. This is then applied to
reflection arrangements in Proposition 3.10. We then apply these methods to the imprimitive reflection groups G(r, 1, ℓ − 1). Combined with the fact that the reflection arrangement
A (G(r, 1, ℓ−1)) itself is is inductively free (cf. Theorem 2.3), we show that (A (G(r, 1, ℓ))′′, κ)
is inductively free in Proposition 3.11. As a further relevant consequence for our classification, we derive that (A (W )′′ , κ) is inductively free, for W of type F4 , H4 , G31 and G32 in
Corollary 3.14.
In Section 4 we present our proofs of Theorems 1.4 and 1.9. The fact that inductively free
multiarrangements are closed under localization proves extremely useful in this context, see
Theorem 2.22. By this means, we can for instance deduce that (A (E6 )′′ , κ) and (A (E7 )′′ , κ)
are inductively free from the fact that (A (E8 )′′ , κ) is so, see Lemma 4.5. The same theorem
in turn allows us to derive that (A (G34 )′′ , κ) is not inductively free from the fact that
(A (G(3, 3, 5))′′, κ) isn’t either, see Lemma 4.10.
Due to its large rank, the proof of inductive freeness for (A (E8 )′′ , κ) is the most challenging
among all groups of exceptional type. It involves delicate arguments of induction, see Lemma
4.5. A crucial ingredient in our proof is the fact that on A (E8 )′′ every multiplicity between
the simple one and Ziegler’s multiplicity κ is free, thus readily providing free filtrations of
(A (E8 )′′ , κ) for our analysis, see Proposition 3.10.
2. Recollections and Preliminaries
2.1. Hyperplane Arrangements. Let V = Kℓ be an ℓ-dimensional K-vector space. A
hyperplane arrangement is a pair (A , V ), where A is a finite collection of hyperplanes in V .
Usually, we simply write A in place of (A , V ). We write |A | for the number of hyperplanes
in A . The empty arrangement in V is denoted by Φℓ .
The lattice L(A ) of A is the set of subspaces of V of the form H1 ∩· · ·∩Hi where {H1 , . . . , Hi }
is a subset of A . For X ∈ L(A ), we have two associated arrangements, firstly AX := {H ∈
A | X ⊆ H} ⊆ A , the localization of A at X, and secondly, the restriction of A to X,
(A X , X), where A X := {X ∩ H | H ∈ A \ AX }. Note that V belongs to L(A ) as the
4
intersection of the empty collection of hyperplanes and A V = A . The lattice L(A ) is a
partially ordered set by reverse inclusion: X ≤ Y provided Y ⊆ X for X, Y ∈ L(A ).
If 0 ∈ H for each H in A , then A is called central. If A is central, then the center
TA := ∩H∈A H of A is the unique maximal element in L(A ) with respect to the partial
order. We have a rank function on L(A ): r(X) := codimV (X). The rank r := r(A ) of
A is the rank of a maximal element in L(A ). The ℓ-arrangement A is essential provided
r(A ) = ℓ. If A is central and essential, then TA = {0}. Throughout, we only consider
central arrangements.
2.2. Free Hyperplane Arrangements. Let S = S(V ∗ ) be the symmetric algebra of the
dual space V ∗ of V . If x1 , . . . , xℓ is a basis of V ∗ , then we identify S with the polynomial
ring K[x1 , . . . , xℓ ]. Letting Sp denote the K-subspace of S consisting of the homogeneous
polynomials of degree p (along with 0), S is naturally Z-graded: S = ⊕p∈Z Sp , where Sp = 0
in case p < 0.
Let Der(S) be the S-module of algebraic K-derivations of S. Using the Z-grading on S,
Der(S) becomes a graded S-module. For i = 1, . . . , ℓ, let Di := ∂/∂xi be the usual derivation
of S. Then D1 , . . . , Dℓ is an S-basis of Der(S). We say that θ ∈ Der(S) is homogeneous of
P
polynomial degree p provided θ = ℓi=1 fi Di , where fi is either 0 or homogeneous of degree
p for each 1 ≤ i ≤ ℓ. In this case we write pdeg θ = p.
Let A be an arrangement in V . Then for H ∈ A we Q
fix αH ∈ V ∗ with H = ker(αH ). The
defining polynomial Q(A ) of A is given by Q(A ) := H∈A αH ∈ S.
The module of A -derivations of A is defined by
D(A ) := {θ ∈ Der(S) | θ(αH ) ∈ αH S for each H ∈ A }.
We say that A is free if the module of A -derivations D(A ) is a free S-module.
With the Z-grading of Der(S), also D(A ) becomes a graded S-module, [OT92, Prop. 4.10].
If A is a free arrangement, then the S-module D(A ) admits a basis of ℓ homogeneous
derivations, say θ1 , . . . , θℓ , [OT92, Prop. 4.18]. While the θi ’s are not unique, their polynomial
degrees pdeg θi are unique (up to ordering). This multiset is the set of exponents of the free
arrangement A and is denoted by exp A .
The fundamental Addition Deletion Theorem due to Terao [T80a] plays a crucial role in the
study of free arrangements, [OT92, Thm. 4.51].
Theorem 2.1. Suppose A 6= Φℓ and let (A , A ′ , A ′′ ) be a triple of arrangements. Then
any two of the following statements imply the third:
(i) A is free with exp A = {b1 , . . . , bℓ−1 , bℓ };
(ii) A ′ is free with exp A ′ = {b1 , . . . , bℓ−1 , bℓ − 1};
(iii) A ′′ is free with exp A ′′ = {b1 , . . . , bℓ−1 }.
Theorem 2.1 motivates the notion of inductively free arrangements, cf. [OT92, Def. 4.53]:
5
Definition 2.2. The class IF of inductively free arrangements is the smallest class of arrangements subject to
(i) Φℓ ∈ IF for each ℓ ≥ 0;
(ii) if there exists a hyperplane H0 ∈ A such that both A ′ and A ′′ belong to IF , and
exp A ′′ ⊆ exp A ′ , then A also belongs to IF .
2.3. Reflection Arrangements. The irreducible finite complex reflection groups were classified by Shephard and Todd, [ST54]. Let W ⊆ GL(V ) be a finite complex reflection group.
For w ∈ W , we write Fix(w) := {v ∈ V | wv = v} for the fixed point subspace of w. For
U ⊆ V a subspace, we define the parabolic subgroup WU of W by WU := {w ∈ W | U ⊆
Fix(w)}.
The reflection arrangement A = A (W ) of W in V is the hyperplane arrangement consisting
of the reflecting hyperplanes of the elements in W acting as reflections on V . By Steinberg’s
Theorem [Ste60, Thm. 1.5], for X ∈ L(A ), the parabolic subgroup WX is itself a complex
reflection group, generated by the unitary reflections in W that are contained in WX . Thus,
we identify the reflection arrangement A (WX ) of WX as a subarrangement of A .
Using the classification and nomenclature of Shephard and Todd [ST54], we recall the main
classification results from [HR15, Thm. 1.1] and [AHR14, Thm. 1.2], respectively.
Theorem 2.3. For a finite complex reflection group W , its reflection arrangement A (W )
is inductively free if and only if W does not admit an irreducible factor isomorphic to a
monomial group G(r, r, ℓ) for r, ℓ ≥ 3, G24 , G27 , G29 , G31 , G33 , or G34 .
Theorem 2.4. Let W be a finite, irreducible, complex reflection group with reflection arrangement A = A (W ) and let X ∈ L(A ). The restricted arrangement A X is inductively
free if and only if one of the following holds:
(i) A is inductively free;
(ii) W = G(r, r, ℓ) and A X ∼
= Apk (r), where p = dim X and p − 2 ≤ k ≤ p;
(iii) W is one of G24 , G27 , G29 , G31 , G33 , or G34 and X ∈ L(A ) \ {V } with dim X ≤ 3.
2.4. Multiarrangements. A multiarrangement is a pair (A , µ) consisting of a hyperplane
arrangement A and a multiplicity function µ : A → Z≥0 associating to each hyperplane H
in A a non-negative integer µ(H). Alternately, the multiarrangement (A , µ) can also be
thought of as the multiset of hyperplanes
(A , µ) = {H µ(H) | H ∈ A }.
The order of the multiarrangement
(A , µ) is the cardinality of the multiset (A , µ); we write
P
|µ| := |(A , µ)| = H∈A µ(H). For a multiarrangement (A , µ), the underlying arrangement
A is sometimes called the associated simple arrangement, and so (A , µ) itself is simple if
and only if µ(H) = 1 for each H ∈ A .
Let (A , µ) be a multiarrangement in V and let X ∈ L(A ). The localization of (A , µ) at X
is defined to be (AX , µX ), where µX = µ|AX .
6
There is a natural partial order on the set of multiplicities for a given simple arrangement,
as follows.
Definition 2.5. For multiplicities µ1 and µ2 on an arrangement A , we define µ1 ≤ µ2
provided µ1 (H) ≤ µ2 (H) for every H in A .
2.5. Freeness of multiarrangements. Following Ziegler [Z89], we extend the notion of
freeness to multiarrangements as follows. The defining polynomial Q(A , µ) of the multiarrangement (A , µ) is given by
Y µ(H)
αH ,
Q(A , µ) :=
H∈A
a polynomial of degree |µ| in S.
The module of A -derivations of (A , µ) is defined by
µ(H)
D(A , µ) := {θ ∈ Der(S) | θ(αH ) ∈ αH
S for each H ∈ A }.
We say that (A , µ) is free if D(A , µ) is a free S-module, [Z89, Def. 6].
As in the case of simple arrangements, D(A , µ) is a Z-graded S-module and thus, if (A , µ)
is free, there is a homogeneous basis θ1 , . . . , θℓ of D(A , µ). The multiset of the unique
polynomial degrees pdeg θi forms the set of exponents of the free multiarrangement (A , µ)
and is denoted by exp(A , µ).
Next we recall Ziegler’s analogue of Saito’s criterion.
Theorem 2.6 ([Z89, Thm. 8]). Let θ1 , . . . , θℓ be in D(A , µ). Then {θ1 , . . . , θℓ } is a basis of
D(A , µ) if and only if
(2.7)
det M(θ1 , . . . , θℓ ) =
˙ Q(A , µ).
In particular, if θ1 , . . . , θℓ are all homogeneous, then {θ1 , . . . , θℓ } is a basis of D(A , µ) if and
only if θ1 , . . . , θℓ are independent over S and
X
(2.8)
pdeg θi = deg Q(A , µ) = |µ|.
Here the notation =
˙ indicates equality up to a non-zero scalar multiple and M(θ1 , . . . , θℓ ) is
the coefficient matrix of {θ1 , . . . , θℓ }, see [OT92, Def. 4.11].
Note that, owing to (2.7), if (A , µ) is free, one can recover µ from the basis {θ1 , . . . , θℓ }.
We recall the construction from [ATW08] for the counterpart of Theorem 2.1 in this more
general setting.
Definition 2.9. Let (A , µ) 6= Φℓ be a multiarrangement. Fix H0 in A . We define the
deletion (A ′ , µ′ ) and restriction (A ′′ , µ∗ ) of (A , µ) with respect to H0 as follows. If µ(H0 ) =
1, then set A ′ = A \ {H0 } and define µ′ (H) = µ(H) for all H ∈ A ′ . If µ(H0 ) > 1, then set
A ′ = A and define µ′ (H0 ) = µ(H0 ) − 1 and µ′ (H) = µ(H) for all H 6= H0 .
Let A ′′ = {H ∩ H0 | H ∈ A \ {H0 } }. The Euler multiplicity µ∗ of A ′′ is defined as follows.
Let Y ∈ A ′′ . Since the localization AY is of rank 2, the multiarrangement (AY , µY ) is
free, [Z89, Cor. 7]. According to [ATW08, Prop. 2.1], the module of derivations D(AY , µY )
admits a particular homogeneous basis {θY , ψY , D3 , . . . , Dℓ }, such that θY ∈
/ α0 Der(S) and
7
ψY ∈ α0 Der(S), where H0 = ker α0 . Then on Y the Euler multiplicity µ∗ is defined to be
µ∗ (Y ) = pdeg θY .
Often, (A , µ), (A ′ , µ′ ) and (A ′′ , µ∗ ) is referred to as the triple of (A , µ) with respect to H0 .
We require some core results from [ATW08].
Theorem 2.10 ([ATW08, Thm. 0.4]). Suppose that (A , µ) 6= Φℓ . Fix H0 in A . If both
(A , µ) and (A ′ , µ′ ) are free, then there exists a basis {θ1 , . . . , θℓ } of D(A ′ , µ′ ) such that
{θ1 , . . . , αk θk , . . . , θℓ } is a basis of D(A , µ) for some 1 ≤ k ≤ ℓ.
Fix H0 = ker α0 in A and let A ′′ be the restriction with respect to H0 . Consider the
projection S → S := S/α0 S, f 7→ f . Following [ATW08], there is a canonical map
D(A , µ) → D(A ′′ , µ∗), θ 7→ θ, where θ(f ) := θ(f ), for f ∈ S.
Theorem 2.11 ([ATW08, Thm. 0.6]). Suppose that (A , µ) 6= Φℓ . Fix H0 in A . Suppose
that both (A , µ) and (A ′ , µ′ ) are free. Let {θ1 , . . . , θℓ } be a basis of D(A ′ , µ′) as in Theorem
2.10. Then {θ1 , . . . , θk−1 , θk+1 , . . . , θℓ } is a basis of D(A ′′ , µ∗ ).
Theorem 2.12 ([ATW08, Thm. 0.8] Addition Deletion Theorem for Multiarrangements).
Suppose that (A , µ) 6= Φℓ . Fix H0 in A and let (A , µ), (A ′ , µ′) and (A ′′ , µ∗) be the triple
with respect to H0 . Then any two of the following statements imply the third:
(i) (A , µ) is free with exp(A , µ) = {b1 , . . . , bℓ−1 , bℓ };
(ii) (A ′ , µ′) is free with exp(A ′ , µ′ ) = {b1 , . . . , bℓ−1 , bℓ − 1};
(iii) (A ′′ , µ∗ ) is free with exp(A ′′ , µ∗ ) = {b1 , . . . , bℓ−1 }.
In order to apply Theorem 2.12 effectively, it is crucial to be able to determine the Euler
multiplicity for the restrictions at hand. The following is part of [ATW08, Prop. 4.1] relevant
for our purposes.
Proposition 2.13. Let H0 ∈ A , A ′′ = A H0 and let X ∈ A ′′ . Let µ be a multiplicity on
A . Let µ0 = µ(H0 ). Further let k = |AX | and µ1 = max{µ(H) | H ∈ AX \ {H0 }}.
(1) If k = 2, then µ∗ (X) = µ1 .
(2) If k = 3, 2µ0 ≤ |µX |, and 2µ1 ≤ |µX |, then µ∗ (X) = ⌊|µX |/2⌋.
(3) If |µX | ≤ 2k − 1 and µ0 > 0, then µ∗ (X) = k − 1.
2.6. Ziegler’s Multiplicity and concentrated Multiplicities. Recall Ziegler’s multiplicity κ from Definition 1.1 and concentrated multiplicities from Definition 1.6. The following
combines [ATW08, Prop. 5.2], parts of its proof and Theorem 1.2. The proof of Proposition
2.14(i) given in [ATW08] depends on Theorem 2.12. In [HRS17, Prop. 1], we presented an
elementary explicit construction for a homogeneous S-basis of D(A , δ).
Proposition 2.14. Let A be a free arrangement with exponents exp A = {1, e2 , . . . , eℓ }.
Fix H0 ∈ A , m0 ∈ Z≥1 and let δ = δH0 ,m0 be as in Definition 1.6. Let (A ′′ , δ ∗ ) be the
restriction with respect to H0 . Then we have
8
(i) (A , δ) is free with exponents exp(A , δ) = {m0 , e2 , . . . , eℓ };
(ii) (A ′′ , δ ∗ ) = (A ′′ , κ) is free with exponents exp(A ′′ , κ) = {e2 , . . . , eℓ }.
Part (i) of the following is a converse of Proposition 2.14(i).
Proposition 2.15. Let A be an arrangement. Fix H0 ∈ A , m0 ∈ Z≥1 and let δ = δH0 ,m0 be
as in Definition 1.6. Let (A ′′ , δ ∗ ) be the restriction with respect to H0 . Suppose that (A , δ)
is free with exponents exp(A , δ) = {m0 , e2 , . . . , eℓ }. Then we have
(i) A is free with exponents exp A = {1, e2 , . . . , eℓ }.
(ii) (A ′′ , δ ∗ ) = (A ′′ , κ) is free with exponents exp(A ′′ , κ) = {e2 , . . . , eℓ }.
Proof. Clearly, part (ii) follows from part (i) and Proposition 2.14(ii).
The proof of (i) is analogous to the one of [OT92, Prop. 4.27]. Let α0 ∈ V ∗ with H0 = ker α0 .
Now let
Ann(H0 ) = {θ ∈ D(A , δ) | θ(α0 ) = 0}
be the annihilator of H0 in D(A , δ) which is a graded S-submodule of D(A , δ). Let θE be
the Euler derivation in Der(S) [OT92, Def. 4.7]. Define
θδ := α0m0 −1 θE .
Then θδ belongs to D(A , δ). Moreover for each θ ∈ D(A , δ), we have
θ(α0 )
θ − m0 θδ ∈ Ann(H0 ).
α0
Therefore, since
Sθδ ∩ Ann(H0 ) = {0},
we derive that
D(A , δ) = Sθδ ⊕ Ann(H0 )
is a direct sum of S-modules. Let {θ2 , . . . , θℓ } be a minimal system of homogeneous generators of the S-module Ann(H0 ). Then {θδ , θ2 , . . . , θℓ } is a minimal system of homogeneous
generators of D(A , δ). It follows from [OT92, Thm. A.19], that {θδ , θ2 , . . . , θℓ } is a homogeneous S-basis of D(A , δ).
Since Ann(H0 ) ⊂ D(A , δ) ⊂ D(A ) are S-submodules, {θ2 , . . . , θℓ } is linearly independent
over S in D(A ). Since |δ| = |A |+m0 −1 and pdeg θδ = m0 , it follows from Ziegler’s analogue
P
of Saito’s criterion (3.13) that ℓi=2 pdeg θi = |A | − 1. Clearly, θE belongs to D(A ) but not
to Ann(H0 ). Consequently, it follows from Saito’s criterion, [OT92, Thm. 4.23], that A is
free with {θE , θ2 , . . . , θℓ } a homogeneous S-basis of D(A ).
Note that the proof of Proposition 2.15 shows that m0 is necessarily an exponent of (A , δ)
if the latter is free (so it does no harm to require this in the hypothesis).
The following is a special case of a general compatibility result of the Euler multiplicity with
localizations from [HRS17, Lem. 2].
Lemma 2.16. Let X ∈ L(A ), H0 ∈ AX , and let A ′′ be the restriction with respect to H0 .
Let δ = δH0 ,m0 be as in Definition 1.6. Then we have
9
(i) ((AX )′′ , (δX )∗ ) = ((A ′′ )X , (δ ∗ )X ), and
(ii) ((AX )′′ , κ) = ((A ′′ )X , κX ) (where κ on the left is the canonical multiplicity resulting
from restriction of AX to H0 ).
Proof. Part (i) follows from loc. cit. Part (ii) follows from part (i) together with the fact
that ((AX )′′ , (δX )∗ ) = ((AX )′′ , κ), cf. Proposition 2.14(ii).
2.7. Inductive Freeness for Multiarrangements. As in the simple case, Theorem 2.12
motivates the notion of inductive freeness.
Definition 2.17 ([ATW08, Def. 0.9]). The class IF M of inductively free multiarrangements
is the smallest class of arrangements subject to
(i) Φℓ ∈ IF M for each ℓ ≥ 0;
(ii) for a multiarrangement (A , µ), if there exists a hyperplane H0 ∈ A such that both
(A ′ , µ′) and (A ′′ , µ∗) belong to IF M, and exp(A ′′ , µ∗ ) ⊆ exp(A ′ , µ′ ), then (A , µ)
also belongs to IF M.
Remark 2.18. As for simple arrangements, if r(A ) ≤ 2, then (A , µ) is inductively free,
[Z89, Cor. 7].
Remark 2.19. In analogy to the simple case, cf. [OT92, §4.3, p. 119], [HR15, Rem. 2.9],
it is possible to describe an inductively free multiarrangement (A , µ) by means of a so
called induction table. In this process we start with an inductively free multiarrangement
(frequently Φℓ ) and add hyperplanes successively ensuring that part (ii) of Definition 2.17
is satisfied. We refer to this process as induction of hyperplanes. This procedure amounts
to choosing an order on consecutive multiplicities µi of the simple arrangement A for i =
1, . . . , n = |µ| (see Definition 2.5), so that |µi | = i, µn = µ, and each restriction (A ′′ , µ∗i ) is
inductively free. As in the simple case, in the associated induction table we record in the i-th
row the information of the i-th step of this process, by listing exp(A ′ , µ′i ) = exp(A , µi−1 ),
the defining form αHi of Hi , as well as exp(A ′′ , µ∗i ), For instance, see Tables 1, 3, and 4
below.
We record an easy consequence of Proposition 2.14(ii) in the context of inductive freeness
which is going to be very useful in our proof of Theorem 1.9.
Corollary 2.20. Suppose that both A and (A ′′ , κ) are inductively free. Then (A , δ) is
inductively free.
Proof. The result follows readily from induction on m0 ≥ 1, Proposition 2.14(ii) and a
repeated application of the addition part of Theorem 2.12.
The following special case of Corollary 2.20 is immediate from [Z89, Cor. 7].
Corollary 2.21. Suppose that the rank 3 arrangement A is inductively free. Then (A , δ)
is inductively free.
10
The following is is a very useful tool for showing that a given multiarrangement is not
inductively free by exhibiting a localization which fails to be inductively free.
Theorem 2.22 ([HRS17, Thm. 3]). The class IF M is closed under taking localizations.
We also require the fact that inductive freeness for multiarrangements behave well with the
product construction.
Theorem 2.23 ([HRS17, Thm. 4]). A product of multiarrangements belongs to IF M if and
only if each factor belongs to IF M.
We close this section with an example which shows that the reverse implication of Corollary
1.5 fails in general.
Example 2.24. Consider the complex 4-arrangement A given by the defining polynomial
Q(A ) := xyzt(x − y)(x − z)(x − y + t)(x − z + t). Then one checks that A Ht is inductively
free, while both A and (A Ht , κ) are not free.
3. Filtrations of Free Multiplicities
In the sequel, we denote by 11 the simple multiplicity on a given arrangement. Recall the
partial order on the set of multiplicities on an arrangement from Definition 2.5.
Definition 3.1. Let A be a free arrangement. Suppose there is a free multiplicity µ > 11
on A such that there is a sequence of free multiplicities µi on A satisfying µi < µi+1 and
|µi+1 | = |µi | + 1, for i = 1, . . . , n − 1, where µ1 := 11 and µn := µ. Then we say that
the sequence of multiarrangements (A , µi ) is a free filtration of (A , µ) or simply that the
sequence µi is a filtration of free multiplicities on A .
Lemma 3.2. Let A be an inductively free arrangement. Suppose there is a free multiplicity
µ > 11 on A along with a free filtration 11 =: µ1 < . . . < µn := µ of (A , µ) as in Definition
3.1. If each restriction along the free filtration is inductively free, then so is (A , µ).
Proof. It follows from Theorem 2.10 that the exponents of (A ′ , µ′i ) = (A , µi−1) and of
(A , µi ) differ in precisely one entry by 1. So that, by Theorem 2.12, the exponents of each
restriction along the free filtration satisfy exp(A ′′ , µ∗i ) ⊆ exp(A ′ , µ′i ). Now, since A as well
as each restriction (A ′′ , µ∗i ) is inductively free, a repeated application of the addition part
of Theorem 2.12 gives the desired result.
We record a special case of Lemma 3.2.
Corollary 3.3. Let A be an inductively free 3-arrangement. Suppose there is a free multiplicity µ ≥ 11 on A along with a free filtration as in Lemma 3.2. Then (A , µ) is inductively
free.
Proof. Since each restriction along the free filtration is of rank 2, it is already inductively
free, by Remark 2.18. So the result is immediate from Corollary 3.3.
11
In our next result we present a mild condition on a free multiplicity µ of a free arrangement
A which implies that every intermediate multiplicity 11 < ν < µ is also free.
Lemma 3.4. Let A be a free arrangement with exponents 1 ≤ e2 ≤ . . . ≤ eℓ . Assume
that there is a free multiplicity µ > 11 on A with exp(A , µ) = {e, e2 , . . . , eℓ }, where e =
1 + |µ| − |A |. Suppose that |µ| − |A | ≥ eℓ . Let ν be a multiplicity satisfying 11 < ν < µ.
Then (A , ν) is free with exp(A , ν) = {ẽ, e2 , . . . , eℓ }, where ẽ = 1 + |ν| − |A |.
Moreover, if θ1 , . . . , θℓ is a homogeneous basis of D(A , µ) with pdeg θi = ei for i = 2, . . . , ℓ
and pdeg θ1 = e, then
!
Y ν(H)−1
(3.5)
αH
θE , θ2 , . . . , θℓ
H∈A
is a basis of D(A , ν).
Proof. Let θ1 , . . . , θℓ be a homogeneous basis of D(A , µ) with pdeg θi = ei for i = 2, . . . , ℓ
and pdeg θ1 = e = 1 + |µ| − |A |. It is enough to show that θE , θ2 , . . . , θℓ is a basis of D(A ).
For, then the derivations given in (3.5) are linearly independent over S and obviously also
members of D(A , ν) (as ν ≤ µ) so that the sum over their polynomial degrees equals |ν|, as
Pℓ
i=2 ei = |A | − 1. We are going to check that the conditions of [OT92, Thm. 4.42] hold.
For that we have to show that
(3.6)
θi 6∈ SθE + Sθ2 + · · · + Sθi−1
holds for i = 2, . . . , ℓ. Assume that (3.6) fails for some i. Then there are polynomials
p, p2 , . . . , pi−1 ∈ S such that
(3.7)
p θE =
i−1
X
pk θk + θi .
k=2
Consequently, since θ2 , . . . , θi are linearly independent over S, p 6= 0. Now applying (3.7) to
µ(H)
αH for H ∈ A , we get that p · αH ∈ αH S. Hence
!
Y µ(H)−1
p∈
αH
S.
H∈A
P
Therefore, deg p ≥ H∈A (µ(H)−1) = |µ|−|A |. By (3.7), we have 1+deg p = pdeg(p θE ) =
pdeg θi . Using the hypothesis |µ| − |A | ≥ eℓ , we get
pdeg θi ≥ 1 + |µ| − |A | > |µ| − |A | ≥ eℓ = pdeg θℓ ≥ pdeg θi ,
which is absurd. Consequently, (3.6) holds for all i = 2, . . . , ℓ and so θE , θ2 , . . . , θℓ is a basis
of D(A ), as claimed, by loc. cit.
We record an important consequence of Lemma 3.4 and Theorem 2.12 which shows that
the multisets of exponents of the restrictions along a free filtration, as in Lemma 3.4, are
constant and do not depend on the multiplicities µ∗i .
12
Corollary 3.8. Let A and (A , µ) be as in Lemma 3.4. Let 11 =: µ1 < . . . < µn :=
µ be a free filtration of (A , µ) as in Definition 3.1. Then for each restriction, we have
(A ′′ , µ∗i ) = (A ′′ , κ), where κ is Ziegler’s canonical multiplicity on A . In particular, each
such restriction along the filtration is free with exp(A ′′ , µ∗i ) = exp(A ′′ , κ) = {e2 , . . . , eℓ },
where exp A = {1, e2 , . . . , eℓ }.
Proof. Let H0 ∈ A and let 11 ≤ ν ′ < ν ≤ µ be free multiplicities as given by Lemma 3.4,
where (A ′ , ν ′ ) is the deletion of (A , ν) with respect to H0 . Let θ1 , θ2 , . . . , θℓ be a free basis
of D(A ′ , ν ′ ). It follows from Lemma 3.4 that α0 θ1 , θ2 , . . . , θℓ is a free basis of D(A , ν), as
given by (3.5), where ker α0 = H0 . Owing to Theorem 2.11, we see that θ2 , . . . , θℓ is a basis
of D(A ′′ , ν ∗ ).
Now let δ be the concentrated multiplicity on A given by δ(H0 ) = 2. Then 11 < δ ≤ µ.
Hence, according to Lemma 3.4, (A , δ) is free with α0 θE , θ2 , . . . , θℓ being a basis of D(A , δ),
by (3.5). Thanks to Proposition 2.14, (A ′′ , δ ∗ ) = (A ′′ , κ).
Since θE , θ2 , . . . , θℓ is a basis of D(A ), it follows again from Theorem 2.11 applied to (A , δ)
and (A ′ , δ ′ ) = (A , 11), that θ2 , . . . , θℓ is also a basis of D(A ′′ , κ). Hence, thanks to (2.7),
ν ∗ = κ.
Next we apply Lemma 3.4 in the context of Ziegler’s canonical multiplicity κ.
Corollary 3.9. Let A be a free arrangement with exponents 1 = e1 ≤ e2 ≤ · · · ≤ eℓ−1 < eℓ
and let H ∈ A such that A H is free with exp(A H ) = {1, e2 , . . . , eℓ−1 }. Then (A H , µ) is
free for every multiplicity 11 < µ < κ with exp(A H , µ) = {1 + |µ| − |A H |, e2 , . . . , eℓ−1 }.
Proof. Since A H is free with exp(A H ) = {1, e2 , . . . , eℓ−1 }, we have |A H | = 1+e2 +. . .+eℓ−1 .
Also note that |κ| = |A | − 1. Therefore,
1 + |κ| − |A H | = |A | − (1 + e2 + . . . + eℓ−1 ) = eℓ .
Consequently, exp(A H , κ) = {eℓ , e2 , . . . , eℓ−1 } = {1 + |κ| − |A H |, e2 , . . . , eℓ−1 }, by Theorem
1.2. Thus |κ| − |A H | = eℓ − 1 ≥ eℓ−1 , by hypothesis. So the result follows from Lemma 3.4
applied to A H and κ.
Corollary 3.9 yields that under rather weak assumptions, every multiplicity between 11 and
the Ziegler multiplicity κ is free. This yields an abundance of free filtrations from 11 to κ on
A H . We demonstrate this in the context of reflection arrangements.
Proposition 3.10. Let W be of type Aℓ , Bℓ , G(r, 1, ℓ), F4 , H4 , G31 , G32 E6 , E7 , or E8 . Let
A = A (W ). Then for any H ∈ A and every multiplicity 11 ≤ µ ≤ κ, the multiarrangement
(A H , µ) is free with exp(A H , µ) = {1 + |µ| − |A H |, e2 , . . . , eℓ−1 }.
Proof. It readily follows from [OT92, Props. 6.73, 6.77] and the tables in [OT92, App. C]
that in the given cases any restriction A H satisfies the condition of Corollary 3.9.
Armed with Corollaries 3.8 and 3.9, we can now prove our first main result of our classification.
13
Proposition 3.11. Let A = A (W ) be the reflection arrangement of the full monomial
group W = G(r, 1, ℓ) for r, ℓ ≥ 2. Then both (A ′′ , κ) and (A , δ) are inductively free.
Proof. Owing to Theorem 2.3 and Corollary 2.20, it is enough to show that (A ′′ , κ) is
inductively free. We prove the latter by induction on ℓ. Clearly, for ℓ ≤ 3, the statement
holds. Thus suppose that ℓ ≥ 4 and that the result is true for smaller ranks.
Note that irrespective of the choice of hyperplane in A , we have
(3.12)
A ′′ ∼
= A (G(r, 1, ℓ − 1)),
by [OS82, Prop. 2.11] (cf. [OT92, Prop. 6.82]). Owing to [OS82, Prop. 2.13] (cf. [OT92,
Prop. 6.77]), we have
(3.13)
exp A = {1, r + 1, . . . , (ℓ − 1)r + 1}.
Thanks to Proposition 3.10, each multiplicity µ on A ′′ with 11 ≤ µ ≤ κ is free. So we pick an
arbitrary free filtration of A ′′ from 11 to κ. Since |κ| −|A ′′ | = (ℓ −1)r > (ℓ −2)r + 1, which is
the highest exponent of A ′′ , by (3.12) and (3.13), the hypotheses of Lemma 3.4 are satisfied
for B := A ′′ . Consequently, Corollary 3.8 implies that for each multiplicity µ along our free
chain we have (B ′′ , µ∗ ) = (B ′′ , κ1 ), where κ1 denotes the Ziegler multiplicity on B ′′ . Since
B ′′ ∼
= A (G(r, 1, ℓ − 2)), by (3.12), it follows from our induction hypothesis that (B ′′ , κ1 ) is
inductively free. Consequently, since the simple arrangement A ′′ is inductively free, thanks
to Theorem 2.3, it follows that (A ′′ , κ) is inductively free, by a repeated application of the
addition part of Theorem 2.12.
Moreover, Corollaries 3.3 and 3.9, imply the following consequence of Theorem 2.4.
Corollary 3.14. Let W be of type F4 , H4 , G31 or G32 . Then (A (W )′′ , κ) is inductively free.
Proof. By Proposition 3.10, in each of the given cases the condition in Corollary 3.9 on the
exponents of A (W )′′ is satisfied, so that every chain of multiplicities between 11 and κ on
A (W )′′ is free. By Theorem 2.4, A (W )′′ is inductively free in each instance. Therefore, it
follows from Corollary 3.3 that also (A (W )′′ , κ) is inductively free.
4. Proofs of Theorems 1.4 and 1.9
Throughout, W denotes a complex reflection group. Owing to [HR15, Prop. 2.10] and
Theorem 2.23, we may assume that W is irreducible. Fix H0 ∈ A and let A ′′ be the
restriction of A with respect to H0 . Moreover, fix m0 ∈ Z≥1 and let δ = δH0 ,m0 as in
Definition 1.6. We prove Theorems 1.4 and 1.9 more less simultaneously making repeated
use of Corollary 2.20.
4.1. Groups of low rank. It is immediate from [Z89, Cor. 7] that the reverse implication
in Theorem 1.4 holds provided W has rank at most 3. Likewise Theorem 1.9 holds for W of
rank at most 2.
14
4.2. Coxeter Groups. Let W be an irreducible Coxeter group. By [BC12], every Coxeter
arrangement is inductively free. So once we have shown that (A (W )′′, κ) is inductively free,
so is (A (W ), δ), thanks to Corollary 2.20.
Lemma 4.1. Let A = A (Aℓ−1 ) be the Coxeter arrangement of type Aℓ−1 . Then both (A ′′ , κ)
and (A , δ) are inductively free.
Proof. For ℓ ≤ 4, the result follows from §4.1 and Corollary 2.21. So suppose that ℓ ≥ 5.
Since the underlying Coxeter group is transitive on A , there is no harm in fixing H0 =
ker(x1 − x2 ). Thanks to [OT92, Prop. 6.73], A ′′ ∼
= A (Aℓ−2 ). One readily checks that
Ziegler’s multiarrangement (A (Aℓ−2 ), κ) has defining polynomial
Y
Y
(xi − xj ).
(x1 − xj )2
Q(A (Aℓ−2 ), κ) =
2≤i<j≤ℓ−2
2≤j≤ℓ−2
It follows from [CR16, Cor. 1.2] that (A (Aℓ−2 ), κ) is inductively free. Since A is inductively
free, it follows from Corollary 2.20 that (A , δ) is inductively free.
The following is the special case r = 2 in Proposition 3.11.
Lemma 4.2. Let A = A (Bℓ ) be the Coxeter arrangement of type Bℓ . Then both (A ′′ , κ)
and (A , δ) are inductively free.
Lemma 4.3. Let A = A (Dℓ ) be the Coxeter arrangement of type Dℓ for ℓ ≥ 3. Then both
(A ′′ , κ) and (A , δ) are inductively free.
Proof. We first prove that (A ′′ , κ) is inductively free. We argue by induction on ℓ ≥ 3. For
ℓ = 3, we have A (D3 ) = A (A3 ) and so then the result follows from Lemma 4.1.
Now let W = W (Dℓ ) for ℓ ≥ 4 and suppose that the result holds for all instances of smaller
rank. Since W is transitive on A , we may fix H0 = ker(x1 − x2 ) without loss. It follows
1
from [OT92, Prop. 6.85] that A ′′ = Aℓ−1
(2).
One readily checks that Ziegler’s multiarrangement (A (Dℓ )′′ , κ) has defining polynomial
Y
Y
1
Q(A (Dℓ )′′ , κ) = Q(Aℓ−1
(2), κ) = x1
(x21 − x2j )2
(x2i − x2j ).
2≤j≤ℓ−1
3≤i<j≤ℓ−1
′′
According to Theorem 1.2, (A (Dℓ ) , κ) is free with
(4.4)
1
exp(A (Dℓ )′′ , κ) = exp(Aℓ−1
(2), κ) = {3, 5, . . . , 2(ℓ − 3) + 1, ℓ − 1}.
By our induction hypothesis and Theorem 2.23, the sub-multiarrangement (A (Dℓ−1 )′′ , κ) ×
Φ1 of (A (Dℓ )′′ , κ) is inductively free. By (4.4), its set of exponents is
1
exp ((A (Dℓ−1 )′′ , κ) × Φ1 ) = exp (Aℓ−2
(2), κ) × Φ1 = {3, 5, . . . , 2(ℓ − 4) + 1, ℓ − 2, 0}.
This forms the start of our induction table.
We first add and restrict to the hyperplanes ker(x1 − xℓ−1 ) and ker(x1 + xℓ−1 ) twice. Here
we use Proposition 2.13 in order to determine the Euler multiplicity on the restrictions.
We consider the special case when ker(x1 + xℓ−1 ) is added and restricted to for the second
time. Then in the notation of Proposition 2.13, when considering the localization AX =
15
{ker(x1 + xℓ−1 ), ker(x1 − xℓ−1 ), ker x1 }, we have |AX | = 3, |µX | = 5 and so Proposition
2.13(2) applies.
1
Note that each of the restrictions is isomorphic to (Aℓ−2
(2), κ) which is inductively free, by
our inductive hypothesis. Consequently, since in each case both (A ′ , µ′ ) and (A ′′ , µ∗ ) are
inductively free and exp(A ′′ , µ∗ ) ⊆ exp(A ′ , µ′ ), it follows that also (A , µ) is inductively
free.
exp(A ′ , µ′ )
αH
exp(A ′′ , µ∗ )
x1 − xℓ−1
1 (2), κ)
exp(Aℓ−2
{3, 5, . . . , 2(ℓ − 4) + 1, ℓ − 2, 1}
x1 + xℓ−1
1 (2), κ)
exp(Aℓ−2
{3, 5, . . . , 2(ℓ − 4) + 1, ℓ − 2, 2}
x1 − xℓ−1
1 (2), κ)
exp(Aℓ−2
{3, 5, . . . , 2(ℓ − 4) + 1, ℓ − 2, 3}
x1 + xℓ−1
1 (2), κ)
exp(Aℓ−2
{3, 5, . . . , 2(ℓ − 4) + 1, ℓ − 2, 4}
..
.
x2 − xℓ−1
..
.
1 (2), κ)
exp(Aℓ−2
..
.
{3, 5, . . . , 2(ℓ − 4) + 1, 2(ℓ − 3) + 1, ℓ − 2}
1 (2), κ)
xℓ−2 + xℓ−1 exp(Aℓ−2
1 (2), κ) × Φ
exp (Aℓ−2
1
= {3, 5, . . . , 2(ℓ − 4) + 1, ℓ − 2, 0}
exp(A (Dℓ )′′ , κ) = {3, 5, . . . , 2(ℓ − 3) + 1, ℓ − 1}
Table 1. Induction Table for (A (Dℓ )′′ , κ).
It follows from Table 1 that (A (Dℓ )′′ , κ) is inductively free. Since A itself is inductively
free, so is (A , δ), according to Corollary 2.20.
Lemma 4.5. Let W be an irreducible Coxeter group of exceptional type of rank at least 3
and let A = A (W ) be the Coxeter arrangement of W . Then both (A ′′ , κ) and (A , δ) are
inductively free.
Proof. For W of type H3 the result follows from Corollary 2.21.
For the remaining instances, we first prove that (A ′′ , κ) is inductively free. For W of type
F4 and H4 , this follows from Corollary 3.14.
Since W (E6 ) and W (E7 ) are parabolic subgroups of W (E8 ), we obtain the result for the
former from the latter, thanks to [OT92, Cor. 6.28], Lemma 2.16(ii), and Theorem 2.22.
The case for W = W (E8 ) is considerably more complicated. Since W is transitive on A ,
we may take H0 = ker x. It follows from Proposition 3.10 that (A ′′ , ν) is free for every
multiplicity ν with 11 ≤ ν ≤ κ. So we pick a filtration 11 = ν1 < ν2 . . . < ν28 < ν29 =
κ of free multiplicities. Since B := A ′′ is inductively free, by [BC12], it follows from
Lemma 3.2 that (A ′′ , κ) is inductively free, once we have established that the corresponding
multirestrictions (B ′′ , νi∗ ) are inductively free. Moreover, each multiplicity νi∗ is just Ziegler’s
canonical multirestriction κ1 stemming from the simple arrangement B, by Corollary 3.8.
Let C := B ′′ . Since C is inductively free, by [BC12], we can build an induction table by
16
starting with the simple arrangement C and can increasing the multiplicities all the way to
(C , κ1 ). This is carried out in Table 2, where we indicate how to increase the multiplicities on
the hyperplanes of the inductively free arrangement C such that each resulting multiplicity
is again a free multiplicity for C . Note that in the first 16 steps of the table the Euler
multiplicity µ∗ of the multirestriction is just Ziegler’s canonical restriction on the simple
arrangement C . The remaining two Euler multiplicities are small extensions of Ziegler’s
canonical multiplicity. In particular, here the order in which the hyperplanes are added
plays a crucial role. There are precisely 3 different rank 5 multiarrangements that occur
as restrictions in Table 2. We have checked that each of them is again inductively free, as
required. However, we omit to include this data. We also do not give linear forms for αH in
the table, but give the number of the induction step at the beginning of each row.
Step exp(A ′ , µ′ )
αH
exp(A ′′ , µ∗ )
...
..
.
{7, 11, 13, 14, 17}
..
.
15 {7, 11, 13, 14, 16, 17}
...
{7, 11, 13, 14, 17}
16 {7, 11, 13, 14, 17, 17}
..
.
...
..
.
{7, 11, 13, 17, 17}
..
.
20 {7, 11, 13, 17, 17, 18}
...
{7, 11, 13, 17, 17}
21 {7, 11, 13, 17, 17, 19}
..
.
...
..
.
{7, 11, 13, 17, 19}
..
.
26 {7, 11, 13, 17, 19, 22}
...
{7, 11, 13, 17, 19}
0 exp(C ) = {1, 7, 11, 13, 14, 17}
..
.
exp(C , κ1 ) = {7, 11, 13, 17, 19, 23}
Table 2. Induction Table for (C , κ1 ).
We mention that the only types of restrictions which do occur throughout this induction
for (A ′′ , κ) are (E8 , A2 ) in rank 6, (E8 , A3 ) in rank 5, (E8 , A4 ) and (E8 , D4 ) in rank 4
and (E8 , A1 × D5 ), (E8 , D4 ) and (E8 , A5 ) in rank 3, where we use the notation to indicate
restrictions following [OS82, §3, App.] (cf. [OT92, §6.4, App. C]).
Finally, since (A ′′ , κ) is inductively free in each instance, it follows again from Corollary 2.20
that also (A , δ) is inductively free.
4.3. The Monomial Groups G(r, 1, ℓ). For W = G(r, 1, ℓ) for r, ℓ ≥ 2, the result follows
from Proposition 3.11.
4.4. Non-real inductively free reflection arrangements of exceptional type. Let W
be irreducible of rank at least 3 of exceptional type so that A = A (W ) is inductively free.
It then follows from Theorem 2.3 that W equals G25 , G26 , or G32 .
17
Lemma 4.6. Let A = A (W ) be the reflection arrangement where W is one of G25 , G26 ,
or G32 . Then both (A ′′ , κ) and (A , δ) are inductively free.
Proof. Since G25 and G26 are both of rank 3, it follows from Corollary 2.21 that then
(A (W ), δ) is inductively free. Since A ′′ is of rank 2, also (A ′′ , κ) is inductively free.
For W = G32 it follows from Corollary 3.14 that (A ′′ , κ) is inductively free. Owing to
Corollary 2.20, also (A , δ) is inductively free.
4.5. Non-inductively free reflection arrangements. Suppose that W is irreducible so
that A = A (W ) is not inductively free. Then, according to Theorem 2.3, either W =
G(r, r, ℓ), for r, ℓ ≥ 3 or W is of exceptional type G24 , G27 , G29 , G31 , G33 , or G34 .
The following is [HRS17, Prop. 2], its proof consists of an application of Theorem 2.22.
Lemma 4.7. Let A = A (W ) be the reflection arrangement of the monomial group W =
G(r, r, ℓ) for r ≥ 3 and ℓ ≥ 5. Then both (A ′′ , κ) and (A , δ) are not inductively free.
Lemma 4.8. Let W be G(r, r, 3), G(r, r, 4), for r ≥ 3 or of exceptional type G24 , G27 , G29 ,
G31 , G33 , or G34 . Then (A (W ), δ) is not inductively free.
Proof. Let A = A (W ) be the reflection arrangement for W as in the statement. We argue
by induction on m0 ≥ 1. For m0 = 1, the result follows from Theorem 2.3. So suppose
that m0 > 1 and that the claim is proved for smaller values for m0 . Pick H ∈ A and let
(A , A ′ , A ′′ ) be the triple of A with respect to H. If H = H0 , then (A ′ , δ ′ ) = (A , δH0 ,m0 −1 )
which is not inductively free by inductive hypothesis.
Now suppose that H 6= H0 . In each case apart from G31 it follows from [HR15, Cor. 1.3,
Cor. 2.4] that A ′ = A \ {H} is not free for any choice of hyperplane. Therefore, in each
of these instances, since A ′ is not free, neither is (A ′ , δ ′ ), by Theorem 1.7. In particular,
(A ′ , δ ′ ) is not inductively free. Thus (A ′ , δ ′ ) is not inductively free for any choice of H and
therefore, neither is (A , δ).
For W of exceptional type G31 , we argue as follows. Suppose that (A , δ) is inductively
free. Let A ′ = A \ {H} and (A ′ , δ ′ ) = (A ′ , δ|A ′ ). If (A ′ , δ ′ ) is free, then so is A ′ ,
again thanks to Theorem 1.7. Continuing in this fashion removing further hyperplanes from
the multiarrangement, we obtain a free chain of consecutive subarrangements of the simple
arrangement A (G31 ) ending at the empty 4-arrangement. But this contradicts the fact that
there is no such free chain of A (G31 ), according to the proof of [HR15, Lem. 3.5].
Lemma 4.9. Let W be G(r, r, 3), G(r, r, 4), for r ≥ 3 or W is exceptional of type G29 or
G31 . Then (A (W )′′ , κ) is inductively free.
Proof. For W = G(r, r, 3) the result is clear, thanks to Remark 2.18, and for W = G31 the
result follows from Corollary 3.14.
If W = G(r, r, 4) or W is of type G29 , then A (W )′′ is inductively free, according to Theorem
2.4 (W is transitive on A (W ) in each of these cases). Consequently, we may initialize an
induction table in these instances with the simple inductively free arrangement A (W )′′ and
add hyperplanes to reach the Ziegler multiplicity κ.
18
Let W = G(r, r, 4). Then, by [OT92, Prop. 6.82], for any choice H0 ∈ A , we have A (W )′′ ∼
=
A31 (r) which is inductively free, according to Theorem 2.4(ii). Fix H0 = ker(x1 −x4 ). Thanks
to [OT92, Cor. 6.86], we have
exp A (G(r, r, 4)) = {1, r + 1, 2r + 1, 3(r − 1)}
and therefore, by Theorem 1.2, (A (G(r, r, 4))′′ , κ) is free with
exp(A (G(r, r, 4))′′ , κ) = {r + 1, 2r + 1, 3(r − 1)}.
So in particular, |κ| = 6r − 1. Moreover, by [OT92, Prop. 6.85], we have
exp A (G(r, r, 4))′′ = exp A31 (r) = {1, r + 1, 2r − 1}.
The defining polynomial for (A (G(r, r, 4))′′ , κ) is given by
Q(A (G(r, r, 4))′′ , κ) = Q(A31 (r), κ) = x1r−1 (xr1 − xr2 )2 (xr1 − xr3 )2 (xr2 − xr3 ) .
Table 3 is the induction table for (A (G(r, r, 4))′′ , κ). Note that since A (G(r, r, 4))′′ is of
rank 3, each of the restrictions in Table 3 is of rank 2 and so is inductively free. Let ζ be a
primitive r th root of unity.
There are two different types of steps in this induction table. The first one consists of adding
a hyperplane ker(x1 −ζ i x2 ) for some i or ker(x1 −ζ j x3 ) for some j and the second one consists
of adding ker x1 .
In the first step, our defining polynomial is of the form
Y
Q(A , µ) = x1 (xr2 − xr3 )
(x1 − ζ i x2 )ki (x1 − ζ i x3 )li
0≤i<r
with ki , li ∈ {1, 2} for all i.
P
We claim that (A , µ) is free with exp(A , µ) = {1 + 0≤i<r (ki + li − 2), r + 1, 2r − 1}. This
is true if ki = li = 1 for all i, since then (A , µ) is the simple arrangement A31 (r). For
each hyperplane in this step the argument is essentially the same, so we only present it for
H = ker(x1 − ζ i x2 ). Fix i and let ki = 1. The defining polynomial of the restriction is
Y
Q(A H , µ∗ ) = xr1
(x1 − ζ j x3 )2 .
0≤j<r
The Euler multiplicities can be computed using Proposition 2.13(2) and (3). A basis of
D(A H , µ∗) is given by
r
r+1
θ1 = rxr+1
1 D1 + ((r + 1)x1 x3 − x3 )D3 ,
θ2 = x1r−1 (rx1 x3r−1 D1 + (xr1 + (r − 1)xr3 ))D3 .
Therefore, we have exp(A H , µ∗ ) = {r + 1, 2r − 1}. The Addition Deletion Theorem 2.12
then yields that (A , µ) is free with the desired exponents.
In the second step we consider defining polynomials of the form
Q(A , µ) = xi1 (xr1 − xr2 )2 (xr1 − xr3 )2 (xr2 − xr3 ),
for 1 ≤ i ≤ r − 2.
19
We claim that (A , µ) is free with exp(A , µ) = {r + 1, 2r + 1, 2r − 1 + i}. We know from the
first step that this is true for i = 1. Let H = ker x1 . The restriction is given by
r+1 r
r
Q(A H , µ∗ ) = xr+1
2 x3 (x2 − x3 ).
Note that (A H , µ∗ ) coincides with (A (G(r, 1, 3))ker x1 , κ). Therefore, thanks to Theorem 1.2,
(A H , µ∗ ) is free and exp(A H , µ∗ ) = {r + 1, 2r + 1}. Since the hypotheses of Theorem 2.12
are fulfilled, we get freeness of (A , µ) along with the desired exponents for each i.
This completes the argument for W = G(r, r, 4). We record the details of the induction in
Table 3.
exp(A ′ , µ′ )
αH
exp(A ′′ , µ∗ )
exp A31 (r) = {1, r + 1, 2r − 1}
x1 − ζx2
{r + 1, 2r − 1}
{2, r + 1, 2r − 1}
..
.
x1 − ζ 2 x2
..
.
{r + 1, 2r − 1}
..
.
{r, r + 1, 2r − 1}
x1 − x2
{r + 1, 2r − 1}
{r + 1, r + 1, 2r − 1}
..
.
x1 − ζx3
..
.
{r + 1, 2r − 1}
..
.
{r + 1, 2r, 2r − 1}
x1 − x3
{r + 1, 2r − 1}
{r + 1, 2r + 1, 2r − 1}
..
.
x1
..
.
{r + 1, 2r + 1}
..
.
{r + 1, 2r + 1, 2r − 1 + (r − 3)}
x1
{r + 1, 2r + 1}
exp(A (G(r, r, 4))′′ , κ) = {r + 1, 2r + 1, 3(r − 1)}
Table 3. Induction Table for (A (G(r, r, 4))′′ , κ).
Next let W be of type G29 . Since W is transitive on A (G29 ), we may take H0 = ker x4 .
Then the defining polynomial for (A (G29 )′′ , κ) is given by
Q(A (G29 )′′ , κ) = (x1 − x2 − ix3 )2 (x1 − x2 )(x2 − x3 )x33 (x1 − ix2 − x3 )2 (x1 − x2 + ix3 )2
(x1 − x3 )(x1 + ix2 − ix3 )2 x32 (x1 + ix2 − x3 )2 (x1 − ix2 + ix3 )2
x31 (x1 − ix2 + x3 )2 (x1 − ix2 − ix3 )2 (x1 + ix2 + ix3 )2 (x1 + ix2 + x3 )2
(x1 + x2 − ix3 )2 (x1 + x2 + ix3 )2 (x2 + x3 )(x1 + x3 )(x1 + x2 ).
We present an induction table for (A (G29 )′′ , κ) in Table 4. Note that since A (G29 )′′ is of
rank 3 each of the restrictions in Table 4 is of rank 2 and so is inductively free.
By the data in Tables 3, 4 and Remark 2.19, (A ′′ , κ) is inductively free in each instance.
Observe it follows from Theorem 2.3 and Lemmas 4.8 and 4.9 that the implication in Corollary 2.20 fails without the assumption that A ′′ is inductively free.
20
exp(A ′ , µ′ )
αH
exp(A ′′ , µ∗ ), µ∗
exp A (G29 )′′ = {1, 9, 11}
x1 − x2 − ix3
{9, 11}, (1, 2, 2, 4, 2, 3, 4, 2)
{2, 9, 11}
x3
{9, 11}, (3, 4, 3, 4, 3, 3)
{3, 9, 11}
x3
{9, 11}, (3, 4, 3, 4, 3, 3)
{4, 9, 11}
x1 − ix2 − x3
{9, 11}, (2, 1, 2, 2, 4, 3, 4, 2)
{5, 9, 11}
x1 − x2 + ix3
{9, 11}, (1, 2, 2, 3, 2, 4, 4, 2)
{6, 9, 11}
x1 + ix2 − ix3 {9, 11}, (2, 2, 3, 2, 1, 4, 2, 4)
{7, 9, 11}
x2
{9, 11}, (3, 4, 3, 4, 3, 3)
{8, 9, 11}
x2
{9, 11}, (3, 4, 3, 4, 3, 3)
{9, 9, 11}
x1 + ix2 − x3
{9, 11}, (2, 2, 2, 3, 2, 1, 4, 4)
{9, 10, 11}
x1 − ix2 + ix3 {9, 11}, (2, 2, 3, 2, 1, 4, 2, 4)
{9, 11, 11}
x1
{9, 11}, (3, 3, 3, 3, 4, 4)
{9, 11, 12}
x1
{9, 11}, (3, 3, 3, 3, 4, 4)
{9, 11, 13}
x1 − ix2 + x3
{9, 13}, (2, 2, 2, 5, 1, 2, 3, 5)
{9, 12, 13}
x1 − ix2 − ix3 {9, 13}, (2, 1, 2, 5, 3, 5, 2, 2)
{9, 13, 13}
x1 + ix2 + ix3 {9, 13}, (2, 1, 2, 5, 3, 5, 2, 2)
{9, 13, 14}
x1 + ix2 + x3
{9, 13}, (2, 5, 2, 5, 2, 3, 1, 2)
{9, 13, 15}
x1 + x2 − ix3
{9, 13}, (2, 2, 3, 1, 2, 5, 5, 2)
x1 + x2 + ix3
{9, 13}, (1, 2, 2, 5, 2, 3, 5, 2)
{9, 13, 16}
exp(A (G29
)′′ , κ)
= {9, 13, 17}
Table 4. Induction Table for (A (G29 )′′ , κ).
Lemma 4.10. Let W be of type G33 or G34 . Then (A (W )′′ , κ) is not inductively free.
Proof. First consider the case of W = G33 . Let A = A (W ). Direct computations have
shown that there is no chain of free multiplicities from the zero multiplicity on A ′′ to κ. The
maximal length of such chains is 22 < 44 = |κ|. Hence (A ′′ , κ) is not inductively free.
Finally, let W = G34 and A = A (W ). Then W admits a parabolic subgroup WX of
type G(3, 3, 5), by [OT92, Table C.15]. Thanks to [OT92, Cor. 6.28], we have A (WX ) =
A (W )X = AX . Fix H0 ∈ AX . Then considering the restriction with respect to H0 , we
have (A ′′ )X = (AX )′′ . So we see that ((A ′′ )X , κX ) = ((AX )′′ , κX ) = (A (G(3, 3, 5))′′, κ), by
Lemma 2.16(ii). According to Lemma 4.7, the latter is not inductively free, thus neither is
(A (G34 )′′ , κ), owing to Theorem 2.22.
Finally, Theorems 1.4 and 1.9 follow from the results from §4.1 to §4.5 above.
Remark 4.11. The fact that (A (G33 )′′ , κ) is not inductively free, proved in Lemma 4.10,
and the data in the induction tables 2 and 4, were obtained by computational means.
Specifically, we retrieved relevant data about reflection groups from some GAP code provided
by J. Michel [M15]; see also [S+ 97] and [GHL+ 96]. Subsequently, we made use of SAGE,
21
cf. [Ste15], and SINGULAR, cf. [GPS09], in order to calculate the various bases of modules
of derivations involved.
Acknowledgments: We acknowledge support from the DFG-priority program SPP1489
“Algorithmic and Experimental Methods in Algebra, Geometry, and Number Theory”. Part
of the research for this paper was carried out while the authors were staying at the Mathematical Research Institute Oberwolfach supported by the “Research in Pairs” programme.
References
[ATW08] T. Abe, H. Terao, and M. Wakefield, The Euler multiplicity and addition-deletion theorems for
multiarrangements. J. Lond. Math. Soc. (2) 77 (2008), no. 2, 335–348.
[AY13]
T. Abe and M. Yoshinaga, Free arrangements and coefficients of characteristic polynomials, Math.
Z. 275 (2013), no. 3-4, 911–919.
[AHR14] N. Amend, T. Hoge and G. Röhrle, On inductively free restrictions of reflection arrangements, J.
Algebra 418 (2014), 197–212.
[BC12]
M. Barakat and M. Cuntz, Coxeter and crystallographic arrangements are inductively free, Adv.
Math 229 (2012), 691–709.
[CR16]
H. Conrad and G. Röhrle, On inductively free multiderivations of braid arrangements,
Ann. Comb., 20, (2016), 719–735
[GHL+ 96] M. Geck, G. Hiß, F. Lübeck, G. Malle, and G. Pfeiffer, CHEVIE — A system for computing and
processing generic character tables, Appl. Algebra Engrg. Comm. Comput. 7 (1996), 175–210.
[GPS09] G.-M. Greuel, G. Pfister, and H. Schönemann, Singular 3-1-1, A Computer Algebra System for
Polynomial Computations, Centre for Computer Algebra, University of Kaiserslautern, 2009.
[HR15]
T. Hoge and G. Röhrle, On inductively free reflection arrangements, J. Reine u. Angew. Math.
701 (2015), 205–220.
[HRS17] T. Hoge, G. Röhrle and A. Schauenburg, Inductive and Recursive Freeness of Localizations of multiarrangements, in: Algorithmic and Experimental Methods in Algebra, Geometry, and Number
Theory, Springer Verlag.
[M15]
J. Michel, The development version of the CHEVIE package of GAP3. J. Algebra, 435 (2015),
308–336.
[OS82]
P. Orlik and L. Solomon, Arrangements Defined by Unitary Reflection Groups, Math. Ann. 261,
(1982), 339–357.
[OT92]
P. Orlik and H. Terao, Arrangements of hyperplanes, Springer-Verlag, 1992.
[S+ 97]
M. Schönert et al., GAP – Groups, Algorithms, and Programming – version 3 release 4, 1997.
[ST54]
G.C. Shephard and J.A. Todd, Finite unitary reflection groups. Canadian J. Math. 6, (1954),
274–304.
[S12]
M. Schulze, Freeness and multirestriction of hyperplane arrangements. Compos. Math. 148
(2012), no. 3, 799–806.
[Ste15]
W. A. Stein, et al, Sage Mathematics Software (Version 7.5.1), The Sage Development Team,
2015, http://www.sagemath.org.
[Ste60]
R. Steinberg, Invariants of finite reflection groups, Canad. J. Math. 12, (1960), 616–618.
[T80a]
H. Terao, Arrangements of hyperplanes and their freeness I, II, J. Fac. Sci. Univ. Tokyo 27 (1980),
293–320.
, Free arrangements of hyperplanes and unitary reflection groups. Proc. Japan Acad. Ser.
[T80b]
A Math. Sci. 56 (1980), no. 8, 389–392.
[Y04]
M. Yoshinaga, Characterization of a free arrangement and conjecture of Edelman and Reiner.
Invent. Math. 157 (2004), no. 2, 449–454.
, On the freeness of 3-arrangements. Bull. London Math. Soc. 37 (2005), no. 1, 126–134.
[Y05]
22
[Y14]
[Z89]
, Freeness of hyperplane arrangements and related topics, Ann. Fac. Sci. Toulouse Math.
(6) 23 (2014), no. 2, 483–512.
G. Ziegler, Multiarrangements of hyperplanes and their freeness. Singularities (Iowa City, IA,
1986), 345–359, Contemp. Math., 90, Amer. Math. Soc., Providence, RI, 1989.
Institut für Algebra, Zahlentheorie und Diskrete Mathematik, Fakultät für Mathematik
und Physik, Leibniz Universität Hannover, Welfengarten 1, 30167 Hannover, Germany
E-mail address: [email protected]
Fakultät für Mathematik, Ruhr-Universität Bochum, D-44780 Bochum, Germany
E-mail address: [email protected]
23
| 4math.GR
|
Deriving Ontologies from XML Schema
Ivan Bedini*, Georges Gardarin**
Benjamin Nguyen**
*Orange Labs, Site of Caen - 14000, France
[email protected]; http://www.orange.com
**Prism, Université de Versailles St-Quentin, 78035 Versailles Cedex 035
<prenom>.<nom>@prism.uvsq.fr; http://www.prism.uvsq.fr
Abstract. In this paper, we present a method and a tool for deriving a skeleton
of an ontology from XML schema files. We first recall what an is ontology and
its relationships with XML schemas. Next, we focus on ontology building
methodology and associated tool requirements. Then, we introduce Janus, a
tool for building an ontology from various XML schemas in a given domain.
We summarize the main features of Janus and illustrate its functionalities
through a simple example. Finally, we compare our approach to other existing
ontology building tools.
1 Introduction
Ontologies appear as useful building blocks in several domains including the Semantic
Web, data source integration, data visualization and zooming, text indexing and mining, etc.
In Data Warehouse and On Line Analysis, ontologies are particularly useful for integrating
multiple data sources and for zooming on data cube dimensions. They provide a deep understanding of data by composing and transforming it automatically. Though helpful and trendy,
ontologies are still hard to build, to maintain and make evolve, notably when large (e.g.,
more than thousands of concepts).
The “nuts and bolts” of ontologies are concepts and attributes with is-a and part of relationships. In that respect, ontologies are similar to object models of specific domains. XML
schemas also have similarities with object models. While there exist many XML schemas in
certain domains (eg., B2B e-commerce are at standardization level), there are only very few
ontologies. In this paper, we propose a methodology and a tool for semi-automatic derivation
of ontologies from XML schemas. We present the main features of our tool called Janus and
illustrate it through a simple example. Janus is unique in the sense that it mixes several technologies (language analysis, text mining, schema validation, graphical representation) into an
incremental methodology of ontology construction and evolution. We further propose a classification of ontology building tools and their comparison in terms of functionalities.
The rest of this paper is organized in four main sections: i) ontology definition and requirements for building tools; ii) overview of our Janus ontology skeleton building tool; iii)
presentation of other tools; and iv) comparisons of tools.
EDA 2008
Deriving Ontologies from XML Schema
2 Ontologies : Definition and Requirements
2.1 Definition
An ontology is a common abstract and simplified view of an application domain, e.g., finance, tourism, medicine, transport, B2B, wine, etc. It defines the concepts, properties and
relationships of the field. It also aims at describing the objects and their relations according
to a common point of view to which all participants agree. Concepts are often assimilated to
object classes and properties to object facets. Among the facets, some are mono-valued while
others are multi-valued. A mono-valued property corresponds to a function having for domain a class and for image a data type or a class.
Formally, according to (Amann et al. 2000), an ontology is a triple (C, S, ISA) where:
(i) C is a set of classes c1, c2,…. cm ;
(ii) S is a set of properties s1, s2,… s3 ; a data property (integer, real, string, date, etc.) is
an attribute denoted by a name and having value in a base type; an object property represents
a binary relation role characterized by a name and a target class ;
(iii) ISA is a set of inheritance links defined between classes.
This definition can be extended with instances of classes, i.e., objects, and with constraints
between classes and between properties.
2.2 Ontology Construction Requirements
By building an ontology, we aim at constituting a knowledge representation integrating
all the information and the suitable rules of a domain. This is hardly possible without a
staged approach: an ontology must be developed progressively without disturbing existing
applications when a new version is adopted; a basic ontology must be enriched by successive
extensions. The construction technique must maintain discovered extensions and use them to
identify quickly similarities between new and old concepts. These characteristics, together
with the definition lead to the following ontology construction tool requirements: efficient
memorization of concepts/relationships for reuse with clear distinction of polyseme, good
support for dynamic evolution and resourceful automation of production.
The foster is an important issue because as pointed out in (Ehrig et al. 2004) and (Sabou
et al. 2006), time performance for computing data integration is a key problem when mapping ontologies. Thus, a solution able to store the best known concepts for a domain, should
generate better reusability of components by machines applying reasoning algorithms.
Also to maintain and enrich an ontology as automatically as possible is a fundamental requirement. Even in a specific field, the concepts handled by the applications can be numerous and the quantity of information which we wish to maintain for each concept is vast.
Solely relying on human management becomes impossible: for example, consider a corpus
made up of a thousand of files with concepts per hundreds in each file: this is a task which
can not be completely manually.
2.3 Ontology and Schema
XML schemas and ontologies in a given domains are somehow related. In general, schemas are built in a domain before ontologies. Consider for example the B2B domain: there
EDA 2008
I. Bedini et al.
exist hundreds of schemas to encode exchanged data but not many ontologies. To benefit
from preexisting schemas, we propose a method and a tool to derive an ontology or at least
an enriched taxonomy (i.e., a concept hierarchy with concepts properties and main relationships) from a set of XML schemas.
Let us give two examples of schemas that can be used in the wine domain. The problem
is how to derive the wine-tasting ontology from these schemas, or at least something resembling a subset of this ontology. From this first version, knowledge could be added (possibly
derived from new schemas) or updated (by an expert) to derive a satisfactory ontology.
Example 1 – Example of Schemas in the wine-tasting domain
FIG. 1- Wine Tasting XSD
FIG. 2- Wine Drinkers XSD
3 Deriving Ontologies from XML Schemas
3.1 Methodology
The aim of this paper is to provide a general view of the automation aspect of the ontology generation; thus, in the rest of this section, we provide some elements that we consider
crucial in order to achieve this result.
EDA 2008
Deriving Ontologies from XML Schema
Several methodologies for building ontologies exist, such as OTK (Sure et al. 2004),
METHONTOLOGY (Corcho et al. 2003) or DILIGENT (Vrandecic et al 2005), but they
target ontology engineers and not machines. We do not develop here a new methodology yet,
but we define the automatic ontology generation life cycle as a process composed of five
main steps necessary to achieve our goal. These steps represent the main tasks of the process
for building ontologies starting from an existing corpus, like XML schemas. We do not focus
on what techniques are available for each task, but mainly describe what we expect from
each task. The whole process is depicted in FIG. 3. The five steps are:
Analysis
information
Sources
Generation
Extraction
Validation
Evolution
FIG. 3 – Automatic ontology generation process
1. Extraction: This step deals with the acquisition of information
needed to generate the ontology (concepts, attributes, relationships and axioms) starting
from an existing corpus. Input resources can be of many types: structured, semi-structured or
unstructured. Techniques for information retrieval and extraction can be of different types,
such as: NLP (Natural Language Processing), word clustering, machine learning, semantic,
morphological, lexical approaches, and more often a combination of them.
2. Analysis: This step focuses on the matching of retrieved information and/or alignment
of two or more existing ontologies, depending on the use case. This step requires techniques
already used in the first stage, as morphological and lexical analysis of labels; a semantic
analysis to detect synonyms, homonyms and other relations of this type; an analysis of concept structures to find hierarchical relationships and identify common attributes; techniques
based on reasoning to detect inconsistencies and induced relations.
3. Generation: This step deals with ontology merging, if appropriate, and the production
of a first version of the target ontology based on the tool formal meta-model, i.e., in a universal language interpretable by other applications, such as OWL and RDF/S.
4. Validation: All previous steps may introduce wrong concepts or relationships, thus a
validation phase is needed. Conversely, a validation task can be introduced at the end of each
previous step. Validation is often done by hand, but can sometimes be automated.
5. Evolution: An ontology is not a static description of a domain, but with the evolution
of applications, in quality and number, the ontology may also require some changes. Some
concepts, as well as properties, relationships, and other parameters, can be added or modified. Incremental construction should be supported efficiently and could integrate a new step
of information extraction, if new resources are not yet integrated in ontology format. It could
also directly provide new matches and alignments by integrating a new analysis step.
3.2 Deriving Concepts and Properties
Unlike simple text documents, XML documents provide likely annotated text with important information about objects and their structures, thus concepts for the ontology to build.
EDA 2008
I. Bedini et al.
Considering the examples presented in FIG. 1, it is pretty simple to imagine equivalences
between OWL classes and XSD elements, like Wine or Drinker; between OWL data-type
properties and XSD element pointing to a simple type, like Year or Name. Also we can retrieve information about relationships like IS A (e.g.: an Owner is a Person), thus provide a
classification of concepts, etc….
We can formalize ontology learning concepts and properties from XSD for as follows:
Definition 1. Given a set of XSD files X as input source, it is possible to retrieve a complete set of concepts and properties O by a surjective mapping m1 (based on XML mining
techniques), m : X → O, that we call domain conceptualization.
Definition 2. A concept is the basic element of O and is defined as a quadruple c = <L,
Hc, Rc, I> where:
• the label L is a common word (simple or compound) that best represents the concept. It is selected from a close set of names extracted from X that can be associated
to the concept (e.g.: owner and person).
• Hc is the set of structural relationship, which correspond to the subsumption hierarchy (fundamentally IS A relationship and propertyOf).
• Rc is the set of relations partitioned to two subsets. One is the set of all assertions in
which the relation is a semantic relation (like synonymy) and the other one is the set
of all assertions in which the relation is a non-semantic relation (like usage for
common label l abbreviations, or proved similarities such as the correspondence
Person.Name => Person.FirstName + Person.LastName).
• I is the set of originating instances of a concept, a link to the source.
Depending on domain usages it is important to know if a concept represents a detail of
the corpus source or a main concept. For example when looking for characteristic concepts
of the wine ontology, we can state that wine and person best represent the domain than status
or address, because in this context the latter are properties. More precisely we are able to
specialize concepts in order to obtain a more grained refinement of concept roles as follows:
Definition 3. Let C represent a set of concepts called the set of concept classes, C = {c1,
…, cm} a finite subset of O. A concept is considered to be a class, thus element of C if it has
more than one property. We say that a concept c∈C if P(c)= {c1, …, cm}, for m > 1, and of
course P(c) <> 0. P(c) is called the set of properties for a given concept c and c is a class.
Definition 4. Let P represent a set of concepts called the set of concepts properties a finite subset of O. A concept cp is considered a concept property if exists at least one super
class of which it is a property. A concept cp ∈ P if ∃ c ∈ C | cp ∈ P(c)
Definition 5. Let D represent a set of concepts called the set of printable concepts, also
called data-type concepts, a finite subset of O. A concept cdt is considered a data-type concept if it has no properties and it is directly related to a printable type (as defined by the
XML Schema built-in Datatype list2).
3.3 Deriving Some Rules
As seen in the previous section, it is possible to retrieve information about concepts from
XSD. For this it is necessary to provide a set of rules that we can summarize with the following principles:
1
A mapping from set A onto B is called surjective (or 'onto') if every member of B is the image of at least one
member of A. f : A → B is surjective if ∀b∈B (∃a∈A (f(a)=b))
EDA 2008
Deriving Ontologies from XML Schema
•
XSD complex types with complex content (i.e., a combination of attributes and sequence of elements like for Wine) produces concepts classes, otherwise printable types;
• XSD elements can assume different facets as simple properties if they point to a simple
type or a printable structure (like Name and Street), as classes if they declare a complex
content (like Drinker), as a specialization of a class or a role, if they are named elements
with declared type (like Owners);
Starting from these macro rules it is possible to build a complete list of rules for the surjective mapping (see Definition 1) as shown in TAB. 1. The first column lists main XSD
structures as defined by XML Schema primer recommendations, while the Mapping column
provides the corresponding domain conceptualization.
XSD Structure
xs:complexType
xs:simpleType
xs:extension et xs:restriction
Xs:union
xs:any
xs:simpleContent and xs:complexType with declared
xs:simpleContent
Element with attribute "ref" to xs:complexType
Named xs:element with attribute "type"
Named xs:element
xs:minOccurs, xs:maxOccurs
xs:sequence, xsd:all
Attributes of xs:element and xs:compleType
xs:choice
TAB. 1 - XSD Mining information
Mapping to O
Concept class
Concept datatype
Datatype property and is a relationship
ComplexType properties
Datatype property of the correspondent concept
Concept datatype
Concept class with propertyOf relationship
Concept class with Is a relationship
Concept class without attribute type
Respective cardinalities
Concept properties
Concept properties
Disjointness concepts
extraction and correspondent mapping
Working on XSD we just target TBox statements, but anyway we are also able to define
rules for detecting subsumption, equivalence, disjunction and classification relationships
between concepts. For example we can observe that the concept Drinker subsumes Person
because it is less general (in fact XSD file declares that it is an extension). Also when tasting,
Coca and Wine are disjoint classes (within the XSD the WineTasting element proposes a
choice between these two classes). We can classify concepts as classes or properties and look
for equivalences like Owner and Person.
When concepts are extracted following rules in TAB. 1, it is common to find different
formalizations of the same set of concepts from different sources. Then, we need to define
other rules to match similar concepts. Similarity between concepts can be of three types. The
first one uses the semantics of labels (synonyms, lexical, common terms for compound
words like PostalAddress or GeographicalAddress), the second is related to properties (like
two concepts with different label but exactly the same consistent set of properties) and the
last looks for common relationships between concepts, considered as the concept context.
To conclude we can state that the derivation of all these rules lets us to obtain a consistent
mapping producing the conceptualisation of a domain, a skeleton of the domain ontology.
3.4 The Janus Prototype
Despite the large adoption of XML files and schemas, only few tools provide advanced
ontology learning from this format. They are mostly limited to language translations (like
XML conversion to OWL), and mining of one file at a time. Real practices, like in the B2B
domain, show that each workgroup often defines several XSD instances for related domains
EDA 2008
I. Bedini et al.
(Bedini et al. 2008). This precludes the usage of existing tools to such use cases. This is the
motivation that brought us to develop a new tool.
The main idea of our approach is: i) for each file to capture concepts from words contained in XML tags; ii) to get relationships from XML structures; iii) to match concepts
similarities and; iv) to merge concepts to produce a general view of input sources. In fact,
this approach can be applied to different kinds of sources (like UML, WSDL, OWL…) because the conceptualization is general enough, but currently we focus on XML schemas.
3.4.1
Architecture
Our prototype, called Janus, follows the generation process defined in section 3.1 and
implements rules defined above in order to generate an ontology from XML heterogeneous
corpora as automatically as possible.
FIG. 4 – Janus overall architecture
FIG. 4 shows the overall architecture of Janus which is composed of three main steps.
The Extraction and Acquisition tasks provide the knowledge retrieval needed to generate
the ontology (concepts, properties and relationships). Implemented techniques for this task
are a combination of different types, such as: NLP (Natural Language Process) for morphological and lexical analysis, association mining for calculating term frequencies and association rules, semantics to detect synonymy, and clustering for grouping semantic and structural
similar concepts. We refer to the adaptation of these techniques as XML Mining. The Analysis step focuses on the matching of retrieved information and alignment of concepts issued
from different sources. This step requires techniques already used in the first stage, as syntax
and semantic measures, to establish the best similarities; it also requires an analysis of concept structures to determine hierarchical relationships and identify common properties. The
last step is composed of Merging and Generation tasks and looks for concepts with evident
affinities (e.g., concept fully included into another) to merge them. It transforms the metamodel used by the tool into a semantic network that can be described in RDFS or OWL. The
tool can derive from the network useful views provided to users, which can also step into the
process to parameterize thresholds for refining results.
3.4.2
Application to the Wine Example
The example above does not show great complexity due to the fact that only one description is provided for each concept and that semantics are well formed (only real words are
used). Nevertheless the tool is able to catch all concepts and relationships, in respect to humans, just one relationship is quite wrong. It is the plural form, provided by drinkers, this
EDA 2008
Deriving Ontologies from XML Schema
because the tool does not make difference between singular and plural yet. The table below
resumes concepts discovered and their relationships.
Classes
6 (wine_taste, wine
person drinker drink
address)
Object Properties of related
classes
12 (quantity, vineyard, year,
zip, status, city, coca, street,
boolean, liquid, name, owner)
TAB. 2 – Tasting Wine XSD
Object Datatypes of
related Properties
7 (string anyURI gYear
token integer byte number)
Relationships
Owner IS A Person
Drinker IS A Person
Coca DISJOINT Wine
Mining concept extraction
4 Overview of Ontology Building Tools
A brief glance at current solutions of automatic ontology building systems, sometimes referred to as Ontology Learning, shows many problems still need solving. In this section, we
cover a large spectrum of existing ontology building tools, starting with general articles.
(Mehrnoush, Abdollahzadeh, 2003) present a complete framework that classifies software and techniques for building ontologies in six main categories (called dimensions). It is a
detailed and interesting classification, but it focuses only on the learning method. (Buitelaar
et al., 2005) provide a comprehensive tutorial on learning ontology from text, which is really
useful, but the considered corpus source does not fit our use case. (Euzenat et al., 2004)
provide a detailed analysis of technical alignment for ontologies and a state of the art on
existing tools, probably the best known matching and alignment software, but they concentrate on the one task of aligning two existing ontologies, without investigating other steps in
the generation process, such as information extraction or the problem of multiple merging.
(Castano et al., 2007) are also limited to two existing ontologies.
Despite these documents, it remains difficult to clearly understand who makes what and
why. This is because frameworks designed for evaluation and analysis tend to organize
methods according to adopted technologies. Our approach wants to facilitate the understanding of what a method does within the ontology generation life cycle. The process defined in
Section 3.1 constitutes our base framework for evaluating software and experiments presented in this Section. It provides us elements to evaluate which part of the process can be
automated and how, as well as which techniques are more appropriate, and if they still require human intervention, thus further research. In this paper, we complement the documents
above, and overlap on tools that are closest to our interests.
4.1 Ontology Generator Classification
Ontology generation is mainly hand-made by domain experts, but this is of no interest to
us. In this paper we group experiences and software in four main categories as follow:
- Conversion or translation for applications that make the hypothesis that an ontology is
already well defined by someone or somewhere. The ontology format representation is wider
than other common knowledge representations, such as XML or UML. Software is produced
to compute this transformation. Experiments show that this approach presents a high degree
of automation, mainly because it does not address the whole problem of ontology generation.
- Mining based for applications implementing mining techniques in order to retrieve
enough information to generate an ontology. Most experiences are focused on unstructured
sources, like text documents and implement Natural Language Processing (NLP) techniques.
These experiences tell us that recovering structured concepts from unstructured documents
EDA 2008
I. Bedini et al.
still requires human assistance and that mining techniques from natural text can be used only
in complement with other existing structured knowledge representations techniques.
- External knowledge based for applications that build or enrich a domain ontology by
using an external resource. This category sometimes overlaps the mining based because
techniques applied to retrieve information can be the same, nevertheless we classify here
experiences with an approach closer to the integration of external dictionaries, existing ontology or from a more general knowledge resource, like WordNet (Miller, 1995) or the Web.
- Frameworks for applications with an approach based on integrating different modules.
As always when creating a classification, the border is not well defined, therefore we
classify works with respect to their automation approach rather than with regards to the techniques they implement. In fact we support the thesis that there is not a single technique to
develop, but that only an appropriate mix of techniques can bring us to our goal.
We will now describe the software and experiences, using our classification.
4.2 Conversion or Translation
The methods presented here assume that the source of the transformation is a complete
and well-defined ontology in XML or UML format.
(Bohring, Auer, 2005) have developed a tool that converts given XML files to OWL It is
based on the idea that items specified in the XSD file can be converted to ontology’s classes,
attributes and so on. Technically they have developed four XSLT instances to transform
XML files to OWL, without any other intervention on semantics and structures during the
transformation. This method has been applied to the Ontowiki platform (Auer et al., 2006).
Similarly, (Gasevic et al., 2004) propose using UML profiles to overcome UML’s limitations, also using XSLT instances to convert to OWL.
(Bertrand, et al., 2006) propose a semi-automatic process, starting from a UML class diagram representation of the ontology domain. A human selected part of the UML model is
then transformed into ODM (Ontology Definition Metamodel) as pivot model before automatically generating an RDFS file.
Within the PICSEL project, (Giraldo, Reynaud, 2002) have developed a semi-automatic
ontology generation software for the tourism industry domain extracting information contained in DTD files. This experience is interesting because it goes further, than the XML to
OWL transformation seen previously, and shows that tags and structure of XML files have
sufficient information to produce an ontology. Human intervention is needed to detect abbreviations or false positives during the ontology validation task. This experience is really close
to our use case but is limited to the sole domain of tourism; therefore the detection of relevant concepts does not produce conflicts between different representations.
4.3 Mining Based Approaches
The general drawback of these methods is the use of a more general ontology to define
concepts for the targeted ontology (for instance WordNet). Our feeling is that these approaches should be used in complement of the automatic ontology generation process.
(Biebow, Szulman, 1999) present the TERMINAE method and tool for building ontological models from text. The method is divided into 4 stages, corpus selection and organisation; linguistic analysis with the help of several NLP tools, such as LEXTER (Bourigault,
1996); normalization according to some structuring principles and criteria; formalization and
EDA 2008
Deriving Ontologies from XML Schema
validation. An expert is called to select the most important notions (concepts) for the targeted
ontology from the list of candidate terms extracted by the tool and to provide a definition of
the meaning of each term in natural language. These terms may or may not be inserted.
(Lonsdale et al., 2002) propose a process to generate domain ontologies from text documents called SALT. Their methodology requires the use of three types of knowledge
sources: one is a more general and well defined ontology for the domain, a dictionary or any
external resource to discover lexical and structural relationships between terms and a consistent set of training text documents. Using these elements they are able to automate the creation of a new sub-ontology of the more general ontology. User intervention is required at the
end of the process to remove false positives. The authors state that with a large set of training
documents their solution can achieve good results.
Similar to (Lonsdale et al., 2002), (Kitz et al., 2000) describe a generic approach for the
creation of an ontology for a domain based on a source with multiple entries which are: a
generic ontology to generate the main structure; a dictionary containing generic terms close
to the domain; and a textual corpus specific to the area to clean the ontology from wrong
concepts. This approach combines several input sources, for greater generality and better
reliability of the result. However, a manual check of the ontology must still be performed.
(Hu, Liu, 2004) have developed an automatic generation based on an analysis of a set of
texts followed by the use of WordNet. The analysis of the corpus retrieves words as concepts. These words are then searched in WordNet to find the concepts. The ontology generation seems to be one of the most automated, but no details of how the terms are extracted
from the body nor any qualitative assessment of the work are provided.
4.4 External Knowledge Retrieval
Also based on WordNet, (Moldova, Girju, 2000) expose a similar method but with the
difference that if a word is not found in WordNet then a supplementary module will look for
it over the Internet. Then linguistic and mining techniques extract new "concepts" to be
added to the ontology. User intervention is necessary here to avoid incongruous concepts.
(Aguirre et al., 2000) have developed a strategy to enrich existing ontologies using the
Web to acquire new information. The method takes as input a word which one wants to “improve” the knowledge of. WordNet is questioned about this word, and the different meanings
are used to generate queries for the web. For each query, that constitutes a “group”, different
search engines are queried. Terms frequencies are then calculated and compared with each
group, and of course the winning group, (i.e. sense), for the concept is the one with the highest frequencies. In addition a statistical analysis is performed on the result, in order to estimate the most common meaning of the concept. This method alone can not be adopted to
build ontologies, but it has the merit to be able to iterate with an external knowledge base to
provide further information that may be used for the validation task of an ontology in absence of human intervention.
(Cho et al. 2006) present the problem of proximity between two ontologies as a choice
between alignment and merging. The first case is limited to establishing links between ontologies while the second creates a single, new ontology. With their experience they directly
merge two ontologies based on WordNet. For this they use two approaches in their method
that they call the horizontal approach and the vertical approach. The horizontal approach first
checks fall the relationships between concepts of the “same level” in the two ontologies and
merges or ties them as defined by WordNet, while the vertical approach completes the merg-
EDA 2008
I. Bedini et al.
ing operation for concepts with “different levels”, but belonging to the same branch of the
tree. In this case they fill the resulting ontology with concepts from both ontologies and do
not make a choice. A similarity measure is calculated in order to define the hierarchy between these concepts in the resulting tree. This method, while not providing an adequate
solution to automation, does provide a purely semantic approach to the merging solution.
4.5 Frameworks
SymOntoX (Missikoff, Taglino, 2003) is an OMS (Ontology Management System) specialized in the e-business domain, which provides an editor, a mediator and a versioning
management system. The creation of the ontology is mainly done by an expert using the
editor, but the framework contains a first step towards an easier generation: it contains highlevel predefined concepts (such as Business Process, Business Object, Business Actor, etc.),
as well as different modules used for ontology mapping and alignment to simplify the work
of the expert. Ontology generation is not automatic, but merely assisted.
Protégé (Noy et al., 2000) is a free, open source, platform to design ontologies. It is supported by a strong community and experience shows that Protégé is one of the most widely
used platforms for ontology development and training. This software has an extensible architecture which makes it possible to integrate plug-ins. Some of these modules are interesting
and relevant in our case, like those from the PROMPT Suite (Noy, Musen, 2003). They automate, or at least assist, in the mapping, merging and managing of versions and changes. Also
the related project Protégé-OWL offers a Java API to manage OWL and RDF formats.
The glue between these pieces of software still remains human, yet program modules and
libraries provide a fundamental basis for developing the automation of ontology generation.
(Maedche, Staab, 2003) are contributors of several interesting initiatives within the ontology design field as well as the automation of this process, like MAFRA Framework, TextTo-Onto and KAON. In this paper we focus on their framework for ontology learning.
They propose a process that includes five steps: import, extraction, pruning, refinement,
and evaluation. This approach offers their framework a flexible architecture that consists of
many extensible parts, such as: a component to manage different input resources, capable of
providing information extraction from a large variety of formats (UML, XML, database
schema, documents text and web); a library of algorithms for acquiring and analyzing ontology concepts; a graphical interface that allows users to modify the generated ontology, but
also to choose which algorithms to apply and treatments to perform.
They bring together many algorithms and methods for ontology learning. Despite their
framework not allowing a completely automatic generation process, they are the only people
to propose a learning process close to a methodology of automatic ontology generation.
(Raghunathan, 2003) introduces the system LOGS (Lightweight universal Ontology Generation and operating architectureS). He states that generating ontology automatically from
text documents is still an open question. Their system is developed with a modular architecture that integrates the core functionality that can be expected by automatic ontology building software. It consists of the following modules: document source parser, NLP engine,
analyser, ontology engine, interface, integrator, ontological database and dictionary. It also
contains other modules able to crawl an intranet, refine the process of ontology design and a
module implementing trial and error iterative analysis of related texts to find patterns. No
qualitative analysis is provided, but the authors argue that they obtained significant results.
Unfortunately this software seems to have not met great consensus within the community.
EDA 2008
Deriving Ontologies from XML Schema
Extraction
- Human
Analysis
-
Generation
C – No merging.
Direct transformation using
XSLT files.
C – No merging.
Direct transformation using
XSLT files.
C – No merging.
Direct transformation using
XSLT files.
C – No standard
ontology representation
C – OWL format
Validation
- Human,
upstream to
the generation
- Human,
upstream to
the generation
- Human,
upstream to
the generation
- Human
Evolution
-
B – Static table of correspondences
-
UML2OWL
B
-
Semi-automatic
Ontology Building
from DTDs
Learning OWL
ontologies from free
texts
TERMINAE
C – automatic extraction
from DTD Sources
B – structure
analysis without
alignment
-
-
-
C – No standard
ontology representation
B – No standard
ontology representation
C – Automatic
merging. No
standard ontology
representation.
C
- Human
-
SALT
D – Text sources. NLP
techniques.
Multi entries.
-
C – Similarity
analysis of concepts
B
B –Limited
human intervention
-
-
-
C – Enrich existing ontology
-
-
-
C – Main concept defined by
a domain expert.
B – Grammatical
analysis of text
C
- Human
-
C – NLP techniques. Multi
entries source.
B – Meaning
analysis of concepts
B
-
C – Matching
analysis
Protégé
(Mainly from plugin)
B – extraction from Relational DB and some XML
format
D – Matching and
Alignment analysis.
B – Cyclic
approach can
manage evolutions
B – Manage
versions, but
still human.
C – Ontology
evolution
detection
LOGS
C – Text source analysis.
C – Similarity
NLP engine. Morphological
based on concepts
and semantic analysis.
and relationships
Machine learning approach
analysis.
for rules.
D – Extraction from several
C – Libraries for
formats (XML, UML, OWL,
clustering, formal
RDF, text…). NLP, Semanconcept analysis
tic and lexical analysis. Multi and associations
entries source.
rules
TAB. 3 – Comparative analysis
B - Provide some
predefined concepts.
B – Assisted
merging. Export
in several ontology formats.
C – Different
format. Internal
ontology structure based on a
lattice.
C - OWL and
RDF/S
B – User
required for
undecidabe
cases
- Human
Generating an
ontology from an
annotated business
model
XML2OWL
A new Method for
Ontology Merging
based on Concept
using WordNet
Enriching Very
Large Ontologies
Using the WWW
Domain-Specific
Knowledge Acquisition and Classification Using WordNet
A Method for SemiAutomatic Ontology
Acquisition from a
Corporate Intranet
SymOntoX
Ontology Learning
C – Text sources. NLP
techniques. WordNet as
resource dictionary/ontology
C – Text sources. NLP
techniques
B – Concept relationships analysis
- Human
-
-
-
B – Validation at the
end of each
module
-
B - Assisted
-
of methods
EDA 2008
I. Bedini et al.
5 Comparative Analysis and Discussion
Section 4 presents only a subset of all the works we have studied, but we have tried to
cover all types of approaches. In this section, we provide a comparative analysis of methods
following the five steps composing the automatic ontology generation process. We present
the strengths and weaknesses of experiments implementing at least one step of this process.
The degree of automation can not be measured exactly. Moreover, qualitative results
were not always available. When conducting this assessment only three tools presented in
this paper were both freely available and able to process XML Schema files (as required by
our use case), and therefore specifically tested by us. These are Protégé, XML2OWL
(Bohring, Auer 2005) and MAFRA (Maedche, et al., 2002). Despite this lack of availability,
the purpose of this study is mainly theoretical, thus information obtained by public material
was enough to perform a qualitative evaluation.
Values are assigned to each step according to the following criteria:
A– when step is not developed (marked by a – symbol)
B – for solutions using a semi-automatic approach
C – for solutions where human intervention is optional
D – for solutions that are, a priori, completely automatic
We draw the following conclusions from the analysis of Table 3: Information extraction
can reach good results. The most studied input corpora are for text documents, a lot of information can be reached from this type of corpus source. Methods based on this type of
resource have the advantage to have a lot of resources, that can be found over Internet or an
Intranet, and that several tools for NLP and mining are available. Nevertheless they require a
most important human validation task and are preferred for defining a high level definition of
concepts. Structures, like classes, attributes and relationships, are mostly provided by other
external resources. Thus mining directly structured documents can reach better results with
less validation, but not so much methods deepening study this approach.
To this end WordNet surely deserves some special attention because we can observe that
it is an essential resource for the automation process. In fact it plays different roles. The first
is that of an electronic dictionary and thesaurus, which is fundamental. The other is that of a
reference ontology, mainly by using its sibling or hierarchical terms discovery, with relationships like hyponym, meronym, holonym and hyperonym. WordNet has the drawback of
being too generic and not adapted to specific domain ontology development. Even so, it
remains an important module to further be developed.
Matching and alignment modules are the most challenging tasks but, as told in (Euzenat
et al., 2006), they are growing and methods and techniques in the future should achieve
valuable results. Such modules should be available as shared libraries.
Merging, which is strictly related to alignment, is currently implemented with two input
ontologies. Multi ontology alignment seems to be an open question yet to be investigated in
detail. This point could be resolved with consecutives merges, but it appears that the final
ontology can be different depending on the sequence in which the ontologies are merged.
Validation is almost always human and only automatic consistency checking has been
implemented. The only solution to improve it, is to limit its range, thus: adopting a bottomup approach, which has shown better results; to use successive refinements and reasoners, in
order to guarantee consistency in the resulting ontology and; by querying external resources
like Watson (d’Aquin et al., 2007) rather than the Web directly, that provides the advantage
EDA 2008
Deriving Ontologies from XML Schema
of returning structured information, which is more suitable for machine interpretation.
Evolution management is still rare. Some methods manage versions and other go further
and provide automatic detection of changes. But in reality what we are really looking for is
an ontology able to grow and not a static adaptation of some knowledge representation.
One important aspect is that most successful solutions integrate different resources for retrieving information and also as reference knowledge for detecting wrong alignments.
Most methods offer automation of only some steps of the generation process. Modular
solutions, rather then monolithic applications should offer a better architecture for covering
the larger part of the ontology life cycle, although integration of steps is mostly manual.
In order to be able to fulfill our goal, there still remains a lot of work that we could divide
into three main actions: i) one is the automatic construction of a dynamic reference ontology;
ii) the second is to build applications able to integrate this new approach (that we could call
semantic method in opposition to the exactness method) and further investigating the new
types of exceptions that it could involve; iii) and more in the semantic web area to further
develop a new methodology for automatic ontology generation.
Références
Agirre, E., Ansa, O., Hovy, E., Martinez, D. (2000). Enriching Very Large Ontologies Using the
WWW. in Proc. of the Ontology Learning Workshop, ECAI.
Amann, B., Fundulaki, I., Scholl, M. (2000). Integrating ontologies and thesauri for RDF schema
creation and metadata querying. Int. J. on Digital Libraries (JODL) 3(3), p221-236.
Aquin (d'), M., Baldassarre, C., Gridinoc, L., Angeletou, S., Sabou, M., Motta, E. (2007). Watson: A
Gateway for Next Generation Semantic Web Applications. Poster session of the ISWC
Auer, S., Dietzold, S., Riechert, T. (2006) OntoWiki – A Tool for Social, Semantic Collaboration. In
Proc. ISWC, p736–749.
Bedini, I., Nguyen, B., Gardarin, G. (2008). B2B Automatic Taxonomy Construction. To Appear in
Proc. of ICEIS. p12 - 16.
Bertrand, F., Faucher, C., Lafaye, M-C., Lafaye, J-Y., Bouju, A. (2006). Génération d'une ontologie à
partir d'un modèle métier annoté. In Proc. of IDM.
Biebow, B., Szulman, S. (1999). TERMINAE: A linguistics-based tool for the building of a domain
ontology. In Proc. of EKAW.
Bohring H, Auer S. (2005). Mapping XML to OWL Ontologies. Leipziger Informatik-Tage. p147-156.
Bourigault, D. (1996). Lexter, a natural laguage processing tool for terminology extraction. In Proc.
EURALEX International Congress.
Buitelaar, P., Cimiano, P., Grobelnik, M., Sintek, M. (2005) Ontology Learning from Text Tutorial. In
ECML/PKDD.
Castano, S., Ferrar. A., Montanelli. S., Hess. G. N., Bruno, S. (2007) State of the Art on Ontology
Coorination and Matching. Deliverable 4.4 Version 1.0. BOEMIE Project
Cho, M.,, Kim, H., Kim, P. (2006). A new method for ontology merging based on concept using wordnet. In Proc. of ICACT, vol 3, p1573-1576.
Corcho, O., Fernandez-Lopez, M., Gomez-Perez, A., (2003). Methodologies, tools, and languages for
building ontologies. Where is their meeting point? Data Knowl. Eng. 46(1): 41-64 (2003)
Ehrig, M., Staab, S. (2004). QOM - Quick Ontology Mapping. In Proc. of ISWC, 2004, p 683-697
Euzenat J., Le Bach T., Barrasa J., Bouquet P., De Bo J., Dieng R., Ehrig M., Hauswirth M., Jarrar M.,
Lara R., Maynard D., Napoli A., Stamou G., Stuckenschmidt H., Shvaiko P., Tessaris S., Van Acker
S., Zaihrayeu I., (2004) State of the Art on Ontology Alignment. Knowledge Web Deliverable
#D2.2.3, INRIA.
EDA 2008
I. Bedini et al.
Euzenat, J., Mochol, M., Shvaiko, P., Stuckenschmidt, H., Svab, O., Svatek, V., van Hage, W., Yatskevich, M. (2006). Results of the ontology alignment evaluation initiative 2006. In Proc. of the
ISWC workshop on Ontology Matching, p73–95.
Gasevic, D., Djuric, D., Devedzic, V., Damjanovic, V. (2004). From UML to ready-to-use OWL ontologies. In Proc. of IEEE-ISC, p485-490
Giraldo, G., Reynaud, C., (2002). Construction semi-automatique d'ontologies à partir de DTDs relatives à un même domaine. 13èmes journées francophones d'Ingénierie des Connaissances, Rouen
Hu, H. Liu, D-Y. (2004). Learning OWL ontologies from free texts. In Machine Learning and Cybernetics, vol 2, p.1233-1237.
Kietz, J., Maedche, A., Volz, R. (2000). A Method for Semi-Automatic Ontology Acquisition from a
Corporate Intranet. In Proc. EKAW.
Lonsdale, D., Ding, Y., Embley, D., Melby, A. (2002). Peppering knowledge sources with SALT :
Boosting conceptual content for ontology generation. AAAI Workshop, June 2002.
Maedche, A., Motik, B., Silva, N., and Volz, R. (2002). MAFRA - Mapping Distributed Ontologies in
the Semantic Web. In Proc. of EKAW. p 235–250.
Maedche, A., Staab, S. (2001). Ontology Learning for the Semantic Web. IEEE Intelligent Systems,
16(2), p72-79.
Mehrnoush, S., Abdollahzadeh, B. (2003) The State of the Art in Ontology Learning: A Framework for
Comparison. The Knowledge Engineering Review, Volume 18, Issue 4.
Miller, G.A. (1995). WORDNET: A lexical database for English. Communications of ACM (11).
Missikoff, M.,Taglino, F. (2003). Symontox : a web-ontology tool for ebusiness domains. In WISE,
p343-346.
Moldovan, D., I.,Girju, R. (2000). Domain-specific knowledge acquisition and classification using
wordnet. In Proc. of the IFAIRSC, p224-228
Noy, N. F.,, Fergerson, R. W., Musen, M. A. (2000). The knowledge model of Protege-2000: Combining interoperability and flexibility. In Proc. of EKAW.
Noy, N. F., Musen, M.A. (2003). The PROMPT Suite: Interactive Tools For Ontology Merging And
Mapping. International Journal of Human-Computer Studies.
Raghunathan, P., (2003). Fast semi-automatic generation of ontologies and their exploitation. Technical Report, Kansas State University,
Sabou, M., D’Aquin, M., Motta E. (2006) Using the Semantic Web as Background Knowledge for
Ontology Mapping. In Proc. of IWOM.
Sure, Y., Staab, S., Studer, R. (2004) On-To-Knowledge Methodology (OTKM). Handbook on Ontologies, p117-132
Vrandecic, D., Pinto., H. S., Sure, Y., Tempich, C. (2005) The DILIGENT Knowledge Processes. Journal of Knowledge Management 9 (5): p85-96.
Résumé
Dans cet article, nous présentons une méthode et un outil pour dériver le noyau d'une ontologie à partir de plusieurs schémas XML. Nous rappelons d'abord ce qu’est une ontologie
et ses rapports avec les schémas XML. Ensuite, nous nous concentrons sur la méthodologie
de construction de l'ontologie et les exigences pour les outils associés. Ensuite, nous introduisons Janus, un outil pour construire une ontologie à partir de schémas XML différents
mais dans un même domaine. Nous résumons les caractéristiques principales de Janus et
illustrons ses fonctionnalités sur un exemple simple. Finalement, nous comparons notre outil
avec ceux existants pour aider à la construction d’ontologie.
EDA 2008
| 6cs.PL
|
Constructing small tree grammars and small circuits for
formulas∗
Moses Ganardi†1 , Danny Hucke‡ 1 , Artur Jeż§2 , Markus Lohrey¶1 , and Eric Noethk1
1
arXiv:1407.4286v3 [cs.DS] 21 Sep 2015
2
University of Siegen, Germany
University of Wroclaw, Poland
February 20, 2018
Abstract
It is shown that every tree of size n over a fixed set of σ different ranked symbols
can
log σ
be decomposed (in linear time as well as in logspace) into O logn n = O nlog
many
n
σ
hierarchically defined pieces. Formally, such a hierarchical decomposition
has
the
form
of
a straight-line linear context-free tree grammar of size O logn n , which can be used as a
σ
compressed representation of the input tree. This generalizes an analogous result for strings.
Previous grammar-based tree compressors were not analyzed for the worst-case size of the
computed grammar,except for the top dag of Bille et al. [6], for which only
the weaker upper
n·log logσ n
n
(which
was
very
recently
improved
to
O
bound of O log0.19
[23])
for unranked
log σ n
n
σ
and unlabelled trees has been derived. The main result is used to show that every arithmetical
formula of size n, in which only m ≤ n different variables occur, can be transformed
(in linear
m
and depth O(log n).
time as well as in logspace) into an arithmetical circuit of size O n·log
log n
This refines a classical result of Brent from 1974, according to which an arithmetical formula
of size n can be transformed into a logarithmic depth circuit of size O(n). A short version of
this paper appeared as [24].
1
Introduction
Grammar-based string compression. Grammar-based compression has emerged to an active
field in string compression during the past 20 years. The idea is to represent a given string s by
a small context-free grammar that generates only s; such a grammar is also called a straight-line
program, briefly SLP. For instance, the word (ab)1024 can be represented by the SLP with the
rules A0 → ab and Ai → Ai−1 Ai−1 for 1 ≤ i ≤ 10 (A10 is the start symbol). The size of this
grammar is much smaller than the size (length) of the string (ab)1024 . In general, an SLP of size
n (the size of an SLP is usually defined as the total length of all right-hand sides of the rules) can
produce a string of length 2Ω(n) . Hence, an SLP can be seen indeed as a succinct representation
of the generated string. The goal of grammar-based string compression is to construct from a
given input string s a small SLP that produces s. Several algorithms for this have been proposed
and analyzed. Prominent grammar-based string compressors are for instance LZ78, RePair, and
BiSection, see [12] for more details.
To evaluate the compression performance of a grammar-based compressor C, two different
approaches can be found in the literature:
∗ The
fourth and fifth author are supported by the DFG research project QUANT-KOMP (Lo 748/10-1).
† [email protected]
‡ [email protected]
§ [email protected]
¶ [email protected]
k [email protected]
1
(a) One can analyze the maximal size of SLPs produced by C on strings of length n over the alphabet Σ (the size of Σ is considered to be a constant larger than one in the further discussion).
Formally, let
σC (n) = maxn |C(x)|,
x∈Σ
where C(x) is the SLP produced by C on input x, and |C(x)| is the size of this SLP. An
information-theoretic argument shows that for almost all binary
strings of length n (up to an
exponentially small part) the smallest SLP has size Ω logn n .1 Explicit examples of strings for
which the smallest SLP has size Ω logn n result from de Bruijn sequences; see Section 3. On
the other hand, for many grammar-based compressors C one gets σC (n) ∈ O logn n and hence
σC (n) ∈ Θ logn n . This holds for instance for the above mentioned LZ78, RePair, and BiSection,
and in fact for all compressors that produce so-called irreducible SLPs [29]. This fact is used
in [29] to construct universal string compressors based on grammar-based compressors.
(b) A second approach is to analyze the size of the SLP produced by C for an input string x
compared to the size of a smallest SLP for x. This leads to the approximation ratio for C,
which is formally defined as
|C(x)|
αC (n) = maxn
,
x∈Σ
g(x)
where g(x) is the size of a smallest SLP for x. It is known that unless P = NP, there is no
polynomial time grammar-based compressor C such that αC (n) < 8569/8568 for all n [12].
The best known polynomial time grammar-based compressors [12, 26, 27, 43, 44] have an
approximation ratio of O(log(n/g)), where g is the size of a smallest SLP for the input string
and each of them works in linear time.
Grammar-based tree compression. In this paper, we want to follow approach (a), but for trees
instead of strings. A tree in this paper is always a rooted ordered tree over a ranked alphabet,
i.e., every node is labelled with a symbol, the rank of this symbol is equal to the number of
children of the node and there is an ordering among the children of a node. In [11], grammarbased compression was extended from strings to trees. For this, linear context-free tree grammars
were used. Linear context-free tree grammars that produce only a single tree are also known
as tree straight-line programs (TSLPs) or straight-line context-free tree grammars (SLCF tree
grammars). TSLPs generalize dags (directed acyclic graphs), which are widely used as a compact
tree representation. Whereas dags only allow to share repeated subtrees, TSLPs can also share
repeated internal tree patterns.
Several grammar-based tree compressors were developed in [2, 8, 11, 28, 37], where the work
from [2] is based on another type of tree grammars (elementary ordered tree grammars). The
algorithm from [28] achieves an approximation ratio of O(log n) (for a constant set of node labels).
On the other hand, for none of the above mentioned compressors it is known, whether for every
input tree with n nodes the size of the output grammar is bounded by O logn n , as it is the case
for many grammar-based string compressors. Recently, it was shown
that the so-called top dag
n
of an unranked and unlabelled tree of size n has size O log0.19
[6]
and this bound has been
n
n·log logσ n
subsequently improved to O
in [23]. The top dag can be seen as a slight variant of a
logσ n
TSLP for an unranked tree.
In this paper, we present a grammar-based tree compressor that transforms a given nodelabelled tree of size n with
σ different node labels, whose rank is bounded by a constant, into
a TSLP of size O logn n and depth O(log n), where the depth of a TSLP is the depth of the
σ
corresponding derivation tree (we always assume
that σ ≥ 2). In particular, for an unlabelled
binary tree we get a TSLP of size O logn n . Our compressor is basically an extension of the
1 If
we do not specify the base of a logarithm, we always mean log2 (n).
2
BiSection algorithm [30] from strings to trees and is therefore called TreeBiSection. It works in two
steps:2
In the first step, TreeBiSection hierarchically decompose in a top-down way the input tree into
pieces of roughly equal size. This is a well-known technique that is also known as the (1/3, 2/3)Lemma [33]. But care has to be taken to bound the ranks of the nonterminals of the resulting
TSLP. As soon as we get a tree with three holes during the decomposition (which corresponds in
the TSLP to a nonterminal of rank three) we have to do an intermediate step that decomposes the
tree into two pieces having only two holes each. This may involve an unbalanced decomposition.
On the other hand, such an unbalanced decomposition is only necessary in every second step. This
trick to bound the number of holes by three was used by Ruzzo [42] in his analysis of space-bounded
alternating Turing machines.
The TSLP produced in the first step can be identified with its derivation tree, which has
logarithmic depth. Thanks to the fact that all nonterminals have rank at most three, we can
encode the derivation tree by a tree with O(σ) many labels. Moreover, this derivation tree is
weakly balanced in the following sense. For each edge (u, v) in the derivation tree such that both
u and v are internal nodes, the derivation tree is balanced at u or v. In a second step, TreeBiSection
computes the minimal dag of the derivation
tree. Due to its balanced shape, we can show that
this minimal dag has size at most O logn n . The nodes of this dag are the nonterminals of our
σ
final TSLP.
We prove that the algorithm sketched above can be implemented so that it works in logarithmic
space, see Section 5.2. Concerning the running time, we show the upper bound of O(n log n). An
alternative algorithm BU-Shrink (for bottom-up shrink) with a linear running time is presented
in Section 5.6. In a first step, it merges nodes of the
input tree in a bottom-up way. Thereby it
constructs a partition of the input tree into O logn n many connected parts of size at most c·logσ n,
σ
where c is a suitably chosen constant. Each such connected part has at most three neighbors, one
at the top and two at the bottom. By associating with each such connected part a nonterminal,
we obtain a TSLP for the input tree consisting of a start rule S → s, where s consists of O logn n
σ
many nonterminals of rank at most two, each having a right-hand side consisting of c · logσ n many
terminal symbols. By choosing the constant c suitably, we can (using the formula for the number
of binary trees of size m, which is given
numbers) bound the number of different
√ by the Catalan
subtrees of these right-hand sides by n ∈ O logn n . This allows to build up the right-hand
σ
sides for the non-start nonterminals using O logn n many nonterminals. A combination of this
σ
algorithm with TreeBiSection – we call the resulting algorithm BU-Shrink+TreeBiSection
– finally
yields a linear time algorithm for constructing a TSLP of size O logn n and depth O(log n).
σ
Let us remark that our size bound O logn n does not contradict any information-theoretic lower
σ
bounds (it actually matches the information
theoretic limit): Consider for instance unlabelled
√
ordered trees. There are roughly 4n / πn3 such trees with n nodes. Hence under any binary
encoding of unlabelled trees, most trees are mapped to bit strings of length at least 2n − o(n). But
when encoding a TSLP of size m into a bit string, another log(m)-factor arises. Hence, a TSLP
of size O logn n is encoded by a bit string of size O(n).
It is also important to note that our size bound O logn n only holds for trees, where the
σ
maximal rank is bounded by a constant. In particular, it does not directly apply to unranked
trees (that are, for instance, the standard tree model for XML), which is in contrast to top
dags. To overcome this limitation, one can transform an unranked tree of size n into its firstchild-next-sibling encoding [31, Paragraph 2.3.2], which is a ranked tree of size
n. Then, the
first-child-next-sibling encoding can be transformed into a TSLP of size O logn n .
σ
Transforming formulas into circuits. Our main result has an interesting application for
the classical problem of transforming formulas into small circuits, which will be presented in
Section 6. Spira [45] has shown that for every Boolean formula of size n there exists an equivalent
2 The following outline works only for binary trees, but it can be easily adapted to trees of higher rank, as long
as the rank is bounded by a constant.
3
Boolean circuit of depth O(log n) and size O(n). Brent [9] extended Spira’s theorem to formulas
over arbitrary semirings and moreover improved the constant in the O(log n) bound. Subsequent
improvements that
mainly concern constant factors can be found in [7, 10]. An easy corollary
of our O logn n bound for TSLPs is that for every (not necessarily commutative) semiring (or
σ
field), every formula of size n, in which only m ≤ n different variables occur, can be transformed
m
into a circuit of depth O(log n) and size O n·log
log n . Hence, we refine the size bound from O(n)
n·log m
to O log n (Theorem 23). The transformation can be achieved in logspace and, alternatively,
in linear time. Another interesting point of our formula-to-circuit conversion is that most of the
construction (namely the construction of a TSLP for the input formula) is purely syntactic. The
remaining part (the transformation of the TSLP into a circuit) is straightforward. In contrast, the
constructions from [7, 9, 10, 45] construct a log-depth circuit from a formula in one step.
Related work. Several papers deal with algorithmic problems on trees that are succinctly represented by TSLPs, see [35] for a survey. Among other problems, equality checking and the
evaluation of tree automata can be done in polynomial time for TSLPs.
It is interesting to compare our O logn n bound with the known bounds for dag compression. A
σ
counting argument shows that for all unlabelled binary
trees of size n, except for an exponentially
small part, the size of a smallest TSLP is Ω logn n , and hence (by our main result) Θ logn n . This
implies that also the average size of the minimal TSLP, where
the average is taken for the uniform
distribution on unlabelled binary trees of size n, is Θ logn n . In contrast, the average size of the
n
[19], whereas the worst-case size of the dag is n.
minimal dag for trees of size n is Θ √log
n
In the context of tree compression, succinct data structures for trees are another big topic.
There, the goal is to represent a tree in a number of bits that asymptotically matches the information theoretic lower bound, and at the same time allows efficient querying (in the best case in time
O(1)) of the data structure. For unlabelled unranked trees of size n there exist representations
with 2n + o(n) bits that support navigation and some other tree queries in time O(1) [25, 39].
This result has been extended to labelled trees, where (log σ) · n + 2n + o(n) bits suffice when σ
is the number of node labels [17]. See [41] for a survey.
In view of Theorem 23 on arithmetical circuits, we should mention the following interesting
difference between commutative and noncommutative semirings: By a classical result of Valiant
et al., every arithmetical circuit of polynomial size and degree over a commutative semiring can
be restructured into an equivalent unbounded fan-in arithmetical circuit of polynomial size and
logarithmic depth [47]. This result fails in the noncommutative case: In [32], Kosaraju gave an
example of a circuit family of linear size and degree over a noncommutative semiring that is not
equivalent to a polynomial size circuit family of depth o(n). In Kosaraju’s example, addition
(multiplication, respectively) is the union (concatenation, respectively) of languages. A similar
example was given by Nisan in [40].
2
Computational models
We will consider time and space bounds for computational problems. For time bounds, we will use
the standard RAM model. We make the assumption that for an input tree of size n, arithmetical
operations on numbers with O(log n) bits can be carried out in time O(1). We assume that the
reader has some familiarity with logspace computations, see e.g. [3, Chapter 4.1] for more details.
A function can be computed in logspace, if it can be computed on a Turing machine with three
tapes: a read-only input tape, a write-only output tape, and a read-write working tape of length
O(log n), where n is the length of the input. It is an important fact that if functions f and g can
be computed in logspace, then the composition of f and g can be computed in logspace as well.
We will use this fact implicitly all over the paper.
4
3
Strings and Straight-Line Programs
Before we come to grammar-based tree compression, let us briefly discuss grammar-based string
compression. A straight-line program, briefly SLP, is a context-free grammar that produces a single
string. Formally, it is defined as a tuple G = (N, Σ, S, P ), where N is a finite set of nonterminals,
Σ is a finite set of terminal symbols (Σ ∩ N = ∅), S ∈ N is the start nonterminal, and P is a finite
set of rules of the form A → w for A ∈ N , w ∈ (N ∪ Σ)∗ such that the following conditions hold:
• There do not exist rules (A → u) and (A → v) in P with u 6= v.
• The binary relation {(A, B) ∈ N × N | (A → w) ∈ P, B occurs in w} is acyclic.
These conditions ensure that every nonterminal A ∈ N produces a unique stringP
valG (A) ∈ Σ∗ .
The string defined by G is val(G) = valG (S). The size of the SLP G is |G| =
(A→w)∈P |w|,
where |w| denotes the length of the string w. SLPs are also known as word chains in the area of
combinatorics on words [5, 15].
A simple induction shows that for every SLP G of size m one has |val(G)| ≤ O(3m/3 ) [12, proof
of Lemma 1]. On the other hand, it is straightforward to define an SLP H of size 2n such that
|val(H)| ≥ 2n . This justifies to see an SLP G as a compressed representation of the string val(G),
and exponential compression rates can be achieved in this way.
Let σ ≥ 2 be the size of the terminal alphabet Σ.3 It is well-known that for every string x ∈ Σ∗
of length n there exists an SLP G of size O logn n such that val(G) = x, see e.g. [29]. On the
σ
other hand, an information-theoretic argument shows that for all
strings of length n, except for
an exponentially small part, the smallest SLP has size Ω logn n . For SLPs, one can, in contrast
σ
to other models like Boolean circuits, construct explicit strings that achieve this worst-case bound
as the following result shows. For σ ≥ 3 this result is also shown in [5], using a slightly different
argument.
Proposition 1. Let Σ be an alphabet of size σ ≥ 2. For every n ≥ σ 2 , one can construct in
time polynomial
in n and σ a string sσ,n ∈ Σ∗ of length n such that every SLP for sσ,n has size
n
Ω log n .
σ
Proof. Let r = ⌈logσ n⌉ ≥ 2. The sequence sσ,n is in fact a prefix of a de Bruijn sequence [14]. Let
x1 , . . . , xσr−1 be a list of all words from Σr−1 . Construct a directed graph by taking these strings
as vertices and drawing an a-labelled edge (a ∈ Σ) from xi to xj if xi = bw and xj = wa for some
w ∈ Σr−2 and b ∈ Σ. This graph has σ r edges and every vertex of this graph has indegree and
outdegree σ. Hence, it has a Eulerian cycle, which can be viewed as a sequence u, b1 , b2 , . . . , bσr ,
where u ∈ Σr−1 is the start vertex, and the edge traversed in the ith step is labelled with bi ∈ Σ.
Define sσ,n as the prefix of ub1 b2 · · · bσr of length n. The construction implies that sσ,n has n−r +1
different substrings of length r. By the so-called mk-Lemma from [12], every SLP for sσ,n has size
at least
n−r+1
n
n
> −1≥
− 1.
r
r
logσ (n) + 1
This proves the proposition.
In [4] a set of n binary strings of length n is constructed such that any concatenation circuit
2
that computes this set has size Ω logn2 n . A concatenation circuit for a set S of strings is simply
an SLP such that every string from S is derived from a nonterminal of the SLP. Using the above
n2
2
construction, this lower bound can be improved to Ω log
n : Simply take the string s2,n and write
n2
it as s1 s2 · · · sn with |si | = n. Then any concatenation circuit for {s1 , . . . , sn } has size Ω log
n .
3 The case σ = 1 is not interesting, since every string of length n over a unary alphabet can be produced by an
SLP of size O(log n).
5
4
Trees and Tree Straight-Line Programs
For every i ≥ 0, we fix a countably infinite
S set Fi of terminals
S of rank i and a countably infinite set
Ni of nonterminals of rank i. Let F = i≥0 Fi and N = i≥0 Ni . Moreover, let X = {x1 , x2 , . . . }
be a countably infinite set of parameters. We assume that the three sets F , N , and X are pairwise
disjoint. A labelled tree t = (V, λ) is a finite, rooted and ordered tree t with node set V , whose
nodes are labelled by elements from F ∪ N ∪ X . The function λ : V → F ∪ N ∪ X denotes the
labelling function. We require that a node v ∈ V with λ(v) ∈ Fk ∪ Nk has exactly k children,
which are ordered from left to right. We also require that every node v with λ(v) ∈ X is a leaf of
t. The size of t is |t| = |{v ∈ V | λ(v) ∈ F ∪ N }|, i.e., we do not count parameters.
We denote trees in their usual term notation, e.g. b(a, a) denotes the tree with root node
labelled by b and two children, both labelled by a. We define T as the set of all labelled trees.
The depth of a tree t is the maximal length (number of edges) of a path from the root to a leaf.
Let labels(t) = {λ(v) | v ∈ V }. For L ⊆ F ∪ N ∪ X we let T (L) = {t | labels(t) ⊆ L}. We write
<t for the depth-first-order on V . Formally, u <t v if u is an ancestor of v or if there exists a
node w and i < j such that the ith child of w is an ancestor of u and the j th child of w is an
ancestor of v. The tree t ∈ T is a pattern if there do not exist different nodes that are labelled
with the same parameter. For example, f (x1 , x1 , x3 ) is not a pattern, whereas f (x1 , x21 , x99 ) and
f (x1 , x2 , x3 ) are patterns. A pattern t ∈ T is valid if (i) labels(t) ∩ X = {x1 , . . . , xn } for some
n ≥ 0 and (ii) for all u, v ∈ V with λ(u) = xi , λ(v) = xj and u <t v we have i < j. For example the
pattern f (x1 , x21 , x99 ) is not valid, whereas f (x1 , x2 , x3 ) is valid. For a pattern t we define valid(t)
as the unique valid pattern which is obtained from t by renaming the parameters. For instance,
valid(f (x21 , x2 , x99 )) = f (x1 , x2 , x3 ). A valid pattern t in which the parameters x1 , . . . , xn occur
is also written as t(x1 , . . . , xn ) and we write rank(t) = n (the rank of the pattern). We say that a
valid pattern p of rank n occurs in a tree t if there exist n trees t1 , . . . , tn such that p(t1 , . . . , tn )
(the tree obtained from p by replacing the parameter xi by ti for 1 ≤ i ≤ n) is a subtree of t.
The following counting lemma will be needed several times:
Lemma 2. The number of trees t ∈ T (F ) with 1 ≤ |t| ≤ n and |labels(t)| ≤ σ is bounded by
4
n
3 (4σ) .
2k
1
k
Proof. The number of different rooted ordered (but unranked) trees with k nodes is k+1
k ≤ 4
th
(the k Catalan number, see e.g. [46]). Hence, the number of trees in the lemma can be bounded
by
n
X
4n+1 − 1
4
σn ·
4k ≤ σ n ·
≤ (4σ)n .
3
3
k=1
We now define a particular form of context-free tree grammars (see [13] for more details on
context-free tree grammars) with the property that exactly one tree is derived. A tree straight-line
program (TSLP) is a tuple G = (N, Σ, S, P ), where N ⊆ N is a finite set of nonterminals, Σ ⊆ F
is a finite set of terminals, S ∈ N ∩ N0 is the start nonterminal, and P is a finite set of rules
of the form A(x1 , . . . , xn ) → t(x1 , . . . , xn ) (which is also briefly written as A → t), where n ≥ 0,
A ∈ N ∩ Nn and t(x1 , . . . , xn ) ∈ T (N ∪ Σ ∪ {x1 , . . . , xn }) is a valid pattern such that the following
conditions hold:
• For every A ∈ N there is exactly one tree t such that (A → t) ∈ P .
• The binary relation {(A, B) ∈ N × N | (A → t) ∈ P, B ∈ labels(t)} is acyclic.
Note that N and Σ are implicitly defined by the rules from P . Therefore, we can (and always
will) write a TSLP as a pair G = (S, P ) consisting of rules and a start nonterminal.
The above conditions ensure that from every nonterminal A ∈ N ∩Nn exactly one valid pattern
valG (A) ∈ T (F ∪{x1 , . . . , xn }) is derived by using the rules as rewrite rules in the usual sense. The
tree defined by G is val(G) = valG (S). Instead of giving a formal definition, we show a derivation
of val(G) from S in an example:
6
Example 3. Let G = (S, P ), S, A, B, C, D, E, F ∈ N , a ∈ F0 , b ∈ F2 and
P = {S → A(B), A(x1 ) → C(F, x1 ), B → E(F ), C(x1 , x2 ) → D(E(x1 ), x2 ),
D(x1 , x2 ) → b(x1 , x2 ), E(x1 ) → D(F, x1 ), F → a}.
A possible derivation of val(G) = b(b(a, a), b(a, a)) from S is:
S → A(B) → C(F, B) → D(E(F ), B) → b(E(F ), B) → b(D(F, F ), B) → b(b(F, F ), B)
→ b(b(a, F ), B) → b(b(a, a), B) → b(b(a, a), E(F )) → b(b(a, a), D(F, F ))
→ b(b(a, a), b(F, F )) → b(b(a, a), b(a, F )) → b(b(a, a), b(a, a))
The size |G| of a TSLP G = (S, P ) is the total size of all trees on the right-hand sides of P :
X
|t|
|G| =
(A→t)∈P
For instance, the TSLP from Example 3 has size 12.
Note that for the size of a TSLP we do not count nodes of right-hand sides that are labelled
with a parameter. To justify this, we use the following internal representation of valid patterns
(which is also used in [28]): For every non-parameter node v of a tree, with children v1 , . . . , vn
we store in a list all pairs (i, vi ) such that vi is a non-parameter node. Moreover, we store for
every symbol (node label) its rank. This allows to reconstruct the valid pattern, since we know
the positions where parameters have to inserted.
A TSLP is in Chomsky normal form if for every rule A(x1 , . . . , xn ) → t(x1 , . . . , xn ) one of the
following two cases holds:
t(x1 , . . . , xn ) =
t(x1 , . . . , xn ) =
B(x1 , . . . , xi−1 , C(xi , . . . , xk ), xk+1 , . . . , xn ) for B, C ∈ N
f (x1 , . . . , xn ) for f ∈ Fn .
(1)
(2)
If the tree t in the corresponding rule A → t is of type (1), we write index(A) = i. If otherwise t is
of type (2), we write index(A) = 0. One can transform every TSLP efficiently into an equivalent
TSLP in Chomsky normal form with a small size increase [38]. We mainly consider TSLPs in
Chomsky normal form in the following.
We define the rooted, ordered derivation tree DG of a TSLP G = (S, P ) in Chomsky normal
form as for string grammars: The inner nodes of the derivation tree are labelled by nonterminals
and the leaves are labelled by terminal symbols. Formally, we start with the root node of DG
and assign it the label S. For every node in DG labelled by A, where the right-hand side t of the
rule for A is of type (1), we attach a left child labelled by B and a right child labelled by C. If
the right-hand side t of the rule for A is of type (2), we attach a single child labelled by f to A.
Note that these nodes are the leaves of DG and they represent the nodes of the tree val(G). We
denote by depth(G) the depth of the derivation tree DG . For instance, the depth of the TSLP
from Example 3 is 4.
A TSLP is monadic if every nonterminal has rank at most one. The following result was shown
in [38]:
Theorem 4. From a given TSLP G in Chomsky normal form such that every nonterminal has
rank at most k and every terminal symbol has rank at most r, one can compute in time O(|G| · k · r)
a monadic TSLP H with the following properties:
• val(G) = val(H),
• |H| ∈ O(|G| · r),
• depth(H) ∈ O(depth(G)).
Moreover, one can assume that every production of H has one of the following four forms:
7
• A → B(C) for A, C ∈ N0 , B ∈ N1 ,
• A(x1 ) → B(C(x1 )) for A, B, C ∈ N1 ,
• A → f (A1 , . . . , An ) for f ∈ Fn , A, A1 , . . . , An ∈ N0 ,
• A(x1 ) → f (A1 , . . . , Ai−1 , x1 , Ai+1 , . . . , An ) for f ∈ Fn , A ∈ N1 , A1 , . . . , An ∈ N0 .
A commonly used tree compression scheme is obtained by writing down repeated subtrees only
once. In that case all occurrences except for the first are replaced by a pointer to the first one.
This leads to a node-labelled directed acyclic graph (dag). It is known that every tree has a unique
minimal dag, which is called the dag of the initial tree. An example can be found in Figure 2,
where the right graph is the dag of the middle tree. A dag can be seen as a TSLP where every
nonterminal has rank zero: The nonterminals are the nodes of the dag. A node v with label f
and n children v1 , . . . , vn corresponds to the rule v → f (v1 , . . . , vn ). The root of the dag is the
start variable. Vice versa, it is straightforward to transform a TSLP, where every variable has
rank zero, into an equivalent dag.
The dag of a tree t can be constructed in time O(|t|) [16]. The following lemma shows that
the dag of a tree can be also constructed in logspace.
Lemma 5. The dag of a given tree can be computed in logspace.
Proof. Assume that the node set of the input tree t is {1, . . . , n}. We denote by t[i] the subtree
of t rooted at node i. Given two nodes i, j of t one can verify in logspace whether the subtrees t[i]
and t[j] are isomorphic (we write t[i] ∼
= t[j] for this) by performing a preorder traversal over both
trees and thereby comparing the two trees symbol by symbol.
The nodes and edges of the dag of t can be enumerated in logspace as follows. A node i of t is
a node of the dag if there is no j < i with t[i] ∼
= t[j]. By the above remark, this can be checked in
logspace. Let i be a node of the dag and let j be the k th child of i in t. Then j ′ is the k th child
of i in the dag where j ′ is the smallest number such that t[j ′ ] ∼
= t[j]. Again by the above remark
this j ′ can be found in logspace.
5
Constructing a small TSLP for a tree
Let t be a tree of size n with σ many different node
labels. In this section we present two algorithms
that each construct a TSLP for t of size O logn n and depth O(log n), assuming the maximal rank
σ
of symbols is bounded by a constant. Our first algorithm TreeBiSection achieves this while only
using logarithmic space, but needs time O(n · log n). The second algorithm first reduces the size of
the input tree and then performs TreeBiSection on the resulting tree, which yields a linear running
time.
5.1
TreeBiSection
TreeBiSection uses the well-known idea of splitting a tree recursively into smaller parts of roughly
equal size, see e.g. [9, 45]. For a valid pattern t = (V, λ) ∈ T (F ∪ X ) and a node v ∈ V we denote
by t[v] the tree valid(s), where s is the subtree rooted at v in t. We further write t \ v for the
tree valid(r), where r is obtained from t by replacing the subtree rooted at v by a new parameter.
If for instance t = h(g(x1 , f (x2 , x3 )), x4 ) and v is the f -labelled node, then t[v] = f (x1 , x2 ) and
t \ v = h(g(x1 , x2 ), x3 ). The following lemma is well-known, at least for binary trees; see e.g. [33].
Lemma 6. Let t ∈ T (F ∪ X ) be a tree with |t| ≥ 2 such that every node has at most r children
(where r ≥ 1). Then there is a node v such that
1
r+1
· |t| ≤ |t[v]| ≤
· |t|.
2(r + 2)
r+2
8
Proof. We start a search at the root node, checking at each node v whether |t[v]| ≤ d+1
d+2 · |t|, where
d is the number of children of v. If the property does not hold, we continue the search at a child
that spawns a largest subtree (using an arbitrary tie-breaking rule). Note that we eventually reach
d+1
· |t|: If |t[v]| = 1, then |t[v]| ≤ 12 |t| since |t| ≥ 2.
a node such that |t[v]| ≤ d+2
So, let v be the first node with |t[v]| ≤ d+1
d+2 · |t|, where d is the number of children of v. We get
|t[v]| ≤
r+1
d+1
· |t| ≤
· |t|.
d+2
r+2
Moreover v cannot be the root node. Let u be its parent node and let e be the number of children
e+1
of u. Since v spans a largest subtree among the children of u, we get e · |t[v]| + 1 ≥ |t[u]| ≥ e+2
· |t|,
i.e.,
|t[v]| ≥
=
≥
=
≥
e+1
1
· |t| −
e(e + 2)
e
1
e+1
· |t|
−
e(e + 2) |t| · e
e+1
1
· |t|
−
e(e + 2) 2e
1
· |t|
2(e + 2)
1
· |t|.
2(r + 2)
For the remainder of this section we refer with split(t) to the unique node in a tree t which is
obtained by the procedure from the proof above. Based on Lemma 6 we now construct a TSLP
Gt = (S, P ) with val(Gt ) = t for a given tree t. It is not the final TSLP produced by TreeBiSection.
For our later analysis, it is important to bound the number of parameters in the TSLP Gt by a
constant. To achieve this, we use an idea from Ruzzo’s paper [42].
We will first present the construction and analysis of Gt only for trees, where every node has at
most two children, i.e., we consider trees from T (F0 ∪ F1 ∪ F2 ). Let us write F≤2 for F0 ∪ F1 ∪ F2 .
In Section 5.5, we will consider trees of larger branching degree. For the case that r = 2, Lemma 6
yields for every tree s ∈ T (F≤2 ∪ X ) with |s| ≥ 2 a node v = split(s) such that
1
3
· |s| ≤ |s[v]| ≤
· |s|.
8
4
(3)
Consider an input tree t ∈ T (F≤2 ) (we assume that |t| ≥ 2). Every nonterminal of Gt will be of
rank at most three. Our algorithm stores two sets of rules, Ptemp and Pfinal . The set Pfinal will
contain the rules of the TSLP Gt and the rules from Ptemp ensure that the TSLP (S, Ptemp ∪ Pfinal )
produces t at any given time of the procedure. We start with the initial setting Ptemp = {S → t}
and Pfinal = ∅. While Ptemp is non-empty we proceed for each rule (A → s) ∈ Ptemp as follows:
Remove the rule from Ptemp . Let A ∈ Nr . If r ≤ 2 we determine the node v = split(s) in s.
Then we split the tree s into the two trees s[v] and s \ v. Let r1 = rank(s[v]), r2 = rank(s \ v) and
let A1 ∈ Nr1 and A2 ∈ Nr2 be fresh nonterminals. Note that r = r1 + r2 − 1. If the size of s[v]
(s \ v, respectively) is larger than 1 we add the rule A1 → s[v] (A2 → s \ v, respectively) to Ptemp .
Otherwise we add it to Pfinal as a final rule. Let k be the number of nodes of s that are labelled
by a parameter and that are smaller (w.r.t. <s ) than v. To link the nonterminal A to the fresh
nonterminals A1 and A2 we add the rule
A(x1 , . . . , xr ) → A1 (x1 , . . . , xk , A2 (xk+1 , . . . , xk+r2 ), xk+r2 +1 , . . . , xr )
to the set of final rules Pfinal .
9
v
x1
v
x2
x3
x1
x2
x3
Figure 1: Splitting a tree with three parameters
S
A
B
C
G
b
D
a
H
K
L
b
a
1
1
E
1
F
I
J
b
a
a
1
a
1
b
1
1
b
1
a
1
a
b
a
b
1
1
a
Figure 2: The derivation tree from Example 7
To bound the rank of the introduced nonterminals by three we handle rules A → s with
A ∈ N3 as follows. Let v1 , v2 , and v3 be the nodes of s labelled by the parameters x1 , x2 , and x3 ,
respectively. Instead of choosing the node v by split(s) we set v to the lowest common ancestor
of (v1 , v2 ) or (v2 , v3 ), depending on which one has the greater distance from the root node (see
Figure 1). This step ensures that the two trees s[v] and s \ v have rank 2, so in the next step each
of these two trees will be split in a balanced way according to (3). As a consequence, the resulting
TSLP Gt has depth O(log |t|) but size O(|t|).
Example 7. If we apply our construction to the binary tree t = b(b(a, a), b(a, a)) we get the TSLP
Gt = (S, P ) with the following rules, where A, B, C, . . . , L ∈ N , a ∈ F0 , and b ∈ F2 :
P = {S → A(B), A(x1 ) → C(D, x1 ), B → E(F ), C(x1 , x2 ) → G(H(x1 ), x2 ), D → a,
E(x1 ) → I(J, x1 ), F → a, G(x1 , x2 ) → b(x1 , x2 ), H(x1 ) → K(L(x1 )),
I(x1 , x2 ) → b(x1 , x2 ), J → a, K(x1 , x2 ) → b(x1 , x2 ), L → a}.
In the next step we want to compact the TSLP by considering the dag of the derivation tree.
For this we first build the derivation tree Dt := DGt from the TSLP Gt as described above. The
derivation tree for the TSLP Gt from Example 7 is shown on the left of Figure 2.
We now want to identify some (but not all) nonterminals that produce the same tree. Note that
if we just omit the nonterminal labels from the derivation tree, then there might exist isomorphic
subtrees of the derivation whose root nonterminals produce different trees. This is due to the fact
that we lost for an A-labelled node of the derivation tree with a left (right, respectively) child
that is labelled with B (C, respectively) the information at which argument position of B the
nonterminal C is substituted. To keep this information we replace every label A in the derivation
tree with index(A) ∈ {0, 1, 2, 3} (the index of a nonterminal of a TSLP in Chomsky normal form
was defined in Section 4). Moreover, we remove every leaf v and write its label into its parent
node. We call the resulting tree the modified derivation tree and denote it by Dt∗ . Note that Dt∗
10
is a full binary tree with node labels from {1, 2, 3} ∪ labels(t). The modified derivation tree for
Example 7 is shown in the middle of Figure 2.
The following lemma shows how to compact our grammar by considering the dag of Dt∗ .
Lemma 8. Let u and v be nodes of Dt labelled by A and B, respectively. Moreover, let u′ and
v ′ be the corresponding nodes in Dt∗ . If the subtrees Dt∗ [u′ ] and Dt∗ [v ′ ] are isomorphic (as labelled
ordered trees), then valGt (A) = valGt (B).
Proof. We prove the lemma by induction over the size of the trees Dt∗ [u′ ] and Dt∗ [v ′ ]. Consider u
and v labelled by A and B, respectively. We have index(A) = index(B) = i. For the induction
base assume that i = 0. Then u′ and v ′ are both leaves labelled by the same terminal. Hence,
valGt (A) = valGt (B) holds. For the induction step assume that i > 0. Let A1 (B1 , respectively)
be the label of the left child of u (v, respectively) and let A2 (B2 , respectively) be the label of
the right child of u (v, respectively). By induction, we get valGt (A1 ) = valGt (B1 ) = s(x1 , . . . , xm )
and valGt (A2 ) = valGt (B2 ) = t(x1 , . . . , xn ). Therefore, rank(A) = rank(B) = m + n − 1 and
valGt (A) = s(x1 , ..., xi−1 , t(xi , ..., xi+n−1 ), xi+n , ..., xn+m−1 ) = valGt (B).
By Lemma 8, if two subtrees of Dt∗ are isomorphic we can eliminate the nonterminal of the
root node of one subtree. Hence, we construct the minimal dag d of Dt∗ to compact our TSLP. The
minimal dag of the TSLP of Example 7 is shown on the right of Figure 2. The nodes of d are the
nonterminals of the final TSLP produced by TreeBiSection. For a nonterminal that corresponds
to an inner node of d (a leaf of d, respectively), we obtain a rule whose right-hand side has the
form (1) ((2), respectively). Let n1 be the number of inner nodes of d and n2 be the number of
leaves. Then the size of our final TSLP is 2n1 +n2 , which is bounded by twice the number of nodes
of d. The dag from Figure 2 gives the TSLP for the tree b(b(a, a), b(a, a)) described in Example 3.
Algorithm 1 shows the pseudocode of TreeBiSection. For a valid pattern s(x1 , . . . , xk ) we denote
by lcas (xi , xj ) the lowest common ancestor of the unique leafs that are labelled with xi and xj .
In Section 5.2 we will analyze the running time of TreeBiSection, and we will present a logspace
implementation. In Sections 5.3 and 5.4 we will analyze the size of the produced TSLP.
5.2
Running time and space consumption of TreeBiSection
In this section, we show that TreeBiSection can be implemented so that it works in logspace, and
alternatively in time O(n · log n). Note that these are two different implementations.
Lemma 9. Given a tree t ∈ T (F≤2 ) of size n one can compute (i) in time O(n log n) and (ii) in
logspace the TSLP produced by TreeBiSection on input t.
Proof. Let t be the input tree of size n. The dag of a tree can be computed in (i) linear time [16]
and (ii) in logspace by Lemma 5. Hence, it suffices to show that the modified derivation tree Dt∗
for t can be computed in time O(n · log n) and in logspace.
For the running time let us denote with Ptemp,i the set of productions Ptemp after i iterations
of the while loop. Moreover, let ni be the sum of the sizes of all right-hand sides in Ptemp,i . Then,
we have ni+1 ≤ ni : When a single rule A → s is replaced with A1 → t1 and A2 → t2 then each
non-parameter node in t1 or t2 is one of the nodes of s. Hence, we have |s| = |t1 | + |t2 | (recall
that we do not count parameters for the size of a tree). We might have ni+1 < ni since rules with
a single terminal symbol on the right-hand side are put into Pfinal . We obtain ni ≤ n for all i.
Hence, splitting all rules in Ptemp,i takes time O(n), and a single iteration of the while loop takes
time O(n) as well. On the other hand, since every second split reduces the size of the tree to
which the split is applied by a constant factor (see (3)). Hence, the while loop is iterated at most
O(log n) times. This gives the time bound.
The inquisitive reader may wonder whether our convention of neglecting parameter nodes for
the size of a tree affects the linear running time. This is not the case: Every right-hand side s in
Ptemp,i has at most three parameters, i.e., the total number of nodes in s is at most |s| + 3 ≤ 4|s|.
This implies that the split node can be computed in time O(|s|). Doing this for all right-hand
sides in Ptemp,i yields the time bound O(n) as above.
11
Algorithm 1: TreeBiSection(t, k)
input
: binary tree t
Ptemp := {S → t}
Pfinal := ∅
while Ptemp 6= ∅ do
foreach (A → s) ∈ Ptemp do
Ptemp := Ptemp \ {A → s}
if rank(s) = 3 then
v := the lower of nodes lcas (x1 , x2 ), lcas (x2 , x3 )
else
v := split(s)
end
t1 := s[v]; t2 := s \ v
r1 := rank(t1 ); r2 := rank(t2 )
Let A1 and A2 be fresh nonterminals.
foreach i = 1 to 2 do
if |ti | > 1 then
Ptemp := Ptemp ∪ {Ai (x1 , . . . , xri ) → ti }
else
Pfinal := Pfinal ∪ {Ai (x1 , . . . , xri ) → ti }
end
end
r := r1 + r2 − 1
Let k be the number of nodes in s labelled by parameters that are smaller than v
w.r.t. <s .
Pfinal := Pfinal ∪ {A(x1 , . . . , xr ) →
A1 (x1 , . . . , xk , A2 (xk+1 , . . . , xk+r2 ), xk+r2 +1 , . . . , xr )}
end
end
Let G be the TSLP (S, Pfinal ).
Construct the modified derivation tree Dt∗ of G.
Compute the minimal dag of Dt∗ and let H be the corresponding TSLP.
return TSLP H
For the logspace version, we first describe how to represent a single valid pattern occurring in
t in logspace and how to compute its split node. Let s(x1 , . . . , xk ) be a valid pattern which occurs
in t and has k parameters where 0 ≤ k ≤ 3, i.e., s(t1 , . . . , tk ) is a subtree of t for some subtrees
t1 , . . . , tk of t. We represent the tree s(x1 , . . . , xk ) by the tuple rep(s) = (v0 , v1 . . . , vk ) where
v0 , v1 , . . . , vk are the nodes in t corresponding to the roots of s, t1 , . . . , tk , respectively. Note that
rep(s) can be stored using O((k + 1) · log(n)) many bits. Given such a tuple rep(s) = (v0 , . . . , vk ),
we can compute in logspace the size |s| by a preorder traversal of t, starting from v0 and skipping
subtrees rooted in the nodes v1 , . . . , vk . We can also compute in logspace the split node v of s: If
s has at most two parameters, then v = split(s). Note that the procedure from Lemma 6 can be
implemented in logspace since the size of a subtree of s can be computed as described before. If
s has three parameters, then v is the lowest common ancestor of either v1 and v2 , or of v2 and
v3 , depending on which node has the larger distance from v0 . The lowest common ancestor of
two nodes can also be computed in logspace by traversing the paths from the two nodes to the
root upwards. From rep(s) = (v0 , . . . , vk ) and a split node v we can easily determine rep(s[v]) and
rep(s \ v) in logspace.
Using the previous remarks we are ready to present the logspace algorithm to compute Dt∗ .
Since Dt∗ is a binary tree of depth O(log n) we can identify a node of Dt∗ with the string u ∈ {0, 1}∗
of length at most c·⌊log n⌋ that stores the path from the root to the node, where c > 0 is a suitable
12
constant. We denote by su the tree (with at most three parameters) described by a node u of Dt∗
in the sense of Lemma 8. That is, if u′ is the corresponding node of the derivation tree Dt and u′
is labelled with the nonterminal A, then su = valGt (A).
To compute Dt∗ , it suffices for each string w ∈ {0, 1}∗ of length c · ⌊log n⌋ to check in logspace
whether it is a node of Dt∗ and in case it is a node, to determine the label of w in Dt∗ . For this, we
compute for each prefix u of w, starting with the empty word, the tuple rep(su ) and the label of
u in Dt∗ , or a bit indicating that u is not a node of Dt∗ (in which case also w is not a node of Dt∗ ).
Thereby we only store the current bit strings w, u and the value of rep(su ), which fit into logspace.
If u = ε, then rep(su ) consists only of the root of t. Otherwise, we first compute in logspace the
size |su | from rep(su ). If |su | = 1, then u is a leaf in Dt∗ with label λ(u) and no longer prefixes
represent nodes in Dt∗ . If |su | > 1, then u is an inner node in Dt∗ and, as described above, we
can compute in logspace from rep(su ) the tuples rep(su0 ) and rep(su1 ), from which we can easily
read off the label of u from {1, 2, 3}. If u = w, then we stop, otherwise we continue with ui and
rep(sui ), where i ∈ {0, 1} is such that ui is a prefix of w.
In view of the above logspace algorithm, it is interesting to remark that Gagie and Gawrychowski
considered in [20] the problem of computing in logspace a small SLP for a given
string. They
p
present a logspace algorithm that achieves an approximation ratio of O(min{g, n/ log n}), where
g is the size of a smallest SLP and n is the length of the input word.
5.3
Size of the minimal dag
In order to bound the size of the TSLP produced by TreeBiSection we have to bound the number
of nodes in the dag of the modified derivation tree. To this end, we prove in this section a general
result about the size of dags of certain weakly balanced binary trees that might be of independent
interest.
Let t be a binary tree and let 0 < β < 1. The leaf size of a node v is the number of leaves
of the subtree rooted at v. We say that an inner node v with children v1 and v2 is β-balanced if
the following holds: If ni is the leaf size of vi , then n1 ≥ βn2 and n2 ≥ βn1 . We say that t is
β-balanced if the following holds: For all inner nodes u and v such that v is a child of u, we have
that u is β-balanced or v is β-balanced.
Theorem 10. If t is a β-balanced binary tree having σ different node labels and n leaves (and
α·n
, where α ∈ O(log 1+β (β −1 ))
hence |t|, σ ≤ 2n − 1), then the size of the dag of t is bounded by log
σ n
only depends on β.
Proof. Let us fix a tree t = (V, λ) as in the theorem with n leaves. Moreover, let us fix a number
k that will be defined later. We first bound the number of different subtrees with at most k leaves
in t. Afterwards we will estimate the size of the remaining top tree. The same strategy is used for
instance in [22, 34] to derive a worst-case upper bound on the size of binary decision diagrams.
Claim 1. The number of different subtrees of t with at most k leaves is bounded by dk with
d = 16σ 2 (since t is a binary trees, we could deduce a more precise bound (4σ 2 )k , but this is not
crucial).
A subtree of t with at most k leaves has at most 2k − 1 nodes, each of which is labelled with one
of σ many labels. Hence, by Lemma 2 we can bound the number of subtrees of t with at most k
1
(4σ)2k ≤ (16σ 2 )k .
leaves by 34 (4σ)2k−1 = 3σ
Let top(t, k) be the tree obtained from t by removing all nodes with leaf size at most k.
Claim 2. The number of nodes of top(t, k) is bounded by c · nk , where c ∈ O(log 1+β (β −1 )) only
depends on β.
The tree top(t, k) has at most n/k leaves since it is obtained from t by removing all nodes with leaf
size at most k. Each node in top(t, k) has at most two children. Therefore, if we can show that
the length of unary chains in top(t, k) is bounded by some c ∈ O(log1+β (β −1 )), then it follows
that top(t, k) has at most 2cn/k many nodes.
13
v5
v4′
v4
v3′
v3
≥ β(1 + β)k
v2′
v2
≥ β(1 + β)k
OR
v1′
v1
≥ βk
≤k
≤k
≥ βk
OR
≥k
Figure 3: A chain within a top tree. The subtree rooted at v1 has more than k leaves.
Let v1 , . . . , vm be a unary chain in top(t, k) where vi is the single child node of vi+1 . Moreover,
let vi′ be the removed sibling of vi in t, see Figure 3. Note that each node vi′ has leaf size at most
k.
We claim that the leaf size of v2i+1 is larger than (1 + β)i k for all i with 2i + 1 ≤ m. For i = 0
note that v1 has leaf size more than k since otherwise it would have been removed in top(t, k). For
the induction step, assume that the leaf size of v2i−1 is larger than (1 + β)i−1 k. One of the nodes
′
′
v2i and v2i+1 must be β-balanced. Hence, v2i−1
or v2i
must have leaf size more than β(1 + β)i−1 k.
Hence, v2i+1 has leaf size more than (1 + β)i−1 k + β(1 + β)i−1 k = (1 + β)i k.
Let ℓ = log1+β (β −1 ). If m ≥ 2ℓ + 3, then v2ℓ+1 exists and has leaf size more than (1 +
ℓ
′
′
β) k = k/β, which implies that the leaf size of v2ℓ+1
or v2ℓ+2
(both nodes exist) is more than
k, which is a contradiction. Hence, we must have m ≤ 2 log1+β (β −1 ) + 2, i.e., we can choose
c = 2 log1+β (β −1 ) + 2. Figure 3 shows an illustration of our argument.
Using Claim 1 and 2 we can now prove the theorem: The number of nodes of the dag of t is
bounded by the number of different subtrees with at most k leaves (Claim 1) plus the number of
nodes of the remaining tree top(t, k) (Claim 2). Let k = 12 logd n. Recall that d = 16σ 2 and hence
log d = 4 + 2 log σ, which implies that logd n ∈ Θ(logσ n). With Claim 1 and 2 we get the following
bound on the size of the dag, where c ∈ O(log1+β (β −1 )) is the bound from Claim 2:
√
n
c·n
n
c·n
n
=O
= n + 2c ·
∈O
dk + c · = d(logd n)/2 + 2c ·
k
logd n
logd n
logd n
logσ n
This proves the theorem.
Obviously, one could relax the definition of β-balanced by only requiring that if (v1 , v2 , . . . , vδ )
is a path down in the tree, where δ is a constant, then one of the nodes v1 , v2 , . . . , vδ must be
β-balanced. Theorem 10 would still hold with this definition (with α depending linearly on δ).
Before we apply Theorem 10 to TreeBiSection let us present a few other results on the size
of dags that are of independent interest. If β is a constant, then a β-balanced binary tree t has
14
Figure 4: Tree t16 from the proof of Theorem 11.
depth O(log |t|). One might think that this logarithmic depth is responsible for the small dag size
in Theorem 10. But this intuition is wrong:
Theorem 11. There is a family of trees tn ∈ T ({a, b, c}) with a ∈ F0 , b ∈ F1 , and c ∈ F2 (n ≥ 1)
with the following properties:4
• |tn | ∈ Θ(n)
• The depth of tn is Θ(log n).
• The size of the minimal dag of tn is at least n.
Proof. Let k = logn n (we ignore rounding problems with log n, which only affect multiplicative
factors). Choose k different binary trees s1 , . . . , sk ∈ T ({a, c}), each having log n many internal
nodes. This is possible: For n large enough there are roughly
4log n
p
n2
p
=
>n
π · log3 n
π · log3 n
many different binary trees with log n many internal nodes. Then consider the trees s′i =
blog n (si ). Each of these trees has size Θ(log n) and depth Θ(log n). Next, let un (x1 , . . . , xk ) ∈
T ({c, x1 , . . . , xk }) be a binary valid pattern (all non-parameter nodes are labelled with c) of depth
Θ(log k) = Θ(log n) and size Θ(k) = Θ( logn n ). We finally take tn = un (s′1 , . . . , s′k ). Figure 4 shows
the tree t16 . We obtain |tn | = Θ( logn n ) + Θ(k · log n) = Θ(n). The depth of tn is Θ(log n). Finally,
in the minimal dag for tn the unary b-labelled nodes cannot be shared. Basically, the pairwise
different trees t1 , . . . , tn work as different constants that are attached to the b-chains. But the
number of b-labelled nodes in tn is k · log n = n.
Note that the trees from Theorem 11 are not β-balanced for any constant 0 < β < 1, and by
Theorem 10 this is necessarily the case. Interestingly, if we assume that every subtree s of a binary
tree t has depth at most O(log |s|), then Hübschle-Schneider and Raman [23] have implicitly shown
logσ n
for the size of the minimal dag.
the bound O n·log
log n
σ
Theorem 12 ([23]). Let α be a constant. Then there is a constant β that only depends on α such
that the following holds: If t is a binary tree of size n with σ many node labels such that every
logσ n
+ β.
subtree s of t has depth at most α log2 n+ α, then the size of the dag of t is at most β·n·log
log n
σ
Interestingly, we can show that the bound in this result is sharp:
Theorem 13. There is a family of trees tn ∈ T ({a, b, c}) with a ∈ F0 , b ∈ F1 , and c ∈ F2 (n ≥ 1)
with the following properties:5
• |tn | ∈ Θ(n)
4 The
unary node label b can replaced by the pattern c(d, x), where d ∈ F0 \ {a} to obtain a binary tree.
the unary node label b can replaced by the pattern c(d, x), where d ∈ F0 \ {a} to obtain a binary tree.
5 Again,
15
• Every subtree s of a tree tn has depth O(log |s|).
log n
• The size of the minimal dag of tn is Ω( n·log
).
log n
Proof. The tree tn is similar to the one from the proof of Theorem 11. Again, let k = logn n . Fix
a balanced binary tree vn ∈ T ({a, c}) with log k ∈ Θ(log n) many leaves. From vn we construct
k many different trees s1 , . . . , sk ∈ T ({a, b, c}) by choosing in vn an arbitrary subset of leaves
(there are k such subsets) and replacing all leaves in that subset by b(a). Note that |si | ∈ Θ(log n).
Moreover, every subtree s of a tree si has depth O(log |s|) (since vn is balanced). Then consider the
trees s′i = blog log n (si ) (so, in contrast to the proof of Theorem 11, the length of the unary chains
is log log n). Clearly, |s′i | ∈ Θ(log n). Moreover, we still have the property that every subtree s of
a tree s′i has depth O(log |s|): This is clear, if that subtree is rooted in a node from si . Otherwise,
the subtree s has the form bh (si ) for some h ≤ log log n. This tree has depth h + Θ(log log n) =
Θ(log log n) and size Θ(log n). Finally we combine the trees s′1 , . . . , s′k ∈ T ({a, b, c}) in a balanced
way to a single tree using the binary valid pattern un (x1 , . . . , xk ) from the proof of Theorem 11,
i.e., tn = un (s′1 , . . . , s′k ). Then, |tn | = Θ(n). Moreover, by the same argument as in the proof
log n
of Theorem 11, the dag for tn has size Ω( n·log
) since the nodes in the k = logn n many unary
log n
chains of length log log n cannot be shared with other nodes. It remains to show that every
subtree s of tn has depth O(log |s|). For the case that s is a subtree of one of the trees s′i , this
has been already shown above. But the case that s is rooted in a node from un (x1 , . . . , xk ) is
also clear: In that case, s is of the form s = u′ (s′i , . . . , s′j ), where u′ (xi , . . . , xj ) is a subtree of
un (x1 , . . . , xk ). Assume that d is the depth of u′ (xi , . . . , xj ). Since un (x1 , . . . , xk ) is a balanced
binary tree, we have |u′ (xi , . . . , xj )| ∈ Ω(2d ) and j − i + 1 ∈ Ω(2d ). Hence, s = u′ (s′i , . . . , s′j ) has
size Ω(2d + 2d · log n) = Ω(2d · log n) and depth d + Θ(log log n). This shows the desired property
since log(2d · log n) = d + log log n.
logσ n
Hübschle-Schneider and Raman [23] used Theorem 12 to prove the upper bound O n·log
logσ n
for the size of the top dag of an unranked tree of size n with σ many node labels. It is not clear
whether this bound can be improved to O logn n . The trees used in Theorem 13 do not seem to
σ
arise as the top trees of unranked trees.
5.4
Size of the TSLP produced by TreeBiSection
Let us fix the TSLP Gt for a tree t ∈ T (F≤2 ) that has been produced by the first part of TreeBiSection. Let n = |t| and σ be number of different node labels that appear in t. For the modified
derivation tree Dt∗ we have the following:
• Dt∗ is a binary tree with n leaves and hence has 2n − 1 nodes.
• There are σ + 3 possible node labels, namely 1, 2, 3 and those appearing in t.
• Dt∗ is (1/7)-balanced by (3). If we have two successive nodes in Dt∗ , then we split at one of
the two nodes according to (3). Now, assume that we split at node v according to (3). Let
v1 and v2 be the children of v, let ni be the leaf size of vi , and let n = n1 + n2 be the leaf
size of v. We get 81 n ≤ n1 ≤ 43 n and 41 n ≤ n2 ≤ 78 n (or vice versa). Hence, n1 ≥ 81 n ≥ 71 n2
and n2 ≥ 14 n ≥ 31 n1 .
Recall that the nodes of the dag of of Dt∗ are the nonterminals of the TSLP produced by TreeBiSection and that this TSLP is in Chomsky normal form. Moreover, recall that the depth of Dt∗ is
in O(log n). Hence, with Lemma 9 we get:
Corollary 14. For a tree from T (F≤2 ) of size n with σ different node labels, TreeBiSection produces a TSLP in Chomsky normal form of size O logn n and depth O(log n). Every nonterminal
σ
of that TSLP has rank at most 3, and the algorithm can be implemented in logspace and, alternatively, in time O(n · log n).
In particular, for an unlabelled tree of size n we obtain a TSLP of size O logn n .
16
5.5
Extension to trees of larger degree
If the input tree t has nodes with many children, then we cannot expect good compression by
TSLPs. The extreme case is tn = fn (a, . . . , a) where fn is a symbol of rank n. Hence, |tn | = n + 1
and every TSLP for tn has size Ω(n). On the other hand, for trees, where the maximal rank is
bounded by a constant r ≥ 1, we can easily generalize TreeBiSection. Lemma 6 allows to find a
splitting node v satisfying
r+1
1
· |t| ≤ |t[v]| ≤
· |t|.
(4)
2(r + 2)
r+2
The maximal arity of of nodes also affects the arity of patterns: we allow patterns of rank up to
r. Now assume that t(x1 , . . . , xr+1 ) is a valid pattern of rank r + 1, where r is again the maximal
number of children of a node. Then we find a subtree containing k parameters, where 2 ≤ k ≤ r:
Take a smallest subtree that contains at least two parameters. Since the root node of that subtree
has at most r children, and every proper subtree contains at most one parameter (due to the
minimality of the subtree), this subtree contains at most r parameters. By taking the root of that
subtree as the splitting node, we obtain two valid patterns with at most r parameters each. Hence,
we have to change TreeBiSection in the following way:
• As long as the number of parameters of the tree is at most r, we choose the splitting node
according to Lemma 6.
• If the number of parameters is r+1 (note that in each splitting step, the number of parameters
increases by at most 1), then we choose the splitting node such that the two resulting
fragments have rank at most r.
As before, this guarantees that in every second splitting step we split in a balanced way. But the
balance factor β from Section 5.3 now depends on r. More precisely, if in the modified derivation
tree Dt∗ we have a node v with children v1 and v2 of leaf size n1 and n2 , respectively, and this
node corresponds to a splitting satisfying (4), then we get
1
·n
2(r + 2)
1
r+1
·n
·n = 1−
r+2
r+2
r+1
· n,
r+2
1
2r + 3
≤
1−
·n =
·n
2(r + 2)
2(r + 2)
≤
n1 ≤
≤
n2
or vice versa. This implies
n1
≥
n2
≥
1
1
·n ≥
· n2 ,
2(r + 2)
2r + 3
1
1
·n ≥
· n1 .
r+2
r+1
Hence, the modified derivation tree becomes β-balanced for β = 1/(2r + 3). Moreover, the label
alphabet of the modified derivation tree now has size σ + r + 1 (since the TSLP produced in the
first step has nonterminals of rank at most r + 1). The proof of Theorem 10 yields the following
bound on the dag of the modified derivation tree and hence the size of the final TSLP:
!
log(2r + 3)
n
n
1
O
· log1+ 2r+3
·
(2r + 3) = O
1
logσ+r (n)
logσ+r (n) log(1 + 2r+3
)
Note that log(1 + x) ≥ x for 0 ≤ x ≤ 1. Hence, we can simplify the bound to
n · log(σ + r) · r · log r
.
O
log n
By Lemma 6, the depth of the produced TSLP can be bounded by 2 · d, where d is any number
that satisfies
d
r+1
≤ 1.
n·
r+2
17
Hence, we can bound the depth by
log n
≤ 2 · ⌈(r + 1) · log n⌉ ∈ O(r · log n).
2·
1
)
log(1 + r+1
Theorem 15. For a tree of size n with σ different node labels, each of rank at most r, TreeBiSection
r
produces a TSLP in Chomsky normal form of size O n·log(σ+r)·r·log
and depth O(r·log n). Every
log n
nonterminal of that TSLP has rank at most r + 1.
For the running time we obtain the following bound:
Theorem 16. TreeBiSection can be implemented such that it works in time O(r · n · log n) for a
tree of size n, where each symbol has rank at most r.
Proof. TreeBiSection makes O(r · log n) iterations of the while loop (this is the same bound as
for the depth of the TSLP) and each iteration takes time O(n). To see the latter, our internal
representation of trees with parameters from Section 4 is important. Using this representation, we
can still compute the split node in a right-hand side s from Ptemp in time O(|s|): We fist compute
for every non-parameter node v of s (i) the size of the subtree rooted at v (as usual, excluding
parameters) and (ii) the number of parameters below v. This is possible in time O(|s|) using a
straightforward bottom-up computation. Using these size informations, we can compute the split
node in s in time O(|s|) for both cases (number of parameters in s is r + 1 or smaller than r + 1)
by searching from the root downwards.
In particular, if r is bounded by a constant, TreeBiSection computes a TSLP of size O logn n
σ
and depth O(log n) in time O(n · log n). Moreover, our logspace implementation of TreeBiSection
(see Lemma 9) directly generalizes to the case of a constant rank.
On the other hand, for unranked trees where the number of children of a node is arbitrary
and not determined by the node label (which is the standard tree model in XML) all this fails:
TreeBiSection only yields TSLPs of size Θ(n) and this is unavoidable as shown by the example
from the beginning of this section. Moreover, the logspace implementation from Section 5.2 no
longer works since nonterminals have rank at most r + 1 and we cannot store anymore the pattern
derived from a nonterminal in space O(log n) (we have to store r + 1 many nodes in the tree).
Fortunately there is a simple workaround for all these problems: An unranked tree can be
transformed into a binary tree of the same size using the well known first-child next-sibling encoding [8, 31]. Then, one can simply apply
TreeBiSection to this encoding to get in logspace and time
O(n · log n) a TSLP of size O logn n .
σ
For the problem of traversing a compressed unranked tree t (which is addressed in [6] for
top dags) another (equally well known) encoding is more favorable. Let c(t) be a compressed
representation (e.g., a TSLP or a top dag) of t. The goal is to represent t in space O(|c(t)|) such
that one can efficiently navigate from a node to (i) its parent node, (ii) its first child, (iii) its next
sibling, and (iv) its previous sibling (if they exist). For top dags [6], it was shown that a single
navigation step can be done in time O(log |t|). Using the right binary encoding, we can prove the
same result for TSLPs: Let r be the maximal rank of a node of the unranked tree t. We define
the binary encoding bin(t) by adding for every node v of rank s ≤ r a binary tree of depth ⌈log s⌉
with s many leaves, whose root is v and whose leaves are the children of v. This introduces at
most 2s many new binary nodes, which are labelled
by a new symbol. We get |bin(t)| ≤ 3|t|. In
particular, we obtain a TSLP of size O logn n for bin(t), where n = |t| and σ is the number of
σ
different node labels. Note that a traversal step in the initial tree t (going to the parent node,
first child, next sibling, or previous sibling) can be simulated by O(log r) many traversal steps in
bin(t) (going to the parent node, left child, or right child). But for a binary tree s, it was recently
shown that a TSLP G for s can be represented in space O(|G|) such that a single traversal step
takes time O(1) [36].6 Hence, we can navigate in t in time O(log r) ≤ O(log |t|).
6 This
generalizes a corresponding result for strings [21].
18
5.6
Linear Time TSLP Construction
Recall that TreeBiSection works in time O(n · log n) for a tree of size n. In this section we present
a linear time algorithm BU-Shrink (for bottom-up shrink) that also constructs a TSLP of size
O logn n and depth O(log n) for a given tree of size n with σ many node labels of constant rank.
σ
The basic idea of BU-Shrink is to merge in a bottom-up way nodes of the tree to patterns of
size roughly k, where k is defined later. This is a bottom-up computation in the sense that we
begin with individual nodes and gradually merge them into larger fragments (the term “bottomup” should not be understood in the sense that the computation is done from the leaves of the
tree towards the root). The dag of the small trees represented by the patterns then yields the
compression.
For a valid pattern p of rank r we define the weight of p as |p| + r. This is the total number of
nodes in p including those nodes that are labelled with a parameter (which are not counted in the
size |p| of p). A pattern tree is a tree in which the labels of the tree are valid patterns. If a node v
is labelled with the valid pattern p and rank(p) = d, then we require that v has d children in the
pattern tree. For convenience, BU-Shrink also stores in every node the weight of the corresponding
pattern. For a node v, we denote by pv its pattern and by w(v) the weight of pv .
Let us fix a number k ≥ 1 that will be specified latter. Given a tree t = (V, λ) of size n such
that all node labels in t are of rank at most r, BU-Shrink first creates a pattern tree (which will be
also denoted with t) by replacing every label f ∈ Fd by the valid pattern f (x1 , . . . , xd ) of weight
d+ 1. Note that the the parameters in these patterns correspond to the edges of the tree t. We will
keep this invariant during the algorithm (which will shrink the tree t). Hence, the total number of
all parameter occurrences in the patterns that appear as labels in the current pattern tree t will
be always the number of nodes of the current tree t minus 1. This allows us to ignore the cost of
handling parameters for the running time of the algorithm.
BU-Shrink then creates a queue Q that contains references to all nodes of t having at most one
child (excluding the root node) in an arbitrary ordering. During the run of the algorithm, the
queue Q will only contain references to non-root nodes of the current tree t that have at most
one child (but Q may not contain references to all such nodes). For each node v of the queue we
proceed as follows. Let v be the ith child of its parent node u. If w(v) > k or w(u) > k, we simply
remove v from Q and proceed. Otherwise we merge the node v into the node u. More precisely,
we delete the node v, and set the ith child of u to the unique child of v if it exists (otherwise, u
loses its ith child). The pattern pu is modified by replacing the parameter at the position of the
ith child by the pattern pv and re-enumerating all parameters to get a valid pattern. We also set
the weight w(u) to w(v) + w(u) − 1 (which is the weight of the new pattern pu ). Note that in this
way both the number of edges of t and the total number of parameter occurrences in all patterns
decreases by 1 and so these two sizes stay equal. For example, let u be a node with pu = f (x1 , x2 )
and let v be its second child with pv = g(x1 ). Then the merged pattern becomes f (x1 , g(x2 )), and
its weight is 4. If the node u has at most one child after the merging and its weight is at most k,
then we add u to Q (if it is not already in the queue). We do this until the queue is empty. Note
that every pattern appearing in the final pattern tree has rank at most r (the maximal rank in
the initial tree).
Now consider the forest consisting of all patterns appearing in the resulting final pattern tree.
We construct the dag of this forest, which yields grammar rules for each pattern with shared
nonterminals. The dag of a forest (i.e., a disjoint union of trees) is constructed in the same way
as for a single tree. This dag has for every subtree appearing in the forest exactly one node.
The parameters x1 , x2 , . . . , xr that appear in the patterns are treated as ordinary constants when
constructing the dag. As usual, the dag can be viewed as a TSLP, where the nodes of the dag are
the nonterminals. In this way we obtain a TSLP in which each pattern is derived by a nonterminal
of the same rank. Finally, we add to the TSLP the start rule S → s, where s is obtained from
the pattern tree by labelling each node v with the unique nonterminal A such that A derives p(v).
Algorithm 2 shows a pseudocode for BU-Shrink.
Example 17. Consider the pattern tree depicted in Figure 5. Assuming no further mergings are
done, the final TSLP is S → A(B(C), B(B(C))), A(x, y) → f (g(x), y), B(y) → f (C, y), C →
19
Algorithm 2: BU-Shrink(t, k)
input
: binary tree t = (V, λ), number k ≤ |t|
Q := ∅
foreach v ∈ V do
let f = λ(v) ∈ Fd be the label of node v
w(v) := 1 + d
(the weight of node v)
pv := f (x1 , . . . , xd )
(the pattern stored in node v)
if d ≤ 1 and v is not the root then
Q := Q ∪ {v}
end
end
while Q 6= ∅ do
choose arbitrary node v ∈ Q and set Q := Q \ {v}
let u be the parent node of v
if w(v) ≤ k and w(u) ≤ k then
d := rank(pv ); e := rank(pu )
let v be the ith child of u
w(u) := w(u) + w(v) − 1
pu := pu (x1 , . . . , xi−1 , pv (xi , . . . , xi+d−1 ), xi+d , . . . , xd+e−1 )
if v has a (necessarily unique) child v ′ then
set v ′ to the ith child of u
end
delete node v
if d + e − 1 ≤ 1 and w(u) ≤ k then
Q := Q ∪ {u}
end
end
end
P := ∅
compute the minimal dag for the forest consisting of all patterns pv for v a node of t
foreach node v of t do
create a fresh nonterminal Av of rank d := rank(pv )
P := P ∪ {Av (x1 , . . . , xd ) → pv (x1 , . . . , xd )}
λ(v) := Av
(the new label of node v)
end
return TSLP (S, P ∪ {S → t})
g(a).
It is easy to see that BU-Shrink runs in time O(n) for a tree of size n. First of all, the number
of mergings is bounded by n, since each merging reduces the number of nodes of the pattern tree
by one. Moreover, if a node is removed from Q (because its weight or the weight of its parent
node is larger than k) then it will never be added to Q again (since weights are never reduced). A
single merging step needs only a constant number of pointer operations and a single addition (for
the weights). For this, it is important that we do not copy patterns, when the new pattern (for
the node u in the above description) is constructed. This size of the forest, for which we construct
the dag has size O(n): The number of non-parameter nodes is exactly n, and the number of
parameters is at most n − 1: initially the forest has n − 1 parameters (as there is a parameter
for each node except the root) and during BU-Shrink we can only decrease the total amount of
parameters.
Let us now analyze the size of the constructed TSLP. In the following, let t be the input tree
of size n and let r be the maximal rank of a label in t. Let Σ = labels(t) and σ = |Σ|.
20
f (x, y)
g(x)
f (g(x), y)
f (x, y)
f (x, y)
f (x, y)
g(x)
g(x)
a
a
g(x)
a
f (x, y)
g(x)
g(x)
a
a
g(a)
g(a)
f (g(x), y)
f (x, y)
f (g(a), y)
f (g(a), y)
g(a) f (x, y)
g(a)
f (g(a), y)
g(a)
g(a)
g(a)
Figure 5: When compressing, we start with the pattern tree left (the weights are omitted to
improve readability). The second tree and third tree depict possible intermediate steps.
Lemma 18. Let tp be the pattern tree resulting from BU-Shrink. Then |tp | ≤
4·r·n
k
+ 2.
Proof. Let us first assume that r ≥ 2. The number of non-root nodes in tp of arity at most one is
at least |tp |/2 − 1. For each of those nodes, either the node itself or the parent node has weight
at least k. We now map each non-root node of arity at most one to a node of weight at least k:
Let u (which is not the root) have arity at most one. If the weight of u is at least k, we map u
to itself, otherwise we map u to its parent node, which then must be of weight at least k. Note
that at most r nodes are mapped to a fixed node v: If v has arity at most one, then v and its
child (if it exists) can be mapped to v; if v has arity greater than one then only its children can
|t |−2
be mapped to v, and v has at most r children. Therefore there must exist at least p2r many
nodes of weight at least k. Because the sum of all weights in tp is at most 2n, this yields
|tp | − 2
· k ≤ 2n,
2r
which proves the lemma for the case r ≥ 2. The case r = 1 can be proved in the same way: Clearly,
the number of non-root nodes of arity at most one is |tp | − 1 and at most 2 nodes are mapped to a
|t |−1
|t |−1
fixed node of weight at least k (by the above argument). Hence, there exist at least p2 = p2r
many nodes of weight at least k.
Note that each node in the final pattern tree has weight at most 2k since BU-Shrink only merges
nodes of weight at most k. By Lemma 2 the number of different patterns in T (Σ ∪ {x1 , . . . , xr })
of weight at most 2k is bounded by 43 (4(σ + r))2k ≤ dk for d = (6(σ + r))2 . Hence, the size of the
dag constructed from the patterns is bounded by dk . Adding the size of the start rule, i.e. the size
of the resulting pattern tree (Lemma 18) we get the following bound for the constructed TSLP:
dk +
4·r·n
+ 2.
k
1
2
· logd n. We get the following bound for the constructed TSLP:
√
n·r
8·n·r
n·r
n · log(σ + r) · r
1
·logd n
2
+
=O
=O
.
d
+2= n+O
logd n
logd n
logσ+r n
log n
Let us now set k =
Theorem 19. BU-Shrink computes for a given treeof size n with σ many node labels of rank at
. Every nonterminal of that TSLP has rank
most r in time O(n) a TSLP of size O n·log(σ+r)·r
log n
at most r.
Clearly, if r is bounded by a constant, we obtain the bound O logn n . On the other hand,
σ
already for r ∈ Ω(log n) the bound O logn·r n is in ω(n). But note that the size of the TSLP
r+σ
produced by BU-Shrink can never exceed n.
21
Combining TreeBiSection and BU-Shrink to achieve logarithmic grammar depth. Recall that TreeBiSection produces TSLPs in Chomsky normal form of logarithmic depth, which will
be important in the next section. Clearly, the TSLPs produced by BU-Shrink are not in Chomsky
normal form. To get a TSLP in Chomsky normal form we have to further partition the right-hand
sides of the TSLP. Let us assume in this section that the maximal rank of symbols appearing in
the input tree is bounded by a constant. Hence, BU-Shrink
produces for an input tree t of size n
with σ many node labels a TSLP of size O logn n . The weight and hence also the depth of the
σ
patterns that appear in tp is O(logd n) = O(log σ n) ≤ O(log n). The productions that arise from
the dag of the forest of all patterns have the form A → f (A1 , . . . , Ar ) and A → x where f is a node
label of the input tree, r is bounded by a constant, and x is one of the parameters (recall that in
the dag construction, we consider the parameters appearing in the patterns as ordinary constants).
Productions A → x are eliminated by replacing all occurrences of A in a right-hand side by the
parameter x. The resulting productions can be split into Chomsky normal form productions in
an arbitrary way (see also [38]). This increases the size and depth only by the maximal rank of
node labels, which is a constant in our consideration.
Recall that
for the start
rule S → s, where s is the tree returned by BU-Shrink, the tree s has
σ
. We want to apply TreeBiSection to balance the tree s. But we cannot
size O logn n = O n·log
log
n
σ
use it directly because the resulting running time would not be linear if σ is not a constant: Since
TreeBiSection needs time O(|s| log |s|) on trees of constant rank (see the paragraph after the proof
of Theorem 16), this yields the time bound
n · log σ
n · log σ
n · log σ
=O
· log
· (log n + log log σ − log log n) = O(n log σ).
O
log n
log n
log n
To eliminate the factor log σ, we apply BU-Shrink again to s with k = log σ ≤ log n. Note that
the maximal rank in s is still bounded by a constant (the same constant
as for the input tree).
By Lemma 18 this yields in time O(|s|) ≤ O(n) a tree s′ of size O logn n on which we may now
use TreeBiSection to get a TSLP for s′ in Chomsky normal form of size O(|s′ |) = O logn n (note
that every node of s′ may be labelled with a different symbol, in which case TreeBiSection cannot
achieve any compression for s′ , when we count the size in bits) and depth O(log |s′ |) = O(log n).
Moreover, the running time of TreeBiSection on s′ is
n
n
n
′
′
O(|s | · log |s |) = O
log
· (log n − log log n) = O(n).
=O
log n
log n
log n
Let us recall this combined algorithm BU-Shrink+TreeBiSection.
Theorem 20. BU-Shrink+TreeBiSection computes for a given tree t of size n with σ manynode
labels of constant rank each in time O(n) a TSLP in Chomsky normal form of size O logn n and
σ
depth O(log n). The rank of every nonterminal of that TSLP is bounded by the maximal rank of
a node label in t (a constant).
6
Arithmetical Circuits
In this section, we present our main application of Corollary 14 and Theorem 20. Let S = (S, +, ·)
be a (not necessarily commutative) semiring. Thus, (S, +) is a commutative monoid with identity
element 0, (S, ·) is a monoid with identity element 1, and · left and right distributes over +.
We use the standard notation of arithmetical formulas and circuits over S: An arithmetical
formula is just a labelled binary tree where internal nodes are labelled with the semiring operations + and ·, and leaf nodes are labelled with variables y1 , y2 , . . . or the constants 0 and 1. An
arithmetical circuit is a (not necessarily minimal) dag whose internal nodes are labelled with +
and · and whose leaf nodes are labelled with variables or the constants 0 and 1. The depth of
a circuit is the length of a longest path from the root node to a leaf. An arithmetical circuit
evaluates to a multivariate noncommutative polynomial p(y1 , . . . , yn ) over S, where y1 , . . . , yn are
22
the variables occurring at the leaf nodes. Two arithmetical circuits are equivalent if they evaluate
to the same polynomial.
Brent [9] has shown that every arithmetical formula of size n over a commutative ring can be
transformed into an equivalent circuit of depth O(log n) and size
O(n) (the proof easily generalizes
m
to semirings). By first constructing a TSLP of size O n·log
, where m is the number of different
log n
variables in the formula,
and
then
transforming
this
TSLP
into
a circuit, we will refine the size
m
.
Moreover,
by
Corollary
14
(Theorem
20,
respectively)
this conversion can
bound to O n·log
log n
be done in logspace (linear time, respectively).
In the following, we consider TSLPs, whose terminal alphabet consists of the binary symbols +
and · and the constant symbols 0, 1, y1 , . . . , ym for some m. Let us denote this terminal alphabet
with Σm . For our formula-to-circuit conversion, it will be important to work with monadic TSLPs,
i.e., TSLPs where every nonterminal has rank at most one.
Lemma 21. From a given tree t ∈ T (Σm ) of size n one can construct in logspace (linear time,
m
and depth O(log n) with val(H) = t and such
respectively) a monadic TSLP H of size O n·log
log n
that all productions are of the following forms:
• A → B(C) for A, C ∈ N0 , B ∈ N1 ,
• A(x) → B(C(x)) for A, B, C ∈ N1 ,
• A → f (B, C) for f ∈ {+, ·}, A, B, C ∈ N0 ,
• A(x) → f (x, B), A(x) → f (B, x) for f ∈ {+, ·}, A ∈ N1 , B ∈ N0 ,
• A → a for a ∈ {0, 1, y1, . . . , ym }, A ∈ N0 ,
• A(x) → B(x), A(x) → x for A, B ∈ N1 .
Proof. The linear time version is an immediate consequence of Theorem 4 and Theorem 20.7 It
remains show the logspace version. We first apply TreeBiSection
(Corollary 14) to get in logspace a
m
TSLP G in Chomsky normal form of size O n·log
and
depth
O(log
n) with val(G) = t. Note that
log n
every nonterminal of G has rank 3. Moreover, for every nonterminal A of rank k ≤ 3, TreeBiSection
computes k + 1 nodes v0 , v1 , . . . , vk of t that represent the pattern valG (A): v0 is the root node
of an occurrence of valG (A) in t and vi (1 ≤ i ≤ k) is the node of the occurrence to which the
parameter xi is mapped, see also the proof of Lemma 9. We can assume that for every nonterminal
A of rank k this tuple sA has been computed.
We basically show that the construction from [38], which makes a TSLP monadic, works
in logspace if all nonterminals and terminals of the input TSLP have constant rank.8 For a
nonterminal A of rank 3 with sA = (v0 , v1 , v2 , v3 ), the pattern valG (A) has two possible branching
structures, see Figure 1. By computing the paths from the three nodes v1 , v2 , v3 up to v0 , we
can compute in logspace, which of the two branching structures valG (A) has. Moreover, we can
compute the two binary symbols f1 , f2 ∈ {+, ·} at which the three paths that go from v1 , v2 ,
and v3 , respectively, up to v0 meet. We finally, associate with each of the five dashed edges in
Figure 1 a fresh unary nonterminal Ai (0 ≤ i ≤ 4) of the TSLP H. In this way we can built up in
logspace what is called the skeleton tree for A. It is one of the following two trees, depending on
the branching structure of valG (A), see also Figure 6:
A0 (f1 (A1 (f2 (A2 (x1 ), A3 (x1 ))), A4 (x3 )))
A0 (f1 (A1 (x1 ), A2 (f2 (A3 (x2 ), A4 (x3 )))))
For a nonterminal A of rank two there is only a single branching structure and hence a single
7 Note that productions of the form A(x) → B(x) and A(x) → x do not appear in Theorem 4. We allow them
in the lemma, since they make the logspace part of the lemma easier to show and do not pose a problem in the
remaining part of this section.
8 We only consider the case that nonterminals have rank at most three and terminals have rank zero or two,
which is the case we need, but the general case, where all nonterminals and all terminals of the input TSLP have
constant rank could be handled in a similar way in logspace.
23
A0
A0
f1
f1
A1
A4
A1
A2
f2
x3
x1
f2
A2
A3
A3
A4
x1
x2
x2
x3
Figure 6: The two possible skeleton trees for a nonterminal A of rank three
skeleton tree A0 (f (A1 (x1 ), A2 (x2 ))) for f ∈ {+, ·}. Finally, for a nonterminal A of rank at most
one, the skeleton tree is A itself (this is in particular the case for the start nonterminal S, which
will be also the start nonterminal of H), and this nonterminal then belongs to H (nonterminals of
G that have rank larger than one do not belong to H). What remains is to construct in logspace
productions for the nonterminals of H that allow to rewrite the skeleton tree of A to valG (A).
For this, let us consider the productions of G, whose right-hand sides have the form (1) and (2).
A production A(x1 , . . . , xk ) → f (x1 , . . . , xk ) with k ≤ 1 is copied to H. On the other hand, if
k = 2, then A does not belong to H and hence, we do not copy the production to H. Instead,
we introduce the productions Ai (x1 ) → x1 (0 ≤ i ≤ 2) for the three nonterminals A0 , A1 , A2 that
appear in the skeleton tree of A. Now consider a production
A(x1 , . . . , xk ) → B(x1 , . . . , xi−1 , C(xi , . . . , xi+l−1 ), xi+l , . . . , xk ),
where k, l, k − l + 1 ≤ 3. We have constructed the skeleton trees tA , tB , tC for A, B, and C,
respectively. Consider the tree tB (x1 , . . . , xi−1 , tC (xi , . . . , xi+l−1 ), xi+l , . . . , xk ). We now introduce
the productions for the nonterminals that appear in tA in such a way that tA (x1 , . . . , xk ) can be
rewritten to tB (x1 , . . . , xi−1 , tC (xi , . . . , xi+l−1 ), xi+l , . . . , xk ). There are several cases depending
on k, l, and i. Let us only consider two typical cases (all other cases can be dealt in a similar
way): If the production is A(x1 , x2 , x3 ) → B(x1 , C(x2 , x3 )) then the trees tA (x1 , x2 , x3 ) and
tB (x1 , tC (x2 , x3 )) are shown in Figure 7. Note that the skeleton tree tA (x1 , x2 , x3 ) must be the
right tree from Figure 6. We add the following productions to H:
A0 (x) → B0 (x), A1 (x) → B1 (x), A2 (x) → B2 (C0 (x)), A3 (x) → C1 (x), A4 (x) → C2 (x).
Let us also consider the case A(x1 , x2 ) → B(x1 , x2 , C). The trees tA (x1 , x2 ) and tB (x1 , x2 , tC ) =
tB (x1 , x2 , C) are shown in Figure 8 (we assume that the skeleton tree for B is the left one from
Figure 6). We add the following productions to H:
A0 (x) → B0 (f1 (B1 (x), B4 (C))), A1 (x) → B2 (x), A2 (x) → B3 (x).
(5)
Other cases can be dealt with similarly. In each case we write out a constant number of productions that clearly can be produced by a logspace machine using the shape of the skeleton
trees. Correctness of the construction (i.e., val(G) = val(H)) follows from valH (tA ) = valG (A),
which can be shown by a straightforward induction, see [38]. Clearly, the size and depth of H
is linearly related to the size and depth, respectively, of G. Finally, productions of the form
A(x) → B(f (C(x), D(E))) (or similar forms) as in (5) can be easily split in logspace into productions of the forms shown in the lemma. For instance, A(x) → B(f (C(x), D(E))) is split into
A(x) → B(F (x)), F (x) → G(C(x)), G(x) → f (x, H), H → D(E). Again, the size and depth of
the the TSLP increases only by a linear factor.
Going from a monadic TSLP to a circuit (or dag) that evaluates over every semiring to the
same noncommutative polynomial is easy:
24
A0
A0
B0
f1
f1
A1
A2
x1
A1
B1
B2
x1
C0
A2
f2
f2
A3
A4
x2
x3
A3
C1
C2
x2
x3
A4
Figure 7: The skeleton tree tA (x1 , x2 , x3 ) and the tree tB (x1 , tC (x2 , x3 ))
B0
A0
A0
f2
A1
A2
x1
x2
A1
f1
B1
B4
f2
C
B2
B3
x1
x2
A2
Figure 8: The skeleton tree tA (x1 , x2 ) and the tree tB (x1 , x2 , tC )
Lemma 22. From a given monadic TSLP G over the terminal alphabet Σm such that all productions are of the form shown in Lemma 21, one can construct in logspace (linear time, respectively)
an arithmetical circuit C of depth O(depth(G)) and size O(|G|) such that over every semiring, C
and val(G) evaluate to the same noncommutative polynomial in m variables.
Proof. Fix an arbitrary semiring S and let R be the polynomial semiring R = S[y1 , . . . , ym ].
Clearly, for a nonterminal A of rank 0, valG (A) is a tree without parameters that evaluates to
an element pA of the semiring R. For a nonterminal A(x) of rank 1, valG (A) is a tree, in which
the only parameter x occurs exactly once. Such a tree evaluates to a noncommutative polynomial
pA (x) ∈ R[x]. Since the parameter x occurs exactly once in the tree val(A), it turns out that pA (x)
is linear and contains exactly one occurrence of x. More precisely, by induction on the structure
of the TSLP G we show that for every nonterminal A(x) of rank 1, the tree valG (A) evaluates in
R[x] to a noncommutative polynomial of the form
pA (x) = A0 + A1 xA2 ,
where A0 , A1 , A2 ∈ R = S[y1 , . . . , ym ]. Using the same induction, one can build up a circuit of size
O(|G|) and depth O(depth(G)) that contains gates evaluating to A0 , A1 , A2 . For a nonterminal A
of rank zero, the circuit contains a gate that evaluates to the semiring element pA ∈ R, and we
denote this gate with A as well.
The induction uses a straightforward case distinction on the rule for A(x). The cases that the
unique rule for A has the form A(x) → x, A(x) → B(x), A(x) → f (x, B), A(x) → f (B, x), A →
25
f (B, C), or A → a is clear. For instance, for a rule A(x) → +(B, x), we have pA (x) = B + 1 · x · 1,
i.e., we set A0 := B, A1 := 1, A2 := 1. Now consider a rule A(x) → B(C(x)) (for A → B(C)
the argument is similar). We have already built up a circuit containing gates that evaluate to
B0 , B1 , B2 , C0 , C1 , C2 , where
pB (x) = B0 + B1 x1 B2 ,
pC (x) = C0 + C1 xC2 .
We get
pA (x)
= pB (pC (x))
= B0 + B1 (C0 + C1 xC2 )B2
= (B0 + B1 C0 B2 ) + B1 C1 x1 C2 B2
and therefore set
A0 := B0 + B1 C0 B2 , A1 := B1 C1 , A2 := C2 B2 .
So we can define the polynomials A0 , A1 , A2 using the gates B0 , B1 , B2 , C0 , C1 , C2 with only 5
additional gates. Note that also the depth only increases by a constant factor (in fact, 2).
The output gate of the circuit is the start nonterminal of the TSLP G. The above construction
can be carried out in linear time as well as in logspace.
Now we can show the main result of this section:
Theorem 23. A given arithmetical formula F of size n having m different variables can be
transformed in logspace
(linear time, respectively) into an arithmetical circuit C of depth O(log n)
m
and size O n·log
such
that over every semiring, C and F evaluate to the same noncommutative
log n
polynomial in m variables.
Proof. Let F be an arithmetical formula of size n and let y1 , . . . , ym be the variables occurring in
F . Fix an arbitrary semiring S and let R be the polynomial semiring R = S[y1 , . . . , ym ]. Using
Lemma 21
we can construct in logspace (linear time, respectively) a monadic TSLP G of size
m
and depth O(log n) such that val(G) = F . Finally, we apply Lemma 22 in order to
O n·log
log n
m
transform G in logspace (linear time, respectively) into an equivalent circuit of size O n·log
and
log n
depth O(log n).
Theorem 23 can also be shown for fields instead of semirings. In this case, the expression is built
up using variables, the constants −1, 0, 1, and the field operations +, · and /.
The proof is similar to
m
the semiring case. Again, we start with a monadic TSLP of size O n·log
and depth O(log n) for
log n
the arithmetical expression. Again, one can assume that all rules have the form A(x) → B(C(x)),
A → B(C), A(x) → f (x, B), A(x) → f (B, x), A → f (B, C), A(x) → B(x), A(x) → x, or
A → a, where f is one of the binary field operations and a is either −1, 0, 1, or a variable. Using
this particular rule format, one can show that every nonterminal A(x) evaluates to a rational
function (A0 + A1 x)/(A2 + A3 x) for polynomials A0 , A1 , A2 , A3 in the circuit variables, whereas a
nonterminal of rank 0 evaluates to a fraction of two polynomials. Finally, these polynomials can
m
and depth O(log n).
be computed by a single circuit of size O n·log
log n
Lemma 22 has an interesting application to the problem of checking whether the polynomial
represented by a TSLP over a ring is the zero polynomial. The question, whether the polynomial
computed by a given circuit is the zero polynomial is known as polynomial identity testing (PIT).
This is a famous problem in algebraic complexity theory. For the case that the underlying ring is Z
or Zn (n ≥ 2) polynomial identity testing belongs to the complexity class coRP (the complement
of randomized polynomial time), see [1]. PIT can be generalized to arithmetic expressions that
are given by a TSLP. Using Lemma 22 and Theorem 4 we obtain:
Theorem 24. Over any semiring, the question, whether the polynomial computed by a given
TSLP is the zero polynomial, is equivalent with respect to polynomial time reductions to PIT.
In particular, if the underlying semiring is Z or Zn , then the question, whether the polynomial
computed by a given TSLP is the zero polynomial, belongs to coRP.
26
7
Future work
In [48] a universal (in the information-theoretic sense) code for binary trees is developed. This code
is computed in two phases: In a first step, the minimal dag for the input tree is constructed. Then,
a particular binary encoding is applied to the dag. It is shown that the average redundancy of the
resulting code converges to zero (see [48] for definitions) for every probability distribution on binary
trees that satisfies the so-called domination property and the representation ratio negligibility
property. Whereas the domination property is somewhat technical and easily established for
many distributions, the representation ratio negligibility property means that the average size of
the dag divided by the tree size converges to zero for the underlying probability distribution. This
n
is, for instance, the case for the uniform distribution, since the average size of the dag is Θ √log
n
[19].
We construct TSLPs that have worst-case size O logn n assuming a constant number of node
labels. We are confident that replacing the minimal dag by a TSLP of worst-case size O logn n
in the universal encoder from [48] leads to stronger results. In particular, we hope that for
the resulting encoder the maximal pointwise redundancy converges to zero for certain probability
distributions. For strings, such a result was obtained in [29] using
the fact that every string of
length n over a constant size alphabet has an SLP of size O logn n .
It would be interesting to know the worst-case output size of the grammar-based tree compressor from [28]. It works in linear time and produces a TSLP that is only by a factor O(log n) larger
than an optimal TSLP. From a complexity theoretic point of view, it would be also interesting
to see, whether TreeBiSection (which can be implemented in logspace) can be even carried out in
NC1 .
In [6] the authors proved that the top dag of a given tree t of size n is at most by a factor log n
larger than the minimal dag of t. It is not clear, whether the TSLP constructed by TreeBiSection
has this property too. The construction of the top dag is done in a bottom-up way, and as a
consequence identical subtrees are compressed in the same way. This property is crucial for the
comparison with the minimal dag. TreeBiSection works in a top-down way. Hence, it is not
guaranteed that identical subtrees are compressed in the same way. In contrast, BU-Shrink works
bottom-up, and one might compare the size of the produces TSLP with the size of the minimal
dag.
Acknowledgment. We have to thank Anna Gál for helpful comments.
References
[1] M. Agrawal and R. Saptharishi. Classifying polynomials and identity testing. Current Trends
in Science — Platinum Jubilee Special, pages 149–162, 2009
[2] T. Akutsu. A bisection algorithm for grammar-based compression of ordered trees. Inf.
Process. Lett., 110(18–19):815–820, 2010.
[3] S. Arora and B. Barak. Computational Complexity — A Modern Approach. Cambridge
University Press, 2009.
[4] V. Arvind, S. Raja, and A. V. Sreejith. On lower bounds for multiplicative circuits and linear
circuits in noncommutative domains. In Proc. CSR 2014, LNCS 8476, pages 65–76. Springer,
2014.
[5] J. Berstel and S. Brlek. On the length of word chains. Inf. Process. Lett., 26(1): 23–28, 1987.
[6] P. Bille, I. L. Gørtz, G. M. Landau, and O. Weimann. Tree compression with top trees. Inf.
Comput., 243: 166–177, 2015.
[7] M. L. Bonet and S. R. Buss. Size-depth tradeoffs for Boolean fomulae. Inf. Process. Lett.,
49(3):151–155, 1994.
27
[8] M. Bousquet-Mélou, M. Lohrey, S. Maneth, and E. Noeth. XML compression via DAGs.
Theor. Comput. Syst., 2014. to appear, DOI 10.1007/s00224-014-9544-x.
[9] R. P. Brent. The parallel evaluation of general arithmetical expressions. J. ACM, 21(2):201–
206, 1974.
[10] N. H. Bshouty, R. Cleve, and W. Eberly. Size-depth tradeoffs for algebraic formulas. SIAM
J. Comput., 24(4):682–705, 1995.
[11] G. Busatto, M. Lohrey, and S. Maneth. Efficient memory representation of XML document
trees. Inf. Syst., 33(4–5):456–474, 2008.
[12] M. Charikar, E. Lehman, A. Lehman, D. Liu, R. Panigrahy, M. Prabhakaran, A. Sahai, and
A. Shelat. The smallest grammar problem. IEEE Trans. Inf. Theory, 51(7):2554–2576, 2005.
[13] H. Comon, M. Dauchet, R. Gilleron, F. Jacquemard, D. Lugiez, C. Löding, S. Tison, and
M. Tommasi. Tree automata techniques and applications. http://tata.gforge.inria.fr/,
2007.
[14] N. de Bruijn. A combinatorial problem. Nederl. Akad. Wet., Proc., 49:758–764, 1946.
[15] A. A Diwan. A new combinatorial complexity measure for languages. Tata Institute, Bombay,
India, 1986.
[16] P. J. Downey, R. Sethi, and R. E. Tarjan. Variations on the common subexpression problem.
J. ACM, 27(4):758–771, 1980.
[17] P. Ferragina, F. Luccio, G. Manzini, and S. Muthukrishnan. Compressing and indexing
labeled trees, with applications. J. ACM, 57(1), 2009.
[18] P. Flajolet and R. Sedgewick. Analytic Combinatorics. Cambridge University Press, 2009.
[19] P. Flajolet, P. Sipala, and J.-M. Steyaert. Analytic variations on the common subexpression
problem. In Proc. ICALP 1990, LNCS 443, pages 220–234. Springer, 1990.
[20] T. Gagie and P. Gawrychowski. Grammar-based compression in a streaming model. In
Proc. LATA 2010, LNCS 6031, pages 273–284. Springer, 2010.
[21] L. Gasieniec, R. M. Kolpakov, I. Potapov, and P. Sant. Real-time traversal in grammar-based
compressed files. In Proc. DCC 2005, page 458. IEEE Computer Society, 2005.
[22] M. A. Heap and M. R. Mercer. Least upper bounds on OBDD sizes. IEEE Trans. Computers,
43(6):764–767, 1994.
[23] L. Hübschle-Schneider and Rajeev Raman. Tree compression with top trees revisited. In
Proc. SEA 2015, LNCS 9125, pages 15–27. Springer, 2015. Long version at arXiv.org, 2015.,
http://arxiv.org/abs/1506.04499.
[24] D. Hucke, M. Lohrey, and E. Noeth. Constructing small tree grammars and small circuits
for formulas. In Proc. FSTTCS 2014, vol. 29 of LIPIcs, pages 457–468. Leibniz-Zentrum für
Informatik, 2014.
[25] G. Jacobson. Space-efficient static trees and graphs. In Proc. FOCS 1989, pages 549–554.
IEEE Computer Society, 1989.
[26] A. Jeż. Approximation of grammar-based compression via recompression. In Theor. Comput.
Sci., 592:115–134, 2015.
[27] A. Jeż. A really simple approximation of smallest grammar. In Proc. CPM 2014, LNCS 8486,
pages 182–191. Springer, 2014.
28
[28] A. Jeż and M. Lohrey. Approximation of smallest linear tree grammars. In Proc. STACS
2014, vol. 25 of LIPIcs, pages 445–457. Leibniz-Zentrum für Informatik, 2014.
[29] J. C. Kieffer and E.-H. Yang. Grammar-based codes: A new class of universal lossless source
codes. IEEE Trans. Inf. Theory, 46(3):737–754, 2000.
[30] J. C. Kieffer, E.-H. Yang, G. J. Nelson, and P. C. Cosman. Universal lossless compression via
multilevel pattern matching. IEEE Trans. Inf. Theory, 46(4):1227–1245, 2000.
[31] D. E. Knuth. The Art of Computer Programming, Vol. I: Fundamental Algorithms. AddisonWesley, 1968.
[32] S. R. Kosaraju. On parallel evaluation of classes of circuits. In Proc. FSTTCS 1990, LNCS
472, pages 232–237. Springer, 1990.
[33] P. M. Lewis II, R. E. Stearns, and J. Hartmanis. Memory bounds for recognition of contextfree and context-sensitive languages. In Proc. 6th Annual IEEE Symp. Switching Circuit
Theory and Logic Design, pages 191–202, 1965.
[34] H.-T. Liaw and C.-S. Lin. On the OBDD-representation of general Boolean functions. IEEE
Trans. Computers, 41(6):661–664, 1992.
[35] M. Lohrey. Algorithmics on SLP-compressed strings: A survey. Groups Complexity Cryptology, 4(2):241–299, 2012.
[36] M. Lohrey.
Traversing grammar-compressed trees with constant
http://www.eti.uni-siegen.de/ti/veroeffentlichungen/14-traversal.pdf
delay.
[37] M. Lohrey, S. Maneth, and R. Mennicke. XML tree structure compression using RePair. Inf.
Syst., 38(8):1150–1167, 2013.
[38] M. Lohrey, S. Maneth, and M. Schmidt-Schauß. Parameter reduction and automata evaluation for grammar-compressed trees. J. Comput. Syst. Sci., 78(5):1651–1669, 2012.
[39] J. I. Munro and V. Raman. Succinct representation of balanced parentheses and static trees.
SIAM J. Comput., 31(3):762–776, 2001.
[40] N. Nisan. Lower bounds for non-commutative computation (extended abstract).
Proc. STOC 1991, pages 410–418. ACM, 1991.
In
[41] R. Raman and S. S. Rao. Succinct Representations of Ordinal Trees. In Space-Efficient Data
Structures, Streams, and Algorithms - Papers in Honor of J. Ian Munro on the Occasion of
His 66th Birthday, LNCS 8066, pages 319–332. Springer, 2013.
[42] W. L. Ruzzo. Tree-size bounded alternation. J. Comput. Syst. Sci., 21:218–235, 1980.
[43] W. Rytter. Application of Lempel-Ziv factorization to the approximation of grammar-based
compression. Theor. Comput. Sci., 302(1–3):211–222, 2003.
[44] H. Sakamoto. A fully linear-time approximation algorithm for grammar-based compression.
J. Discrete Algorithms, 3(2-4):416–430, 2005.
[45] P. M. Spira. On time-hardware complexity tradeoffs for Boolean functions. In Proc. 4th
Hawaii Symp. on System Sciences, pages 525–527, 1971.
[46] R. P. Stanley. Catalan Numbers. Cambridge University Press, 2015.
[47] L. G. Valiant, S. Skyum, S. Berkowitz, and C. Rackoff. Fast parallel computation of polynomials using few processors. SIAM J. Comput., 12(4):641–644, 1983.
[48] J. Zhang, E.-H. Yang, and J. C. Kieffer. A universal grammar-based code for lossless compression of binary trees. IEEE Trans. Inf. Theory, 60(3):1373–1386, 2014.
29
| 8cs.DS
|
Best Practices for Applying Deep Learning to Novel Applications
Leslie N. Smith
Navy Center for Applied Research in Artificial Intelligence
U.S. Naval Research Laboratory, Code 5514
Washington, DC 20375
[email protected]
ABSTRACT
This report is targeted to groups who are subject matter experts in their application but
deep learning novices. It contains practical advice for those interested in testing the
use of deep neural networks on applications that are novel for deep learning. We
suggest making your project more manageable by dividing it into phases. For each
phase this report contains numerous recommendations and insights to assist novice
practitioners.
Introduction
Although my focus is on deep learning (DL) research, I am finding that more and more frequently I am
being asked to help groups without much DL experience who want to try deep learning on their novel
(for DL) application. The motivation for this NRL report derives from noticing that much of my advice
and guidance is similar for all such groups. Hence, this report discusses the aspects of applying DL that
are more universally relevant.
While there are several useful sources of advice on best practices for machine learning [1-5], there are
differences relevant to DL that this report addresses. Still, I recommend the reader read and become
familiar with these references as they contain numerous gems. In addition, there are many sources on
best practices on the topic of software engineering and agile methodologies that I assume the reader is
already familiar with (e.g., [6, 7]). The closest reference to the material in this report can be found in
Chapter 11 of “Deep Learning” [8] on “Practical Methodology” but here I discuss a number of factors
and insights not covered in this textbook.
You can see below that a deep learning application project is divided into phases. However, in practice
you are likely to find it helpful to return to an earlier phase. For example, while finding an analogy in
phase 3, you might discover new metrics that you hadn’t considered in phase 1. All of these best
practices implicitly include iteratively returning to a phase and continuous improvement as the project
proceeds.
Phase 1: Getting prepared
In this report I assume you are (or have access to) a subject matter expert for your application. You
should be familiar with the literature and research for solving the associated problem and know the
state-of-the-art solutions and performance levels. I recommend you consider here at the beginning if a
deep learning solution is a worthwhile effort. You must consider the performance level of the state-ofthe-art and if it is high, whether it is worthwhile to put in the efforts outlined in this report for an
incremental improvement. Don’t jump into deep learning only because it seems like the latest and
1
Version 1.0
2/27/17
greatest methodology. You should also consider if you have the computer resources since each job to
train a deep network will likely take days or weeks. I have made ample use of DoD’s HPC systems in my
own research. In addition, you should consider if machine learning is appropriate at all – remember
training a deep network requires lots of labeled data, as described in phase 2.
The first step is quantitatively defining what success looks like. What will you see if this is successful,
whether it is done by human or machine? This helps define your evaluation metrics. Which metrics are
important? Which are less important? You need to specify all quantitative values that play a role in the
success of this project and determine how to weigh each of them. You also need to define objectives for
your metrics; is your goal surpass human level performance? Your objectives will strongly influence the
course of the project. Knowing quantitatively what human performance is on this task should guide
your objectives; how does the state-of-the-art compare to human performance? Also, knowing how a
human solves this task will provide valuable information on how the machine might solve it.
Some of these metrics can also lead to the design of the loss function, which is instrumental in guiding
the training of the networks. Don’t feel obligated to only use softmax/cross entropy/log loss just
because that is the most common loss function, although you should probably start with it. Your
evaluation metrics are by definition the quantities that are important for your application. Be willing to
test these metrics as weighted components of the loss function to guide the training (see phase 6).
Although you are likely considering deep learning because of its power, consider how to make the
network’s “job” as easy as possible. This is anti-intuitive because it is the power of deep networks that
likely motivates you to try it out. However, the easier the job that the networks must perform, the
easier it will be to train and the better the performance. Are you (or the state-of-the-art) currently
using heuristics/physics that can be utilized here? Can the data be preprocessed? While the network
can learn complex relationships, remember: “the easier the network’s job, the better it will perform”.
So it is worthwhile to spend time considering what you can leverage from previous work and what the
network needs to do for you. Let’s say you want to improve on a complex process where the physics is
highly approximated (i.e., a “spherical cow” situation); you have a choice to input the data into a deep
network that will (hopefully) output the desired result or you can train the network to find the
correction in the approximate result. The latter method will almost certainly outperform the former.
On the other hand, do not rely on manual effort to define potential heuristics – the scarcest resource is
human time so let the network learn its representations rather than require any fixed, manual
preprocessing.
In addition, you might want to write down any assumptions or expectations you have regarding this
state-of-the-art process as it will clarify them for yourself.
Phase 2: Preparing your data
Deep learning requires a great deal of training data. You are probably wondering “how much training
data do I need?” The number of parameters in the network is correlated with the amount of training
data. The number of training samples will limit your architectural choices in phase 6. The more training
data, the larger and more accurate the network can be. So the amount of training data depends on the
objectives you defined in phase 1.
2
Version 1.0
2/27/17
In addition to training data, you will need a smaller amount of labeled validation or test data. This test
data should be similar to the training data but not the same. The network is not trained on the test data
but it is used to test the generalization ability of the network.
If the amount of training data is very limited, consider transfer learning [9] and domain adaptation [10,
11]. If this is appropriate, download datasets that are closest to your data to use for pre-training. In
addition, consider creating synthetic data. Synthetic data has the advantages that you can create plenty
of samples and make it diverse.
The project objectives also guides the choosing of the training data samples. Be certain that the training
data is directly relevant to the task and that it is diverse enough that it covers the problem space. Study
the statistics of each class. For example, are the classes balanced? An example of a balanced class is
cats versus dogs while an unbalanced class with be cats versus all other mammals (if your problem is
inherently unbalanced, talk with a deep learning expert).
What preprocessing is possible? Can you zero mean and normalize the data? This makes the network’s
job easier as it removes the job of learning the mean. Normalization also makes the network’s job
easier by creating greater similarity between training samples.
As discussed above, investigate if there are ways to lower the dimensionality of the data using a priori
knowledge or known heuristics. You don’t need to spend time to manually determine heuristics
because the goal is to save human time and you can let the network learn its own representations. Just
know that the more irrelevant data the network has to sift through, the more training data is needed
and the more time it will take to train the network. So leverage what you can from prior art.
Phase 3: Find an analogy between your application and the closest deep learning applications
Experts know not to start from scratch for every project. This is what makes them experts. They reuse
solutions that have worked in the past and they search the deep learning literature for solutions from
other researchers. Even if no one has ever done what you are trying to do, you still need to leverage
whatever you can from the experts.
Deep learning has been applied to a variety of applications. In order to create your baseline model –
your starting point – you need to find the applications that are in some ways similar to your application.
You should search the DL literature and consider the “problems” various applications are solving to
compare with the “problem” you need to solve in your application. Find similarities and analogies
between these problems. Also, note of the differences between your new application and the deep
learning application because these differences might require changing the architecture in phase 6.
When you find the closest application, look for code to download. Many researchers make their code
available when they publish in a wide spread effort to release reproducible research. Your first aim is to
replicate the results in the paper of the closest application. Later you should modify various aspects to
see the effects on the results in a “getting to know it” stage. If you are lucky, there will be several codes
available and you should replicate the results for all of them. This comparison will provide you with
enough information so you can create a baseline in phase 4.
There are a few “classic” applications of deep learning and well known solutions. These include image
classification/object recognition (convolutional networks), processing sequential data (RNN/LSTM/GRU)
3
Version 1.0
2/27/17
such as language processing, and complex decision making (deep reinforcement learning). There are
also a number of other applications that are common, such as image segmentation and super-resolution
(fully convolutional networks) and similarity matching (Siamese networks). Appendix A lists a number of
recent deep learning applications, the architecture used, and links to the papers that describe this
application. This can give you some ideas but should not be your source for finding deep learning
applications. Instead you should carefully search https://scholar.google.com and https://arxiv.org for
the deep learning applications.
Phase 4: Create a simple baseline model
Always start simple, small, and easy. Use a smaller architecture than you anticipate you might need.
Start with a common objective function. Use common settings for the hyper-parameters. Use only part
of the training data. This is a good place to adopt some of the practices of agile software
methodologies, such as simple design, unit testing, and short releases. Only get the basic functionality
now and improve on it during phase 6. That is, plan on small steps, continuous updates, and to iterate.
Choose only one of the common frameworks, such as Caffe, TensorFlow, or MXnet. Plan to only use one
framework and one computer language to minimize errors from unnecessary complexity. The
framework and language choice will likely be driven by the replication effort you performed in phase 3.
If the network will be part of a larger framework, here is a good place to check that the framework APIs
are working properly.
Phase 5: Create visualization and debugging tools
Understanding what is happening in your model will affect the success of your project. Carpenters have
an expression “measure twice, cut once”. You should think “code once, measure twice”. In addition to
evaluating the output, you should visualize your architecture and measure internal entities to
understand why you are getting the results you are obtaining. Without diagnostics, you will be shooting
in the dark to fix problems or improve performance.
You should have a general understanding of problems related to high bias (converging on the wrong
result) versus high variance (not converging well) because there are different solutions for each type of
problem; for example, you might fix high bias problems with a larger network but you would handle high
variance problems by increasing the size of your training dataset.
Set up visualizations so you can monitor as much as possible while the architecture evolves. When
possible, set up unit tests for all of your code modifications. You should compare training error to test
error and both to human level performance. You might find your network will behave strangely and you
need ways to determine what is going on and why. Start debugging the worst problems first. Find out if
the problems are with the training data, aspects of the architecture, or the loss function.
Keep in mind that error analysis tries to explain the difference between current performance and
perfect performance. Ablative analysis tries to explain the difference between some baseline
performance and current performance. One or the other or both can be useful.
One motivation for using TensorFlow as your framework is that it has a visualization system called
TensorBoard that is part of the framework. One can output the necessary files from TensorFlow and
TensorBoard can be used to visualize your architecture, monitor the weights and feature maps, and
4
Version 1.0
2/27/17
explore the embedded space the network creates. Hence, the debugging and visualization tools are
available in the framework. With other frameworks, you need to find these tools (they are often
available online) or create your own.
Phase 6: Fine tune your model
This phase will likely take the most time. You should experiment extensively. And not just with factors
you believe will improve the result but try changing every factor just to learn what happens when it
changes. Change the architecture design, depth, width, pathways, weight initialization, loss function,
etc. Change each hyper-parameter to learn what the effect of increasing or decreasing it is. I
recommend using the learning rate range test [12] to learn about the behavior of your network over a
large range of learning rates. A similar program can be made to study the effect of other hyperparameters.
Try various regularization methods, such as data augmentation, dropout, and weight decay.
Generalization is one of the key advantages of deep networks so be certain to test regularization
methods in order to maximize this ability to generalize to unseen cases.
You should experiment with the loss function. You used a simple loss function in the baseline but you
also created several evaluation metrics that you care about and define success. The only difference
between the evaluation metrics and the loss function is that the metrics apply to the test data and the
loss function is applied to the training data in order to train the network. Can a more complicated loss
function produce a more successful result? You can add weighted components to the loss function to
reflect the importance of each metric to the results. Just be very careful to not complicate the loss
function with unimportant criterion because it is the heart of your model.
Earlier you found analogies between your application and existing deep learning applications and chose
the closest to be your baseline. Now compare to the second closest application. Or the third. What
happens if you follow another analogy and use that architecture? Can you imagine a combination of the
two to test?
In the beginning, you should have successes from some low hanging fruit. As you go on it will become
more difficult to improve the performance. The objectives you defined in phase 1 should guide how far
you want to pursue the performance improvements. Or you might want to now revise the objectives
your defined earlier.
Phase 7: End-to-end training, ensembles and other complexities
If you have the time and budget, you can investigate more complex methods and there are worlds of
complexities that are possible. There exists a huge amount of deep learning literature and more papers
are appearing daily. Most of these papers declare new state-of-the-art results with one twist or another
and some might provide you a performance boost. This section alone could fill a long report because
there are so many architectures and other options to consider but if you are at this stage, consider
talking with someone with a great deal of deep learning expertise because advice at this stage is likely to
be unique to your application. .
However, there are two common methods you might consider; end-to-end training and ensembles.
5
Version 1.0
2/27/17
As a general rule, end-to-end training of a connected system will outperform a system with multiple
parts because a combined system with end-to-end training allows each of the parts to adapt to the task.
Hence, it is useful to consider combining parts, if it is relevant for your application.
Ensembles of diverse learners (i.e., bagging, boosting, stacking) can also improve the performance over
a single model. However, this will require you to train and maintain all the members of the ensemble. If
your performance objectives warrant this, it is worthwhile to test an ensemble approach.
Summary
This report lays out many factors for you to consider when experimenting with deep learning on an
application where it hasn’t been previously used. Not every item here will be relevant but I hope that it
covers most of the factors you should consider during a project. I wish you much luck and success in
your efforts.
6
Version 1.0
2/27/17
References:
1. Martin Zinkevich, “Rules of Machine Learning: Best Practices for ML Engineering”,
http://martin.zinkevich.org/rules_of_ml/rules_of_ml.pdf
2. Brett Wujek, Patrick Hall, and Funda Güneș, “Best Practices for Machine Learning Applications”,
https://support.sas.com/resources/papers/proceedings16/SAS2360-2016.pdf
3. Jason Brownlee, “How to Use a Machine Learning Checklist to Get Accurate Predictions,
Reliably”, http://machinelearningmastery.com/machine-learning-checklist/
4. Domingos, Pedro. "A few useful things to know about machine learning." Communications of the
ACM 55.10 (2012): 78-87.
5. Grégoire Montavon, Geneviève Orr, Klaus-Robert Müller, “Neural Networks: Tricks of the
Trade”, Springer, 2012
6. Fergus Henderson, “Software engineering at Google”, CoRR, arXiv:1702.01715, 2017.
7. Gamma, Erich. Design patterns: elements of reusable object-oriented software. Pearson
Education India, 1995.
8. Goodfellow, I., Bengio, Y., Courville, A., “Deep Learning”, MIT Press, 2016
9. Weiss, Karl, Taghi M. Khoshgoftaar, and DingDing Wang. "A survey of transfer learning." Journal
of Big Data 3.1 (2016): 1-40.
10. Patel, Vishal M., Raghuraman Gopalan, Ruonan Li, and Rama Chellappa. "Visual domain
adaptation: A survey of recent advances." IEEE signal processing magazine 32, no. 3 (2015): 5369.
11. Gabriela Csurka “Domain Adaptation for Visual Applications: A Comprehensive Survey”, CoRR,
arXiv:1702.05374, 2017.
12. Leslie N. Smith. Cyclical learning rates for training neural networks. In Proceedings of the IEEE
Winter Conference on Applied Computer Vision, 2017.
Appendix A: Table of various deep learning applications
The following table lists some recent applications of deep learning, the architecture used for this
application and a few references to papers in the literature that describe the application in much more
detail.
Application
Architecture
Comments
Colorization of Black and White
Images.
Large, fully convolutional
http://www.cs.cityu.edu.hk/~qiyang/publicatio
ns/iccv-15.pdf
http://arxiv.org/pdf/1603.08511.pdf
7
Version 1.0
2/27/17
Adding Sounds To Silent Movies.
CNN + LSTM
http://arxiv.org/pdf/1512.08512.pdf
Automatic Machine Translation.
Stacked networks of large
LSTM recurrent neural
networks
http://www.nlpr.ia.ac.cn/cip/ZongPublications/
2015/IEEE-Zhang-8-5.pdf
https://arxiv.org/abs/1612.06897
https://arxiv.org/abs/1611.04558
Object Classification in Photographs.
Residual CNNs, ResNeXt,
Densenets
http://papers.nips.cc/paper/5207-deep-neuralnetworks-for-object-detection.pdf
Automatic Handwriting Generation.
RNNs
http://arxiv.org/pdf/1308.0850v5.pdf
Character Text Generation.
RNNs
http://arxiv.org/pdf/1308.0850v5.pdf
Image Caption Generation.
CNN + LSTM
http://arxiv.org/pdf/1505.00487v3.pdf
Automatic Game Playing.
Reinforcement learning +
CNNs
http://www.nature.com/nature/journal/v529/n
7587/full/nature16961.html
https://arxiv.org/abs/1612.00380
Generating audio
WaveNet = Dilated PixelCNN
https://arxiv.org/abs/1609.03499
https://arxiv.org/abs/1612.07837
https://arxiv.org/abs/1610.09001
Object tracking
CNN + hierarchical LSTMs
https://arxiv.org/abs/1701.01909
https://arxiv.org/abs/1611.06878
https://arxiv.org/abs/1611.05666
Lip reading
CNN + LSTMs
https://arxiv.org/abs/1701.05847
https://arxiv.org/abs/1611.05358
Modifying synthetic data into
labeled training data
GAN
https://arxiv.org/abs/1701.05524
Single image super-resolution
Deep, fully convolutional
networks
https://arxiv.org/abs/1612.07919
8
Version 1.0
2/27/17
https://arxiv.org/abs/1611.03679
https://arxiv.org/abs/1511.04587
https://arxiv.org/abs/1611.00591
Speech recognition
LSTMs
https://arxiv.org/abs/1701.03360
https://arxiv.org/abs/1701.02720
Generate molecular structures
RNNs
https://arxiv.org/abs/1701.01329
Time series analysis
Resnet + RNNs
https://arxiv.org/abs/1701.01887
https://arxiv.org/abs/1611.06455
Intrusion detection
RNNs or CNNs
https://arxiv.org/abs/1701.02145
Autonomous Planning
Predictron architecture
https://arxiv.org/abs/1612.08810
https://arxiv.org/abs/1612.08242
Object detection
Multi-modality classification
CNN + GAN
https://arxiv.org/abs/1612.07976
https://arxiv.org/abs/1612.00377
https://arxiv.org/abs/1611.06306
Health monitoring
All are used
https://arxiv.org/abs/1612.07640
Robotics
CNN (perception)
https://arxiv.org/abs/1612.07139
RL (control)
https://arxiv.org/abs/1611.00201
https://arxiv.org/abs/1612.06897
Domain adaptation
Self-driving
CNNs
https://arxiv.org/abs/1612.06573
https://arxiv.org/abs/1611.08788
https://arxiv.org/abs/1611.05418
Visual question answering
Resnet
https://arxiv.org/abs/1612.05386
9
Version 1.0
2/27/17
https://arxiv.org/abs/1611.01604
https://arxiv.org/abs/1611.05896
https://arxiv.org/abs/1611.05546
Weather prediction
Graphical RNN
https://arxiv.org/abs/1612.05054
Detecting cancer
RBM
https://arxiv.org/abs/1612.03211
Genomics
Multiple NNs
https://arxiv.org/abs/1611.09340
Semantic segmentation
Fully conv Densenet
https://arxiv.org/abs/1611.09326
https://arxiv.org/abs/1611.06612
Hyperspectral classification
CNN
https://arxiv.org/abs/1611.09007
Natural Language Processing (NLP)
LSTM, GRU
https://arxiv.org/abs/1606.0673
Face detection
CNN
https://arxiv.org/abs/1611.00851
10
Version 1.0
2/27/17
| 9cs.NE
|
arXiv:1601.03472v4 [math.AT] 26 Apr 2017
A GOLOD COMPLEX WITH NON-SUSPENSION MOMENT-ANGLE
COMPLEX
KOUYEMON IRIYE AND TATSUYA YANO
Abstract. It could be expected that the moment-angle complex associated with a Golod
simplicial complex is homotopy equivalent to a suspension space. In this paper, we provide a
counter example to this expectation. We have discovered this complex through the studies of
the Golod property of the Alexander dual of a join of simplicial complexes, and that of a union
of simplicial complexes.
1. Introduction
The Stanley-Reisner ring (or face ring) of a simplicial complex K over an index set [m] =
{1, · · · , m} is defined as the quotient graded algebra
k[K] = k[v1 , · · · , vm ]/IK ,
where k is a commutative ring with unit and IK = (vi1 · · · vik | {i1 , · · · , ik } 6∈ K) is the StanleyReisner ideal of K. K is called Golod over a field k if its Stanley-Reisner ring k[K] is Golod
over k. That is, the multiplication and all higher Massey products in
Tork[v1 ,··· ,vm ] (k[K], k) = H(Λ[u1, · · · , um ] ⊗ k[K], d)
are trivial, where the Koszul differential algebra (Λ[u1, · · · , um ] ⊗ k[K], d) is the bigraded differential algebra with deg ui = (1, 2), deg vi = (0, 2), and dui = vi for i = 1, · · · , m. Originally,
the algebra k[K] or the ideal IK was defined to be Golod if the following equation holds:
X
(1 + tz)m
k[K]
,
dimk Torj,2i (k, k)tj z i =
P
k[v ,··· ,v ]
1 − t i≥0; j≥1 dimk Torj,2i1 m (k[K], k)tj z i
i≥0; j≥0
k[K]
k[v ,··· ,v ]
where Torj,2i (k, k) and Torj,2i1 m (k[K], k) denote the homogeneous components of degree
2i. Golod [8] proved the equivalence of the two conditions, and thereafter his name has been
used to refer a ring that satisfies the condition. The reader may also refer to Gulliksen and
Levin [11] or Avramov [1].
Baskakov, Buchstaber, and Panov [3] and Franz [7] independently demonstrated that the
torsion algebra Tork[v1 ,··· ,vm ] (k[K], k) is isomorphic to the cohomology ring of the moment-angle
complex ZK associated with K.
2010 Mathematics Subject Classification. 55P15(13F55).
Key words and phrases. Stanley-Reisner ring, Golod, Massey product, Hopf map.
1
2
KOUYEMON IRIYE AND TATSUYA YANO
Theorem 1.1([3, 7]). For a commutative ring k with unit, the following isomorphisms of
algebras hold:
M
1 ,··· ,vm ]
H̃ ∗ (KI ; k),
H ∗ (ZK ; k) ∼
(k[K], k) ∼
= Tork[v
=
∗
I⊂[m]
where H̃ ∗ (KI ; k) denotes the reduced cohomology of the full subcomplex KI of K on I, and
H̃ ∗ (K∅ ; k) = 0 for ∗ =
6 −1 and = k for ∗ = −1. The last isomorphism is the sum of isomorphisms given by
M
H p (ZK ; k) ∼
H̃ p−|I|−1(KI ; k),
=
I⊂[m]
and the ring structure is given by the maps
H̃ p−|I|−1(KI ; k) ⊗ H̃ q−|J|−1(KJ ; k) → H̃ p+q−|I|−|J|−1(KI∪J ; k)
that are induced by the canonical inclusion maps ιI,J : KI∪J → KI ∗ KJ for I ∩ J = ∅ and zero
otherwise, where KI ∗ KJ denotes the join of two simplicial complexes KI and KJ .
Here, we recall that if the moment-angle complex ZK is homotopy equivalent to a suspension,
then the multiplication and all higher Massey products in H ∗ (ZK ; k) are trivial. For example,
see Corollary 3.11 of [23]. That is, the following implication holds:
(1.1)
ZK is homotopy equivalent to a suspension =⇒ K is Golod,
where K is Golod if K is Golod over any field k. This observation enables us to investigate the
Golod property through the study of moment-angle complexes. One of the first studies in this
direction was introduced by Grbić and Theriault [10]. They demonstrated that the momentangle complex associated with a shifted simplicial complex is homotopy equivalent to a wedge
of spheres. In [15], Kishimoto and the first author extended this result to dual sequentially
Cohen-Macaulay complexes, and provided some new Golod complexes. In these studies, the
following theorem concerning the decomposition of polyhedral products (see Definition 2.1), as
introduced by Bahri, Bendersky, Cohen, and Gitler [2], plays an essential role.
Theorem 1.2([2]). Let K be a simplicial complex on [m] and let (CX, X) = {(CXi , Xi )}i∈[m] ,
where each Xi is a based space and CX is the reduced cone of a based space X. Then, the following homotopy equivalence holds:
_
bI,
ΣZK (CX, X) ≃ Σ
Σ|KI | ∧ X
I⊂[m]
b I = ∧i∈I Xi and X
b ∅ = ∗.
where X
We call this decomposition of polyhedral products the BBCG decomposition for K. If this
decomposition is desuspendable, i.e., if the homotopy equivalence
_
bI
Σ|KI | ∧ X
ZK (CX, X) ≃
I⊂[m]
A GOLOD COMPLEX WITH NON-SUSPENSION MOMENT-ANGLE COMPLEX
3
holds for any sequence of based CW-complexes X, then we say that the BBCG decomposition is
desuspendable for K. In particular, this implies that ZK is homotopy equivalent to a suspension.
In this paper, we study the Golod properties of the Alexander dual of K ∗ L and K ∪α L,
where α is a common face of K and L. The precise statements of the results are given in the
next section.
By Theorem 1.1, the multiplicative structure of H ∗ (ZK ; k) is trivial if and only if the maps
ιI,J : KI∪J → KI ∗ KJ for I ∩ J = ∅ induce the trivial maps on the reduced cohomology
theory. By strengthening this condition, K is said to be (stably) homotopy Golod [15] if the
maps |ιI,J | : |KI∪J | → |KI ∗ KJ | for I ∩ J = ∅ are (stably) null homotopic and H ∗(ZK ; k) has
trivial higher Massey products for any fields k. By definition, the following implication holds:
stably homotopy Golod =⇒ Golod.
The second purpose of this paper is to prove that this implication is strict.
Theorem 1.3. There is a Golod simplicial complex K such that K is not stably homotopy
Golod. Moreover, ZK can be chosen to be torsion free.
Here, a space or a simplicial complex X is called torsion free if its integral homology groups
H∗ (X; Z) are torsion free. In general, ZK can have torsion in its integral homology groups even
if K is Golod. The 6 vertex triangulation of RP 2 is such an example, see for example [9] or
Example 10.10 of [15].
It could be expected that the converse of the implication (1.1) is also true. Theorem 1.3
provides a counter example to this expectation. In fact, if ZK is homotopy equivalent to a
suspension then the fat wedge filtration of ZK is trivial, by Theorem 1.3 of [15]. By Theorem
6.9 of the same paper, we see that K is stably homotopy Golod, which contradicts our result.
Thus, ZK is not homotopy equivalent to a suspension.
In the next section, we state the main results of this paper. The subsequent sections are
devoted to their proofs.
Acknowledgments We are grateful to Daisuke Kishimoto for providing useful discussions
and pointing out the paper [19], which was applied in the construction of K in Theorem 1.3.
We also thank Lukas Katthän for his comments on the first draft of this paper. He kindly
pointed out the papers [13] and [6], and provided us with ideas for simplifying the arguments
in the subsection 5.3. We would also like to thank the anonymous referee for his/her careful
reading of our manuscript and his/her many constructive comments and suggestions, which
helped us to improve the manuscript. Finally, the first author thanks Japan Society for the
Promotion of Science (KAKENHI grant number 26400094) for funding.
4
KOUYEMON IRIYE AND TATSUYA YANO
2. Results
In this section, we state our main results. We begin by setting notation regarding simplicial
complexes.
Let K be a simplicial complex on an index set V . In this paper, we only consider finite
simplicial complexes. A facet of a simplicial complex is a maximal face with respect to the
inclusion, and a subset σ ⊂ V is called a minimal non-face if σ 6∈ K but σ − v ∈ K for every
v ∈ σ. The subset V (K) of V defined by V (K) = ∪σ∈K σ is called the vertex set of K, and an
element of V − V (K) is called a ghost vertex. In this paper we allow a simplicial complex to
have a ghost vertex. But we assume that a simplicial complex has the simplex ∅.
For a finite set V , we denote the full simplex on V by ∆V . Its boundary is denoted by
∂∆V . We also use the symbol ∆n to denote an n-dimensional simplex. |K| denotes a geometric
realization of K. The link and star of a face σ of K are denoted by linkK (σ) and starK (σ),
respectively:
linkK (σ) = {τ ⊂ σ c = V − σ | τ ∪ σ ∈ K}, starK (σ) = {τ ⊂ V | τ ∪ σ ∈ K}.
For a subset I ⊂ V , KI denotes the full subcomplex of K indexed by I; that is,
KI = {σ ⊂ I | σ ∈ K}. Then, KI is called the restriction of K to I. For a vertex v of K, the
deletion of v from K is denoted by K − v = KV −v . For simplicial complexes K and L with
disjoint index sets V and W , the simplicial join K ∗ L on the index set V ⊔ W is defined by
K ∗ L = {σ ⊔ τ ⊂ V ⊔ W | σ ∈ K, τ ∈ L}, where ⊔ always denotes the disjoint union of sets.
We write K ∗ c to stand for K ∗ ∆{c} . For disjoint subsets I, J ⊂ V the canonical inclusion map
ιI,J : KI⊔J → KI ∗ KJ mentioned in Theorem 1.1 is defined as ιI,J (σ) = (σ ∩ I) ⊔ (σ ∩ J).
The (simplicial) Alexander dual KV∗ for a simplicial complex K on an index set V such that
K 6= ∆V is defined by KV∗ = {σ ⊂ V | σ c 6∈ K}. If V = V (K) or V is clear from the context,
we simply write K ∗ for KV∗ . It is easy to see that for a subset I of V , I is a facet of K if and
only if I c is a minimal-non face of K ∗ . The restriction of K ∗ and the dual to a link of K are
related by the following formula: (K ∗ )I = (linkK (I c ))∗ .
Next we review polyhedral products, which are a generalization of moment-angle complexes.
Definition 2.1. Let K be a simplicial complex on [m], and (X, A) be a sequence of pairs of
based spaces {(Xi , Ai )}i∈[m] . The polyhedral product ZK (X, A) is defined by
[
ZK (X, A) =
(X, A)σ (⊂ X1 × · · · × Xm ),
σ∈K
where (X, A)σ = Y1 × · · · × Ym , with Yi = Xi for i ∈ σ and Ai for i 6∈ σ. If (Xi , Ai ) = (X, A)
for all i ∈ [m], then we write ZK (X, A) for ZK (X, A).
The polyhedral product ZK (D 2 , S 1) is the moment-angle complex of K, and is written simply
as ZK . We refer the reader to [15, 5] for further examples of polyhedral products. In this paper,
we are interested in the homotopy types of polyhedral products ZK (CX, X).
A GOLOD COMPLEX WITH NON-SUSPENSION MOMENT-ANGLE COMPLEX
If K has a ghost vertex, then
ZK (X, A) = ZKV (K) (X V (K) , AV (K) ) ×
Y
5
Ai ,
i∈V (K)c
where X I is the sub-sequence of X indexed by I ⊂ [m]. In particular, the moment-angle
c
complex ZK = ZKV (K) × (S 1 )|V (K) | . If |V (K)c | > 1 or ZKV (K) is not contractible, K is nonGolod. When we consider Golodness of simplicial complexes, therefore, we usually restrict
ourselves to simplicial complexes without ghost vertices. Nevertheless, in this paper we allow a
simplicial complex to have a ghost vertex. The reason for this is that we mainly consider dual
simplicial complexes as in Theorems 2.3, 2.5 and 2.7. These dual simplicial complexes do not
have ghost vertices even if the initial simplicial complexes have.
Example 2.2. In [22], Porter proved the following homotopy equivalence:
m
[
CX1 × · · · × CXi−1 × Xi × CXi+1 × · · · × CXm
Z∂∆[m] (CX, X) =
i=1
≃ X1 ∗ · · · ∗ Xm
≃ Σm−1 X1 ∧ · · · ∧ Xm .
He used this homotopy equivalence to define a higher order Whitehead product.
Theorem 2.3. Let K and L be simplicial complexes on disjoint index sets V and W ,
respectively. If K =
6 ∆V and L 6= ∆W , then the BBCG decomposition is desuspendable for
(K ∗ L)∗ .
Corollary 2.4. Let K and L be simplicial complexes on disjoint index sets V and W ,
respectively. If K 6= ∆V and L 6= ∆W , then (K ∗ L)∗ is Golod by (1.1).
Because the Stanley-Reisner ideal of (K ∗ L)∗ is the product of those of K ∗ and L∗ , this
corollary provides a topological proof of a classical result given in [13]. The reader may also
refer to [6] and [17]. We also remark that (∆V ∗ L)∗ is Golod if and only if L∗ is Golod, because
(∆V ∗ L)∗ = ∆V ∗ L∗ (see Lemma 3.1).
Theorem 2.5. Let K and L be simplicial complexes of non-negative dimension on index sets
V and W , respectively. Assume that α = V ∩ W is a common face of K and L. If K 6= ∆V or
L 6= ∆W , and α is neither a facet of K nor L, then the BBCG decomposition is desuspendable
for (K ∪α L)∗ .
Corollary 2.6. Let K and L be simplicial complexes that satisfy the same conditions as
stated in Theorem 2.5. If α is neither a facet of K nor L, then (K ∪α L)∗ is Golod by (1.1).
It is natural to ask whether (K ∪α L)∗ is still Golod even when α is a facet of K or L in
Corollary 2.6. In general, the answer is that this does not hold. In fact, we can construct many
6
KOUYEMON IRIYE AND TATSUYA YANO
non-Golod simplicial complexes (∆V ∪α L)∗ , such as in Example 5.3.4. A necessary and sufficient
condition for (K∪α L)∗ being non-Golod is given by Corollary 5.3.3. Incidentally, (∆V ∪α ∆W )∗ is
a non-Golod simplicial complex if α 6= V and α 6= W , since (∆V ∪α ∆W )∗ = ∂∆V −α ∗∆α ∗∂∆W −α
(see Lemma 4.1).
To construct a simplicial complex satisfying Theorem 1.3, we first need to fix some notation.
Let K be a simplicial complex on an index set V , with facets F1 , · · · , Fk . We take new
vertices v1 , · · · , vk , and define a new simplicial complex F (K) on the index set V ⊔ {v1 , · · · , vk }
with facets F1 + v1 , · · · , Fk + vk , where Fi + vi = Fi ⊔ {vi } is the disjoint union of two sets.
Then, K is a subcomplex of F (K) and |K| is a deformation retract of |F (K)|.
For two simplicial complexes K and L, we define a “product” K ⊠L as follows. Define a liner
order ≤ on the vertex sets of K and L. The vertex set of K ⊠ L is V (K) × V (L). An n-simplex
is a set {(x0 , y0 ), · · · , (xn , yn )} such that x0 ≤ · · · ≤ xn , y0 ≤ · · · ≤ yn , {x0 , · · · , xn } is a simplex
of K, and {y0, · · · , yn } is a simplex of L. It is well-known that |K ⊠ L| is homeomorphic to
|K| × |L|. If v is a vertex of L, then the subcomplex K ⊠ ∆{v} of K ⊠ L is abbreviated as K ⊠ v.
By Skn we denote a triangulation of an n-sphere S n with k-vertices. It follows from the
simplicial approximation theorem that there is a simplicial map ηk : Sk3 → S42 for sufficiently
large k whose geometrical realization |ηk | : |Sk3 | → |S42| is homotopic to the Hopf map η : S 3 →
S 2 . In fact, we can choose k = 12 in this case, by [19].
We consider the simplicial set ∆1 as the full simplex on [2]. By Sk3 ⊠ ∆1 ∪ηk S42 we denote the
simplicial complex obtained from the disjoint union of two simplicial complexes Sk3 ⊠∆1 and S42 ,
given by identifying (v, 2) ∈ V (Sk3 ) × [2] with ηk (v) ∈ V (S42 ). We embed Sk3 into Sk3 ⊠ ∆1 ∪ηk S42
by applying the map v 7→ (v, 1).
We set V to be the vertex set of the union of two simplicial complexes Sk3 ⊠ ∆1 ∪ηk S42 and
F (Sk3 ) along Sk3 . Finally, we take new vertices v0 , w1 , w2 and set
K = ∆V +v0 ∪ ((Sk3 ⊠ ∆1 ∪ηk S42 ) ∪ F (Sk3)) ∗ ∆{w1 } ∪ Sk3 ∗ ∆{w1 ,w2} ,
which is the union of ∆V +v0 and ∆V ∪ ((Sk3 ⊠ ∆1 ∪ηk S42 ) ∪ F (Sk3 )) ∗ ∆{w1 } ∪ Sk3 ∗ ∆{w1 ,w2 } .
Theorem 2.7. K ∗ is a Golod simplicial complex that is not stably homotopy Golod. More3
→ S42 defined in [19].
over, ZK ∗ is torsion free if K is constructed from the map η12 : S12
3. Proof of Theorem 2.3.
In this section, we prove Theorem 2.3. We begin by stating some elementary lemmas, for
which the proofs are omitted.
Lemma 3.1. Let K and L be simplicial complexes with disjoint index sets V and W ,
respectively. Then, (K ∗ L)∗V ⊔W = KV∗ ∗ ∆W ∪ ∆V ∗ L∗W .
A GOLOD COMPLEX WITH NON-SUSPENSION MOMENT-ANGLE COMPLEX
7
Lemma 3.2. Let K and L be simplicial complexes with disjoint index sets, and let Ki and
Li for i = 1, 2 be subcomplexes of K and L, respectively. Then,
(K1 ∗ L1 ) ∩ (K2 ∗ L2 ) = (K1 ∩ K2 ) ∗ (L1 ∩ L2 ).
Lemma 3.3. Let K be a simplicial complex with two subcomplexes K1 and K2 . If
K = K1 ∪ K2 , then
ZK (CX, X) = ZK1 (CX, X) ∪ ZK2 (CX, X)
and
ZK1 (CX, X) ∩ ZK2 (CX, X) = ZK1 ∩K2 (CX, X).
Lemma 3.4. Let K and L be simplicial complexes with disjoint index sets V and W ,
respectively. Then,
ZK∗L (CX, X) = ZK (CX V , X V ) × ZL (CX W , X W ).
Proposition 3.5. Let K be a simplicial complex on [m] and X be a sequence of based
CW-complexes. If ZK (CX, X) is a simply connected co-H-space, then
_
bI.
ZK (CX, X) ≃
Σ|KI | ∧ X
I⊂[m]
W
b I is also simply connected. By Theorem 1.2
Proof. First, we show that I⊂[m] Σ|KI | ∧ X
W
W
b I is a
bI) ∼
we have that H1 ( I⊂[m] Σ|KI | ∧ X
= H1 (ZK (CX, X)) = 0. Because I⊂[m] Σ|KI | ∧ X
W
b I ) = 0.
suspension, its fundamental group is a free group. Thus, π1 ( I⊂[m] Σ|KI | ∧ X
Q
Q
For a subset I ⊂ [m], the canonical projection pI : i∈[m] CXi → i∈I CXi induces a map
ZK (CX, X) → ZKI (CX I , X I ), which is also denoted by pI . Let I = {i1 , · · · , ik } ⊂ [m]. We
define a subset ZKI (CX I , X I )′ of ZKI (CX I , X I ) by the equation
ZKI (CX I , X I )′ = {(xi1 , · · · , xik ) ∈ ZKI (CX I , X I ) | xij = ∗ for some j}.
b I . Now, we consider the
In [15], it is shown that ZKI (CX I , X I )/ZKI (CX I , X I )′ ≃ Σ|KI | ∧ X
composite of maps
m
f : ZK (CX, X) →
2
_
∨I⊂[m] pI
ZK (CX, X) −−−−−→
→
_
I⊂[m]
_
ZKI (CX I , X I )
I⊂[m]
ZKI (CX I , X I )/ZKI (CX I , X I )′ ≃
_
I⊂[m]
bI,
Σ|KI | ∧ X
8
KOUYEMON IRIYE AND TATSUYA YANO
where the first map is the iterated co-multiplication of ZK (CX, X). Because ZK (CX, X) and
W
bI
I⊂[m] Σ|KI |∧ X are simply connected CW-complexes, to prove that f is homotopy equivalent
it suffices to show that f induces a homology isomorphism. In [16], it is shown that Σf is a
homotopy equivalence. In particular, f induces a homology isomorphism, and thus we complete
the proof.
Proof of Theorem 2.3. By Lemmas 3.1 and 3.2, we have that (K ∗L)∗ = K ∗ ∗∆W ∪∆V ∗L∗
and K ∗ ∗ ∆V (L) ∩ ∆V (K) ∗ L∗ = K ∗ ∗ L∗ . Therefore, from Lemma 3.3 we obtain the following
push-out diagram of spaces:
ZK ∗ ∗L∗ (CX, X) −−−→ ZK ∗ ∗∆W (CX, X)
y
y
Z∆V ∗L∗ (CX, X) −−−→ Z(K∗L)∗ (CX, X).
Here, we remark that K ∗ ∗ L∗ and K ∗ ∗ ∆W are non-void simplicial complexes, because we
assume that K 6= ∆V and L 6= ∆W . By Lemma 3.4, the above push-out diagram is equivalent
to the following push-out diagram:
Q
ZK ∗ (CX V , X V ) × ZL∗ (CX W , X W ) −−−→ ZK ∗ (CX V , X V ) × w∈W CXw
.
y
y
Q
−−−→
Z(K∗L)∗ (CX, X).
v∈V CXv × ZL∗ (CX W , X W )
Q
Q
Because v∈V CXv and w∈W CXw are contractible, the above diagram yields the following
homotopy equivalences:
Z(K∗L)∗ (CX, X) ≃ ZK ∗ (CX V , X V ) ∗ ZL∗ (CX W , X W )
≃ ΣZK ∗ (CX V , X V ) ∧ ZL∗ (CX W , X W ).
By Theorem 1.2, ΣZK ∗ (CX V , X V ) is a double suspension, which implies that Z(K∗L)∗ (CX, X)
is also a double suspension. By invoking Proposition 3.5, we complete the proof.
4. Proof of Theorem 2.5.
In this section, we prove Theorem 2.5. Again, we begin by stating some elementary lemmas.
Lemma 4.1. Let K and L be simplicial complexes without ghost vertices on index sets V
and W , respectively. Let α = V ∩ W be a common face of K and L. Then,
(K ∪α L)∗ = (∂∆V −α ∗ ∆α ∗ ∂∆W −α ) ∪ (K ∗ ∗ ∆W −α ) ∪ (∆V −α ∗ L∗ ).
Proof. For u ∈ V − α and v ∈ W − α, {u, v} is a minimal non-face of K ∪α L, and a minimal
non-face of K or L is also a minimal non-face of K ∪α L. For any u ∈ V − α and v ∈ W − α,
A GOLOD COMPLEX WITH NON-SUSPENSION MOMENT-ANGLE COMPLEX
9
V ∪ W − {u, v} is a facet of (K ∪α L)∗ . Furthermore, for any minimal non-face σ of K or L,
V ∪ W − σ is a facet of (K ∪α L)∗ . This implies the desired equality of the two complexes.
Lemma 4.2. If K is a simplicial complex with a ghost vertex, then the BBCG decomposition
for K ∗ is desuspendable. In particular, K ∗ is Golod.
Proof. Let K be a simplicial complex on [m] and v be a ghost vertex of K. Then, [m] − v
is a facet of K ∗ , and thus dim K ∗ ≥ m − 2. Then, it follows from Theorem 1.2 and Proposition
3.5 of [15] that the BBCG decomposition for K ∗ is desuspendable.
Lemma 4.3. Let α be a face of a simplicial complex K on an index set V . If K 6= ∆V and
α is not a facet of K, then the inclusion map
ZK ∗ (CX, X) → Z(∆α )∗ (CX, X) = Z∂∆V −α ∗∆α (CX, X)
is null homotopic.
Proof. Because α is not a facet of K, there is a face β of K such that α ( β. Then,
∆ ( ∆β ⊂ K, which implies that K ∗ ⊂ (∆β )∗ ( (∆α )∗ . That is,
K ∗ ⊂ ∂∆V −β ∗ ∆β ( ∂∆V −α ∗ ∆α . Therefore, the inclusion ZK ∗ (CX, X) ֒→ Z∂∆V −α ∗∆α (CX, X)
factors as ZK ∗ (CX, X) → Z∂∆V −β ∗∆β (CX, X) → Z∂∆V −α ∗∆α (CX, X). To show that the inclusion ZK ∗ (CX, X) → Z∂∆V −α ∗∆α (CX, X) is null homotopic, it is sufficient to show that
Z∂∆V −β ∗∆β (CX, X) → Z∂∆V −α ∗∆α (CX, X) is null homotopic. We have the following homotopy
commutative diagram
Q
≃
∗i∈V −β Xi ,
Z∂∆V −β ∗∆β (CX, X) = Z∂∆V −β (CX V −β , X V −β ) × j∈β CXj −−−→
y
y
Q
≃
Z∂∆V −α ∗∆α (CX, X) = Z∂∆V −α (CX V −α , X V −α ) × j∈α CXj −−−→ (∗i∈V −β Xi ) ∗ (∗i∈β−α Xi ),
α
where the horizontal maps are homotopy equivalences by Example 2.2 and the right vertical
map is the canonical inclusion X → X ∗ Y . Since this map is null homotopic,
Z∂∆V −β ∗∆β (CX, X) → Z∂∆V −α ∗∆α (CX, X) is also null homotopic, and thus we complete the
proof.
In addition to the above, we require the following lemma to prove Theorem 2.5.
Lemma 4.4 (Lemma 3.2 of [14] with B = ∗). Define Q as the push-out
ι×1
A × C −−−→ CA × C
1×∗
y
y
j
A × D −−−→
Q,
where ι : A → CA is the inclusion. Then, the homotopy equivalence
≃
Q−
→ (A ⋉ D) ∨ Σ(A ∧ C)
10
KOUYEMON IRIYE AND TATSUYA YANO
holds, which is natural with respect to A, C, and D, where X ⋉Y = (X ×Y )/(X ×∗). Moreover,
the inclusion map j : A × D → Q is homotopic to the following composite of maps
A × D → A ⋉ D → (A ⋉ D) ∨ Σ(A ∧ C) ≃ Q,
where the first map is the collapsing map and the second map is the inclusion.
Proof of Theorem 2.5. If K or L has a ghost vertex, then K ∪α L also has a ghost vertex.
In this case, the BBCG decomposition is desuspendable, by Lemma 4.2. Therefore, we assume
that K and L do not have any ghost vertices. In the following proof, ZK (CX, X) is abbreviated
as ZK .
First, we will show that if K = ∆V and L 6= ∆W , then
Z(∆V ∪α L)∗ ≃ (Z∂∆V −α ⋉ Z∂∆W −α ) ∨ Σ(Z∂∆V −α ∧ ZL∗ ).
Here we remark that we need the assumption that L 6= ∆W and that α is not a facet of L to
apply Lemma 4.3.
It follows from Lemma 4.1 that (∆V ∪α L)∗ = (∂∆V −α ∗ ∆α ∗ ∂∆W −α ) ∪ (∆V −α ∗ L∗ ). Then,
by Lemma 3.3 we have the push-out diagram of spaces
// Z∆V −α ∗L∗
Z∂∆V −α ∗L∗
Z∂∆V −α ∗∆α ∗∂∆W −α
// Z(∆V ∪
∗
α L)
,
which by Lemma 3.4 is equivalent to the following push-out diagram:
−−−→ Z∆V −α × ZL∗
Z∂∆V −α × ZL∗
id×incly
y
Z∂∆V −α × Z∆α × Z∂∆W −α −−−→
Z(∆V ∪α L)∗ .
Since we assumed that L 6= ∆W and that α is not a facet of L, by Lemma 4.3 the inclusion
map ZL∗ ֒→ Z∆α × Z∂∆W −α is null-homotopic. Therefore, Z(∆V ∪α L)∗ is homotopy equivalent to
the push-out P of the following diagram:
Z∂∆V −α × ZL∗
−−−→ Z∆V −α × ZL∗
id×∗
y
y
(4.1)
j
Since Z∆V −α
Z∂∆V −α × Z∆α × Z∂∆W −α −−−→
P.
= CZ∂∆V −α , by Lemma 4.4 P is homotopy equivalent to
(Z∂∆V −α ⋉ (Z∆α × Z∂∆W −α )) ∨ Σ(Z∂∆V −α ∧ ZL∗ ) ≃ (Z∂∆V −α ⋉ Z∂∆W −α ) ∨ Σ(Z∂∆V −α ∧ ZL∗ ),
and j in the diagram (4.1) can be identified with the following composite of canonical maps:
(4.2) Z∂∆V −α × Z∆α × Z∂∆W −α → Z∂∆V −α × Z∂∆W −α →
Z∂∆V −α ⋉ Z∂∆W −α ֒→ (Z∂∆V −α ⋉ Z∂∆W −α ) ∨ Σ(Z∂∆V −α ∧ ZL∗ ).
A GOLOD COMPLEX WITH NON-SUSPENSION MOMENT-ANGLE COMPLEX
11
Thus, the following homotopy equivalence holds:
Z(∆V ∪α L)∗ ≃ (Z∂∆V −α ⋉ Z∂∆W −α ) ∨ Σ(Z∂∆V −α ∧ ZL∗ ).
This homotopy equivalence induces the following homotopy equivalences:
Z(∆V ∪α L)∗ ≃ (Z∂∆V −α ⋉ Z∂∆W −α ) ∨ Σ(Z∂∆V −α ∧ ZL∗ )
≃ Z∂∆W −α ∨ (Z∂∆V −α ∧ Z∂∆W −α ) ∨ Σ(Z∂∆V −α ∧ ZL∗ )
_
bJ
ΣZ∂∆V −α ∧ Σ|(L∗ )J | ∧ X
≃ Z∂∆W −α ∨ (Z∂∆V −α ∧ Z∂∆W −α ) ∨
J⊂W
≃ Z∂∆W −α ∨ (Z∂∆V −α ∧ Z∂∆W −α ) ∨
≃
_
I⊂[m]
_
J⊂W
b V −α+J
Σ3 |∂∆V −α | ∧ |(L∗ )J | ∧ X
bI,
Σ|((∆ ∪α L) )I | ∧ X
V
∗
where in the third homotopy equivalence we used the BBCG decomposition for ΣZL∗ . To show
that the last homotopy equivalence holds we need the homotopy type of |((∆V ∪α L)∗ )I | for
I ⊂ [m]. Since (∆V ∪α L)∗ = (∂∆V −α ∗ ∆α ∗ ∂∆W −α ) ∪ (∆V −α ∗ L∗ ), we have the following
push-out diagram for I = J ′ ⊔ J with J ′ ⊂ V − α and J ⊂ W :
|(∂∆V −α )J ′ ∗ (L∗ )J |
y
−−−→ |(∆V −α )J ′ ∗ (L∗ )J |
y
|(∂∆V −α )J ′ ∗ (∆α ∗ ∂∆W −α )J | −−−→ |((∆V ∪α L)∗ )I |.
Since α is not a facet of L, there is a face β of L and we have inclusions ∆α ( ∆β ⊂ L.
This implies that |(L∗ )J | ⊂ |(∆β ∗ ∂∆W −β )J | ⊂ |(∆α ∗ ∂∆W −α )J |. Since the inclusion map
|(∆β ∗ ∂∆W −β )J | ֒→ |(∆α ∗ ∂∆W −α )J | is null homotopic, the left vertical map of the diagram
above is null homotopic. Thus we have
|((∆V ∪α L)∗ )I | ≃ |(∂∆V −α )J ′ ∗ (∆α ∗ ∂∆W −α )J | ∨ |(∆V −α )J ′ ∗ (L∗ )J |/|(∂∆V −α )J ′ ∗ (L∗ )J |.
If J ′ 6= ∅, V −α, then |(∂∆V −α )J ′ | is contractible, and |((∆V ∪α L)∗ )I | is also contractible. If J ′ =
∅, then |((∆V ∪α L)∗ )I | = |(∆α ∗ ∂∆W −α )J | which is |∂∆W −α | if J = W − α, and is contractible
otherwise. Finally we obtain the following homotopy equivalence, where |∂∆V −α ∗ (∂∆W −α )J |
is contractible unless J = W − α.
|((∆V ∪α L)∗ )I | ≃
W −α
|
for I = W − α,
|∂∆
V −α
W −α
V −α
∗
|∂∆
∗ (∂∆
)J | ∨ Σ|∂∆
∗ (L )J | for I = (V − α) ⊔ J for some J ⊂ W ,
∗
otherwise.
Here, we remark that we can apply Proposition 3.5 if |W − α| ≥ 3.
12
KOUYEMON IRIYE AND TATSUYA YANO
Similarly, if K 6= ∆V and L = ∆W then we have
Z(K∪α ∆W )∗ ≃ (Z∂∆V −α ⋊ Z∂∆W −α ) ∨ Σ(ZK ∗ ∧ Z∂∆W −α ) ≃
_
I⊂[m]
bI.
Σ|((K ∪α ∆W )∗ )I | ∧ X
Next, we consider the case that K 6= ∆V and L 6= ∆W . Then, we have the push-out diagram
(∂∆V −α ∗ ∆α ∗ ∂∆W −α )
y
−−−→ (∂∆V −α ∗ ∆α ∗ ∂∆W −α ) ∪ (∆V −α ∗ L∗ )
y
(∂∆V −α ∗ ∆α ∗ ∂∆W −α ) ∪ (K ∗ ∗ ∆W −α ) −−−→
(K ∪α L)∗ ,
which induces the following push-out diagram of spaces by Lemmas 3.3 and 3.4:
−−−→ Z(∂∆V −α ∗∆α ∗∂∆W −α )∪(∆V −α ∗L∗ )
Z∂∆V −α × Z∆α × Z∂∆W −α
y
y
.
Z(∂∆V −α ∗∆α ∗∂∆W −α )∪(K ∗ ∗∆W −α ) −−−→
Z(K∪α L)∗ .
Because the top arrow of the diagram above is a closed cofibration, Z(K∪α L)∗ is homotopy
equivalent to the homotopy push-out of the following diagram:
−−−→ Z(∂∆V −α ∗∆α ∗∂∆W −α )∪(∆V −α ∗L∗ )
Z∂∆V −α × Z∆α × Z∂∆W −α
y
.
Z(∂∆V −α ∗∆α ∗∂∆W −α )∪(K ∗ ∗∆W −α )
Since we have the following homotopy equivalences
Z(∂∆V −α ∗∆α ∗∂∆W −α )∪(∆V −α ∗L∗ ) ≃ (Z∂∆V −α ⋉ (Z∆α × Z∂∆W −α )) ∨ Σ(Z∂∆V −α ∧ ZL∗ )
≃ (Z∂∆V −α ⋉ Z∂∆W −α ) ∨ Σ(Z∂∆V −α ∧ ZL∗ ),
Z(∂∆V −α ∗∆α ∗∂∆W −α )∪(K ∗ ∗∆W −α ) ≃ ((Z∂∆V −α × Z∆α ) ⋊ Z∂∆W −α ) ∨ Σ(ZK ∗ ∧ Z∂∆W −α )
≃ (Z∂∆V −α ⋊ Z∂∆W −α ) ∨ Σ(ZK ∗ ∧ Z∂∆W −α ),
and homotopy push-out preserves homotopy equivalences, Z(K∪α L)∗ is homotopy equivalent to
the homotopy push-out of the following diagram:
Z∂∆V −α × Z∆α × Z∂∆W −α
−−−→ (Z∂∆V −α ⋉ Z∂∆W −α ) ∨ Σ(Z∂∆V −α ∧ ZL∗ )
j1
j
y2
(Z∂∆V −α ⋊ Z∂∆W −α ) ∨ Σ(ZK ∗ ∧ Z∂∆W −α )
where j1 and j2 are as in (4.2). It follows that
Z(K∪α L)∗ ≃ Q ∨ Σ(Z∂∆V −α ∧ ZL∗ ) ∨ Σ(ZK ∗ ∧ Z∂∆W −α ),
where Q is the homotopy push-out of the following diagram:
Z∂∆V −α × Z∂∆W −α −−−→ Z∂∆V −α ⋉ Z∂∆W −α
.
y
Z∂∆V −α ⋊ Z∂∆W −α
A GOLOD COMPLEX WITH NON-SUSPENSION MOMENT-ANGLE COMPLEX
13
where the vertical and horizontal maps are the canonical collapsing maps. Then, it is easy to
see that Q ≃ Z∂∆V −α ∧ Z∂∆W −α , and we have obtained a homotopy equivalence
(4.3)
Z(K∪α L)∗ ≃ (Z∂∆V −α ∧ Z∂∆W −α ) ∨ Σ(Z∂∆V −α ∧ ZL∗ ) ∨ Σ(ZK ∗ ∧ Z∂∆W −α ).
To complete the proof, we now apply Proposition 3.5. Clearly, (4.3) implies that Z(K∪αL)∗ is
a suspension space. Because α is neither a facet of K nor L, we have that |V − α| ≥ 2 and
|W − α| ≥ 2. Then, by Example 2.2 we have that Z∂∆V −α = Z∂∆V −α (CX, X) and
Z∂∆W −α = Z∂∆W −α (CX, X) are suspensions. Thus, (4.3) implies that Z(K∪α L)∗ is a double
suspension, and by Proposition 3.5 we have the desired homotopy equivalence
_
bI
Σ|((K ∪α L)∗ )I | ∧ X
Z(K∪αL)∗ (CX, X) ≃
I⊂[m]
and we complete the proof.
5. Preliminaries for the proof of Theorem 2.7
In this section, we review the Alexander duality and elementary collapses of simplicial complexes, and the Massey products of the Koszul homology.
5.1. Alexander duality. First, we review the Alexander duality of simplicial complexes. Although an elementary proof has been derived [4], we follow the classical argument presented in
chapter six of Spanier’s book [24].
Let M be a simplicial complex that is a subcomplex of ∂∆n+1 , and let L be a subcomplex
of M. We denote the Alexander dual of L and M in V (∂∆n+1 ) by L∗ , M ∗ . Then, M ∗ is a
subcomplex of L∗ , and the following duality is well-known, where we write S n = |∂∆n+1 |:
γ : Hq (|L∗ |, |M ∗ |) ∼
= Hq (|L̄|, |M̄|)
= Hq (|SdL∗ |, |SdM ∗ |) ∼
∼
= H n−q (|M|, |L|)
= Hq (S n − |L|, S n − |M|) ∼
where SdK denotes the barycentric subdivision of K, L̄ is the supplement of L in ∂∆n+1 defined
in Definition 2.5.18 of [20], and the last isomorphism is the (topological) duality
γU : Hq (S n − |L|, S n − |M|) ∼
= H n−q (|M|, |L|)
induced by an orientation class U ∈ H n (S n × S n , S n × S n − ∆(S n )) of S n , where ∆(S n ) is the
diagonal set of S n × S n .
Now, we consider the cohomology theory with coefficient Z2 = Z/2Z, and the relation between the duality map γ : Hq (|L∗ |, |M ∗ |; Z2 ) → H n−q (|M|, |L|; Z2 ) and the Steenrod squaring
operations Sq i . For a pair of finite complexes (X, A), the Steenrod squaring operations Sq∗i on
the homology theory are defined as follows:
< x, Sq∗i a >=< Sq i x, a >
for x ∈ H q (X, A; Z2 ) and a ∈ Hq+i (X, A; Z2 ),
where < −, − > denotes the Kronecker product.
14
KOUYEMON IRIYE AND TATSUYA YANO
Lemma 5.1.1. For a ∈ Hq (|L∗ |, |M ∗ |) and i > 0, we have that
γ(Sq∗i a)
=
i−1
X
Sq i−k (γ(Sq∗k a)).
k=0
Proof. In this proof, we omit the coefficient ring Z2 in the (co)homology theory. First, we
remark that Sq i (U) = 0 for i > 0. This follows from the fact that the natural restriction map
H n (S n × S n , S n × S n − ∆(S n )) → H̃ n (S n × S n ) is monomorphic. In the definition of the duality
map γ, all maps except for γU commute with the squaring operations. Therefore, it suffices to
prove the corresponding formula for γU .
Because γU (Sq∗i a) ∈ H n−q+i (|M|, |L|), we take an element b ∈ Hn−q+i (|M|, |L|) and compute
the Kronecker product < γU (Sq∗i a), b >. Let
j : (|M|, |L|) × (S n − |L|, S n − |M|) → (S n × S n , S n × S n − ∆(S n ))
be the inclusion map. Then,
< γU (Sq∗i a), b > = < j ∗ (U)/Sq∗i a, b >
= < j ∗ (U), b × Sq∗i a >
= < j ∗ (U), Sq∗i (b × a) −
i−1
X
Sq∗i−k b × Sq∗k a >
k=0
i ∗
= < Sq j (U), b × a > +
i−1
X
< j ∗ (U), Sq∗i−k b × Sq∗k a >
k=0
=
i−1
X
< j ∗ (U)/Sq∗k a, Sq∗i−k b >
k=0
= <
i−1
X
Sq i−k (γU (Sq∗k a)), b >,
k=0
where in the fifth equation we use the fact that Sq i j ∗ (U) = j ∗ Sq i (U) = 0. Thus, the proof is
completed.
5.2. Elementary collapse. In this subsection, we briefly review the notion of elementary
collapse, which is necessary to prove Theorem 2.7.
Definition 5.2.1. A non-empty face σ in a simplicial complex K is a free face if it is not a
facet of K and is contained in exactly one facet of K.
An elementary collapse of K is a simplicial complex K ′ obtained from K by the removal of
a free face σ, along with all faces that contain σ. If there is a sequence of elementary collapses
leading from K to K ′ , we say that K is collapsible onto K ′ , and we use the notation K ց K ′ .
Then |K ′ | is a deformation retract of |K|.
A GOLOD COMPLEX WITH NON-SUSPENSION MOMENT-ANGLE COMPLEX
15
Example 5.2.2. Let ∆m denote the full simplex on the vertex set [m + 1]. There exists a
sequence of elementary collapses of ∆m ⊠ ∆1 given by removing pairs of faces from the top to
bottom in the following list:
{1 · · · m + 1, 1 · · · m + 1m + 1},
···
{1 · · · ii + 1 · · · m + 1, 1 · · · ii i + 1 · · · m + 1},
···
{12 · · · m + 1, 11 2 · · · m + 1},
where i = (i, 1), i = (i, 2), and 1 · · · ii + 1 · · · m + 1 denotes, for example, the face
{1, · · · , i, i + 1, · · · , m + 1} of ∆m ⊠ ∆1 . Thus, we see that ∆m ⊠ ∆1 ց (∂∆m ⊠ ∆1 ) ∪ (∆m ⊠ 2).
By repeating the same process for smaller simplicies, we see that ∆m ⊠ ∆1 ց ∆m ⊠ 2. This
can be applied to any simplicial complex K, and we see that K ⊠ ∆1 is collapsible onto K ⊠ 2.
5.3. Massey product in Tork[v1 ,··· ,vm ] (k[K], k). Finally, we review the Massey products in
Tork[v1 ,··· ,vm ] (k[K], k), according to Section 3.2 of Buchstaber and Panov’s book [5]. Here, we
remark that we adopt the different conventions for the grading of Tork[v1 ,··· ,vm ] (k[K], k).
Recall that the torsion algebra Tork[v1 ,··· ,vm ] (k[K], k) is defined as the homology group of the
Koszul differential algebra (Λ[u1 , · · · , um ] ⊗ k[K], d), where their bigrading and the differentials
are defined by deg ui = (1, 2), deg vi = (0, 2), and dui = vi , for i = 1, · · · , m.
A multigrading of the torsion algebra Tork[v1 ,··· ,vm ] (k[K], k) is defined by setting
im
= (ε1 + · · · + εm , 2(i1 + ε1 ), · · · , 2(im + εm )),
mdeg uε11 · · · uεmm v1i1 · · · vm
and it is easy to see that
1 ,··· ,vm ]
Tork[v
(k[K], k) =
∗
M
k[v ,··· ,vm ]
Tor∗,2a1
(k[K], k).
a∈Nm
A subset I ⊂ [m] may be viewed as a (0, 1)-vector in Nm , whose i-th coordinate is 1 if i ∈ I
and 0 otherwise. Then, the following multigraded version of Hochster’s formula holds.
Theorem 5.3.1(Theorem 3.2.9 of [5]). For any subset I ⊂ [m], we have that
k[v ,··· ,vm ]
Tori,2I1
k[v ,··· ,vm ]
and Tori,2a1
(k[K], k) ∼
= H̃ |I|−i−1(KI ; k),
(k[K], k) = 0 unless a is a (0, 1)-vector.
We recall the definition of the Massey product. Let n be an integer greater than 1. Define ā
to be (−1)deg a+1 a. Then, the n-fold Massey product ha1 , · · · , an i is defined to be the set of all
homology classes represented by elements of the form
a1,n = ā1,1 a2,n + ā1,2 a3,n + · · · + ā1,n−1 an,n ,
16
KOUYEMON IRIYE AND TATSUYA YANO
for all solutions of the equations
ai = [ai,i ] for 1 ≤ i ≤ n,
dai,j = āi,i ai+1,j + āi,i+1 ai+2,j + · · · + āi,j−1 aj,j
for 1 ≤ i < j ≤ n, (i, j) 6= (1, n).
k[v ,··· ,vm ]
1
By inductive argument it is easy to see that if ai ∈ Torki ,2a
i
(k[K], k), then
k[v ,··· ,v ]
m
1
ai,j ∈ Torki +···+k
(k[K], k) for 1 ≤ i ≤ j ≤ n, (i, j) 6= (1, n),
j +(j−i),2(ai +···+aj )
and therefore,
k[v ,··· ,v ]
m
1
ha1 , · · · , an i ⊂ Tork1 +···+k
(k[K], k).
n +(n−2),2(a1 +···+an )
The Massey product ha1 , · · · , an i is said to be trivial if it contains the zero-element.
Proposition 5.3.2. If an n-fold Massey product ha1 , · · · , an i is defined for
k[v ,··· ,v ]
ai ∈ Tor∗,2I1i m (k[K], k), and Ik ∩ Iℓ 6= ∅ for some k 6= ℓ, then ha1 , · · · , an i is trivial.
Proof. Let n ≥ 2, and consider an n-fold Massey product. As is observed above we have
k[v1 ,··· ,vm ]
ha1 , · · · , an i ⊂ Tor∗,2(I
(k[K], k). If Ik ∩ Iℓ 6= ∅ for some k 6= ℓ, then I1 + · · · + In is not a
1 +···+In )
k[v ,··· ,v ]
m
1
(0, 1)-vector. Therefore, it follows from Theorem 5.3.1 that Tor∗,2(I
(k[K], k) = 0. Thus,
1 +···+In )
we complete the proof.
Corollary 5.3.3. Let K = K1 ∪α K2 be a simplicial complex on [m], which is obtained from
two simplicial complexes K1 and K2 by gluing along a common simplex α that is neither equal
to V (K1 ) nor V (K2 ). Then, K ∗ is non-Golod over a field k if and only if there are faces σ and
τ of K satisfying the following conditions:
(1) V (K1 ) − α ⊂ σ, V (K2 ) − α ⊂ τ , and σ ∪ τ = [m],
(2) the inclusion map
starlinkK (σ∩τ ) (σ − σ ∩ τ ) ∪ starlinkK (σ∩τ ) (τ − σ ∩ τ ) → linkK (σ ∩ τ )
induces a non-trivial map in the homology theory with coefficients k.
Proof. First, we show that Massey products for n > 2 in Tor∗k[v1 ,··· ,vm ] (k[K ∗ ], k) are trivial for
any simplicial complex K = K1 ∪α K2 . This reduces the problem to proving that a non-trivial
cup product exists if and only if the conditions (1) and (2) hold.
k[v ,··· ,v ]
Let ai ∈ Tor∗,2I1i m (k[K ∗ ], k) for i = 1, · · · , n, where n > 2. Then, we want to prove that
the n-fold Massey product ha1 , · · · , an i is trivial. If J is a face of K ∗ , then
k[v ,··· ,v ]
Tor∗,2J1 m (k[K ∗ ], k) ∼
= H̃ ∗ ((K ∗ )J ; k) = 0. By this fact and the proposition above, we may
assume that Ii = σic for some simplex σi ∈ K for i = 1, · · · , n, and Ii ∩ Ij = ∅ for i 6= j. This
implies that σi ∪ σj = [m]. Thus, we may also assume that v0 ∈ I1 , where v0 ∈ V (K1 ) − α.
Then, v0 6∈ Ij = σjc , i.e., v0 ∈ σj for j > 1. Because a face of K containing the vertex v0 is a
subset of V (K1 ), we see that V (K2 ) − α = V (K1 )c ⊂ σjc = Ij for j > 1, which contradicts the
A GOLOD COMPLEX WITH NON-SUSPENSION MOMENT-ANGLE COMPLEX
17
assumption that Ij ∩ Ik = ∅ for j 6= k > 1. Thus, we have proved that Massey products for
1 ,··· ,vm ]
n > 2 in Tork[v
(k[K ∗ ], k) are trivial.
∗
It follows from Theorem 1.1 that K ∗ is non-Golod if and only if there are disjoint subsets I
and J of [m] such that
ιI,J : (K ∗ )I⊔J = (linkK (I c ∩ J c ))∗I⊔J → (K ∗ )I ∗ (K ∗ )J = (linkK (I c ))∗I ∗ (linkK (J c ))∗J
induces a non-trivial map in the cohomology theory. Then, I c and J c must be faces of K,
and I c ∪ J c = [m], because I ∩ J = ∅. Thus, we may assume that V (K1 ) − α ⊂ I c = σ and
V (K2 ) − α ⊂ J c = τ . By the Alexander duality and its naturality, it follows that ιI,J induces
a non-trivial map in the cohomology theory if and only if the dual map
ι∗I,J : (linkK (σ)∗ ∗ linkK (τ )∗ )∗I⊔J
= linkK (σ) ∗ ∆J ∪ ∆I ∗ linkK (τ )
= starlinkK (σ∩τ ) (σ − σ ∩ τ ) ∪ starlinkK (σ∩τ ) (τ − σ ∩ τ )
→ linkK (σ ∩ τ )
induces a non-trivial map in the homology theory, where the first equality follows from Lemma
3.1. Thus, we complete the proof.
Example 5.3.4. Let L be a simplicial complex on {2, 3, 4, 5, 6, 7} with facets
267, 367, 467, 567, 236, 456, 2345.
Put α = 2345 and consider the simplicial complex K obtained from ∆α+1 and L by gluing
along a common simplex α. Then, the Alexander dual of K = ∆α+1 ∪α L is not Golod. In fact,
for σ = 12345 and τ = 67, the inclusion map
starK (σ) ∪ starK (τ ) = ∆α+1 ∪ {∅, 2, 3, 4, 5} ∗ ∆{6,7} → K
induces a non-trivial map in one dimensional homology groups. Needless to say, α is a facet of
L.
6. Proof of Theorem 2.7
In this section, we prove Theorem 2.7. The proof is divided into three parts. In the first part,
we show that K ∗ is Golod. Next, we show that K ∗ is not stably homotopy Golod. Finally, we
demonstrate that ZK ∗ is torsion free for a particular K.
Before beginning the proof, we study the topology of the space |(Sk3 ⊠ ∆1 ) ∪ηk S42 |. As
remarked at the end of Example 5.2.2, there exists a sequence of elementary collapses that
collapses Sk3 ⊠ ∆1 onto Sk3 ⊠ 2 = Sk3 . This collapsion induces a deformation retraction |Sk3 ⊠ ∆1 |
onto |Sk3 ⊠ 2| = |Sk3 |, and the composite
|Sk3 | = |Sk3 ⊠ 1| ֒→ |Sk3 ⊠ ∆1 | → |Sk3 ⊠ 2| = |Sk3 |
is homotopic to the identity map. This deformation retraction induces a deformation retraction
π : |(Sk3 ⊠ ∆1 ) ∪ηk S42 | → |S42 |
18
KOUYEMON IRIYE AND TATSUYA YANO
such that
|i|
π
π ◦ |i| ≃ |ηk | : |Sk3 | −
→ |(Sk3 ⊠ ∆1 ) ∪ηk S42 | −
→ |S42 |,
where i is the inclusion map.
6.1. K ∗ is Golod. Recall that
K = ∆V +v0 ∪ ((Sk3 ⊠ ∆1 ∪ηk S42 ) ∪ F (Sk3)) ∗ ∆{w1 } ∪ Sk3 ∗ ∆{w1 ,w2} ,
where F (Sk3) is the simplicial complex defined in §2. We apply Corollary 5.3.3 to K = K1 ∪α K2 ,
where
K1 = ∆V +v0 ,
K2 = ∆V ∪ ((Sk3 ⊠ ∆1 ∪ηk S42 ) ∪ F (Sk3 )) ∗ ∆{w1 } ∪ Sk3 ∗ ∆{w1 ,w2 } ,
α = V , V (K1 ) = V + v0 , and V (K2 ) = V ⊔ {w1 , w2 }. Thus, by Corollary 5.3.3 we only have to
prove that the map
starlinkK (σ∩τ ) (σ − σ ∩ τ ) ∪ starlinkK (σ∩τ ) (τ − σ ∩ τ ) → linkK (σ ∩ τ )
induces the trivial map in the homology theory, where σ and τ are faces of K such that v0 ∈ σ,
{w1 , w2 } ⊂ τ , and σ ∪ τ = V (K).
First, we consider the case with σ ∩ τ = ∅. That is, we prove that the map
starK (σ) ∪ starK (τ ) = ∆V +v0 ∪ starK (τ ) → K
induces the trivial map in the homology theory. Because this map factors as
∆V +v0 ∪ starK (τ ) → ∆V +v0 ∪ Sk3 ∗ ∆{w1 ,w2 } → K,
it suffices to prove that the inclusion map
∆V +v0 ∪ Sk3 ∗ ∆{w1 ,w2 } → K
induces the trivial map in the homology theory. Because |∆V +v0 ∪ Sk3 ∗ ∆{w1 ,w2 } | ≃ Σ|Sk3 | = S 4
and |K| ≃ Σ|S42 | = S 3 , the above inclusion map induces the trivial map in the homology theory.
Next, we consider the case that ρ = σ ∩ τ is a non-empty face of Sk3 . Let L = L1 ∪α L2 be a
simplicial complex obtained from L1 and L2 by gluing along a common face α, and let β be a
face of α. Then linkL (β) = linkL1 (β) ∪α−β linkL2 (β). By this observation we see that
linkK (ρ) = ∆(V +v0 )−ρ ∪ (linkSk3 ⊠∆1 ∪ηk S42 (ρ) ∪ F (linkSk3 (ρ)) ∗ ∆w1 ∪ linkSk3 (ρ) ∗ ∆{w1 ,w2 } .
The situation is, therefore, similar to the case with ρ = ∅. Here, we are only required to prove
that the map
f : |linkSk3 (ρ)| → |linkSk3 ⊠∆1 ∪ηk S42 (ρ)|
induced by the inclusion induces the trivial map in the homology theory. f factors through
a contractible space |linkSk3 ⊠∆1 (ρ)|. In fact, because Sk3 ⊠ ∆1 is a triangulation of S 3 × I and
ρ is a non-empty simplex in the boundary, it follows that linkSk3 ⊠∆1 (ρ) is a triangulation of a
hemisphere. Thus, f clearly induces the trivial map in the homology theory, and we complete
the proof.
A GOLOD COMPLEX WITH NON-SUSPENSION MOMENT-ANGLE COMPLEX
19
6.2. K ∗ is not stably homotopy Golod. Next, we show that the map
(6.1)
|ι{w1 ,w2 },V +v0 | : |K ∗ | → |(K ∗ ){w1 ,w2 } ∗ (K ∗ )V +v0 |
is stably non-trivial. To show this, we consider the mapping cone of |ι{w1 ,w2},V +v0 |, which is
denoted by C|ι{w1 ,w2 },V +v0 | , and show that Sq 2 acts non-trivially on its mod-2 cohomology groups.
Let m = |V + {v0 , w1 , w2 }|. Then we have the following isomorphisms of (co)homology groups:
H̃ p (C|ι{w1 ,w2 },V +v0 | ) ∼
= H p (|(K ∗ ){w1 ,w2 } ∗ (K ∗ )V +v0 |, |K ∗|)
∼
= Hm−2−p (K, ∆V +v0 ∪ S 3 ∗ ∆{w1 ,w2 } )
k
∼
= Hm−2−p (∆V +v0 ∪ (Sk3 ⊠ ∆1 ∪ηk S42 ) ∗ ∆{w1 } , ∆V +v0 ∪ Sk3 ∗ ∆{w1 } )
∼
= Hm−2−p (Σ|Sk3 ⊠ ∆1 ∪ηk S42 |, Σ|Sk3|)
∼
= H̃m−2−p (ΣCP 2 ),
where the second isomorphism follows from the Alexander duality and the fact that
((K ∗ ){w1 ,w2 } ∗ (K ∗ )V +v0 )∗ = ((K ∗ ){w1 ,w2 } )∗{w1 ,w2 } ∗ ∆V +v0 ∪ ∆{w1 ,w2 } ∗ ((K ∗ )V +v0 )∗V +v0 )
= linkK (V + v0 ) ∗ ∆V +v0 ∪ ∆{w1 ,w2 } ∗ linkK ({w1 , w2 })
= ∆V +v0 ∪ Sk3 ∗ ∆{w1 ,w2 } .
In the third isomorphism we deformed Sk3 ∗ ∆{w1 ,w2 } onto Sk3 ∗ ∆{w1 } , and after that we deformed
F (Sk3 ) ∗ ∆{w1 } onto Sk3 ∗ ∆{w1 } , in the fourth isomorphism we collapses |∆V +v0 | to the point,
and in the fifth isomorphism we used the fact that |(Sk3 ⊠ ∆1 ) ∪ηk S42 | is the mapping cylinder
of η : |Sk3 | → |S42 | and CP 2 ≃ S 2 ∪η e4 where e4 denotes a 4-dimensional cell. Since Sq∗2
acts non-trivially on H̃5 (ΣCP 2 ), it follows from Lemma 5.1.1 that Sq 2 acts non-trivially on
H̃ m−7 (C|ι{w1 ,w2 },V +v0 | ), which implies that the map (6.1) is stably non-trivial.
6.3. ZK ∗ is torsion free. Finally, we will demonstrate that ZK ∗ is torsion free if K is con3
structed using the simplicial map η12 : S12
→ S42 described in [19]. We believe that ZK ∗ is
torsion free in general, but we are currently unable to prove this.
By Theorem 1.1 and the Alexander duality, we have the following isomorphisms:
M
M
H̃2|I|−p−2(linkK (I c ) : Z),
H̃ p−|I|−1((K ∗ )I : Z) ∼
H p (ZK ∗ ; Z) ∼
=
=
I⊂[m]
I⊂[m]
where we have used the fact that (K ∗ )I = (linkK (I c ))∗ . Thus, to show that ZK ∗ is torsion free,
it suffices to show that every linkK (σ) is torsion free for all faces σ of K. The longest part of
the computation is the following, and the remaining parts are omitted.
3
1
2
3 ⊠∆1 ∪
2 (σ) is torsion free for any simplex σ in S
Lemma 6.3.1. linkS12
12 ⊠ ∆ ∪η12 S4 .
η12 S4
3
Proof. Set L = S12
⊠ ∆1 ∪η12 S42 . If σ = ∅, then clearly linkL (σ) = L is torsion free.
20
KOUYEMON IRIYE AND TATSUYA YANO
3
Let v be a vertex of S12
. The Mayer-Vietoris sequence associated with the decomposition
L = (L − v) ∪ starL (v) reduces to the long exact sequence
· · · → H̃i (linkL (v)) → H̃i (L − v) → H̃i (L) → · · · .
3
3
The sequence of elementary collapses of S12
⊠ ∆1 onto S12
⊠ 2 given in Example 5.2.2 induces
3
1
2
a sequence of elementary collapses of S12 ⊠ ∆ ∪η12 S4 onto S42 . Moreover, this sequence of
elementary collapses induces a collapse of L − v onto S42 . This means that the inclusion map
|L − v| → |L| is homotopy equivalent, and that H̃∗ (linkL (v)) = 0.
3
To proceed further with the computation, we need to know the concrete structure of L. S12
has twelve vertices {ai , bi , ci , di }i=0,1,2 , and the following is the list of its facets:
a0 b0 c0 c1
a0 a2 b0 d1
a0 b1 d0 d2
b0 c0 c2 d2
a0 b0 b1 c1
a0 b0 b1 d1
b1 c2 d0 d2
b0 c0 c1 d2
a0 a1 b1 c1
b0 b1 c1 d1
a0 c2 d0 d2
a0 c0 c2 d2
a1 a2 b1 c1
b1 c1 c2 d1
a0 a1 b1 d2
a0 c0 c1 d2
a2 b1 c1 c2
a2 c1 c2 d1
a1 a2 b1 d2
a0 a1 c1 d2
a2 b1 b2 c2
a0 a2 c2 d1
a2 b1 b2 d2
a1 a2 c1 d2
a2 b0 b2 c2
a0 b1 d0 d1
a2 b0 b2 d2
a2 b0 d1 d2
a0 a2 b0 c2
b1 c2 d0 d1
b1 b2 c2 d2
b0 c1 d1 d2
a0 b0 c0 c2
a0 c2 d0 d1
b0 b2 c2 d2
a2 c1 d1 d2
We order the vertices as a0 < a1 < a2 < b0 < · · · < c0 < · · · < d0 < d1 < d2 . The vertices of
3
are a, b, c, d, and are ordered such that a < b < c < d. Furthermore, η12 : S12
→ S42 is the
map defined by the mappings ai 7→ a, bi 7→ b, ci 7→ c, and di 7→ d, for i = 0, 1, 2. Then, the list
of the facets of L is as follows:
a0 b0 c0 c1 c
a0 a2 b0 d1 d a0 a2 b0 bd a0 a2 abd a0 b1 d0 d2 d
a0 b0 b1 c1 c a0 b0 b1 bc
a0 b0 b1 d1 d a0 b0 b1 bd
b1 c2 d0 d2 d
a0 a1 b1 c1 c a0 a1 b1 bc a0 a1 abc b0 b1 c1 d1 d b0 b1 c1 cd b0 b1 bcd a0 c2 d0 d2 d
a1 a2 b1 c1 c a1 a2 b1 bc a1 a2 abc b1 c1 c2 d1 d b1 c1 c2 cd
a0 a1 b1 d2 d a0 a1 b1 bd a0 a1 abd
a2 b1 c1 c2 c
a2 c1 c2 d1 d a2 c1 c2 cd
a1 a2 b1 d2 d a1 a2 b1 bd a1 a2 abd
a2 b1 b2 c2 c a2 b1 b2 bc
a0 a2 c2 d1 d a0 a2 c2 cd a0 a2 acd a2 b1 b2 d2 d a2 b1 b2 bd
a2 b0 b2 c2 c a2 b0 b2 bc
a0 b1 d0 d1 d
a2 b0 b2 d2 d a2 b0 b2 bd
a0 a2 b0 c2 c a0 a2 b0 bc a0 a2 abc b1 c2 d0 d1 d
b1 b2 c2 d2 d b1 b2 c2 cd b1 b2 bcd
a0 b0 c0 c2 c
a0 c2 d0 d1 d
b0 b2 c2 d2 d b0 b2 c2 cd b0 b2 bcd
b0 c0 c2 d2 d b0 c0 c2 cd
b0 c0 c1 d2 d b0 c0 c1 cd
a0 c0 c2 d2 d a0 c0 c2 cd
a0 c0 c1 d2 d a0 c0 c1 cd
a0 a1 c1 d2 d a0 a1 c1 cd a0 a1 acd
a1 a2 c1 d2 d a1 a2 c1 cd a1 a2 acd
a2 b0 d1 d2 d
b0 c1 d1 d2 d
a2 c1 d1 d2 d
S42
We see that L = L1 ∪L2 , where L1 is the subcomplex generated by the facets on the first to third
columns from the left in the above list, and L2 is the subcomplex generated by the other facets.
Then, L = starL (c)∪starL (d). If σ∩{c, d} = ∅, then linkL (σ) = linkL1 (σ+c)∗c∪linkL2 (σ+d)∗d,
which implies that |linkL (σ)| is homotopy equivalent to a suspension space. Therefore, if
linkL (σ) ≤ 2, i.e., dim σ ≥ 1, then it follows that linkL (σ) is torsion free, by dimensional
A GOLOD COMPLEX WITH NON-SUSPENSION MOMENT-ANGLE COMPLEX
21
3
reasoning. We consider a face with dim σ ≤ 0. For a vertex or the empty face in S12
, we have
already proved that linkL (σ) is torsion free. If v = a or v = b, then we see that |linkL (v)| is
homotopy equivalent to S 3 . Thus, we have completed the proof in this case.
If c ∈ σ or d ∈ σ, then we see by direct computation that |linkL (c)| and |linkL (d)| are
homotopy equivalent to S 3 . Moreover, linkL (ac), linkL (bc), linkL (cd), linkL (xc), linkL (ad),
3
linkL (bd), and linkL (xd) are triangulations of S 2 , where x is a vertex of S12
. If dim linkL (σ) ≤ 1,
then linkL (σ) is torsion free by dimensional reasoning, and the proof is complete for all cases.
References
[1] L.L. Avramov, Golod homomorphisms, Algebra, algebraic topology and their interactions (Stockholm,
1983), 59–78, Lecture Notes in Math. 1183, Springer, Berlin, 1986.
[2] A. Bahri, M. Bendersky, F.R. Cohen, and S. Gitler, The polyhedral product functor: a method of
decomposition for moment-angle complexes, arrangements and related spaces, Advances in Math. 225
(2010), 1634–1668.
[3] I.V. Baskakov, V.M. Buchstaber, and T.E. Panov, Algebras of cellular cochains, and torus actions,
(Russian) Uspekhi Mat. Nauk 59 (2004), no. 3(357), 159–160; translation in Russian Math. Surveys 59
(2004), no. 3, 562–563.
[4] A. Björner and M. Tancer, Note: Combinatorial Alexander duality: a short and elementary proof,
Discrete Comput. Geom. 42 (2009), no.4, 586–593.
[5] V.M. Buchstaber and T.E. Panov, Toric topology, Mathematical Surveys and Monographs 204, American Mathematical Society, Providence, Rhode Island, 2015.
[6] A. De Stefani, Products of ideals may not be Golod, J. Pure Appl. Algebra 220 (2016), no.6, 2289–2306.
[7] M. Franz, The integral cohomology of toric manifolds, Tr. Mat. Inst. Steklova 252 (2006), Geom. Topol.,
Diskret. Geom. i Teor. Mnozh., 61–70; translation in Proc. Steklov Inst. Math. 252 (2006), no.1, 53–62.
[8] E.S. Golod, Homologies of some local rings (Russian) Dokl. Akad. Nauk SSSR 144(1962), 479–482.
translation in Soviet Math. Dokl. 3 (1962), no.3, 745–748.
[9] J. Grbić, T. Panov, S. Theriault, and J. Wu, Homotopy types of moment-angle complexes for flag
complexes, Trans. Amer. Math. Soc. 368(2016), no.9, 6663-6682.
[10] J. Grbić and S. Theriault, The homotopy type of the complement of a coordinate subspace arrangement,
Topology 46 (2007), no.4, 357–396.
[11] T.H. Gulliksen and G. Levin, Homology of local rings, Queen’s Paper in Pure and Applied Mathematics,
No. 20, Queen’s University, Kingston, Ontario, 1969.
[12] J. Herzog, V. Reiner, and V. Welker, Componentwise linear ideals and Golod rings, Michigan Math. J.
46 (1999), no.2, 211–223.
[13] J. Herzog and M. Steurich, Golodideale der Gestalt a ∩ b, J. of Algebra 58 (1979), no.1, 31–36.
[14] K. Iriye and D. Kishimoto, Decompositions of polyhedral products for shifted complexes, Advances in
Math. 245 (2013), 716–736.
[15] K. Iriye and D. Kishimoto, Fat wedge filtrations and decomposition of polyhedral products, to appear in
Kyoto Journal of Mathematics; preprint available in arXiv:1412.4866.
[16] K. Iriye and D. Kishimoto, Decompositions of suspensions of spaces involving polyhedral products, Algebraic & Geometric Topology 16 (2016), no.2, 825–841.
[17] L. Katthän, A non-Golod ring with a trivial product on its Koszul homology, J. of Algebra 479 (2017),
244–262.
[18] D. Kraines, Massey higher products, Trans. Amer. Math. Soc. 124 (1966), 431–449.
[19] K.V. Madahar and K.S. Sarkaria, A minimal triangulation of the Hopf map and its application, Geometriae Dedicata, 105 (2000), 105-114.
[20] C.R.F. Maunder, Algebraic topology, Cambridge University Press, Cambridge-New York, 1980.
[21] J.P. May, Matric Massey products, J. of Algebra 12 (1969), 533–568.
22
KOUYEMON IRIYE AND TATSUYA YANO
[22] G.J. Porter, Higher order Whitehead products, Topology 3 (1965), 123–135.
[23] G.J. Porter, Higher products, Trans. Amer. Math. Soc. 148 (1970), 315–345.
[24] E.H. Spanier, Algebraic Topology, Corrected reprint. Springer-Verlag, New York-Berlin, 1981.
Department of Mathematics and Information Sciences, Osaka Prefecture University, Sakai,
599-8531, Japan
E-mail address: [email protected]
Department of Mathematics and Information Sciences, Osaka Prefecture University, Sakai,
599-8531, Japan
E-mail address: [email protected]
| 0math.AC
|
WILLIS THEORY VIA GRAPHS
arXiv:1711.09610v1 [math.GR] 27 Nov 2017
TIMOTHY P. BYWATERS AND STEPHAN TORNIER
Abstract. We study Willis’ theory of totally disconnected locally compact
groups and their endomorphisms in a geometric framework using graphs. This
leads to new interpretations of tidy subgroups and the scale function. Foremost,
we obtain a geometric tidying procedure which applies to endomorphisms as
well as a geometric proof of the fact that tidiness is equivalent to being minimizing for a given endomorphism. Our framework also yields an endomorphism
version of the Baumgartner-Willis tree representation theorem. We conclude
with a construction of new endomorphisms of totally disconnected locally compact groups from old via HNN-extensions.
1. Introduction
The study of totally disconnected locally compact (t.d.l.c.) groups has its origin
in the desire to understand general locally compact groups, which are ubiquitous
in mathematics. Any locally compact group G may be viewed as an extension of
its connected component G0 by the totally disconnected quotient G/G0 .
Every connected locally compact group is an inverse limit of Lie groups by
the solution of Hilbert’s fifth problem due to Gleason [Gle52], Yamabe [Yam53],
Montgomery–Zippin [MZ52] and others. Therefore, totally disconnected locally
compact groups have become the new focus.
Little was known beyond van Dantzig’s existence theorem of an identity neighbourhood basis consisting of compact open subgroups [vD31] until Willis raised
the hope for a general structure theory of t.d.l.c. through the introduction of new
concepts around their automorphisms [Wil94], [Wil01]. Aside from developing the
structure theory of totally disconnected locally compact groups [Wil04], [BW06],
[Wil07], [BMW12], this theory has had many unexpected applications, including to
the fields of random walks and ergodic theory [DSW06], [JRW96], [PW03], arithmetic groups [SW13] and Galois theory [CH09].
It is hence natural to search for the largest natural framework of Willis theory.
In [Wil15], the foundations of Willis theory have been extended from automorphisms to endomorphisms. Some work has been done to generalize existing results
accordingly, see [BGT16]. The present article contributes to this scheme: We characterize tidy subgroups and the scale function for endomorphisms of t.d.l.c. groups
via graphs and permutations and thereby generalize work of Möller [Möl02] for the
case of automorphisms. We use these characterizations to give alternate proofs of
fundamental results concerning the scale and tidy subgroups. The reader is referred
to Section 2 for an introduction to Willis theory, including the concepts of tidy
subgroups and the scale function, as well as fundamentals about directed graphs.
Let G be a t.d.l.c. group. Further, let α be a continuous endomorphism of G
and U a compact open subgroup of G. Using a certain graph associated to the
data (G, α, U ) we give a geometric proof of existence of a subgroup of U which is
tidy above for α ([Wil15, Proposition 3]), as well as the tidiness below condition
([Wil15, Proposition 8]). Combining both yields the following characterization of
Date: November 28, 2017.
1
2
TIMOTHY P. BYWATERS AND STEPHAN TORNIER
the scale and tidiness, resembling [Möl02, Lemma 3.1] and [Möl02, Theorem 3.4],
see Lemma 3.1 and Theorem 3.11.
For i ∈ N0 , define v−i := α−i (U ) ∈ P(G) as well as a rooted, directed graph Γ+ by
V (Γ+ ) = {uv−i | u ∈ U++ , i ∈ N0 },
E(Γ+ ) = {(uv−i , uv−i−1 ) | u ∈ U, i ∈ N0 }.
Theorem. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G compact open.
(i) If {v−i | i ∈ N0 } is finite then there is a compact open subgroup U of G with
α(U ) ≤ U and which is tidy for α and s(α) = 1.
(ii) If {v−i | i ∈ N0 } is infinite then U is tidy for α if and only if the graph Γ+ is a
directed tree, rooted at v0 with contant in-valency (excluding the root) equal
to 1 and constant out-valency. In this case, s(α) equals said out-valency.
We use this theorem to establish a new, geometric tidying procedure for the case
of endomorphisms, see Theorem 4.14. It features yet another graph defined in terms
of the data (G, α, U ) which admits an action of U++ , a fundamental subgroup of
G associated to α and U , see Section 2.1. Most of the work goes into showing that
this graph admits a quotient with a connected component isomorphic to a regular
rooted tree. The stabilizer of its root turns out to be tidy for α.
Theorem 4.14 and associated constructions result in a geometric proof of the
fact [Wil15, Theorem 2] that tidiness is equivalent to being minimizing, see Theorem 5.6. Using the aforementioned ideas, we obtain a tree representation theorem for a certain natural subsemigroup of End(G) associated to α, analogous to
[BW04, Theorem 4.1] for the case of automorphisms.
Finally, we give a simple way to construct endomorphisms of non-compact t.d.l.c
groups from certain endomorphisms of compact groups.
1.1. Structure. In Section 2 we collect the necessary preliminaries about Willis
theory, graphs and the permutation topology which appears in this context. We
then move on to characterize the notions of being tidy above, tidy below and tidy
in terms of directed graphs in Section 3. These characterizations are used in Section
4 to develop a geometric tidying procedure. In return, this tidying procedure is
employed in Section 5 to recover fundamental results about the scale function.
Section 6 contains the tree representation theorem. Finally, Section 7 contains a
construction of new endomorphisms of t.d.l.c. groups from old.
1.2. Acknowledgments. The authors owe thanks to George Willis for suggesting a graph-theoretic study of the scale and tidiness for endomorphisms in the
spirit of Möller, and for later discussions surrounding this work. Both authors also
thank Jacqui Ramagge for helpful conversations and suggesting improvements on
an early draft. This research was conducted whilst the first author was supported
by the Australian Government Research Training Program Scholarship. The second author thanks The University of Newcastle and The University of Sydney for
their hospitality, and gratefully acknowledges financial support through the SNSF
Doc.Mobility fellowship 172120.
2. Preliminaries
In this section, we collect the necessary prerequisites for Willis theory, graphs
and the permutation topology. Given a t.d.l.c. group G, we denote by End(G) the
semigroup of continuous homomorphisms from G to itself.
2.1. Willis Theory. Throughout, G denotes a t.d.l.c. group, α ∈ End(G) an endomorphism of G and U ≤ G a compact open subgroup. We begin by defining
essential subgroups associated to this data.
WILLIS THEORY VIA GRAPHS
3
Set U0 := U and define inductively for n ∈ N0 :
Un+1 := U ∩ α(Un )
and U−n−1 := U−n ∩ α−n−1 (U ) =
n+1
\
α−k (U ).
k=0
Using these, we further make the following definitions:
\
[
\
U+ :=
Un , U++ :=
αn (U+ ), U− :=
U−n ,
n∈N
n∈N
U−− :=
n∈N
[
α−n (U− ).
n∈N
Both (Un )n∈N0 and (U−n )n∈N0 are descending sequences of compact subgroups of
U . Whereas U−n is open for all n ∈ N0 , the groups Un (n ∈ N0 ) need not be so.
There is an important alternative description of the above subgroups in terms
of trajectories and regressive sequences: Let x ∈ G. The α-trajectory of x is the
sequence (αn (x))n∈N0 in G. An α-regressive trajectory of x is a sequence (xn )n∈N0
in G such that x0 = x and α(xn ) = xn−1 for all n ∈ N. Consequently, we have the
following verbal descriptions of U± and U±± .
U+ =
U++
elements of U which admit an
α-regressive trajectory contained in U
,
elements of G which admit an α-regressive
=
trajectory that is eventually contained in U
elements of U whose
U− =
,
α-trajectory is contained in U
elements of G whose α-trajectory
U−− =
.
is eventually contained in U
,
The following properties of compact open subgroups in relation to a given endomorphism lie at the heart of Willis theory.
The subgroup U is tidy above for α if U = U+ U− , and tidy below for α if U−− is
closed. It is tidy for α if it is both tidy above and tidy below for α.
Remark 2.1. Our definition of being tidy below deviates from [Wil15, Definition 9].
In the approach given there, the natural definition for being tidy below is that
U++ be closed and the sequence ([αn+1 (U+ ) : αn (U+ )])n∈N be constant. Assuming
tidiness above, the two definitions are equivalent by [Wil15, Proposition 9]. However,
the definition given here is the one which appears naturally in our geometric context.
The scale of α is
s(α) = min{[α(U ) : U ∩ α(U )] | U ≤ G compact open}.
A compact open subgroup U ≤ G is minimizing for α if [α(U ) : α(U ) ∩ U ] = s(α).
Remark 2.2. The scale is well-defined as the index of an open subgroup inside a
compact subgroup is finite: the latter can be covered by cosets of the open subgroup.
One of the major achievements of [Wil94],[Wil01] and [Wil15] is a proof showing
that being minimizing and being tidy for a given endomorphism are equivalent conditions on compact open subgroups [Wil15, Theorem 2]. It relies on an algorithm,
known as a tidying procedure, which transforms, in a finite number of steps, an arbitrary compact open subgroup of G into one that is tidy for a given endomorphism.
This algorithm allows one to prove many desirable properties of the scale function.
In [Möl02], a geometric interpretation of the scale and tidy subgroups as well as
a new tidying procedure were given for the case of automorphisms. In the present
work we generalize this framework to the setting of endomorphisms.
4
TIMOTHY P. BYWATERS AND STEPHAN TORNIER
For the reader’s convenience, we include the statement of [Wil15, Lemma 2]
below. It constitutes an endomorphism version of the equality
!
n
n+k
\
\
k
i
α
α (U ) =
αi (U )
i=m
i=m+k
which holds for an automorphism α and k, m, n ∈ Z with m ≤ n.
Lemma 2.3 ([Wil15, Lemma 2]). Retain the above notation. For all n, m ∈ N:
(i) U−n−m = (U(−n )−m , and
Uk ∩ Uk−n 0 ≤ k ≤ n
, and
(ii) αk (U−n ) =
αk−n (Un ) k ≥ n
(iii) (U−n )k = Uk ∩ U−n for all k ≥ 0 and (U−n )+ = U+ ∩ U−n .
2.2. Directed and Undirected Graphs. We largely follows Möller’s notation for
directed graphs. A directed graph Γ is a tuple (V (Γ), E(Γ)) consisting of a vertex
set V (Γ) and an edge set E(Γ) ⊆ V (Γ) × V (Γ) \ {(u, u) | u ∈ V (Γ)}. We denote
by o, t : E(Γ) → V (Γ) the projections onto the first and second factor, the origin
and terminus of an edge. Let Γ be a directed graph. An arc of length k ∈ N from
v ∈ V (Γ) to v 0 ∈ V (Γ) is a tuple (v = v0 , . . . , vk = v 0 ) of distinct vertices of Γ such
that (vi , vi+1 ) in an edge in Γ for all i ∈ {0, . . . , k − 1}. Two vertices v, w ∈ Γ(V )
are adjacent if either (v, w) ∈ E(Γ) or (w, v) ∈ E(Γ). A path of length k ∈ N from
v ∈ V (Γ) to v 0 ∈ V (Γ) is a tuple (v = v0 , . . . , vk = v 0 ) of distinct vertices of Γ
such that either (vi , vi+1 ) or (vi+1 , vi ) is an edge in Γ for all i ∈ {0, . . . , k − 1}. The
directed graph Γ is connected if for all v, w ∈ V (Γ) there is a path from v to w.
It is a tree if it is connected and has no non-trivial cycles, i.e. tuples (v0 , . . . , vk )
with k ≥ 3 and such that (v0 , . . . , vk−1 ) and (vk−1 , vk ) are both paths and vk = v0 .
Two infinite paths in Γ are equivalent if they intersect in an infinite path. When
Γ is a tree, this is an equivalence relation on infinite paths and the boundary ∂Γ
of Γ is the set of these equivalence classes. For the following, let v ∈ V (Γ). Set
inΓ (v) := {w ∈ V (Γ) | (w, v) ∈ E(Γ)} and outΓ (v) := {w ∈ V (Γ) | (v, w) ∈ E(Γ)}.
The in-valency of v ∈ V (Γ) is the cardinality of inΓ (v) and the out-valency of
v ∈ V (Γ) is the cardinality of outΓ (v). The directed graph Γ is locally finite if all
its vertices have finite in- and out-valency and regular if |inΓ (v)| and | outΓ (v)| are
constant for v ∈ V (Γ).
A directed line in Γ is a sequence (vi )i∈Z of distinct vertices such that either
(vi , vi+1 ) is an edge for every i ∈ Z, or (vi , vi−1 ) is an edge for every i ∈ Z.
For a subset A ⊆ V (Γ), the subgraph of Γ spanned by A is the directed graph
with vertex set A and edge set {(v, w) ∈ E(Γ) | v, w ∈ A}.
The set of descendants of v ∈ V (Γ)
S is descΓ (v) := {w ∈ V (Γ) | ∃ arc from v to w}.
For A ⊆ V (Γ), set descΓ (A) := v∈A descΓ (v). A directed tree Γ is rooted at
v0 ∈ V (Γ) if Γ = desc(v0 ), in which case | inΓ (v)| = 1 for all vertices v 6= v0 and
| inΓ (v0 )| = 0. The definition of being regular is altered for rooted trees: A directed
tree rooted at v0 is regular if | out(v)| is constant for v ∈ V (Γ).
A morphism between directed graphs Γ1 = (V1 , E1 ) and Γ2 = (V2 , E2 ) is a pair
(αV , αE ) of maps αV : V1 → V2 and αE : E1 → E2 preserving the graph structure,
i.e. αV (o(e)) = o(αE (e)) and αV (t(e)) = t(αE (e)) for all e ∈ E1 . An automorphism
of a directed graph Γ = (V, E) is a morphism α = (αV , αE ) from Γ to itself such
that αV and αE are bijective and α admits an inverse morphism.
An undirected graph Γ, after Serre [Ser03], is a tuple (V, E) consisting of a vertex
set V and an edge set E, together with a fixed-point-free involution of E, denoted
by e 7→ e, and maps o, t : E → V , providing the origin and terminus of an edge,
such that o(e) = t(e) and t(e) = o(e) for all e ∈ E. Given e ∈ E, the pair {e, e} is a
WILLIS THEORY VIA GRAPHS
5
geometric edge. For x ∈ V , we let E(x) := o−1 (x) = {e ∈ E | o(e) = x} be the set
of edges issuing from x. The valency of x ∈ V is |E(x)|. In particular, a directed
graph (V, E) yields an undirected graph by passing to the symmetric closure of
E ⊆ V × V . The valency of a vertex in this undirected graph is the sum of the outand in-valency of the same vertex in the directed graph. Undirected trees and their
boundary are defined analogous to the case of directed trees. So are morphisms and
automorphisms.
2.3. Permutation Topology. The results of this article involve groups acting
on directed graphs. In this situation, the acting group may be equipped with the
permutation topology for its action on the vertex set of said graph. This topology
is naturally introduced in the following, more general setting. See e.g. [Möl10].
Let X be a set and consider G := Sym(X). The basic open sets for the permutation topology on G are
Ux,y := {g ∈ G | ∀i ∈ {1, . . . , n} : g(xi ) = yi }
with n ∈ N and x = (x1 , . . . , xn ), y = (y1 , . . . , yn ) ∈ X n . The permutation topology
turns G into a topological group which is Hausdorff and totally disconnected. It is
locally compact if and only if the orbits of Ux := Ux,x are finite for all x ∈ X. Also,
it makes the action map Sym(X) × X → X given by (g, x) 7→ g(x) continuous.
3. Characterization of Tidy Subgroups
Let G be a totally disconnected, locally compact group and let α ∈ End(G). In
this section, we characterize the compact open subgroups U of G which are tidy
for α in terms of certain directed graphs. In doing so we generalize several results
of [Möl02] from conjugation automorphisms to general endomorphisms.
Frequently, we restrict to the case where the set {α−i (U ) | i ∈ N0 } is infinite
and hence all α−i (U ) (i ∈ N0 ) are distinct. The finite case corresponds to Möller’s
periodicity case [Möl02, Lemma 3.1] and is covered by the following lemma.
Lemma 3.1. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G
TNcompact open. If
{α−i (U ) | i ∈ N0 } is finite then there is N ∈ N0 such that V := k=0 α−k (U ) = U−
satisfies α(V ) ≤ V and is tidy for α.
T
Proof. If {α−i (U ) | i ∈ N0 } is finite, then U−T = k∈N0 α−i (U ) is an intersection
N
of finitely many open subgroups. Say U− = k=0 α−k (U ) =: V . Then V ≤ G is
compact open and α(V ) ≤ V . We conclude V = V− . Hence V is tidy above for α.
Since V = V− ≤ V−− we also deduce that V−− is open and hence closed. Thus V
is also tidy below for α.
3.1. Tidiness Above. We recover the T
fact that for every compact open subgroup
n
U ≤ G there is n ∈ N0 such that U−n = k=0 α−n (U ) is tidy above for α.
Consider the graph Γ defined as follows: Set v−i := α−i (U ) ∈ P(G) for i ∈ N0 ,
where P(G) denotes the power set of G. Now set
V (Γ) := {gv−i | g ∈ G, i ∈ N0 }
and E(Γ) := {(gv−i , gv−i−1 ) | g ∈ G, i ∈ N0 }.
Note that G acts on Γ by automorphisms via left multiplication. For this action,
we compute the stabilizer Gv−i = α−i (U ) (i ≥ 0), as well as
\
G{v−m |m≥0} =
α−m (U ) = U− .
m≥0
We now reprove [Wil15, Lemma 4] in terms of the graph Γ.
Lemma 3.2. Retain the above notation. Suppose that UN v−1 = U+ v−1 for some
N ∈ N. Then U−n v−n−1 = (U−n )+ v−n−1 for all n ≥ N .
6
TIMOTHY P. BYWATERS AND STEPHAN TORNIER
Proof. By definition, (U−n )+ v−n−1 ⊆ U−n v−n−1 . Now, let w ∈ U−n v−n−1 . Then
there is u ∈ U−n such that w = uv−n−1 . We obtain αn (u) ∈ αn (U−n ) which equals
Un by Lemma 2.3 and is contained in UN since n ≥ N . Hence, by assumption,
there is u+ ∈ U+ such that αn (u)v−1 = u+ v−1 . By definition of U+ , we may pick
u0+ ∈ U+ ∩ U−n such that u+ v−1 = αn (u0+ )v−1 . Then u0+ ∈ (U−n )+ as by Lemma
2.3 we have U+ ∩ U−n = (U−n )+ . We conclude that u0+ v−n−1 = uv−n−1 since
u0+ u−1 ∈ U−n−1 ≤ Gv−n−1 by the following argument: We have u0+ u−1 ∈ U−n by
definition and u0+ u−1 ∈ α−n−1 (U ) by the subsequent computation:
αn+1 (u0+ u−1 ) = αn+1 (u0+ )αn+1 (u−1 ) = α(u+ αn (u−1 )) ∈ U
since, by construction, u+ αn (u)−1 ∈ Gv−1 = α−1 (U ).
The following Lemma will be used to prove analogues of Theorems 2.1 and 2.3
from [Möl02].
Lemma 3.3. Retain the above notation. Fix N ∈ N and consider the following:
(i) UN v−1 = U+ v−1 .
(ii) For every u ∈ U−N there is u+ ∈ U+ ∩ U−N with u+ vi = uvi for all i ≤ 0.
(iii) The subgroup U−N is tidy above for α.
Then (i) implies (ii), and (ii) implies (iii).
Proof. To see (i) implies (ii) let u ∈ U−N . By induction, we construct a sequence
(un )n∈N contained in U+ ∩ U−N such that un vi = uvi for all i ∈ {−N − n, . . . , 0}.
Then, as U+ ∩ U−N is compact, (un )n∈N has an accumulation point u+ ∈ U+∩U−N .
We conclude that for any given n ∈ N, we have
−n
u−1
(U )
k u+ ∈ Gv−n = α
for large enough k ∈ N because α−n (U ) is open. That is, given n ∈ N we have
u+ (v−n ) = uk (v−n ) = u(v−n ).
for sufficiently large k ∈ N.
Now, by (i), Lemma 3.2 and Lemma 2.3, we may pick u1 ∈ U+ ∩ U−N such that
u1 v−N −1 = uv−N −1 . Next, assume that un has been constructed for some n ∈ N.
Then uu−1
n (vi ) = vi for all i ∈ {−N − n, . . . , 0}. That is,
u−1
n u∈
n+N
\
α−i (U ) = U−N −n .
i=0
By Lemma 3.2, there exists x ∈ (U−N −n )+ such that u−1
n uv−N −n−1 = xv−N −n−1 .
By assumption, un ∈ U+ ∩U−N and, by Lemma 3.2, x ∈ (U−N −n )+ = U+ ∩U−N −n .
Hence un x ∈ U+ ∩ U−N . Also, un x(vi ) = u(vi ) for all i ∈ {−N − n − 1, . . . , 0}. We
may therefore set un+1 := un x.
To see that (ii) implies (iii) we use that, by assumption, for every u ∈ U−N there
is u+ ∈ U+ ∩ U−N such that u and u+ agree on vi for all i ≤ 0. Set u− := u−1
+ u.
Then u− vi = vi for all i ≤ 0. Hence u− ∈ G{vm |m≤0} = U− and
U−N = (U+ ∩ U−N )U− = (U−N )+ (U−N )−
by Lemma 2.3 as required.
Theorem 3.4. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G compact open.
Then there is N ∈ N such that UN v−1 = U+ v−1 , and U−N is tidy above for α.
Proof. First note that U+ v−1 ⊆ Um v−1 ⊆ Un v−1 for all 0 ≤ n ≤ m since the sets
Un (n ∈ N0 ) are nested. Thus it suffices to show that UN v−1 ⊂ U+ v−1 for some
N ∈ N. Towards a contradiction, assume that U+ v−1 ( Un v−1 for all n ∈ N, i.e.
there is wn ∈ V (Γ) such that wn ∈ Un v−1 for all n ∈ N but wn 6∈ U+ v−1 . Then
WILLIS THEORY VIA GRAPHS
7
there is a sequence (un )n∈N contained in U such that un ∈ Un and un v−1 = wn .
Since U is compact, the sequence (un )n∈N has an accumulation point u+ in U . This
accumulation point has to be contained in U+ : Indeed, pick a subsequence (unk )k∈N
of (un )n∈N converging to u+ . Then for any given m ∈ N, we have unk ∈ Um for
almost all k. Since Um is closed we conclude that u+ ∈ Um for every m ∈ N. Hence
\
u+ ∈
Um = U+ .
m∈N
Furthermore, if u+ v−1 = w, then because u+ u−1
nk is contained in the open set
Gv−1 for large enough k ∈ N we must have w = wk for sufficiently large k ∈ N.
We conclude that wk ∈ U+ v−1 for sufficiently large k ∈ N and thus we have a
contradiction. Now, U−N is tidy above for α by Lemma 3.3.
Theorem 3.5. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G compact open.
Then the following statements are equivalent.
(i) U v−1 = U+ v−1 .
(ii) For every u ∈ U there is u+ ∈ U+ such that u+ vi = uvi for all i ≤ 0.
(iii) The subgroup U is tidy above for α.
Proof. Note that (i) implies (ii) and (ii) implies (iii) by Lemma 3.3 for N = 0. Now,
if (iii) holds, then U v−1 = U+ U− v−1 = U+ v−1 as U− ≤ Gv−1 .
Proposition 3.6. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G compact open
as well as tidy above for α. Then
[U−n : U−n−1 ] = [U : U−1 ] = [α−n (U ) : α−n−1 (U ) ∩ α−n (U )]
for all n ∈ N.
Proof. Let u ∈ U−n \U−n−1 . Then αn (u) ∈ U \U−1 . Hence [U−n : U−n−1 ] ≤ [U : U−1 ].
Conversely, if u ∈ U \U−1 then u admits a representative in U+ by Theorem 3.5.
Let (un )n be an α-regressive sequence of u contained in U . Then un ∈ U−n−1 \U−n .
Hence equality holds. The same argument applies to the right hand equality.
The following equality is used in Section 6.
Lemma 3.7. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G compact open as
well as tidy above for α. Then [α(U ) : U ∩ α(U )] = [α(U+ ) : U+ ].
Proof. Note that
α(U )(U ∩ α(U )) = α(U+ )α(U− )(U ∩ α(U )) = α(U+ )(U ∩ α(U ))
as α(U− ) ≤ U . Thus
[α(U ) : U ∩ α(U )] = [α(U+ ) : U ∩ α(U ) ∩ α(U+ )] = [α(U+ ) : U ∩ α(U+ )].
Since U ∩ α(U+ ) = U+ , the desired equality follows.
3.2. Tidiness Below. In this section we present a geometric proof for the commonly used criterion that identifies a compact open and tidy above subgroup U ≤ G
as tidy below if U−− ∩ U =SU− , cf. [Wil15, Proposition
S 8].
First, recall that U++ = i∈N0 αi (U+ ) and U−− = i∈N0 α−i (U− ). In terms of the
graph Γ introduced in Section 3.1, we have
[
U−− =
G{v−m |m≤−n} .
n∈N
Lemma 3.8. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G be compact open
as well as tidy above for α. Then
(i) the group U−− ≤ G is closed if and only if U−− ∩ U = U− , and
(ii) if U−− is closed then U++ ∩ U = U+ .
8
TIMOTHY P. BYWATERS AND STEPHAN TORNIER
Proof. For (i), first assume that U−− ∩ U = U− . Then U−− ∩ U is closed. Since U
is closed, this implies that U−− is closed, see [HR12, 5.37].
Now suppose that U−− ∩ U 6= U− . By definition, U− ⊆ U−− ∩ U . Hence there
exists u ∈ U = Gv0 with u ∈ G{vm |m≤−n} for some n ∈ N but u 6∈ U− = G{vm |m≤0} .
Then there is l ∈ N with 0 < l < n and such that uv−l 6= v−l . Since U is tidy above,
we may decompose u = u+ u− for some u+ ∈ U+ and u− ∈ U− . Hence, replacing u
with uu−1
− , we may assume that u ∈ U+ .
Choose an α-regressive trajectory (uj )j∈N of u contained in U+ . Define a sequence
(xi )i∈N contained in U−− ∩ U+ ≤ U as follows: Set x1 := u and xi+1 := xi uin . We
collect the relevant properties of the sequences (uj )j∈N and (xi )i∈N in the following
lemma, see below for an illustration of the second sequence.
Lemma 3.9. The sequences (uj )j∈N and (xi )i∈N have the following properties.
(a) For all j ∈ N: uj ∈ G{vm |m≤−n−j} ∩ G{vm |−j≤m≤0} ∩ U+ ≤ U−− ∩ U+ .
(b) For all i ∈ N: xi ∈ G{vm |m≤−in} ∩ U+ ≤ U−− ∩ U+ .
(c) For all j ∈ N: uj 6∈ Gv−l−j .
(d) For all i ∈ N and 0 ≤ j ≤ i − 1: xi 6∈ Gv−l−jn and xi+1 v−l−jn = xi v−l−jn .
T
Proof. For (a), note that αj (uT
= k≥n α−k (U ) by assumpj ) = u ∈ G{v
m |m≤−n}
T
−k
tion and therefore uj ∈ α−j
(U ) = k≥n+j α−k (U ) = G{vm |m≤−n−j} .
k≥n α
For the second part, simply recall that (uj )j is an α-regressive trajectory of u
contained in U+ ; in particular, uj ∈ U+ and αm (uj ) ∈ U+ ≤ U for all 0 ≤ m ≤ j.
Therefore, uj ∈ α−m (U ) = Gv−m for all 0 ≤ m ≤ j.
Part (b) follows from (a) given that xi+1 = xi uin = uun · · · u(i−1)n uin .
For part (c), recall that we have u 6∈ α−l (U ) = Gv−l by assumption and therefore
uj 6∈ α−l−j (U ) = Gv−l−j .
In order to prove part (d), we argue by induction: The element x1 = u satisfies
−1
x1 6∈ Gv−l by part (c). Also x2 v−l = x1 v−l because x−1
uun = un and
1 x2 = u
un ∈ G{vm |−n≤m} by part (a). Now assume the statement holds true for i ∈ N and
consider xi+1 = xi uin . Then xi+1 6∈ Gv−l−in because uin 6∈ Gv−l−in by part (a)
whereas xi ∈ Gv−l−in by part (b). Also, xi+1 v−l−jn = xi v−l−jn for all 0 ≤ j ≤ i − 1
since xi+1 = xi uin and uin ∈ G{vm |−in≤m} by part (a).
By Lemma 3.9, the sequence (xi )i∈N ⊆ U−− ∩ U+ ⊆ U has the following shape,
analogous to [Möl02, Figure 1].
x1
b
b
b
n
l
0
l
0
l
0
l
0
x2
b
b
b
b
2n
b
n
xk
b
b
b
b
b
b
b
b
n
2n
kn
k(n + 1)
x
b
k(n + 1)
b
b
kn
b
2n
b
b
b
n
b
Now, since U is compact, the sequence (xi )i∈N ⊆ U−− ∩U+ ⊆ U has an accumulation
point x ∈ U . However, x 6∈ U−− and hence U−− is not closed.
WILLIS THEORY VIA GRAPHS
9
For part (ii), note that U+ ⊆ U++ ∩ U by definition. Hence, towards a contradiction, we assume that there is u ∈ (U++ ∩ U )\U+ . Since U is tidy above
we may decompose u = u+ u− with u+ ∈ U+ and u− ∈ U− . Replacing u with
u−1
+ u ∈ (U++ ∩ U )\U+ we may hence assume u ∈ U− .
Now, since u ∈ U++ , there is an α-regressive trajectory (un )n∈N of u in G
such that for some N ∈ N we have un ∈ U+ for all n ≥ N and uN −1 6∈ U .
Consider the element uN ∈ U . For n ≥ N we have αn (uN ) = αn−N (u) ∈ U− .
Hence uN ∈ U−− ∩ U . However, uN 6∈ U− : Indeed, uN 6∈ Gv−1 = α−1 (U ) because
uN −1 6∈ U . Therefore, by part (i), U−− is not closed.
3.3. Tidiness. Finally, we combine the previous sections in order to characterize
tidiness in terms of a subgraph of the graph Γ introduced above. As before, let G
be a t.d.l.c. group, α ∈ End(G) and U ≤ G compact open. Recall the definition
v−i := α−i (U ) ∈ P(G) for i ∈ N0 . We consider the subgraph Γ+ of Γ defined by
V (Γ+ ) := {uv−i | u ∈ U, i ∈ N0 },
E(Γ+ ) := {(uv−i , uv−i−1 ) | u ∈ U, i ∈ N0 }.
Note that the action of U ≤ G on Γ preserves Γ+ ⊆ Γ and that Γ+ = desc(v0 ).
Lemma 3.10. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G. If U is tidy above
for α then U acts transitively on arcs of a given length issuing from v0 ∈ V (Γ+ ).
Proof. Given that outΓ+ (v−n+1 ) = [α−n+1 (U ) : α−n+1 (U ) ∩ α−n (U )] as well as
U{v−k |k≤n−1} = U−n+1 , the assertion follows by induction from Proposition 3.6.
We are now ready to characterize tidiness of U in terms of Γ+ when the set
{v−i | i ∈ N0 } is infinite. Concerning the case where {v−i | i ∈ N0 } is finite,
Theorem 3.11 is complemented by Lemma 3.1.
Theorem 3.11. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G compact open.
Assume {v−i | i ∈ N0 } is infinite. Then U is tidy for α if and only if Γ+ is a directed
tree with constant in-valency 1, excluding v0 , as well as constant out-valency.
Proof. First, assume that U is tidy for α. Notice that for a given i ≤ 0, the in- and
out-valency is constant among the collection of vertices {uvi | u ∈ U } given that U
acts on Γ+ by automorphisms.
Concerning in-valencies it therefore suffices to show that each v−i for i ≥ 1 has
in-valency equal to one. Suppose otherwise, that is in(v−i ) ≥ 2 for some i ≥ 1.
Then there is u ∈ Uv−i ⊆ α−i (U ) such that uv−i+1 6= v−i+1 . By Theorem 3.5 we
may assume that u ∈ U+ . Now consider u0 := αi (u) ∈ U++ ∩ U . Since U is tidy
below, Lemma 3.8 shows that u0 ∈ U+ = U++ ∩ U . But u 6∈ α−i+1 (U ) and hence
u0 = αi (u) 6∈ α(U ) ⊇ U+ , a contradiction. Thus Γ+ is a directed tree.
Concerning out-valencies, we may also restrict our attention to {v−i | i ∈ N0 }.
Note that out(v0 ) = |U+ v−1 | by Theorem 3.5 as U is tidy above. Furthermore,
out(v−i ) = |(U ∩ α−i (U ))v−i−1 | = |(U+ ∩ α−i (U ))v−i−1 | by the same theorem.
Now, since Γ+ is a tree and U+ fixes v0 , we obtain
out(v−i ) = |(U+ ∩ U−i )v−i−1 | = |(U−i )+ v−i−1 | = |U−i v−i−1 |
by Lemma 3.2. We conclude the argument by showing that
|U−i v−i−1 | = |U v−1 | = |U+ v−1 |.
On the one hand, we have |U−i v−i−1 | ≤ |U v−1 |: Indeed, suppose u ∈ U−i does not
fix v−i−1 . Then αi (u) does not fix v−1 . If it did, we would have αi (u) ∈ α−1 (U )
and hence u ∈ α−i−1 (U ). On the other hand, |U−i v−i−1 | ≥ |U v−1 |: Indeed, assume
u ∈ U does not fix v−1 , i.e. u 6∈ α−1 (U ). By Theorem 3.5, we may assume u ∈ U+ .
Pick an α-regressive trajectory (uj )j∈N0 of u in U . Then αi+1 (ui ) = α(u) 6∈ U and
hence ui 6∈ α−i−1 (U ), i.e. ui does not fix v−i−1 .
10
TIMOTHY P. BYWATERS AND STEPHAN TORNIER
Now assume that Γ+ has all the stated properties. Since Γ+ is a tree, we have
U−− ∩ U ⊆ U− while the reverse inclusion holds by definition. Hence U−− is closed
by Lemma 3.8 and U is tidy below. Combining the constant out-valency assumption
with the fact that Γ+ is a tree we obtain the equality |U v−1 | = |U−i v−i−1 |. Next,
|U−i v−i−1 | = |Ui v−1 | since |Ui v−1 | ≤ |U v−1 | and due to the following observation:
If u ∈ U−i is such that uv−i−1 6= v−i−1 then αi (u) ∈ αi (U−i ) = Ui by Lemma 2.3
and αn (u)v−1 6= v−1 . Thus |Ui v−1 | ≥ |U−i v−i−1 |. Overall, |U v−1 | = |Ui v−1 |.
Finally, to see that the above implies |U v−1 | = |U+ v−1 |, let u ∈ U . Then for
every i ∈ N there is ui ∈ Ui with uv−1 = ui v−1 . The sequence (ui )i∈N is contained
in U and hence admits
T a convergent subsequence. Any such subsequence converges
to an element u+ ∈ i≥0 Ui = U+ which coincides with u on v−1 . Theorem 3.5 now
implies that U is tidy above.
The following Lemma is a useful test of tidiness as it relies only on calculating
inverse images and indices. It is, in a sense, an algebraic way to see if Γ+ satisfies
the requirements of Theorem 3.11. We apply it multiple times in upcoming sections.
Lemma 3.12. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G compact open.
Then U is tidy for α if and only if [U : U∩α−n (U )] = [U : U∩α−1 (U )]n for all n ∈ N.
Proof. First, assume that U is tidy for α. If {v−i | i ∈ N0 } is finite, then for some
N ∈ N0 we have [U−N : U−N −1 ] = 1 by Lemma 3.1 and Proposition 3.6 shows that
1 = [U : U ∩ α−1 (U )] which implies α−1 (U ) ⊇ U . Therefore α−n (U ) ⊇ U for all
n ∈ N and the assertion follows. Now assume that {v−i | i ∈ N0 } is infinite. Then
Γ+ is a rooted directed tree with constant out-valency d and we obtain
[U : U ∩ α−n (U )] = [Uv0 : Uv0 ∩ Uv−n ] = |U v−n | = dn = [U : U ∩ α−1 (U )]n
by the orbit-stabilizer theorem as desired.
Conversely, assume that [U : U ∩ α−n (U )] = [U : U ∩ α−1 (U )]n for all n ∈ N and
consider the graph Γ+ . We have d := out(v0 ) = [Uv0 : Uv0 ∩Uv−1 ] = [U : U ∩α−1 (U )]
as before. By definition of Γ+ , the out-valency of any other vertex is at most d. But
|U v−n | = [Uv0 : Uv0 ∩ Uv−n ] = [U : U ∩ α−n (U )] = [U : U ∩ α−1 (U )]n = dn
by assumption. Therefore, every vertex has out-valency equal to d. Hence Γ+ is a
tree of constant in-valency 1, excluding v0 , and U is tidy for α by Theorem 3.11.
4. A Graph-Theoretic Tidying Procedure
Let G be a totally disconnected, locally compact group and let α ∈ End(G). We
show that there is a compact open subgroup of G which is tidy for α.
The proof is algorithmic: Starting from an arbitrary compact open subgroup we
construct a locally finite graph Γ++ . A certain quotient, inspired by [Möl00], of this
graph has a connected component isomorphic to a regular rooted tree which admits
an action of a subgroup of G. The stabilizer of the root in this tree is the desired
tidy subgroup.
For the remainder of the section, fix U ≤ G compact open. Refering to Lemma
3.1, we shall assume throughout that {α−i (U ) | i ∈ N0 } is infinite. By Theorem 3.4
we may also assume that U is tidy above for α.
4.1. The Graph Γ++ . Consider the graph Γ++ defined by
V (Γ++ ) = {uv−i | u ∈ U++ , i ∈ N0 }, and
E(Γ++ ) = {(uv−i , uv−i−1 ) | u ∈ U++ , i ∈ N0 }.
The following remark will be used in the proof of Theorem 4.14.
WILLIS THEORY VIA GRAPHS
11
Remark 4.1. Note that Γ++ is a subgraph of Γ. Also, if U is tidy above for α,
the graphs Γ+ and Γ have the same out-valency by Theorem 3.5. Consequently,
descΓ (v0 ) = Γ+ ⊆ descΓ++ (v0 ) ⊆ descΓ (v0 ) = Γ+ . Hence descΓ++ (v0 ) = Γ+ .
The following Lemma will help to identify vertices in Γ++ as (un)equal. It is
immediate from the assumption that {α−i (U ) | i ∈ N0 } is infinite and the fact that
left cosets of distinct subgroups are distinct.
Lemma 4.2. Retain the above notation and let u0 v−i , u1 v−j ∈ V (Γ++ ) ⊆ P(G). If
u0 v−i = u1 v−j then i = j.
Note that U++ acts on Γ++ by automorphisms. We now define an injective
graph endomorphism of Γ++ that appears frequently. Let uvi ∈ V (Γ++ ) where
u ∈ U++ . Since α(U++ ) = U++ , there exists u0 ∈ U++ such that α(u0 ) = u. Define
ρ(uvi ) = u0 vi−1 . The following proposition summarizes the properties of ρ and
includes justification that ρ is a well-defined.
Proposition 4.3. Retain the above notation. The map ρ is a graph isomorphism
from Γ++ to ρ(Γ++ ) where
V (ρ(Γ++ )) = {uv−i | u ∈ U++ , i ∈ N}, and
E(ρ(Γ++ )) = {(uv−i , uv−i−1 ) | u ∈ U++ , i ∈ N}.
Proof. We first show ρ is well-defined. Suppose u0 v−i , u1 v−i ∈ V (Γ++ ) represent
−i
the same vertex. Then u−1
0 u1 ∈ α (U ). Choose w0 , w1 ∈ U++ with α(wi ) = ui for
−1
−1
i ∈ {0, 1}. Then α(w0 w1 ) = u0 u1 ∈ α−i (U ) and so w0−1 w1 ∈ α−i−1 (U ). This
implies w0 v−i−1 = w1 v−i−1 . By Lemma 4.2, this is enough to show that setting
ρ(u0 v−i ) = w0 v−i−1 is well-defined.
To see that ρ is injective suppose that ρ(u0 v−i ) = ρ(u1 v−i ). Then there are w0
and w1 such that w0 v−i−1 = w1 v−i−1 and α(wi ) = ui (i ∈ {0, 1}). In particular,
−i
w0−1 w1 ∈ α−i−1 (U ) and so α(w0−1 w1 ) = u−1
0 u1 ∈ α (U ). Thus u0 v−1 = u1 v−i .
As to V (ρ(Γ++ )) we have, V (ρ(Γ++ )) ⊇ {uv−i | u ∈ U++ , i ∈ N} by definition
as α(U++ ) = U++ . Equality follows from Lemma 4.2.
To see that ρ preserves the edge relation, let (uv−i , uv−i−1 ) ∈ E(Γ++ ). Choose
u0 ∈ U++ with α(u0 ) = u. Then (ρ(uv−i ), ρ(uv−i−1 )) = (u0 v−i−1 , u0 v−i−2 ) ∈ E(Γ++ ).
Thus ρ is a graph morphism.
Again, we have E(ρ(Γ++ )) ⊇ {(uv−i , uv−i−1 ) | u ∈ U++ , i ∈ N} by definition as
α(U++ ) = U++ and equality by Lemma 4.2.
The following two results capture arc-transitivity of the action of U++ on Γ++ .
Lemma 4.4. Retain the above notation. Let γ0 and γ1 be arcs of equal length in
Γ++ and with origin uv0 (u ∈ U++ ). Then there is g ∈ U++ such that gγ0 = γ1 .
Proof. Note that u−1 γi (i ∈ {0, 1}) is an arc with origin v0 and thus is contained
in descΓ++ (v0 ). Remark 4.1 and Lemma 3.10 show that there exists u0 ∈ U+ such
that u0 u−1 γ0 = u−1 γ1 . Then uu0 u−1 ∈ U++ and g := uu0 u−1 serves.
In the following, we write [v0 , v−k ] for the arc (v0 , . . . , v−k ).
Proposition 4.5. Retain the above notation. Let γ0 and γ1 be arcs in Γ++ of equal
length. Then there are u ∈ U++ and n ∈ N0 with either uρn γ0 = γ1 or uρn γ1 = γ0 .
If γ0 and γ1 both terminate at v−i (i ∈ N), we may choose n = 0 and u ∈ U++ ∩U−− .
Proof. Suppose γ0 originates at ui v−i0 and γ1 originates at u1 v−i1 . Without loss
of generality assume i0 ≥ i1 . Then ρi0 −i1 (γ1 ) originates at u01 v−i0 = ρi0 −i1 (u1 v−i1 )
for some u01 ∈ U++ . For the first assertion it therefore suffices to show that for any
two arcs γ0 and γ1 originating at vertices u0 v−i and u1 v−i (u0 , u1 ∈ U++ ), there
12
TIMOTHY P. BYWATERS AND STEPHAN TORNIER
exists u ∈ U++ with uγ0 = γ1 . Further still, by considering the image of γ1 under
0
multiplication by u0 u−1
1 , we can assume the u0 = u1 . Now we can extend γj to γj
by concatenating on the left with the path (u0 v0 , . . . , u0 v−i ). By Lemma 4.4, there
exists u ∈ U++ such that uγ00 = γ10 . We must necessarily have uγ0 = γ1 .
For the second assertion, let γ be an arc terminating in v−k . It suffices to show
that there is g ∈ U++ ∩ U−− such that gγ ⊆ [v0 , v−k ]. Extending γ if necessary, we
can assume without loss of generality that γ originates at some uv0 where u ∈ U++ .
We now construct g ∈ U++ ∩ U−− such that gγ = [v0 , v−k ]. By Lemma 4.4, there
exists u0 ∈ U++ such that u0 γ = [v0 , v−k ]. Applying Lemma 4.4, for each n ∈ N0
there exist wn ∈ U++ such that
wn (v0 , . . . , v−k , u0 v−k−1 , . . . , u0 v−k−n ) = [v0 , v−k−n ].
The sequence (wn )n∈N is contained in U as each element fixes v0 . It hence admits a
subsequence converging to some w0 ∈ U . Put g := w0 u0 ∈ U++ . Since the permutation topology is coarser than the topology on G, we get g(v−l ) = v−l for all l ≥ k.
That is, g ∈ U−− and gγ = [v0 , v−k ].
Remark 4.6. Restricting Proposition 4.5 to the case where γ0 and γ1 are single
vertices we conclude that for any two vertices u0 , u1 ∈ V (Γ++ ), there are n ∈ N0
and u ∈ U++ such that either uρn (u0 ) = u1 or uρn (u1 ) = u0 .
We now show that Γ++ is locally finite. We will need the following Lemma which
is a consequence of [Wil15, Proposition 4] given that LU , see [Wil15, Definition 5],
is precisely U++ ∩ U−− .
Lemma 4.7. The closure of U++ ∩ U−− is compact.
The last assertion of the following proposition will be used to show that Γ++
admits a well-defined “depth” function.
Proposition 4.8. Retain the above notation. The graph Γ++
(i) has constant out-valency,
(ii) has constant in-valency among the vertices {uv−i | u ∈ U++ , i ∈ N},
(iii) satisfies that the in-valency of uv0 (u ∈ U++ ) is 0,
(iv) is locally finite, and
(v) satisfies that every arc from uv−i to u0 v−i−k (u, u0 ∈ U++ ; i, k ∈ N0 ) has
length k.
Proof. If u0 , u1 ∈ V (Γ++ ), then by Remark 4.6 and swapping u0 with u1 if necessary, there are g ∈ U++ and n ∈ N0 such that gρn (u0 ) = u1 . Proposition 4.3, shows
that | out(u1 )| = | out(ρn (u0 ))|, hence (i). Similarly, in(u) = in(gρn (u0 )) if neither
u0 and u1 are of the form uv0 for some u ∈ U++ and therefore (ii) holds.
The assertion that | in(uv0 )| = 0 follows since for every edge (u0 v−i , u0 v−i−1 ) we
have u0 v−i−1 6= uv0 by Lemma 4.2.
For local finiteness it now suffices to show that both out(v0 ) and in(v−1 ) are
finite. Note that by Remark 4.1 we have
| out(v0 )| = |U v−1 | = [U : U ∩ α−1 (U )]
which is finite by compactness of U and continuity of α. To see that in(v−1 ) is finite,
note that by Proposition 4.5 each vertex of in(v−1 ) can be written as uv0 where
u ∈ U++ ∩ U−− ∩ α−1 (U ). Conversely, any such u yields a vertex in in(v0 ). Thus
| in(v−1 )| = [U++ ∩ U−− ∩ α−1 (U ) : U++ ∩ U−− ∩ α−1 (U ) ∩ U ].
−1
(U )
If u0 , u1 ∈ U++ ∩U−− ∩α−1 (U ) with u0 u−1
1 6∈ U then u0 , u1 ∈ U++ ∩ U−− ∩α
−1
a fortiori and u0 u1 6∈ U . Thus
| in(v−1 )| ≤ [U++ ∩ U−− ∩ α−1 (U ) : U++ ∩ U−− ∩ α−1 (U ) ∩ U ].
WILLIS THEORY VIA GRAPHS
13
Applying Lemma 4.7 and noting that α−1 (U ) is closed, U++ ∩ U−− ∩ α−1 (U ) is
compact. Furthermore, since U is open, we derive that U++ ∩ U−− ∩ α−1 (U ) ∩ U
is open in U++ ∩ U−− ∩ α−1 (U ). Thus in(v−1 ) is finite.
For part (v), let γ be an arc from uv−i to uv−i−k . Note that by Proposition
4.5, there is g ∈ U++ with gγ ⊆ (v0 , v−1 , . . .). By Lemma 4.2, guv−i = v−i and
gu0 v−i−k = v−i−k . Thus gγ = (v−i , . . . , v−k ) has length k and so does γ because
U++ acts by automorphisms.
4.2. The quotient T . The tidying procedure relies on identifying a certain quotient T of Γ++ as a forest of regular rooted trees. To define this quotient, we first
introduce a “depth” function ψ : V (Γ++ ) → N on Γ++ as follows: For v ∈ V (Γ++ ),
choose an arc γ originating from some uv0 (u ∈ U++ ) and terminating at v. Set
ψ(v) to be the length of γ. The following is immediate from Proposition 4.8.
Lemma 4.9. Retain the above notation. The map ψ is well-defined and ψ(uv−i ) = i
for all u ∈ U++ and i ∈ N0 .
By virtue of Lemma 4.9 we may define the level sets Vk := ψ −1 (k) ⊆ V (Γ++ )
for k ≥ 0 and the edge sets Ek := {(w, w0 ) ∈ E(Γ++ ) | ψ(w0 ) = k} for k ≥ 1. It
is a consequence of Lemma 4.9 and Lemma 4.2 that (w, w0 ) ∈ Ek if and only if
there is u ∈ U++ such that (w, w0 ) = (uv−k+1 , uv−k ). On Vk (k ≥ 1) we introduce
an equivalence relation by w ∼ w0 if w and w0 belong to the same connected
component of Γ++ \Ek . Similarly, for w, w0 ∈ V0 we put w ∼ w0 if they belong to
the same connected component of Γ++ . Write [w] for the collection of vertices w0
with w ∼ w0 . Note that for every g ∈ U++ and k ∈ N0 we have gVk = Vk and
gEk = Ek . Since the action of U++ on Γ++ preserves connected components we see
that w ∼ w0 if and only if gw ∼ gw0 . The following Lemma extends this to ρ.
Lemma 4.10. Retain the above notation and let k ∈ N0 . Then ρ(Vk ) = Vk+1 and
ρ(Ek ) = Ek+1 . Hence, for w, w0 ∈ V (Γ++ ) we have w ∼ w0 if and only if ρ(w) ∼ ρ(w0 ).
Proof. The assertions ρ(Vk ) = Vk+1 and ρ(Ek ) = Ek+1 are immediate from the
definitions. Suppose now that w, w0 ∈ Vk are in the same connected component of
Γ++ \ Ek . By Proposition 4.3, this can occur if and only if ρ(w), ρ(w0 ) ∈ Vk+1 are
in the same connected component of ρ(Γ++ ) \ Ek+1 . By Proposition 4.8 and the
definition of Ek+1 , the embedding ρ(Γ++ ) → Γ++ maps connected components of
ρ(Γ++ )\Ek+1 to connected components of Γ++\Ek+1 and is surjective on Vk+1 .
Lemma 4.11. Retain the above notation. There is N ∈ N such that for every
v ∈ descΓ++ (v0 ) with ψ(v) ≥ N we have in(v) ⊆ descΓ++ (v0 ).
Proof. By Proposition 4.8, we can choose u0 , . . . , uk ∈ U++ ∩ α−1 (U ) such that
in(v−1 ) = {u0 v0 , . . . , uk v0 }. Since ui ∈ U++ for all i ∈ {0, . . . , k}, we may pick
α-regressive trajectories (wji )j∈N0 and Ni ∈ N such that w0i = ui and wni ∈ U for
all n ≥ Ni . Set N = max{Ni | i ∈ {0, . . . , k}} + 1.
Suppose n ≥ N . To see that in(v−n ) ⊆ descΓ++ (v0 ) note that by Proposition 4.8
i
we have in(v−n ) = ρn−1 (in(v−1 )) = {wn−1
v−N +1 | i ∈ {0, . . . , k}}. Since n−1 ≥ Ni
i
i
v0 , . . . , wn−1
v−n+1 ) is contained in descΓ++ (v0 ).
for all i ∈ {0, . . . , k}, the path (wn−1
This shows in(v−n ) ⊆ descΓ++ (v0 ).
In general, let v ∈ descΓ++ (v0 ) with ψ(v) = n ≥ N . Applying Proposition 4.5 to
the arc (v0 , . . . , v−n ) and any arc connecting v0 to v, there is u ∈ U ∩ U++ such that
uv−n = v. Furthermore, u descΓ++ (v0 ) = descΓ++ (v0 ) as uv0 = v0 and it follows
that in(v) = u in(v−n ) ⊆ descΓ++ (v0 ).
Lemma 4.12. Retain the above notation. Then the equivalence classes on Γ++
induced by ∼ have finite constant size.
14
TIMOTHY P. BYWATERS AND STEPHAN TORNIER
Proof. By Proposition 4.5 and Lemma 4.10, it suffices to show that a single equivalence is finite. Using Lemma 4.11, choose N ∈ N such that for every v ∈ descΓ++ (v0 )
with ψ(v) ≥ N we have in(v) ⊂ desc(v0 ). We show that [v−N ] ⊆ descΓ++ (v0 ). Since
desc(v0 ) ∩ Vk is finite for all k ∈ N, this assertion will follow.
Suppose v ∈ [v−N ]. Then v−N and v are in the same connected component of
Γ++\EN . Hence there is a path from v−N to v contained in Γ++ \EN . Choosing arcs
within this path and extending them to VN if necessary, we see that there are vertices u0 , . . . , un ∈ VN with u0 = v−N , un = v and descΓ++ (ui ) ∩ descΓ++ (ui+1 ) 6= ∅.
We use induction to show that ui ∈ descΓ++ (v0 ). Clearly, u0 = v−N ∈ descΓ++ (v0 ).
Suppose uk ∈ descΓ++ (v0 ) and let (w0 , . . . , wl ) be an arc such that w0 = uk+1 and
wl ∈ descΓ++ (uk )∩descΓ++ (uk+1 ). Then wl ∈ descΓ++ (v0 ) and ψ(w−l ) = N +l > N .
This implies wl−1 ∈ in(wl ) ⊆ descΓ++ (v0 ) by the choice of N . Repeating this process until we have uk+1 = w0 ∈ in(w1 ) ⊆ descΓ++ (v0 ) completes the induction.
Now define a directed graph T as the quotient of Γ++ by the vertex equivalence
relation introduced above. In particular, ([w], [w0 ]) is an edge in T if and only if there
are representatives w ∈ [w] and w0 ∈ [w0 ] such that (w, w0 ) is an edge in Γ++ . The
following result collects properties of T . For the statement, we let d+ = | outΓ++ (v0 )|
and d− = | inΓ++ (v−1 )|. We let ϕ : Γ++ → T denote the quotient map.
Lemma 4.13. Retain the above notation. The quotient T is a forest of regular rooted
trees of degree d+ /d− . The map ρ and the action of U++ on Γ++ descend to T .
Furthermore, we have the following.
(i) The map ρ is a graph morphism from T onto ρ(T ) where
V (ρ(T )) = {[uv−i ] | u ∈ U++ , i ∈ N}, and
E(ρ(T )) = {([uv−i ], [uv−i−1 ]) | u ∈ U++ , i ∈ N}.
(ii) For every v ∈ V (T ), the stabilizer (U++ )v acts transitively on outT (v).
Proof. It is clear that if v ∈ V (Γ++ ) ∩ V0 , then | inT ([v])| = 0 since | inΓ++ (u)| = 0
for all u ∈ V0 . We now show that if v ∈ Γ++ \ V0 , then | inT ([v])| = 1. Since
| inΓ++ (v)| ≥ 1, we have | inT ([v])| ≥ 1. Suppose now that (u0 , [v]) and (u1 , [v])
are edges in T . Then there are representatives u0i , wi0 ∈ V (Γ++ ) such that u0i ∈
[ui ], wi ∈ [v] and (u0i , wi0 ) ∈ E(Γ++ ) for i ∈ {0, 1}. In particular, w0 is in the
same connected component of Γ++ \ Eψ(w0 ) as w1 . Consequently, u00 is in the same
connected component of Eψ(w0 )−1 as u01 . As ψ(u00 ) = ψ(w0 )−1 = ψ(w1 )−1 = ψ(u01 ),
this shows that u0 = [u00 ] = [u01 ] = u1 and so (u0 , [v]) = (u1 , [v]). Hence | in([v])| = 1.
The map ρ and the action of U++ on Γ++ descend to T by Lemma 4.10 and
the preceding paragraph. The assertions concerning ρ and ρ(T ) are immediate from
Proposition 4.3. It follows from Proposition 4.3, that uρn (inT (v)) = inT (uρn (v)).
Proposition 4.5 shows that an analogue of Remark 4.6 also holds for T . Hence T is
a forest of regular rooted trees and has constant out-valency.
Let d denote the out-valency of T . As in [Möl00, Lemma 5], we argue that
d = d+ /d− . By Lemma 4.12, equivalence classes of vertices in Γ++ have constant
finite order k ∈ N. Given v ∈ V (T ), let A := ϕ−1 (v). The d edges issuing from v
end in vertices w1 , . . . , wd ∈ V (T ). Put B := ϕ−1 ({w1 , . . . , wd }). Then all edges in
Γ++ ending in B originate in A because T has in-valency 1. The number of edges
issuing from A, which is kd+ , and the number of edges terminating in B, which is
kdd− , are thus equal. Hence d = d+ /d− .
For (ii), let v ∈ V (T ) and u0 , u1 ∈ outT (v). Pick representatives w0 , w00 , w1 , w10
in V (Γ++ ) such that ([wi ], [wi0 ]) = (v, ui ) for i ∈ {0, 1} and choose g ∈ U++ such
that g(w0 , w00 ) = (w1 , w10 ) by Proposition 4.5. Then gv = v and gu0 = u1 .
Theorem 4.14. Let G be a t.d.l.c. group and α ∈ End(G). Then there exists a
compact open subgroup V ≤ G which is tidy for α.
WILLIS THEORY VIA GRAPHS
15
Proof. By Lemma 3.1 we may assume that {v−i | i ∈ N} is infinite. Furthermore,
by Theorem 3.4, we may assume that U is tidy above for α.
For i ∈ N0 , let vi0 := ϕ(vi ) ∈ V (T ). In view of the fact that Γ++ ⊆ Γ, consider
V := G{X0 } where X0 := [v0 ] ⊆ V (Γ++ ) is the equivalence class of v0 in Γ++ . Then
V is open in the permutation topology coming from Γ as GX0 ≤ V = G{X0 } and
hence also open (and closed) in G. Since X0 is finite by Lemma 4.12 we conclude
that V is compact as it contains the compact group U as a finite index subgroup.
By Remark 4.1, we have descΓ++ (X0 ) = descΓ (X0 ). As V descΓ (X0 ) = descΓ (X0 ),
the group V acts on descΓ++ (X0 ) by automorphisms.
It is clear that V preserves Vk , Ek and connected components. So the action
of V descends to T and V stabilizes v00 ∈ V (T ). Note that (U++ )v00 ≤ V and so
iterated application of Lemma 4.13 shows that V acts transitively on vertices of
fixed depth in T . Also, Vvi0 = V ∩ α−i (V ): Suppose g ∈ V and gvi = uvi , where
u ∈ U++ . Then g −1 u ∈ α−i (U ). Thus αi (g −1 u) ∈ U and so αi (g)v0 = αi (u)v0 .
Applying Lemma 4.10, we see that gvi ∼ vi if and only if αi (g)v0 ∼ v0 . Finally,
applying the orbit-stabilizer theorem and Lemma 4.13 we have
0
0
[V : V ∩ α−n (V )] = |V v−n
| = (d+ /d− )n = |V v−1
|n = [V : V ∩ α−1 (V )]n .
for all n ∈ N. Hence V is tidy for α by Lemma 3.12
Remark 4.15. Retain the above notation and assume that U is tidy. We argue
that in this case Γ++ and T coincide: It suffices to show that |in(v)| = 1 for some
v = uv−i with i > 0 as Proposition 4.8 shows that the relation ∼ on Γ++ is trivial.
By Remark 4.1 and Theorem 3.11, the graph descΓ++ (v0 ) = Γ+ is already a tree.
Lemma 4.11 shows that there exists a vertex v with in(v) ⊂ Γ+ . Thus |in(v)| = 1.
The following lemma will be used in Section 6.
Lemma 4.16. Suppose U is tidy for α. Then U++ ∩ U−− ≤ U+ ∩ U− ≤ U .
Proof. Since U is tidy for α, the graph Γ++ is a forest of rooted trees by Remark
4.15 . Note that for each u ∈ U++ ∩ U−− , there exists i ∈ N0 such that uv−i = v−i .
Hence U++ ∩U−− preserves descΓ++ (v0 ). Since this is a tree with root v0 , U++ ∩U−−
is contained within stabG (v0 ) = U . The claim now follows from Lemma 3.8.
5. The Scale Function and Tidy Subgroups
In this section we link the concept of tidy subgroups to the scale function and
thereby recover results of [Wil15] in a geometric manner. First, we make a preliminary investigation into the intersection of tidy subgroups. Let G be a t.d.l.c. group,
α ∈ End(G) and U (1) , U (2) ≤ G compact open as well as tidy for α.
Proposition 5.1. Retain the above notation. Then
[U (1) : U (1) ∩ α−1 (U (1) )] = [U (2) : U (2) ∩ α−1 (U (2) )]
To prove Proposition 5.1, we need some preparatory lemmas concerning inverse
images of U (1) and U (2) . The first one complements Lemma 3.1.
Lemma 5.2. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G compact open and
tidy above for α. If {α−n (U ) | n ∈ N0 } is finite then α(U ) = U = α−1 (U ).
T∞
Proof. By assumption, the intersection k=0 α−k (U ) has only finitely many terms
and hence stabilizes eventually. For sufficiently large n ∈ N0 we therefore have
[U−n : U−n−1 ] = 1. By Proposition 3.6, we get for all m ∈ N0 that
1 = [U−n : U−n−1 ] = [U : U−1 ] = [U−m : U−m−1 ] = [α−m (U ) : α−m (U ) ∩ α−m−1 (U )].
For m = 1, we obtain [U : U−1 ] = [U : U ∩ α−1 (U )] = 1 = [α−1 (U ) : U ∩ α−1 (U )].
That is, α−1 (U ) ⊇ U and U ⊇ α−1 (U ) which yields the assertion.
16
TIMOTHY P. BYWATERS AND STEPHAN TORNIER
The next lemma settles Proposition 5.1 when both {α−n (U (1) ) | n ∈ N0 } and
{α (U (2) ) | n ∈ N0 } are finite.
−n
Lemma 5.3. Retain the above notation. If {α−n (U (i) ) | n ∈ N0 } is finite for both
i ∈ {1, 2} then [U (1) : U (1) ∩α−1 (U (1) )] = [U (2) : U (2) ∩α−1 (U (2) )] and U (1) ∩U (2) is
tidy for α.
Proof. The first assertion follows from Lemma 5.2. By the same Lemma we have
α−1 (U (1) ∩ U (2) ) = α−1 (U (1) ) ∩ α−1 (U (2) ) = U (1) ∩ U (2) . Lemma 3.1 now entails
that (U (1) ∩ U (2) )− = U (1) ∩ U (2) is tidy for α.
Retain the above notation and set V := U (1) ∩ U (2) . Consider the graph Γ+
associated to V .
Lemma 5.4. Retain the above notation. Then either Γ+ is a directed infinite tree,
rooted at v0 , with constant in-valency 1 excluding the root, or there exists n ∈ N0
such that α−n (V ) = α−n−k (V ) for all k ∈ N0 .
Proof. Note that if α−n (V ) = α−n−1 (V ) then α−n (V ) = α−n−k (V ) for all k ∈ N0 .
Suppose instead that α−n (V ) 6= α−n−1 (V ) for all n ∈ N0 . By Lemma 5.3 we
may assume, without loss of generality, that {α−n (U (1) ) | n ∈ N0 } is infinite. In
(1)
particular, we may consider the graph Γ+ associated to U (1) which is an infinite
rooted tree by Theorem 3.11.
We have to show that Γ+ does not contain a cycle, the in-valency of v0 ∈ V (Γ+ )
is 0 and the in-valency of every other vertex in Γ+ is precisely 1. Note that every
vertex excluding v0 has in-valency at least 1: By assumption, v−i 6= v−i−1 for all
i ∈ N. In particular v−i ∈ in(v−i−1 ) for all i ∈ N.
Now, suppose there is a cycle (u0 v−i , . . . , un v−i−n = u0 v−i ) in Γ+ , where uj ∈ V
for all j ∈ {0, . . . , n}. Then α−i (V ) = α−i−n (V ) and so (v−i , . . . , v−i−n ) is a nontrivial cycle. We aim to show that v−i has in-valency at least 2 in this case. We
can choose u ∈ α−i−1 (V )\α−i (V ): If α−i−1 (V ) ⊆ α−i (V ) then iterated applications
of α−1 show α−i (V ) ⊇ α−i−1 (V ) ⊇ α−i−n (V ) = α−i (V ), in contradiction to the
assumption. Since α−i−1 (V ) = α−1 α−i (V ) = α−1 α−i−n (V ), also u ∈ α−i−n−1 (V )\
α−n−i (V ). This implies that (uv−i−n , v−i−n−1 ) is an edge in Γ+ which is distinct
from (v−n−i , v−n−i−1 ).
Noting that if v0 has non-zero in-valency then we have a cycle, it remains to show
that no vertex has in-valency at least 2. We split into two cases: First, consider the
case where {α−n (U (2) ) | n ∈ N0 } is finite. Then α−n (U (2) ) = U (2) for all n ∈ N0 by
Lemma 5.2 and
| inΓ+ (vi )| = [α−i (V ) : α−i ∩ α−i+1 (V )]
= [α−i (U (1) ) ∩ U (2) : α−i (U (1) ) ∩ α−i+1 (U (1) ) ∩ U (2) ]
(1)
≤ [α−i (U (1) ) : α−i (U (1) ) ∩ α−i+1 (U (1) )] = | inΓ(1) (v−i )| = 1
+
for all i ∈ N which suffices.
In the case where {α−n (U (2) ) | n ∈ N0 } is infinite, suppose for the sake of a
contradiction that uv−n ∈ V (Γ+ ) (n ∈ N) has in-valency at least 2. Choose vertices wv−n+1 , zv−n+1 ∈ V (Γ+ ) such that (wv−n+1 , uv−n ) and (zv−n+1 , vv−n ) are
(i)
distinct edges in Γ+ . Let ϕi : Γ+ → Γ+ (i ∈ {1, 2}) be the graph morphism given
(i)
by ϕi (uv−j ) = uv−j for all j ∈ N0 and u ∈ V ⊆ U (i) . Since each vertex excluding
(i)
the root in Γ+ has in-valency 1, we have ϕi (wv−n+1 ) = ϕi (zv−n+1 ). This im−1
plies w z ∈ α−n+1 (U (1) ) ∩ α−n+1 (U (2) ) = α−n+1 (V ). Thus wv−n+1 = zv−n+1 in
contradiction to the assumption.
Set ki = [U (i) : V ] and di = [U (i) : U (i) ∩ α−1 (U (i) )].
WILLIS THEORY VIA GRAPHS
17
Lemma 5.5. Retain the above notation. We have ki dni ≥ |V v−n | ≥ dni /ki . Also, if
{α−i (V ) | i ∈ N0 } is finite then d1 = 1 = d2 .
(i)
Proof. Since U (i) is tidy, either the graph Γ+ is a tree with out-valency di by
Theorem 3.11, or {α−i (U (i) ) | i ∈ N0 } is finite and α(U (i) ) = U (i) = α−1 (U (i) ) by
(i)
Lemma 5.2, whence di = 1. In both cases, ki dni = ki |U (i) v−n |, as the following
arguments show: In the former case this follows from Lemma 3.10, in the latter we
(i)
(i)
(i)
have v−n = v0 whence |U (i) v−n | = 1. Next, we have
(i)
ki |U (i) v−n | = [U (i) : V ][U (i) : U (i) ∩ α−n (U (i) )].
Since [α−n (U (i) ) : α−n (V )] ≤ [U (i) : V ] we obtain
(i)
ki |U (i) v−n | ≥ [U (i) : U (i) ∩ α−n (U (i) )][α−n (U (i) ) : α−n (V )]
≥ [U (i) : U (i) ∩ α−n (U (i) )][α−n (U (i) ) ∩ U (i) : U (i) ∩ α−n (V )]
= [U (i) : U (i) ∩ α−n (V )]
= |U (i) v−n |
where U (i) v−n is the orbit of v−n under the action U (i) in P(G). Since V ≤ U (i) ,
we have ki dni ≥ |U (i) v−n | ≥ |V v−n | which is the first inequality.
(i)
Since α−n (V ) = α−n (U (1) ) ∩ α−n (U (2) ) ≤ α−n (U (i) ), we have |V v−n | ≥ |V v−n |
when considered as orbits in P(G). The orbit-stabilizer theorem now implies
(i)
(i)
|V v−n | =
(i)
≥
(i)
[U (i) : V ][V : stabV (v−n )]
[U (i) : stabV (v−n )]
=
ki
[U (i) : V ]
(i)
[U (i) : stabU (i) (v−n )]
|U v−n |
dn
=
= i,
ki
ki
ki
as required. Finally, if {α−i (V ) | i ∈ N0 } is finite, then α−n (V ) = α−n−k (V )
for n sufficiently large and k ∈ N0 by Lemma 5.4. Thus (|V v−n |)n∈N0 eventually
stabilizes. This implies di = 1.
Proof. (Proposition 5.1). By Lemma 5.5, we may assume that {α−i (V ) | i ∈ N0 } is
infinite. In this case, Lemma 5.4 shows that Γ+ is a rooted tree with root v0 . Let
tn = | outΓ+ (v−n )| for n ∈ N0 . Since Γ+ is a rooted tree, tn = [V−n : V−n−1 ].
The sequence (tn )n∈N0 is non-increasing: Indeed, we have
tn−1 = [V−n+1 : V−n ] ≥ [V−n : V−n−1 ] = tn
for all n ∈ N by the following argument: If u, u0 ∈ V−n with uV−n−1 6= u0 V−n−1 ,
then α(u) ∈ α(V−n ) ≤ V−n+1 by Lemma 2.3. Similarly α(u0 ) ∈ V−n+1 . However
since u−1 u0 6∈ α−n−1 (U ), α(u−1 )α(u0 ) 6∈ V−n .
Since the sequence (tn )n∈N0 is non-negative, non-increasing and takes integer
values it isQeventually constant equal to some integer t. Since Γ+ is a tree, we have
n−1
|V v−n | = i=1 ti . Given that ti = t for almost all i ∈ N0 there is a constant l ∈ Q
such that |V v−n | = ltn for sufficiently large n. Then
ki dni ≥ |V v−n | = ltn ≥
dni
ki
for large enough n ∈ N and i ∈ {1, 2} by the first claim. As a consequence, we have
t = di for i ∈ {1, 2} which implies the overall assertion.
The following theorem links the concept of being tidy to the scale function.
Theorem 5.6. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G compact open. Then
U is tidy for α if and only if U is minimizing for α. In this case, s(α) = | outΓ+ (v0 )|.
18
TIMOTHY P. BYWATERS AND STEPHAN TORNIER
Proof. Suppose that U is minimizing for α. If {α−k (U ) | k ∈ N0 } is finite then
s(α) = 1 by Lemma 3.1. Consequently, α(U ) ≤ U . Therefore, we have U = U− and
U−− ≥ U− = U is open and hence closed.
Assume now that {α−k (U ) | k ∈ N} is infinite. First, we show that U is tidy
above for α. Suppose otherwise. Then by Theorem 3.4 and Lemma 3.3 there is
n ∈ N such that with v−1 ∈ V (Γ) we have |Un v−1 | = |U+ v−1 | |U v−1 | and so that
U−n is tidy above for α. Then
[α(U−n ) : α(U−n ) ∩ U−n ] = [U−n : U−n ∩ α−1 (U−n )] = [Un : Un ∩ α−1 (U )]
= |Un v−1 |
|U v−1 | = [U : U ∩ α−1 (U )] = [α(U ) : α(U ) ∩ U ].
where the equalities follow by applying the appropriate power of α to the respective
quotient, using Lemma 2.3. This contradicts the assumption that U is minimizing.
Now consider the graph Γ++ associated to U with out-valency d+ , and in-valency
−
d , excluding all v ∈ V (Γ++ ) with ψ(v) = 0. Since U is tidy above, Theorem 3.5
and Remark 4.1 imply that d+ = |U v−1 | = [U : U ∩ α−1 (U )] = [α(U ) : α(U ) ∩ U ].
Let V denote the tidy subgroup constructed from the graph Γ++ associated to U
by Theorem 3.11. Then the quotient T of Γ++ has out-valency
d = [V : V ∩ α−1 (V )] = [α(V ) : α(V ) ∩ V ].
Furthermore, d = d+ /d− by Lemma 4.13. The fact that U is minimizing now implies
d− = 1. It follows that Γ+ is already a tree. Thus U is tidy by Theorem 4.14.
Conversely, assume that U is tidy for α. Let V ≤ G be a compact open subgroup
which is minimizing. Then V is tidy by the above and Proposition 5.1 implies that
s(α) = [α(V ) : α(V ) ∩ V ] = [V : V ∩ α−1 (V )] = [U : U ∩ α−1 (U )] = [α(U ) : α(U ) ∩ U ].
That is, U is minimizing.
Corollary 5.7. Let G be a t.d.l.c. group and α ∈ End(G). Then s(αn ) = s(α)n .
Proof. By Theorem 4.14 there is a compact open subgroup U ≤ G which is tidy
for α. Following Theorem 5.6 the group U is minimizing and therefore
s(α) = [α(U ) : α(U ) ∩ U ] = [U : U ∩ α−1 (U )].
Since U is also tidy for αn by Lemma 3.12 we conclude, using the same lemma, that
s(αn ) = [αn (U ) : αn (U ) ∩ U ] = [U : U ∩ α−n (U )] = [U : U ∩ α−1 (U )]n = s(α)n .
Möller’s spectral radius formula [Möl02, Theorem 7.7] for the scale may be proven
as in [Wil15, Proposition 18] but with reference to Theorem 4.14 for the existence
of tidy subgroups.
Theorem 5.8. Let G be a t.d.l.c. group, α ∈ End(G) and U ≤ G compact open.
Then s(α) = limn→∞ [αn (U ) : αn (U ) ∩ U ]1/n .
6. The Tree-Representation Theorem
In this section, we prove an analogue of the following tree representation theorem
for automorphisms due to Baumgartner and Willis [BW04], see also [Hor15].
Theorem 6.1 ([BW04, Theorem 4.1]). Let G be a t.d.l.c. group, α ∈ Aut(G) of
infinite order and U ≤ G compact open as well as tidy for α. Then there is a regular
tree T of degree s(α) + 1 and a homomorphism ϕ : U++ o hαi → Aut(T ) such that
(i) ϕ(U++ o hαi) fixes an end ω ∈ ∂T and is transitive on ∂T \ {ω},
(ii) the stabilizer of each end in ∂T \ {ω} is conjugate to (U+ ∩ U− ) o hαi,
(iii) ker(ϕ) is the largest compact normal subgroup N E U++ with α(N ) = N ,
(iv) ϕ(U++ ) is the set of elliptic elements in ϕ(U++ o hαi).
WILLIS THEORY VIA GRAPHS
19
To prove an analogous statement for endomorphisms, we let α ∈ End(G) have
infinite order and U ≤ G compact open as well as tidy for α. Let S := U++ o hαi
be the topological semidirect product semigroup of the (semi)group U++ ≤ G and
the semigroup hαi ≤ End(G), where End(G) is equipped with the compact-open
topology and hαi acts continuously on U++ by endomorphisms as α(U++ ) = U++ ,
see [CHK83, Theorem 2.9, Theorem 2.10]. In particular:
(1) Elements of S have the form (u, αk ) for some u ∈ U++ and k ∈ N0 . We
identify (U++ , id) with U++ , and (id, hαi) with hαi.
(2) Composition in S is given by (u0 , αk0 )(u1 , αk1 ) = (u0 αk0 (u1 ), αk0 +k1 ).
(3) The topology on S is the product topology on the set U++ × hαi.
(4) The subsemigroup of S generated by (id, α) is isomorphic to (N, +) because
α ∈ End(G) has infinite order.
We split the construction of the desired tree into the cases s(α) = 1 and s(α) > 1.
First, assume s(α) > 1. Recall that v−i := α−i (U ) ∈ P(G) for i ≥ 0. We extend
this definition to positive indices by setting vi := αi (U ) ∈ P(G) for all i ∈ Z. The
following lemma shows that these vertices are all distinct.
Lemma 6.2. Retain the above notation. In particular, assume s(α) > 1. Suppose
αm (U ) = αn (U ) for some n, m ∈ Z. Then m = n.
Proof. For m, n ≤ 0, an equality α−m (U ) = α−n (U ) with m 6= n implies that the
set {α−k (U ) | k ∈ N0 } is finite and hence s(α) = 1 by Lemma 3.1.
Now, let 0 ≤ m < n. Then Lemma 3.7, Lemma 3.12 and Corollary 5.7 show that
s(α)n = [αn (U+ ) : U+ ]
= [αn (U+ ) : αm (U+ )][αm (U+ ) : U+ ]
= [αn (U+ ) : αm (U+ )]s(α)m .
Since m < n and s(α) > 1, we get [αn (U+ ) : αm (U+ )] 6= 1. Hence there exists
u ∈ αn (U+ )\αm (U+ ) ⊆ αn (U ). For the sake of a contradiction, suppose u ∈ αm (U ).
Since U is tidy above, there exists u± ∈ U± such that u = αm (u+ )αm (u− ). It
follows that αm (u+ )−1 u ∈ αn (U+ ) ≤ U++ since αm (U+ ) ≤ αn (U+ ). Also, we have
αm (u− ) ∈ αm (U− ) ≤ U− ≤ U−− , and so applying Lemma 4.16,
αm (u+ )−1 u ∈ U++ ∩ U−− ≤ U+ ∩ U− ≤ αm (U+ ).
It follows that u ∈ αm (U+ ), a contradiction. Thus u 6∈ αm (U ) and αn (U ) 6= αm (U ).
Finally, suppose m < 0 < n and αm (U ) = αn (U ). Then αm (U ) is a compact
open subgroup which is stabilized by αn−m . This shows s(αn−m ) = 1 which implies
s(α) = 1 by Corollary 5.7. This contradicts the assumption s(α) > 1.
We define a directed graph Γ++ by setting
V (Γ++ ) = {uvi | i ∈ Z, u ∈ U++ } and E(Γ++ ) = {(uvi , uvi−1 | i ∈ Z, u ∈ U++ }.
Note that Γ++ is a subgraph of Γ++ and that U++ acts on Γ++ by automorphisms. We will show that the map ρ, defined in the paragraph preceding Proposition 4.3, extends to an automorphism of Γ++ . To do so, consider the following
subgroups associated to α:
par− (α) := {x ∈ G | there exists a bounded α-regressive trajectory for x},
bik(α) := {x ∈ par− (α) | αn (x) = e for some n ∈ N}.
It follows from [Wil15, Proposition 20], [Wil15, Definition 12] and Theorem 5.6 that
bik(α) ≤ U . The same proposition implies that for u1 , u2 ∈ U++ ≤ par− (α) with
α(u1 ) = α(u2 ) we have u−1
1 u2 ∈ bik(α) ≤ U .
Now define ρ : Γ++ → Γ++ as follows: Given uvi ∈ V (Γ++ ), choose u0 ∈ U++
such that α(u0 ) = u and set ρ(uvi ) = u0 vi−1 .
20
TIMOTHY P. BYWATERS AND STEPHAN TORNIER
Proposition 6.3. Retain the above notation. The map ρ is an automorphism of Γ++ .
Proof. We first show that ρ is well-defined: By Lemma 6.2, it suffices to suppose
u0 , u1 , u00 , u01 ∈ U++ and i ∈ Z are such that u0 vi = u1 vi , α(u00 ) = u0 and
i
0 −1 0
α(u01 ) = u1 . Then u−1
u1 ∈ α−1 (αi (U )) ∩ U++ . For any
0 u1 ∈ α (U ) and (u0 )
−1
i−1
u3 ∈ α (U ) with α(u3 ) = u0 u1 we get ((u00 )−1 u1 )−1 u3 ∈ bik(α) ≤ αi−1 (U )
as bik(α) ≤ U and α(bik(α)) = bik(α). Hence (u00 )−1 u1 ∈ αi−1 (U ). This shows
u00 vi−1 = u01 vi−1 , hence ρ is well-defined. To see that ρ is a bijection on V (Γ++ ) note
ρ(α(u)vi+1 ) = uvi and that ρ−1 defined by uvi 7→ α(u)vi+1 is well-defined by the
following argument: If uvi = u0 vi , then u−1 u0 ∈ αi (U ) and α(u)−1 α(u0 ) ∈ αi+1 (U ).
Thus α(u)vi+1 = α(u0 )vi+1 .
Note that Γ++ contains Γ++ as a subgraph and Γ++ is a forest of rooted regular
trees by Remark 4.15. For v ∈ V (Γ++ ), there is n ∈ N0 such that ρn (v) ∈ V (Γ++ ).
This shows that the in-valency of v is 1. We find that Γ++ is a regular tree with
constant out-valency s(α) by Theorem 3.11 and Remark 4.1. Since ρ is a translation
in Aut(Γ++ ) we see that the subsemigroup generated by ρ−1 is isomorphic to (N, +).
Define ϕ : U++ t hαi → Aut(Γ++ ) by ϕ(u)(u0 vi ) = uu0 vi for all u, u0 ∈ U++ and
ϕ(αk ) = ρ−k for all k ∈ N0 .
Lemma 6.4. Retain the above notation. The map ϕ extends to a continuous semigroup homomorphism ϕ : S → Aut(Γ++ ).
Proof. Note that ϕ extends separately both to a semigroup homomorphism of
U++ , and the semigroup generated by α. To show that it extends to a semigroup homomorphism of S it suffices to show that ϕ(α)ϕ(u) = ϕ(α(u))ϕ(α). Then
ϕ(u, αn ) := ϕ(u)ϕ(αn ) is well-defined for all u ∈ U++ and n ∈ N0 . Given a vertex
u0 vi ∈ V (Γ++ ), we obtain as required:
ϕ(α)ϕ(u)u0 vi = ρ−1 (uu0 vi ) = α(uu0 )vi+1 = α(u)ρ−1 (u0 vi ) = ϕ(α(u))ϕ(α)uvi .
To see that ϕ is continuous it suffices to show that {x ∈ S | ϕ(x)w = w0 } is open in
S for all w, w0 ∈ V (Γ++ ). This follows from the fact that the stabilizer V of w0 in
U++ is an open subgroup of U++ , so x is contained in the open subset (V, id)x ⊆ S
and ϕ((V, id)x)w = w0 .
We are now in a position to prove an analogue of Theorem 6.1 for endomorphisms.
Theorem 6.5. Let G be a t.d.l.c. group, α ∈ End(G) of infinite order, U ≤ G
compact open as well as tidy for α, and S := U++ o hαi. Then there is a tree T
and a continuous semigroup homomorphism ϕ : S → Aut(T ) such that
(i) T has constant valency s(α) + 1,
(ii) ϕ(S) fixes an end ω ∈ ∂T and is transitive on ∂T \ {ω},
(iii) ker(ϕ) is the largest compact normal subgroup N E U++ with α(N ) = N ,
(iv) ϕ(U++ ) is the set of elliptic elements of ϕ(S).
Proof. First, assume s(α) > 1. Let T be the undirected graph underlying Γ++ ,
i.e. the graph with vertex set V (Γ++ ) and edge-relation the symmetric closure
of E(Γ++ ) ⊆ V (Γ++ ) × V (Γ++ ). The continuous semigroup homomorphism ϕ
from S to Aut(Γ++ ) defined above induces a continuous semigroup homomorphism
S → Aut(T ) for which we use the same letter.
Part (i) is now immediate from the fact that every vertex in Γ++ has out-valency
s(α) and in-valency 1.
For part (ii), let ω ∈ ∂T be the end associated to the sequence (vi )i∈N0 . Then
ρ(ω) = ω. If u ∈ U++ , then there exists an α-regressive trajectory for u eventually
contained in U . That is u ∈ αn (U ) for all sufficiently large n ∈ N whence uvn = vn
for sufficiently large n. This shows that uω = ω. Overall, we conclude ϕ(S)ω = ω.
WILLIS THEORY VIA GRAPHS
21
Now consider the end −ω ∈ ∂T associated to the sequence (v−i )i∈N0 . Given
another end ω 0 ∈ ∂T defined by (uk−i vk−i )i∈N0 for k ∈ Z and a sequence (uk−i )i∈N0
k 0
00
in U++ , the sequence u−1
k ρ ω represents an end ω ∈ ∂T originating from v0 and
it suffices to show that there is an element u ∈ U++ which maps the sequence of
−ω to that of ω 00 . This is a consequence of Lemma 4.4 by picking a convergent
subsequence inside the compact set U ∩ U++ .
As to (iii), the kernel of ϕ consists of those elements s ∈ S such that ϕ(s) fixes
every vertex of T . That is,
\ \
ker(ϕ) = U++ ∩
uαi (U ).
i∈Z u∈U++
In particular, ker(ϕ) is compact and satisfies α(ker(ϕ)) = ker(ϕ) as α(U++ ) = U++ .
Now, let N be any compact normal subgroup of U++ with α(N ) = N . Then
ϕ(N ) ≤ Aut(Γ++ )v for some v ∈ V (Γ++ ) because ϕ(N ) is compact by Lemma 6.4.
Since N is normal in U++ , we conclude that
ϕ(N ) = ϕ(u)ϕ(N )ϕ(u)−1 ≤ ϕ(N ) ∩ Aut(Γ++ )ϕ(u)v ≤ Aut(Γ++ )v,ϕ(u)v
for all u ∈ U++ . Similarly, given that α(N ) = N we have
ϕ(N ) = ϕ(α(N ))ϕ(α)ϕ(α)−1 = ϕ(α(N ) ◦ α)ϕ(α)−1
= ϕ(α ◦ N )ϕ(α)−1 = ρ−1 ϕ(N )ρ ≤ Aut(Γ++ )v,ρ−1 (v) .
as well as
ϕ(N ) = ϕ(α)−1 ϕ(α)ϕ(N ) = ϕ(α)−1 ϕ(α ◦ N )
= ϕ(α)−1 ϕ(α(N ))ϕ(α) = ρϕ(N )ρ−1 ≤ Aut(Γ++ )v,ρ(v) .
As a consequence, ϕ(N ) fixes every vertex in the orbit of v under the action of the
group generated by ϕ(S). This group acts vertex-transitively as it contains ϕ(U++ )
and both ρ and ρ−1 . This shows that ϕ(N ) fixes T , i.e. ϕ(N ) ≤ ker ϕ.
For part (iv), write s = (u, αk ) (u ∈ U++ , k ∈ N) for elements of S. Given that
ϕ(α) = ρ−1 , we necessarily have k = 0 in order for ϕ(s) to fix a vertex, so s ∈ U++ .
Conversely, every element u ∈ U++ is contained in αn (U ) for all sufficiently large
n ∈ N, so ϕ(u) fixes vn for the same values of n.
Now, assume s(α) = 1. Then α(U+ ) = U+ by Lemma 3.7. This shows that
U++ = U+ is a compact subgroup with α(U++ ) = U++ . Let T be the (undirected)
tree with vertex set Z and i, j ∈ V (T ) connected by an edge whenever |i − j| = 1.
Define ϕ : S → Aut(T ) by setting ϕ(α) to be the translation of length 1 in the
direction of ω := (i)i∈N0 ∈ ∂T , and ϕ(u) to be the identity automorphism of T for
all u ∈ U++ . Then ϕ satisfies all the conclusions of Theorem 6.5.
Remark 6.6. The action in Theorem 6.5 relates to Theorem 6.1 in the following manner: Results from [Wil15, Section 9] show that if U is tidy for α, then
bik(α) E U++ and the endomorphism α of U++ / bik(α) induced by α|U++ is an
automorphism. Let q : U++ → U++ / bik(α) be the quotient map. Then q(U+ ) is
tidy for α, (q(U+ ))++ = q(U++ ) and s(α) = s(α). Extend q to a semigroup homomorphism from S to q(U++ ) o hαi by setting q(α) = α. Also, let ϕ : S → Aut(T )
be as in Theorem 6.5 and ϕ0 : q(U++ ) o hαi → T 0 as in Theorem 6.1. Then there
exists a graph isomorphism ψ : T 0 → T such that the diagram
S
/ Aut(T )
O
ϕ
q
q(U++ ) o hαi
where ψe is conjugation by ψ, commutes.
e
ψ
ϕ0
/ Aut(T 0 ),
22
TIMOTHY P. BYWATERS AND STEPHAN TORNIER
7. New Endomorphisms From Old
We conclude with a construction that produces new endomorphisms of totally
disconnected, locally compact groups from old, inspired by [Wil15, Example 5].
Let G1 and G2 be totally disconnected compact groups. Assume that there are
isomorphisms ϕi : Gi → Hi ∼
= Gi ≤ Gi (i ∈ {1, 2}) of Gi onto compact open
subgroups Hi ≤ Gi . Consider the HNN-extension G of G1 × G2 which makes the
isomorphic subgroups H1 × G2 ∼
= G1 × G2 ∼
= G1 × H2 conjugate:
G := hG1 ×G2 , t | {t−1 (h1 , g2 )t = (ϕ−1
1 (h1 ), ϕ2 (g2 )) | (h1 , g2 ) ∈ H1 ×G2 }i.
Set U := G1 × G2 ≤ G. Given that G commensurates U , it admits a unique
group topology which makes the inclusion of U into G continuous and open, see
[Bou98, Chapter III, §1.2, Proposition 1]. Then G is a non-compact t.d.l.c. group
which contains U := G1 ×G2 as a compact open subgroup. Define β ∈ End(G) by
setting β(t) = t and β(g1 , g2 ) = (ϕ1 (g1 ), g2 ) for all (g1 , g2 ) ∈ G1 × G2 . Then
β(t−1 (h1 , g2 )t) = t−1 (ϕ1 (h1 ), g2 )t = (h1 , g2 ) = β(ϕ−1
1 (h1 ), g2 ).
for all (h1 , g2 ) ∈ H1 ×G2 and hence β indeed extends to G. Note that β is continuous:
Let V ≤ G be open. Then so is V ∩ (H1 ∩ G2 ) and
β −1 (V ) ⊇ β −1 (V ∩ (H1 ∩ G2 )) ∩ U
which is open in U and therefore in G since ϕ1 is continuous. Observe that s(β) = 1
as β(U ) ≤ U . Let α := ct ◦ β ∈ End(G) where ct : G → G, g 7→ tgt−1 is conjugation
by t. For (g1 , h2 ) ∈ G1 ×H2 we have
(E)
α(g1 , h2 ) = tβ(g1 , h2 )t−1 = t(ϕ1 (g1 ), h2 )t−1 = (ϕ21 (g1 ), ϕ−1
2 (h2 ))
We proceed to show that U is tidy for α and compute s(α).
Lemma 7.1. Retain the above notation. Then U is tidy for α and s(α) = [G2 : H2 ].
Proof. We proceed via Lemma 3.12. First, we show that α−n (U )∩U = G1 ×ϕn2 (G2 ).
The inclusion G1 × ϕn2 (G2 ) ≤ α−n (U ) ∩ U follows from equation (E). Suppose
g 6∈ G1 × ϕn2 (G2 ). We will show g 6∈ α−n (U ) ∩ U . If g 6∈ U , then we are done
and so we may write g = (g1 , g2 ) ∈ G1 × (G2 \ ϕn2 (H2 )). By equation (E), there
exists 0 ≤ m < n such that αm (g1 , g2 ) ∈ G1 × (G2 \ H2 ). We therefore show that
αl (g10 , g20 ) 6∈ U for all l ∈ N whenever (g10 , g20 ) ∈ G1 × (G2 \H2 ). Indeed, αl (g1 , g2 ) =
tl (ϕl1 (g1 ), g2 )t−l is not contained in U : If tl (ϕl1 (g1 ), g2 )t−l = (g10 , g20 ) ∈ U then
t · · · t(ϕl1 (g1 ), g2 )t−1 · · · t−1 (g10−1 , g20−1 ) = 1,
contradicting Britton’s Lemma on words in HNN-extensions, see [Bri63, Lemma 4]
or [LS15, Theorem 2.1].
We have shown that α−n (U ) ∩ U = G1 × ϕn2 (G2 ). Since ϕn2 (G2 ) is a nested series
of subgroups for n ∈ N, we have
[U : U ∩ α−n (U )] = [G1 × G2 : G1 × ϕn2 (G2 )] = [G2 : ϕn2 (G2 )]
=
n−1
Y
[ϕi2 (G2 ) : ϕi+1 (G2 )] = [G2 : H2 ]n .
i=0
Lemma 3.12 shows that U is tidy. By Lemma 7.1, we have
s(α) = [U : U ∩ α−1 (U )] = [G1 ×G2 : G1 ×H2 ] = [G2 : H2 ].
WILLIS THEORY VIA GRAPHS
23
References
[BGT16] T. P. Bywaters, H. Glöckner, and S. Tornier, Endomorphisms of totally disconnected
locally compact groups, arXiv preprint 1612.06958 (2016).
[BMW12] U. Baumgartner, R. Möller, and G. A. Willis, Hyperbolic groups have flat-rank at most
1, Israel J. Math. 190 (2012), 365–388. MR 2956246
[Bou98] N. Bourbaki, General topology. chapters 1–4,, Springer 6 (1998), 1–3.
[Bri63]
J. L. Britton, The word problem, Annals of Mathematics (1963), 16–32.
[BW04]
U. Baumgartner and G. A. Willis, Contraction groups and scales of automorphisms of
totally disconnected locally compact groups, Israel Journal of Mathematics 142 (2004),
no. 1, 221–248.
[BW06]
, The direction of an automorphism of a totally disconnected locally compact
group, Math. Z. 252 (2006), no. 2, 393–428. MR 2207803 (2007a:22005)
[CH09]
Z. Chatzidakis and E. Hrushovski, An invariant for difference field extensions, arXiv
preprint 0902.0844 (2009).
[CHK83] J. H. Carruth, J. A. Hildebrant, and R. J. Koch, The theory of topological semigroups,
Pure and Applied Mathematics, M. Dekker, 1983.
[DSW06] S. Dani, N. Shah, and G. A. Willis, Locally compact groups with dense orbits under Zd actions by automorphisms, Ergodic Theory and Dynamical Systems 26 (2006), no. 5,
1443–1465.
[Gle52]
A. Gleason, Groups without small subgroups, Annals of mathematics (1952), 193–212.
[Hor15]
D. Horadam, Tree actions associated with the scale of an automorphism of a totally
disconnected locally compact group, The University of Newcastle, 2015.
[HR12]
E. Hewitt and K. A. Ross, Harmonic Analysis I, vol. 115, Springer, 2012.
[JRW96] W. Jaworski, J. Rosenblatt, and G. A. Willis, Concentration functions in locally compact groups, Mathematische Annalen 305 (1996), no. 1, 673–691.
[LS15]
R. C. Lyndon and P. E. Schupp, Combinatorial group theory, Springer, 2015.
[Möl00]
R. G. Möller, Descendants in highly arc transitive digraphs, Discrete Mathematics 247
(2000), 147–157.
[Möl02]
, Structure theory of totally disconnected locally compact groups via graphs and
permutations, Canadian Journal of Mathematics 54 (2002), no. 4, 795–827.
[Möl10]
, Graphs, permutations and topological groups, arXiv preprint 1008.3062 (2010).
[MZ52]
D. Montgomery and L. Zippin, Small subgroups of finite-dimensional groups, Proceedings of the National Academy of Sciences 38 (1952), no. 5, 440–442.
[PW03]
W. Previts and T. Wu, Dense orbits and compactness of groups, Bulletin of the Australian Mathematical Society 68 (2003), no. 1, 155–159.
[Ser03]
J.-P. Serre, Trees, Springer, 2003.
[SW13]
Y. Shalom and G. A. Willis, Commensurated subgroups of arithmetic groups, totally
disconnected groups and adelic rigidity, Geometric and Functional Analysis 23 (2013),
no. 5, 1631–1683.
[vD31]
D. van Dantzig, Studien over topologische algebra, University of Groningen, 1931.
[Wil94]
G. A. Willis, The structure of totally disconnected locally compact groups, Mathematische Annalen 300 (1994), no. 1, 341–363.
, Further properties of the scale function on a totally disconnected group, Jour[Wil01]
nal of Algebra 237 (2001), no. 1, 142–164.
, Tidy subgroups for commuting automorphisms of totally disconnected groups:
[Wil04]
an analogue of simultaneous triangularisation of matrices, New York J. Math. 10
(2004), 1–35 (electronic). MR 2052362 (2005a:20041)
[Wil07]
, Compact open subgroups in simple totally disconnected groups, J. Algebra 312
(2007), no. 1, 405–417. MR 2320465 (2008d:22005)
[Wil15]
, The scale and tidy subgroups for endomorphisms of totally disconnected locally
compact groups, Mathematische Annalen 361 (2015), no. 1-2, 403–442.
[Yam53] H. Yamabe, A generalization of a theorem of gleason, Annals of Mathematics (1953),
351–365.
Timothy P. Bywaters, The University of Sydney, School of Mathematics and
Statistics, NSW 2006, Australia.
e-mail: [email protected]
Stephan Tornier, ETH Zürich, D-MATH, Rämistrasse 101, 8092 Zürich, Switzerland and The University of Newcastle, School of Mathematical and Physical Sciences, University Drive, Callaghan, NSW 2308, Australia.
e-mail: [email protected]
| 4math.GR
|
On resilient control of dynamical flow networks
Giacomo Como
G. Como is with the Department of Automatic Control, Lund University, BOX118, SE-22100 Lund, Sweden and with the
Lagrange Department of Mathematical Sciences, Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129, Torino, Italy.
Email: [email protected], [email protected].
arXiv:1701.03614v3 [cs.SY] 7 Mar 2017
Abstract
Resilience has become a key aspect in the design of contemporary infrastructure networks. This comes
as a result of ever-increasing loads, limited physical capacity, and fast-growing levels of interconnectedness
and complexity due to the recent technological advancements. The problem has motivated a considerable
amount of research within the last few years, particularly focused on the dynamical aspects of network flows,
complementing more classical static network flow optimization approaches.
In this tutorial paper, a class of single-commodity first-order models of dynamical flow networks is
considered. A few results recently appeared in the literature and dealing with stability and robustness
of dynamical flow networks are gathered and originally presented in a unified framework. In particular,
(differential) stability properties of monotone dynamical flow networks are treated in some detail, and the
notion of margin of resilience is introduced as a quantitative measure of their robustness. While emphasizing
methodological aspects —including structural properties, such as monotonicity, that enable tractability and
scalability— over the specific applications, connections to well-established road traffic flow models are made.
Keywords: dynamical flow networks, transportation networks, compartmental systems, distributed
control, network resilience, network flow control, dynamical routing, robust control, monotone systems.
1. Introduction
As critical infrastructure networks, such as transport and energy, are being utilized closer and closer
to their capacity limits, the complex interaction between physical systems, cyber layers, and human decision makers has created new challenges in simultaneously achieving efficiency and reliability. The recent
technological advancements in terms of smart sensors, high-speed communication, and real-time decision
capabilities have exacerbated the large-scale interconnected nature of these systems, and increased both the
potential gains associated to their optimization and their inherent systemic risks. In fact, while designed to
perform well under normal operation conditions, such complex systems tend to exhibit critical fragilities in
response to unforeseen disruptions. Even if simply started from small local perturbations, such disruptions
have the potential to build up through cascading mechanisms driven by the interconnected dynamics of the
infrastructure network, possibly leading to detrimental systemic effects. The term resilience refers to the
ability of these systems “to plan and prepare for, absorb, respond to, and recover from disasters and adapt
to new conditions” (definition by the US National Academy of Sciences [1]).
Whilst static network flow optimization has long been regarded as a fundamental design paradigm for
infrastructure systems and represents a central area of mathematical programming [2, 3, 4], there is an
increasing awareness that the full potential of the emerging technologies and the nature of the associated
systemic risks can only be understood by developing systems modeling, robustness analysis, and control
I The author is a member of the excellence centres LCCC and ELLIIT. His research has been supported by the Swedish
Research Council through a Project Research Grant.
Preprint submitted to Annual Reviews in Control
March 8, 2017
synthesis within a dynamical framework. This recognition is stimulating considerable interest in the control
systems community. In this tutorial paper, based on a semi-plenary lecture given by the author at the
22nd International Symposium on the Mathematical Theory of Networks and Systems, a few recent results
on stability and robustness of dynamical flow networks are presented. While emphasizing methodological
aspects —in particular, structural properties enabling tractability and scalability of the considered models—
over the specific applications, this paper also makes connections to well-established road traffic flow models.
Our focus is on first-order models of dynamical flow networks, describing the flow of mass among a finite
set of interconnected cells. Such dynamical systems have sometimes been referred to as compartmental
systems in some of the literature [5, 6]. Our main interest is on nonlinear dynamical flow networks, with
nonlinearities especially accounting for congestion effects. Special attention is devoted to demand and
supply constraints limiting, respectively, the maximum outflow from and the maximum inflow in the cells
as a function of their current mass, as in Daganzo’s cell transmission model for road traffic network flows
[7, 8]. We introduce a class of monotone dynamical flow networks, characterized by structural properties of
the dependence of the flow variables on the network state. We show that, while this class is large enough to
encompass many examples of applicative interest, the system structure of monotone dynamical flow networks
is such that their dynamic behaviors, and especially stability and robustness properties, are analyzable in a
tractable and scalable way. In particular, this paper: (i) presents results relating the (differential) stability
of (nonlinear) monotone dynamical flow networks to graph-theoretical properties; (ii) introduces the notion
of margin of resilience as a measure of their robustness against exogenous perturbations; and (iii) studies a
class of locally responsive feedback routing and flow control policies that are able to achieve the maximum
possible margin of resilience for a given network topology in spite of relying on local information only and
requiring no global knowledge of the network.
The remainder of this paper is organized as follows. In Section 2, we first introduce dynamical flow
networks. In Section 3, we focus on their simplest instance, affine dynamical flow networks, for which we
gather some results relating their (global, exponential) stability to outflow-connectivity properties of the
network topology. In Section 4, we define the notion of monotone dynamical flow networks, present stability
results that can be deduced for this system structure, and show how they can be applied to dual ascent
dynamics for static convex network flow optimization. In Section 5, we introduce nonlinear dynamical flow
networks with demand and supply constraints and show how several examples of network flow dynamics
from the literature that can be fit in this framework also belong to the class of monotone dynamical flow
networks (either globally or locally), so that the stability results of Section 4 can be successfully applied.
In Section 6, we study robustness of nonlinear dynamical flow networks with respect to perturbations of
the demand functions (hence, of the flow capacities) as well as of the external inflows. We introduce the
notion of margin of resilience as a quantitative measure of robustness and compute the margin of resilience
of different classes of distributed routing and flow control policies.
We end this introductory section by gathering some notational conventions to be adopted throughout
the paper. The sets of real numbers and of nonnegative real numbers are denoted by R and R+ , respectively.
The all-one vector is denoted by 1, the all-zero vector simply by 0, the identity matrix by I = diag (1), and
the transpose of a matrix M by M T . Inequalities between vectors are meant to hold true entrywise, i.e., if
a, b ∈ Rn , then a ≥ b means that ai ≥ bi for all i = 1, . . . , n. A square matrix M is called: nonnegative if
all its entries are nonnegative; Metzler if all its non-diagonal
entries are nonnegative, i.e., Mij ≥ 0 for all
P
i 6= j [9]; (row) diagonally dominant if |Mii | ≥ j6=i |Mij |; compartmental if it is Metzler and M 1 ≤ 0;
Hurwitz if all its eigenvalues have negative real part; and substochastic if it is nonnegative and such that
M 1 ≤ 1, i.e., its rows all sum up to less than or equal to 1. A directed graph, shortly digraph, is the pair
(V, E) of a finite node set V and a link set E ⊆ V × V, whereby links (i, j) ∈ E are interpreted as pointing
from node i to node j. A length-l path from a node i to a node j in a digraph G = (V, E) is a sequence of
nodes {v0 , v1 , . . . , vl } ⊆ V such that v0 = i, vl = j, vh 6= vk for all 0 ≤ h < k ≤ l, and (vh−1 , vh ) ∈ E for all
h = 1, . . . , l. The gradient of a function f : Rn → Rm in a point x of its domain is the matrix ∇f (x) ∈ Rm×n
whose entries are the partial derivatives [∇f (x)]ij = ∂fi (x)/∂xj : if m = h + l and the variable is explicitly
written as x = (y, z) where y ∈ Rh and z ∈ Rl , then ∇y f (y, z) ∈ Rh×n and ∇z f (y, z) ∈ Rl×n are the left
and right blocks of ∇f (x) ∈ Rm×n , so that ∇f (x) = [∇y f (y, z), ∇z f (y, z)]. Finally, we use the notation
a ∨ b = max{a, b} for the maximum between two scalar values a, b ∈ R.
2
2. Dynamical flow networks
This paper focuses on single-commodity, first-order models of dynamical flow networks, also referred to
as compartmental systems in some of the literature [5, 6]. These are dynamical systems with n-dimensional
state vector x = x(t) that belongs to the nonnegative orthant Rn+ at any time t ≥ 0. The entries xi = xi (t) of
the state vector represent the mass in each cell i ∈ I, where I := {1, . . . , n} is a finite set of interconnected
cells. The network flow dynamics can be compactly expressed as
ẋ = u + F T 1 − F 1 − w
(1)
where:
(i) u ∈ Rn+ is a nonnegative vector supported on a subset R ⊆ I whose entries ui model the external
inflows in the cells i ∈ R;
(ii) F ∈ Rn×n
is a nonnegative matrix supported on a subset A ⊆ I × I of ordered pairs of adjacent cells
+
whose nonzero entries Fij represent the flow from cell i to cell j for all pairs (i, j) ∈ A;1
(iii) and w ∈ Rn+ is a nonnegative vector supported on a subset S ⊆ I whose entries wi model the outflows
from the cells i ∈ R towards the external environment.
Hence, in particular, we have
ui = 0 ,
i∈
/ R,
Fij = 0 ,
(i, j) ∈
/ A,
wi = 0 ,
i∈
/ S.
(2)
While in typical applications the external inflows are —either constant or time-varying— exogenous inputs,
the flow variables F and w in general may depend both on the state x (thus allowing for feedback) and
directly on the time t (thus allowing for exogenous time variability).
Entrywise rewriting of the dynamics (1) reads
X
X
ẋi = ui +
Fji −
Fij − wi ,
i∈I,
j∈I
j∈I
which is physically interpreted as a mass conservation law : the rate of change of the mass in cell i equals the
imbalance between the total inflow in it and the total outflow from it,
Pthe former coinciding with the sum of
the external inflow ui and the aggregate inflow
from
the
other
cells
j∈I Fji , and the latter being given by
P
the aggregate outflow towards other cells j∈I Fij and the outflow towards the external environment wi .
Invariance of the nonnegative orthant Rn+ for the state vector x is guaranteed by the additional constraint
xi = 0
=⇒
wi +
X
j
Fij = 0 ,
(3)
i.e., the outflow from an empty cell is always 0.
An equivalent form of (1) that will often prove convenient is
ẋ = u − (I − RT )z ,
(4)
where
z = F1 + w
P
is a nonegative n-dimensional vector whose entries zi = j∈I Fij + wi represent the total outflows from
the cells i ∈ I, and R ∈ Rn×n is a routing matrix whose entries Rij , sometimes referred to as split ratios,
represent the fraction of outflow from cell i directed towards cell j, i.e., they satisfy
X
Fij = Rij zi ,
wi = 1 −
Rij zi ,
i, j ∈ I .
j∈I
3
i
Ei
R
S
h
j
Figure 1: A network topology T = (I, A, R, S) whose nodes i ∈ I represent cells and links (i, j) ∈ A represent possible
flows between adjacent cells. The subsets of cells R and S is where external inflow and, respectively, outflow towards the
external environment, is possible. This network topology is both inflow-connected, as every cell i ∈ I is reachable from some
h ∈ R by a directed path, and outflow-connected, as from every cell i some cell j ∈ S can be reached by a directed path. The
outneighborhood Ei of a cell i is also highlighted. Note that, in this specific network topology, the out-neighborhoods of two
nodes either coincide or are disjoint. In fact, G = (I, A) is the line-digraph (see Remark 1) of the digraph N = (V, I) displayed
in Figure 2.
The routing matrix R is supported on
/ A and it is substochastic,
P A, i.e., such that Rij = 0 for all (i, j) ∈
i.e., nonnegative and with row sums j Rij ≤ 1, with strict inequality possible only at the cells i ∈ S from
which direct outflow towards the external environment is allowed.
The sets I, A, R, and S determine the network topology, to be denoted by T = (I, A, R, S). In particular,
it is natural to interpret the sets I and A, respectively, as the node set and the link set of a directed graph
G = (I, A): see Figure 1. We refer to a cell i ∈ I as outflow-connected if either i ∈ S or there exists a path
in G starting in i and ending in some j ∈ S. The network topology T is referred to as outflow-connected if
every cell i ∈ I is outflow-connected. Analogously, a cell i ∈ I is referred to as inflow-connected if either
i ∈ R or there exists a path in G starting in some j ∈ R and ending in i, while the network topology is
called inflow-connected if every cell i is inflow-connected. Throughout the paper, we will use the notation
Ei := {k ∈ I : (i, k) ∈ A} ,
i∈I,
for the out-neighborhoods of the cells in T .
Remark 1. In some applications, especially in road traffic network flows, the cells i ∈ I are often represented as links of a different digraph N = (V, I) whose node set V = {v0 , v1 , . . . , vm } includes a node v0
representing the external environment and m nodes v1 , . . . , vm representing either junctions or interfaces
between consecutive cells. (See Figure 2.) In this case, the sets
R = {i ∈ I : i = (v0 , vk ) for some 1 ≤ k ≤ m} ,
S = {i ∈ I : i = (vk , v0 ) for some 1 ≤ k ≤ m}
represent on-ramps and off-ramps, respectively, and
A = {(i, j) ∈ I × I : i = (vh , vk ) and j = (vk , vl ) for some 0 ≤ h, l ≤ m and 1 ≤ k ≤ m}
is the set of ordered pairs of adjacent cells. Clearly, to any such digraph N = (V, I), we can associate a
network topology T = (I, A, R, S): in particular, G = (I, A) turns out to be the so-called line-digraph [10]
of N = (V, I). Observe that the network topology T = (I, A, R, S) corresponding to the line-digraph of
some N = (V, I) has the following properties: (a) cells r ∈ R are sources, i.e., they have no incoming links;
(b) cells s ∈ S are sinks, i.e., they have no outgoing links; and (c) the out-neighborhoods Ei , Ej of two cells
1 Throughout,
we will always assume that (i, i) ∈
/ A, so that Fii = 0, for all i ∈ I.
4
i
R
S
h
j
Figure 2: A digraph N = (V, I) representing a road traffic network: nodes v ∈ V stand for junctions and links i ∈ I
represent cells that correspond to (portions of) roads. The subsets of cells R and S represent, respectively, the on-ramps and
the off-ramps. The line digraph associated to N = (V, I) displayed above is the network topology T = (I, A, R, S) displayed
in Figure 1.
i, j ∈ I either coincide (if i and j point towards the same junction v ∈ V \ {v0 }) or they are disjoint (if they
do not).2 Hence, our choice of formulating results for the network topology T = (I, A, R, S) in which cells
i ∈ I are represented as nodes is in the interest of both clarity and greater generality, as the aforementioned
conventional representation of road traffic networks can be recovered as a special case.
3. Affine Dynamical Flow Networks
The simplest special case of dynamical flow networks is the class of affine dynamical flow networks,
characterized by the property that the flows Fij and wi are linear functions of the state vector x. In fact,
nonnegativity of the flow variables and dynamic invariance of the nonnegative orthant Rn+ imply that such
linear relationships necessarily take the form
Fij = αij xi ,
wi = βi xi ,
i, j ∈ I ,
where the α ∈ Rn×n
and β ∈ Rn+ are such that αij > 0 for all pairs of adjacent cells (i, j) ∈ A, αij = 0 for
+
all i 6= j such that (i, j) ∈
/ A, while βi > 0 for i ∈ S and βi = 0 for i ∈ I \ S. By putting
X
Lij = −αij ,
Lii = βi +
αij ,
i 6= j ∈ I ,
j6=i
and assembling the coefficients Lij in an n × n matrix L, one gets the following form for affine dynamical
flow networks
ẋ = u − LT x .
(5)
Observe that −Lij = αij ≥ 0 for all i 6= j ∈ I, i.e., −L is Metzler, and L1 = β ≥ 0. Hence, −L is a
compartmental matrix and, in particular, equation (5) describes a positive affine system [11]. We refer to
the matrix L as outflow-connected if the corresponding network topology T is. Observe that the diagonal
d = (Lii )i∈I of an outflow-connected compartmental matrix L is strictly positive and, upon defining the
substochastic routing matrix R = I − D−1 L, where D = diag (d), equation (5) can be equivalently rewritten
as
ẋ = u − (I − RT )Dx .
(6)
Some stability properties of affine dynamical flow networks are gathered in the following proposition.
2 On the other hand, properties (a), (b), and (c) are not sufficient conditions for G = (I, A) to be the line-digraph of
N = (V, I): necessary and sufficient conditions can be found in [10].
5
Proposition 2. Every affine dynamical flow network (5) is non-expansive in the l1 -distance, i.e., the l1 distance between any two of its solutions is never increasing in time. Moreover, if the compartmental matrix
−L is outflow-connected, then it is also Hurwitz stable, and, furthermore, if the inflow vector
u is constant,
P
then x∗ = (LT )−1 u is a globally exponentially stable equilibrium of (5) and V (x) = i∈I |xi − x∗i | is a
Lyapunov function.
Proof For any two solutions x(t) and x̃(t) of (5), let y(t) = x(t) − x̃(t) be their difference, and note that
ẏ = −LT y. For all t, s ≥ 0, one has that
Z t+s
||x(t + s) − x̃(t + s)||1 − ||x(t) − x̃(t)||1 = ||y(t + s)||1 − ||y(t)||1 =
−y(τ )T L sgn(y(τ ))dτ , (7)
t
where, sgn(y) ∈ {−1, 0, 1}n stands for the sign vector of y. Now, for any y ∈ Rn , one can partition the cell
set as I = I− ∪ I0 ∪ I+ according to the sign of y’s entries. Let L(s,r) and y(s) for s, r ∈ {−, 0, +} be the
corresponding blocks of L and y, respectively, and notice that
T
T
T
T
L(−,−) 1 .
L(−,+) 1 − y(−)
L(+,−) 1 + y(−)
L(+,+) 1 − y(+)
y T L sgn(y) = y(+)
(8)
T
Now, note that, since −L is Metzler, the blocks L(+,−) and L(+,−) are nonpositive, so that y(+)
L(+,−) 1 ≤ 0
T
and y(−) L(−,+) 1 ≥ 0. On the ther hand, since −L is Metzler and L1 ≥ 0, one has that L(+,+) 1 ≥ 0 and
T
T
L(−,−) 1 ≥ 0. In turn, these inequalities imply, respectively, that y(+)
L(+,+) 1 ≥ 0 and y(−)
L(−,−) 1 ≤ 0.
Hence, all the addends in the right-hand side of (8) are nonnegative, so that
y T L sgn(y) ≥ 0
and (8) implies that
||x(t + s) − x̃(t + s)||1 ≤ ||x(t) − x̃(t)||1 ,
so that the first claim is proved.
That Hurwitz stability of a compartmental matrix −L is equivalent to its outflow-connectivity is a
classical result following from the Perron-Frobenius theory, see, e.g., [5, Theorem 3]. Since the dynamical
system (5) is affine, Hurwitz stability of −L is equivalent to global exponential stability of the equilibrium
x∗ . Finally, that the l1 -distance from equilibrium ||x − x∗ ||1 is a Lyapunov function follows from the first
claim by letting one of the two solutions start from the equilibrium x∗ itself.
Remark 3. Observe that, using the decomposition L = D(I − R) the equilibrium point x∗ = (LT )−1 u of an
outflow-connected affine dynamical flow network (5) may be rewritten as
x∗ = D−1 z ∗ ,
z ∗ = (I − RT )−1 u = u + RT u + (R2 )T u + (R3 )T u + . . . .
(9)
The rightmost side of equation (9) states that the entries of the equilibriumP
point satisfy x∗i = zi∗ /Lii where
∗
zi is the aggregate of the external inflow ui in cell i, plus the weighted sum j Rji uj of the external inflows
P
in the cells j immediately upstream of i, plus the weighted sum j (R2 )ji uj of the external inflows in the
P
cells j two hops upstream of i, and so on. Here, convergence of the nonnegative term series k≥0 (RT )k u
is guaranteed by the fact that the spectral radius of R is strictly less than 1. Recall that, for every inflowconnected cell i ∈ I there exists some cell h ∈ R from which i is reachable by a directed path (c.f. Figure
1) in the digraph G = (I, A), so that there exists some k ≥ 0 such that (Rk )hi uh > 0. It then follows from
(9) that, if the network is both inflow- and outflow-connected, then the equilibrium x∗ has strictly positive
entries.
P
Remark 4. While never increasing, the l1 -distance from equilibrium i∈I |xi − x∗i | of the state vector of a
general outflow-connected affine dynamical flow network, does not necessarily decay to 0 exponentially fast
as time grows large (it does so in the special case
P when S = I). Exponential stability of (5) can instead be
certified by weighted l1 -distances of the form i ωi |xi − x∗i |, for positive values of the ωi ’s that depend on
6
the specific values of the entries of L. In fact, existence of such sum- as well as max-separable Lyapunov
functions is equivalent to stability for positive (not only compartmental) systems [12, Proposition 1], and
this is at the heart of the scalability properties of these systems. (See [13] for extensions of this result to
nonlinear monotone systems.) In contrast, the unweighted l1 -distance can be used to prove global asymptotic
stability of compartmental systems (via LaSalle’s theorem) using structural properties only, i.e., properties
that are independent of the specific values of the entries of the compartmental matrix −L, but depend just
on their sign pattern. As we shall see in the Section 4, such structural properties of the l1 -distance carry
over to a signifcant class of nonlinear dynamical flow networks for which they are very useful in establishing
stability and robustness results.
We conclude this section by noticing that affine dynamical flow network models cannot account for
congestion effects as they prescribe that the total outflow from a cell zi = Lii xi is a linear function of the
mass xi on it. Stated in other terms, in such affine dynamical flow network models, the rate zi /xi at which
single units of mass flow out of a cell i is a constant Lii that does not decrease when the mass xi in that
cell, or the mass xj in other neighbor cells j increase.
4. Monotone dynamical flow networks and their stability
In this and the next sections, we study a class of nonlinear dynamical flow networks and analyze their
stability and robustness properties. While it is known that nonlinear dynamical flow networks can exhibit
arbitrarily complex (including chaotic) dynamic behaviors [5], we focus on a class of nonlinear dynamical
flow networks which is rich enough to encompass many relevant examples and at the same time has enough
additional system structure to allow for tractable analysis and synthesis. Such dynamical system structure
proving particularly relevant in this context is monotonicity.
Recall that a dynamical system is referred to as monotone if its trajectories preserve the partial order
between initial states and inputs. [14, 15, 16] A classical result known as Kamke’s theorem [17, 18] establishes
that, for dynamical systems of the form ẋ = f (x, u) with Lispchitz-continuous f , monotonicity is equivalent
to ∇u f (x, u) being a nonnegative matrix and ∇x f (x, u) being a Metzler matrix for almost every x and u.
Considering their structure (1), we propose the following definition of monotone dynamical flow networks.
Definition 5. A dynamical flow network
ẋ = u + f (x) ,
f (x) = F T (x)1 − F (x)1 − w(x) ,
with Lipschitz-continuous flow functions F (x) and w(x), is monotone in a domain D ⊆ Rn+ if (∇f (x))T is
a compartmental matrix, i.e., if
X ∂
X ∂
∂
∂
Fki (x) −
Fik (x) −
wi (x) ≥ 0 ,
wk (x) ≥ 0 ,
∀i 6= j ∈ I ,
(10)
∂xj
∂xj
∂xj
∂xj
k∈I
k∈I
for almost every x ∈ D.
Observe that equation (10) requires that the transpose of the Jacobian matrix of f (x) be compartmental
rather than simply Metzler, hence it is more restrictive than Kamke’s conditions for monotone dynamical
systems, also referred to as cooperative systems [19, 20]. In fact, we have already pointed out in Section 3
that the matrix −LT determining the dynamics of affine dynamical flow networks (5) is a compartmental
matrix. I.e., affine dynamical flow networks are always monotone: this can be traced back to the fact
that, for linear systems, preservation of the nonnegative orthant is equivalent to preservation of the partial
ordering between trajectories. On the other hand, for nonlinear dynamical flow networks, monotonicity is
an additional property that further constraints their dynamical behavior. However, as it turns out, many
nonlinear dynamical flow networks of interest are monotone either globally (i.e., on the whole state space)
or locally (i.e., in a subdomain of their state space typically including the origin 0).
We now present a first example of nonlinear dynamical flow network here, while five more examples will
be discussed in Section 5.
7
Example 1. For a network topology T = (I, A, R, S) that is both inflow- and outlflow-connected, and a
constant inflow vector u supported on R, consider the following static network flow optimization problem
[3, 4]:
X
X
min
ψij (Fij ) +
ψk (wk ) ,
(11)
F ∈Rn×n
,w∈Rn
+:
+
(2) and
u+F T 1=F 1+w
k∈S
(i,j)∈A
where ψij ( · ), for (i, j) ∈ A, and ψk ( · ), for k ∈ S, are nonnegative-real-valued, twice differentiable, strictly
convex, and increasing cost functions, such that
0
lim ψij
(y) = lim ψk0 (y) = +∞ ,
y→+∞
(i, j) ∈ A ,
y→+∞
k∈S.
The Lagrangian of the minimization problem (11) reads
X
X
X
X
X
L(F, w, x) =
ψij (Fij ) +
ψk (wk ) +
xi ui +
Fji −
Fij − wi
=
X
i∈I
k∈S
(i,j)∈A
(ψij (Fij ) − (xi − xj )Fij ) +
X
j∈I
j∈I
(ψk (wk ) − wk xk ) +
xi ui ,
i∈R
k∈S
(i,j)∈A
X
where xi is the Lagrange multiplier associated to the mass conservation constraint at cell i. We then get the
first-order conditions for optimality
(
(
0
0
if xi − xj < ψij
(0)
0
if xk < ψk0 (0)
Fij =
w
=
(12)
k
0 −1
0
(ψij
) (xi − xj ) if xi − xj ≥ ψij
(0) ,
(ψk0 )−1 (xk ) if xk ≥ ψk0 (0) .
Observe that strict convexity of the cost functions guarantees invertibility of their derivatives in the range
0
[ψij
(0), +∞) and [ψk0 (0), +∞), for all (i, j) ∈ A and k ∈ S. The continuous-time dual ascent dynamics for
the optimization problem (11) is then given by (1) where the flow variables F = F (x) and w = w(x) are
0
given by (2) and (12). Observe that, if xi = 0 and xj ≥ 0 for some (i, j) ∈ A, then xi − xj ≤ 0 ≤ ψij
(0)
0
so that Fij = 0. Similarly, xk = 0 for k ∈ S implies that xk ≤ ψk (0), so that wk = 0. Hence, the sufficient
condition (3) for invariance of nonnegative orthant Rn+ is satisfied, so that (1), (2), and (12) actually define
a dynamical flow network. To verify that it is a monotone dynamical flow network, simply observe that
∂
Fji ≥ 0 ,
∂xj
∂
Fij ≤ 0 ,
∂xj
∂
∂
Fik =
Fki = 0 ,
∂xj
∂xj
∂
wj ≥ 0 ,
∂xj
∂
wk = 0 ,
∂xj
for all choices of three different cells i, j, k ∈ I. Hence, for f (x) = F T (x)1 − F (x)1 − w(x)
∂
∂
∂
fi (x) =
Fji −
Fij ≥ 0 ,
∂xj
∂xj
∂xj
X ∂
∂
wk =
wj ≥ 0 ,
∂xj
∂xj
k
so that (10) is satisfied. Therefore, the dual ascent dynamics defined by (1), (2), and (12) is a monotone
dynamical flow network on its whole state space Rn+ .
In the following theorem we gather a few stability properties that are enjoyed by monotone dynamical flow
networks. For the sake of simplicity, we state the results for dynamical flow networks that are monotone on
the whole state space, whereas they can be adapted to the cases where the monotonicity property is satisfied
in a subdomain of the state space only.
Theorem 6. For a monotone dynamical flow network
ẋ = u + F T (x)1 − F (x)1 − w(x) ,
(i) the partial order between external inflows and initial states is preserved by the state at any time;
8
(ii) the l1 -distance between any two solutions is never increasing.
Moreover, if the external inflows u are constant in time, then:
(iii)
(iv)
(v)
(vi)
every equilibrium x∗ is stable;
an equilibrium x∗ is globally asymptotically stable if and only if it is locally asymptotically stable;
an equilibrium x∗ is asymptotically stable if the compartmental matrix (∇f (x∗ ))T is outflow-connected;
the trajectory started from x(0) = 0 is always entrywise monotonically nondecreasing in time, hence
convergent to a (possibly infinite) limit.
Proof Claim (i) follows from Kamke’s theorem [17] and its generalization to systems with input [18].
Claim (ii) is a generalization of the first claim of Proposition 2 and follows from [21, Lemma 1]. Claim (iii)
is an immediate consequence of (ii): when one of the two trajectories is an equilibrium x∗ , the l1 -distance
||x(t) − x∗ ||1 is non-increasing in time, so that x∗ is necessarily stable. Claim (iii) is also a consequence
of (ii) as proved in [22, Lemma 6]. Claim (v) is a consequence of the Lyapunov linearization method and
Proposition 2. Finally, to prove claim (vi), let Φ(t) (x) denote the associated semi-flow and let x(t) be the
solution started with x(0) = 0: then, x(t + s) = Φ(t) (x(s)) ≥ Φ(t) (0) = Φ(t) (x(0)) = x(t), where the
inequality follows from the fact that x(s) ≥ 0 (since the nonnegative orthant is dynamically invariant) and
system monotonicity.
Quite remarkably, Theorem 6 relies solely on structural properties of the dynamical system, as it applies
to monotone dynamical flow networks irrespective of their specific form. In particular, the non-expansiveness
of the l1 -distance is an extension of the analogous result for affine dynamical flow networks (Proposition 2).
The related issue of existence of sum-separable Lyapunov functions for stable monotone systems has recently
attracted significant attention [23]. Non-expansiveness and contractivity of properly selected metrics, also
known as differential stability, in dynamical systems is an even stronger property than Lyapunov stability and
has been investigated, e.g., in [24, 25]. Finally, monotonicity and l1 -contraction for dynamical flow networks
have analogies with properties of some PDE models of dynamical flows, see, e.g., Kruzkov’s Theorem [26,
Proposition 2.3.6] for entropy solutions of scalar conservation laws.
As a first illustration of the applicability range of Theorem 6, we prove that the dual ascent dynamics
introduced in Example 1 have a globally asymptotically stable equilibrium that corresponds to a minimizer
of the convex network flow optimization (11).
Example 1. (continued) Consider the dual ascent dynamics defined by (1), (2), (12), with initial state
x(0) = 0. Since this is a monotone dynamical flow network, by point (vi) of Theorem 6, xi (t) converges to
a possibly infinite value x∗i , for all cells i ∈ I. We now prove that x∗i < +∞ for all i ∈ I. To see this, let
J = {i ∈ I : x∗i = +∞} ,
K = {i ∈ I : x∗i < +∞} ,
and observe that the limit flows
F ∗ = lim∗ F ,
w∗ = lim∗ w
x→x
x→x
satisfy
∗
Fjk
= +∞ ,
j ∈ J , k ∈ K : (j, k) ∈ A ,
∗
Fkj
= 0,
wh∗ = +∞ ,
j ∈ J,k ∈ K,
h ∈ S ∩ J . (13)
Notice that, if J contains at least one outflow-connected cell, then either J ∩ S is nonempty, or there must
exist some j ∈ J and k ∈ K such that (j, k) ∈ A. It then follows from (13) that
X
X
XX
XX
X
∗
∗
lim∗
ẋj =
uj +
Fkj
−
Fjk
−
wj∗ = −∞ ,
x→x
j∈J
j∈J
j∈J k∈K
j∈J k∈K
P
j∈J
which contradicts the fact limx→x∗ j∈J xj = +∞. Since the network is outflow-connected, then necessarily
J is empty so that x∗ is finite, and hence x∗ is an equilibrium of (1), (2), (12). Hence x∗ , F ∗ , and w∗
satisfy the mass conservation constraint u + (F ∗ )T 1 = F ∗ 1 + w∗ as well as (12), so that F ∗ , and w∗ are
9
Ci
ϕi (xi )
0
σi (xi )
xi
0
xi
xi
Figure 3: A demand function (on the left) and a supply function (on the right).
minimizers in the static network flow optimization (11). Observe that the equilibrium x∗ is necessarily stable
by point (iii) of Theorem 6. To prove its global asymptotic stability one can build on the strict convexity of
the cost functions and inflow- and outflow-connectivity of the network topology in order to prove that the
compartmental matrix (∇f (x∗ ))T is outflow-connected and then use points (v) and (iii) of Theorem 6. We
observe that dynamical flow network structures similar to the ones considered in this example can be found
in the context of dissipative flow networks [27].
The arguments developed in Example 1 can be extended to cover the case of finite capacities constraints
Fij < cij on the cell-to-cell flows for (i, j) ∈ A and wk < ck on the external outflows for k ∈ S, by
allowing the considered convex costs to be extended-real-valued with ψij (y) = +∞ if and only if y ≥ cij
and ψk (y) = +∞ if and only if y ≥ ck . We do not include details on such extention here, as we will deal
with finite capacities bounds zi ≤ Ci on the total outflows from the different cells within a different setting
in the next sections.
5. Nonlinear dynamical flow networks with demand and supply constraints
Whilst linear relations between the flow variables F and w and the state x studied in Section 3 fail to
capture congestion effects in dynamical flow networks, such effects can be modeled by nonlinearities. A
class of such nonlinearities has already been considered in Example 1 of Section 4 within the framework
of dual ascent algorithms in network flow optimization. In this section, we introduce another class of such
nonlinear dynamical flow networks whereby the outflows from, and the inflows in, the cells are bounded
by demand and, respectively, supply functions of their current state. Such class of models is particularly
relevant for road traffic flow networks (where the demand and supply functions represent, respectively, the
rising and decreasing parts of the fundamental diagram [28, 29]), as well as production networks (where
demand functions represent clearing functions [30]), and data networks (where demand functions can be
understood as representing communication link throughput). We present five examples of these models,
show that they all fit within the framework of monotone dynamical flow networks either globally, or in a
relevant subdomain of their state space, and finally use Theorem 6 to establish their stability properties.
We start by introducing demand functions ϕi (xi ) that characterize the maximum possible outflow zi
from the cells i ∈ I given the current mass xi , so that 0 ≤ zi ≤ ϕi (xi ). Throughout this paper, such demand
functions ϕi (xi ) are assumed to be Lipschitz-continuous, non-decreasing (the more mass, the more demand
of outflow from the cell), concave (the more mass the less outflow increment), with ϕi (0) = 0 (no outflow
from cell i), and such that the cell flow capacity
Ci := sup ϕi (xi )
xi ≥0
10
is strictly positive and finite for every cell i (see the leftmost plot in Figure 3). We stack up the demand
functions in an n-dimensional vector ϕ(x) = (ϕi (xi ))i∈I and compactly write the demand constraints as
0 ≤ z ≤ ϕ(x) .
(14)
We may equivalently rewrite the demand constraint (14) by introducing a vector γ ∈ Rn whose entries
γi ∈ [0, 1] for i ∈ I are flow control parameters, and put
z = Γϕ(x) ,
Γ = diag (γ) .
With this notation, equation (4) for nonlinear dynamical flow networks with demand constraints reads
ẋ = u − (I − RT )Γϕ(x) .
(15)
In the above, the external inflow vector u, the substochastic routing matrix R, and the flow control parameters matrix Γ might be either constant or time-varying, in the latter case depending on either exogenous
factors or feedback information. Recall that the routing matrix isPalways substochastic and needs to satisfy
the network topology constraints Rij = 0 for all (i, j) ∈
/ A and j∈I Rij = 1 for all cells i ∈
/ S. Observe
that invariance of the nonnegative orthant for the dynamics (15) is implied by the demand constraint (14)
and the fact that ϕi (0) = 0 so that there can be no outflow from an empty cell. We gather below a few
examples discussed in the literature that fit in this framework.
Example 2. The simplest instance of a nonlinear dynamical flow network with demand constraints is in
the form
ẋ = u − (I − RT )ϕ(x) ,
(16)
with constant external inflow vector u ∈ Rn+ and constant routing matrix R ∈ Rn×n
+ . Here, the demand
constraints (14) are met with equality, i.e., z = ϕ(x), so that Γ = I. The network topology is T =
(I, A, R, S), where
n
o
X
I = {1, . . . , n} ,
A = {(i, j) : Rij > 0} ,
R = {i ∈ I : ui > 0} ,
S= i∈I:
Rij < 1 .
j
Provided that T is outflow-connected, it can be easily verified that a necessary and sufficient condition for
the existence of an equilibrium of the dynamics (16) is that the vector
z ∗ = (I − RT )−1 u
satisfies
zi∗ < Ci ,
i∈I.
(17)
In fact, if this condition is met, and the demand functions are strictly increasing, then there exists a unique
equilibrium
x∗ = ϕ−1 (z ∗ ) .
(18)
On the other hand, it is not hard to verify that (16) is a monotone dynamical flow network, since
∂
Fki (x) = Rki ϕ0k (xk ) ≥ 0 ,
∂xk
∂
wi (x) = 0 ,
∂xj
∂
Fik (x) = 0 ,
∂xj
i, j, k ∈ I ,
X
∂
wi (x) = 1 −
Rik ϕ0i (xi ) ≥ 0 ,
k
∂xi
i, j ∈ I ,
j 6= i
j 6= i .
Therefore, when (17) holds true, the equilibrium (18) is stable by point (iii) of Theorem 6, while its global
asymptotic stability then follows from points (iv) and (v) of Theorem 6.
11
Example 3. In the works [31, 32], nonlinear dynamical flow networks are considered of the form
ẋ = u − (I − RT (x))ϕ(x) ,
(19)
with constant external inflow vector u and increasing demand functions ϕ(x). As in (16) of Example 2, the
demand constraints met with equality, i.e., z = ϕ(x). The main difference with respect to (16) is the locally
responsive feedback routing matrix R(x). By locally responsive feedback it is meant that:
(a) the split ratios Rij (x) depend on local information only, consisting of the mass xk in the cells k ∈ Ei ,
where Ei is the outneighborhood of cell i in the network topology;
(b) the dependence on such local information is monotone in the sense that
∂ X
Rih (x) ≤ 0 ,
∂xk
h∈Ei
∂
Rij (x) ≥ 0 ,
∂xk
i∈I,
j, k ∈ Ei ,
j 6= k ;
(20)
(c) if xj → +∞ for some but not all j ∈ Ei , or if i ∈ S and xj → +∞ for all j ∈ Ei , then
lim Rij (x) = 0 .
(21)
Property (a) above guarantees that the routing policy R(x) is decentralized, in the sense that it depends on
local information only. This in particular implies that the network flow dynamics (19) is distributed as the
righthand side of
X
ẋi = ui +
Rji (x)ϕj (xj ) − ϕi (xi )
j
depends only on the external inflow ui and the states of cell i, of its immediately upstream cells (i.e., such
j such that i ∈ Ej ), as well as their immediately downstream cells (i.e., those cells k such that i, k ∈ Ej
for the same j). Property (b) is a relatively natural monotonicity condition: it states that, if the mass in
an out-neighbor cell k ∈ Ei increases, then the fraction Rij (x) of outflow from cell i that is routed towards
anyP
other out-neighbor cell j ∈ Ei \ {k} cannot decrease and, if i ∈ S, the same holds true for the fraction
1− h∈Ei Rih (x) that is routed towards the external environment directly: as a consequence, this implies that
the fraction routed towards cell k itself cannot increase. Finally, property (c) ensures that, if the mass xj in
a cell j ∈ Ei grows large, and either i ∈ S (so that flow can be routed directly to the external environment)
or there are other cells k ∈ Ei where the mass stays bounded, then the fraction Rij (x) of flow routed from i
to j vanishes. Specific forms of such locally responsive feedback routing policies include the i-logit function
Rij (x) = X
eαj −βj xj
eαk −βk xk + χS (i)
,
j ∈ Ei ,
(22)
k∈Ei
where α, β ∈ Rn are vectors such that β ≥ 0, and χS is the indicator function of S in I, i.e., χS (i) = 1 if
i ∈ S and χS (i) = 0 if i ∈
/ S. As shown in [31], the dynamical flow network (19) with locally responsive
feedback routing R(x) is globally monotone on Rn+ . Points (iv) and (ii) of Theorem 6 then imply a dichotomy:
either the dynamical flow network (19) admits a stable equilibrium x∗ (that is the limit of the solution started
in x(0) = 0) or it is unstable, i.e., all its solutions grow unbounded in time. Globally asymptotic stability of
x∗ can then be established in the latter case (c.f. [31]).
Example 4. In the work [21], routing policies such as those in Example 3 are coupled with flow control
policies whose role consists in reducing the flow towards downstream cells when these are more congested than
the upstream ones. Specifically, locally responsive feedback routing policies R(x) satisfying properties (a),
(b), and (c) of Example 3 are considered together with flow controls Γ(x) = diag (γ(x)) with the following
properties:
(d) each γi (x) depends on local information only, consisting of the mass xk in the cells k ∈ Ei ∪ {i};
12
(e) the dependence on such local information is monotone in the sense that
∂
γi (x) ≤ 0 ,
∂xj
∂
(γi (x)Rij (x)) ≥ 0 ,
∂xk
j, k ∈ Ei ,
j 6= k;
(f ) if xj → +∞ for all j ∈ Ei and xi remains bounded, then lim γi (x) = 0;
(g) if xi → +∞ and xj remains bounded for some j ∈ Ei , then lim γi (x) = 1.
Specific forms include the i-logit routing policies (22) coupled with the locally responsive flow controls
X
eαk −βk xk + χS (i)
γi (x) =
k∈Ei
e
αi −βi xi
+
X
eαk −βk xk + χS (i)
.
(23)
k∈Ei
As shown in [21], conditions (ii) and (v), along with the fact that the demand functions ϕ(x) are nondecreasing imply that (15) is a monotone dynamical flow network. Using arguments analogous to those summarized
in Theorem 6, a dichotomy similar to that mentioned in Example 3 is then proved in [21]: either all solutions
of (15) grow unbounded in time, or they all converge to a globally asymptotically stable equilibrium.
Dynamical flow network models with demand constraints on the outflows such as the ones listed above
do capture some congestion effects, however they do not account for spillbacks and upstream propagation of
perturbations. These can be brought
P into the model through the introduction of supply functions σi (xi ) that
limit the maximum inflow ui + j fji possible in cells i ∈ I given their current mass xi . Throughout the
paper, the supply functions σi (xi ) are assumed to be Lipschitz continuous, non-increasing (the more mass,
the less inflow can be accommodated in the cell), and concave (the more mass the more inflow decrement).
We also denote by
xi = sup{xi ≥ 0 : σi (xi ) > 0}
the (possibly infinite) buffer capacity of cell i. (See the rightmost plot in Figure 3.) Upon stacking up the
supply functions in an n-dimensional vector σ(x) = (σi (xi ))i∈I the supply constraints can be written in the
compact form
u + RT z ≤ σ(x) .
(24)
Dynamical flow network models with both demand and supply constraints were introduced by Daganzo
[7, 8], as discretizations of the Lighthill-Whitham-Richards PDE model of road traffic flow [28, 29]. Observe
that the supply constraint (24) together with the demand constraint (14) imply that the set
X = {x ∈ Rn : 0 ≤ xi ≤ xi , ∀i ∈ I}
is invariant for the network flow dynamics (4).
Example 5. In the continuous-time version of Daganzo’s cell transmission model studied, e.g., in [33, 34,
35, 36, 37], the network flow dynamics read
ẋ = u − (I − RT )Γ(x)ϕ(x) ,
(25)
where the inflow vector u and the routing matrix R are exogenous and possibly time-varying, while Γ(x) =
diag (γi (x)) with the flow control parameters
(
!
)
X
γi (x) = sup α ∈ [0, 1] : max α
Rhk ϕh (xh ) − σk (xk ) ≤ 0 ,
i∈I.
(26)
k∈Ei
h∈I
Such flow control parameters ensure that both supply and demand constraints are met and that a first-in
first-out (FIFO) rule is followed at junctions so that, in particular, the outflow from any cell i is split among
13
its immediately downstream cells j ∈ Ei exactly according to the split ratios Rij in every circumstances.
Observe that in the free-flow region
F := {x ∈ Rn+ : u + RT ϕ(x) ≤ σ(x)} ,
(27)
one has that γi (x) = 1 for every cell i, so that the dynamics (25) coincide with (16) of Example 2. This
is not the case outside the free-flow region, where the supply constraints (24) make γi (x) < 1 for at least
one cell i. In fact, observe that the cell-transmission model (25) with FIFO diverge rule (26) is a monotone
dynamical flow network in a domain D that includes the free-flow region F, where its dynamics coincide
with the ones of Example 2, while it is not necessarily monotone on the whole state space Rn+ , except for
special network topologies (c.f. [36]). Provided that sufficient conditions for the existence of an equilibrium
x∗ in the free-flow region analogous to those in Example 2 are satisfied, local asymptotic stability can be
established using point (v) of Theorem 6 (c.f., [35]). For special network topologies for which (25)-(26) is
globally monotone on Rn+ (essentially, networks with no diverge junctions) global asymptotic stability can
be deduced from point (iv) of Theorem 6. We refer the interested reader to [36] for stability results of the
cell-transmission model with FIFO diverge rule on a certain class of network topologies (polytree networks)
where the resulting dynamics are not generally globally monotone.
Example 6. A variant of Daganzo’s cell transmission model with non-FIFO diverge rule was considered,
e.g., in [22, 38]. Here, the dynamics is more easily expressed in the form
ẋ = u + F T (x)1 − F (x)1 − w(x) ,
where, for a given exogenous routing matrix R, the flows are given by
)
(
X
Rhj ϕh (xh ) ≤ σj (xj ) ,
Fij (x) = γij (x)ϕi (xi ) ,
γij (x) = sup ξ ∈ [0, 1] : ξ ·
(28)
(29)
h∈I
P
and wi (x) = (1− j Rij )ϕi (xi ). Observe that, in the free-flow region (27), the dynamics above coincide with
the ones of Examples 2 and 5, while this is not the case elsewhere. In fact, it can be proved that (28)-(29) is
a monotone dynamical flow network on the whole state space: from this, under certain sufficient conditions,
global stability results can be deduced, as illustrated in [22].
Interestingly, it has been pointed out that, while not necessarily monotone outside the free-flow region,
the cell transmission model with FIFO diverge rule (25)-(26) of Example 5 enjoys a mixed monotonicity
property, i.e., the entries of its Jacobian have signs that are constant throughout the state space. In fact,
such mixed monotonicity property carries over to versions of the cell-transmission model that combine FIFO
with non-FIFO diverge rules as in Example 1. (C.f. [36, 39].)
6. Robustness of dynamical flow networks and their margin of resilience
In the series of papers [31, 32, 21, 40], robustness of nonlinear dynamical flow networks with decentralized
routing and flow controls is analyzed. The main idea consists in studying the impact that perturbations of
the external inflows and of the cell flow capacities have on the dynamic behavior of the flow network. In
order to quantitatively measure such robustness, the notion of margin of resilience is introduced. This is
defined as the smallest magnitude of a perturbation that drives the dynamical flow network to instability,
such magnitude being measured as the sum of the aggregate loss of cell flow capacities and the aggregate
increase of external inflows. In this section, we briefly review some versions of these results.
In the interest of clarity, we will consider dynamical flow networks with constant external inflows and
demand constraints only, while referring to the works [32, 21, 40] for some results with finite buffer capacities
that can be interpreted as special cases of supply constraints, as well as some discussion of extensions to
time-varying external inflows. We start with a dynamical flow network
ẋ = u − (I − RT (x))Γ(x)ϕ(x) ,
14
(30)
Ci
ϕi (xi )
C̃i
ϕ̃i (xi )
xi
0
Figure 4: Considered perturbation affect the external inflow vector and the demand functions. Above a nominal demand
function ϕi (xi ) in a cell i is plotted along with the perturbed demand function ϕ̃i (xi ) in the same cell. The magnitude of a
perturbation is measured as in (31).
with topology T = (I, A, R, S), constant external inflows u, demand functions vector ϕ(x), routing matrix
R(x), and flow control Γ(x) = diag (γ(x)). We then consider both perturbations ũ of the inflow vector u
and perturbations ϕ̃(x) of the demand functions ϕ(x) (see Figure 4) and measure the magnitude of such
perturbations as
X
X
δ :=
|ũi − ui | +
||ϕi ( · ) − ϕ̃i ( · )||∞ .
(31)
i∈R
i∈I
We then propose the following quantitative measure of robustness for dynamical flow networks.
Definition 7. The margin of resilience ν of a dynamical flow network (30) with topology T = (I, A, R, S),
constant external inflows u, demand functions ϕ(x), routing matrix R(x), and flow control Γ(x), is the
infimum of the magnitudes δ of all perturbations (ũ, ϕ̃(x)) of (u, ϕ(x)) that make the perturbed dynamical
flow network
ẋ = ũ + (I − RT (x))Γ(x)ϕ̃(x)
(32)
unstable, i.e., such that the state x(t) is unbounded in time for at least one initial condition x(0) ∈ Rn+ .
A few comments are in order. First, Definition 7 captures a worst case notion of robustness. In fact,
one might think of the margin of resilience ν as the minimum effort required by a hypothetical adversary
in order to modify the external inflows and demand functions so as to make the dynamical flow network
unstable, provided that such effort is measured in terms of the perturbation magnitude (31).
Second, notice from the righthand side of (32) that the (feedback) routing R(x) and flow control Γ(x)
are required to stabilize the system for any perturbation (ũ, ϕ̃(x)) affecting the external inflows and demand
functions of magnitude smaller than the margin of resilience, without knowing the perturbation itself, but
simply by measuring its effect on the state x. This is in line with classical robust control formulations. On
the other hand, the considered notion of stability requires the state to remain bounded in time for every
initial condition x(0), not necessarily to converge to a pre-specified equilibrium. In fact, as we have seen
in Section 5, most of the considered routing and flow controls are such that the dynamical flow network is
monotone and, when stable, it has a globally asymptotically stable equilibrium. Hence, in these cases, when
the perturbed dynamical flow network remains stable, it admits a globally asymptotically stable equilibrium:
the specific value of such equilibrium x̃∗ depends on the perturbation itself as well as on the routing and
flow controls, but it does not affect the notion of margin of resilience.
Observe that the margin of resilience depends on several factors: the network topology T , the unperturbed demand functions ϕ(x) (hence, in particular, the maximum flow capacities C), the external inflows u,
as well as the routing R(x) and flow control Γ(x). In fact, an upper bound on such margin of resilience can
be obtained that depends only on the network topology, the maximum flow capacities, and the unperturbed
15
KJ
k
J
S
R
Figure 5: The residual capacity of a cut J ⊆ I is the aggregate capacity of the cells j ∈ J minus the total external inflow
in the cells k ∈ KJ where KJ is the set of cells that either belong to J or are no longer outflow-connected if the cells in J
are removed. In the network topology displayed above, J consists of two cells (colored in light grey), while KJ consists of five
cells. In this case, since KJ ∩ R = {k}, where k is the cell colored in dark grey, the total external inflow in KJ corresponds to
the inflow uk in cell k.
external inflows: such upper bound holds true for every choice of routing and flow control policies. In order
to state it, some further graph-theoretic notation is needed. For a subset of cells J ⊆ I, let KJ ⊆ I be
the set of cells that either belong to J or are no longer outflow-connected in the network topology obtained
from T by removing the cells in J . In other terms, the set KJ includes J as well as all those cells that
are connected to the set S only though paths passing through J . (See Figure 5.) Then, define the min-cut
residual capacity as
nX
o
X
Cj −
uk ∨ 0 .
(33)
Cmin-cut = min
J ⊆I
j∈J
k∈KJ
As stated in the following theorem, the min-cut residual capacity provides an upper bound
ν ≤ Cmin-cut
(34)
on the margin of resilience of any dynamical flow network with demand constraints (30) with arbitrary
routing and flow control policies. The inequality (34) can be proved using a mass conservation argument
that follows from the inherent structure of dynamical flow networks (1) and can be thought of as the
analogue of a classical result for static network flows. In fact, for such static network flows, tightness
of the min-cut bound is guaranteed by the celebrated max-flow min-cut theorem [41, 42], which can be
interpreted as a linear-programming duality result. For dynamical flow networks, the dynamic constraints
pose additional challenges to the achievement of the maximum margin of resilience, i.e., tightness of the
upper bound (34). The following theorem states that such maximum margin of resilience is achieved on
arbitrary network topologies by the combined locally responsive feedback routing and flow control policies
introduced in Example 4. In contrast, it is proved that not only the fixed routing policies of Example 2
but also the locally responsive feedback routing without flow control of Example 3 fall short of achieving
maximum margin of resilience on arbitrary network topologies.
Theorem 8. Let T = (I, A, R, S) be an outflow-connected network topology, u a constant external inflow
vector supported on S, and ϕ(x) a vector of strictly increasing demand functions. Consider a dynamical flow
network (30) with routing policy R(x) and flow control Γ(x), and let ν be its margin of resilience. Then,
(i) the upper bound (34) holds true.
Moreover:
(ii) for locally responsive routing R(x) and flow control Γ(x) as in Example 4,
ν = Cmin-cut ,
i.e., the margin of resilience coincides with the min-cut residual capacity.
16
On the other hand, if the network topology T is such that G = (I, A) is the line-digraph of an acyclic digraph
N = (V, I) (as per Remark 1), then
(iii) if the routing matrix R is constant and there is no flow control, i.e., Γ = I, then the margin of resilience
coincides with the minimum cell residual capacity
ν = min{(Ci − zi∗ ), ∨0} ,
i∈I
z ∗ = (I − RT )−1 u ;
(35)
(iv) if R(x) is a locally responsive feedback routing policy with no flow control as in Example 3, the margin
of resilience satisfies
X
ν = min
(Cj − zj∗ ) ,
(36)
i∈I
j∈Ei
where z ∗ = ϕ(x∗ ) with x∗ being the equilibrium if the unperturbed network is globally asymptotically
stable, and z ∗ = C otherwise.
Proof Part (i) can be proved as in [31]. Part (ii) follows from [21, Theorem 1(i)]. Part (iv) is proved in
[32], while part (iii) can be proved with a similar (and in fact simpler) argument.
While Part (i) of Theorem 8 captures the physically intuitive fact that a dynamical flow network cannot
be stable if the external inflow violates some cut capacity constraint, Part (ii) states quite a remarkable
result in that a global objective (maximum margin of resilience) is achievable by completely decentralized
routing and flow control policies that use no global knowledge whatsoever of the network state or structure.
That this is nontrivial is confirmed by Parts (iii) and (iv) of Theorem 8 stating that, if there is no flow
control and the routing is either fixed or locally responsive, then the margin of resilience coincides with,
respectively, the minimal cell residual capacity (35), and the minimum out-neighborhood resilience capacity
(36), quantities that have been proved to be arbitrarily far away from the min-cut residual capacity (33):
see [31, 32]. As explained in [21], the reason why the combination of locally responsive feedback flow control
and routing is able to achieve maximal resilience resides in the ability to properly propagate congestion
—and hence, implicitly, information— in a controlled way, both upstream and downstream, thanks to the
flow control, and to reroute flow towards less congested paths (thanks to the routing). In fact, Part (ii)
of Theorem 8 should be compared with results available for packet-switched data networks, e.g., maximal
throughput and stability of the back-pressure algorithm [43]. In contrast, locally responsive routing by itself,
as in Example 3 is not sufficient to achieve maximal margin of resilience since it is not able to propagate
information upstream but just to reroute flow in a locally optimal way, whereas fixed routing as in Example
2 does even worse because it fails to reroute flow towards less congested paths.
7. Conclusion
In this paper, we have surveyed stability and robustness properties of dynamical flow networks. These
are a family of dynamical systems derived from mass conservation laws on directed graphs with exogenous
inflows and outflows towards the external environment. After reviewing some structural stability properties
of linear dynamical flow networks, we have focused on an important class of nonlinear dynamical flow
networks characterized by an additional monotonicity property. While rich enough to encompass many
examples of dynamical flow networks studied in the literature, including some well-established models for
road traffic networks, the class of monotone dynamical flow networks has enough structure to allow for
tractable analysis and scalable design. In particular, we have presented a non-expansiveness result in the
l1 -metric for solutions of mononotone dynamical flow networks and applied it to their (differential) stability
analysis. We have then focused on dynamical flow networks with cell demand constraints and introduced
the margin of resilience as a quantitative measure of their robustness. We have shown that maximal margin
of resilience on a given capacitated network flow topology coincides with a well-known graph-theoretical
quantity, the min-cut residual capacity, and that it can be achieved by a combination of decentralized
17
locally responsive feedback flow control and routing policies that use local information only and require no
global knowledge of the network.
It is worth emphasizing that the treatment in this paper has been focused on single-commodity firstorder models of dynamical flow networks and specifically on their stability and robustness properties. Other
aspects of dynamical flow networks of current interest that have not been discussed here include: optimal
dynamical flow network control (see, e.g., [37] for convex formulations with demand and supply constraints
generalizing ideas in [44, 45]); multi-scale models coupling dynamical flow networks with game-theoretic
learning dynamics [46] modeling the selfish route choice behaviors of the drivers; extensions of the margin
of resilience framework to multi-commodity dynamical flow networks for which only preliminary results are
currently available [47].
Acknowledgements
The author is grateful to Professor Ketan Savla of the University of Southern California for a longstanding collaboration on the topics surveyed in this paper. He would also like to thank his other coauthors
Enrico Lovisari, Munther A. Dahleh, Daron Acemoglu, Emilio Frazzoli, and Anders Rantzer. Finally,
the author wishes to thank Professor Tryphon Georgiou for inviting him to present these results at the
22nd International Symposium on Mathematical Theory of Networks and Systems, and Professor Francoise
Lamnabhi-Lagarrigue for soliciting a contribution in this venue.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
Disaster Resilience: a National Imperative, The National Academies Press, 2012.
R. Ahuja, T. Magnanti, J. Orlin, Network Flows: Theory, Algorithms, and Applications, Prentice Hall, 1993.
D. Bertsekas, Network optimization: Continuous and discrete models, Athena Scientific, 1998.
P. Whittle, Networks: Optimisation and evolution, Cambridge University Press, 2007.
J. A. Jacquez, C. P. Simon, Qualitative theory of compartmental systems, SIAM Review 35 (1) (1993) 43–79.
G. G. Walter, M. Contreras, Compartmental modeling with networks, Birkhäuser, 1999.
C. Daganzo, The cell transmission model: A dynamic representation of highway traffic consistent with the hydrodynamic
theory, Transportation Research B 28B (4) (1994) 269–287.
C. Daganzo, The cell transmission model, Part II: Network traffic, Transportation Research B 29B (2) (1995) 79–93.
A. Berman, R. Plemmons, Nonnegative matrices in the mathematical sciences, SIAM, 1994.
J. B. Orlin, Line-digraphs, arborescences, and theorems of tutte and knuth, Journal of Combinatorial Theory, Series B 25
(1978) 187–198.
L. Farina, S. Rinaldi, Positive linear systems: Theory and applications, Wiley, 2000.
A. Rantzer, Scalable control of positive systems, European Journal of Control 24 (2015) 72–80.
S. Coogan,
Separability of Lyapunov functions for contractive monotone systems,
available at
https://arxiv.org/abs/1609.06258 (2016).
M. W. Hirsch, Differential equations and convergence almost everywhere in strongly monotone semiflows, Contemporary
Mathematics 17 (1983) 267–285.
H. L. Smith, Monotone dynamical systems: An introduction to the theory of competitive and cooperative systems,
American Mathematical Society, 1995.
M. Hirsch, H. Smith, Monotone dynamical systems, Vol. 2 of Handbook of differential equations: Ordinary differential
equations, Chapter 4, North-Holland, 2006, pp. 239 – 357.
E. Kamke, Zur theorie der systeme gewoknlicher differentialgliechungen, II, Acta Mathematica 58 (1932) 57–85.
D. Angeli, E. Sontag, Monotone control system, IEEE Transactions on Automatic Control 48 (10) (2003) 1684 – 1698.
M. Hirsch, Systems of differential equations that are competitive or cooperative II: Convergence almost everywhere, SIAM
Journal on Mathematical Analysis 16 (3) (1985) 423–439.
M. Hirsch, H. Smith, Competitive and cooperative systems: A mini-review, Positive systems. Lecture notes in control and
information sciences 294 (2004) 183–190.
G. Como, E. Lovisari, K. Savla, Throughput optimality and overload behavior of dynamical flow networks under monotone
distributed routing, IEEE Transactions on Control of Network Systems 2 (1) (2015) 57–67.
E. Lovisari, G. Como, K. Savla, Stability of monotone dynamical flow networks, in: Proceedings of 53rd Conference on
Decision and Control, CDC’14, Los Angeles, CA, 2014.
G. Dirr, H. Ito, A. Rantzer, B. Rüffer, Separable Lyapunov functions for monotone systems: Constructions and limitations,
Discrete and Continuous Dynamical Systems - Series B 20 (8) (2015) 2497–2526.
W. Lohmiller, J.-J. Slotine, On contraction analysis for nonlinear systems, Automatica 34 (6) (1998) 683–696.
E. D. Sontag, Contractive systems with inputs, in: Perspectives in Mathematical System Theory, Control, and Signal
Processing, Springer, 2010, pp. 217–228.
18
[26] D. Serre, Systems of conservation Laws 1. Hyperbolicity, entropies, shock waves, Cambridge University Press, 1999.
[27] A. Zlotnik, S. Misra, M. Vuffray, M. Chertkov, Monotonicity of actuated flows on dissipative transport networks, available
at arXiv:1511.07585 (2015).
[28] M. Lighthill, G. Whitham, On kinematic waves. ii. a theory of traffic flow on long crowded roads, Philosophical Transactions
of the Royal Society A 229 (1178) (1955) 317–345.
[29] I. Richards, Shockwaves on the highway, Operations Research 4 (1956) 42–51.
[30] U. S. Karmarkar, Capacity loading and release planning with work-in-progress (WIP) and leadtimes, Journal of Manufacturing and Operations Management 2 (1989) 105–123.
[31] G. Como, K. Savla, D. Acemoglu, M. A. Dahleh, E. Frazzoli, Robust distributed routing in dynamical networks - Part I:
Locally responsive policies and weak resilience, IEEE Transactions on Automatic Control 58 (2) (2013) 317–332.
[32] G. Como, K. Savla, D. Acemoglu, M. A. Dahleh, E. Frazzoli, Robust distributed routing in dynamical networks - Part
II: Strong resilience, equilibrium selection and cascaded failures, IEEE Transactions on Automatic Control 58 (2) (2013)
333–348.
[33] S. E. Jabari, H. X. Liu, A stochastic model of traffic flow: Theoretical foundations, Transportation Research Part B:
Methodological 46 (1) (2012) 156–174.
[34] P. Grandinetti, C. C. de Wit, F. Garin, An efficient one-step-ahead optimal control for urban signalized traffic networks
based on an averaged cell-transmission model, in: Proceedings of the European Control Conference (ECC), 2015, pp.
3478–3483.
[35] S. Coogan, M. Arcak, A compartmental model for traffic networks and its dynamical behavior, IEEE Transactions on
Automatic Control 60 (10) (2015) 2698–2703.
[36] S. Coogan, M. Arcak, Stability of traffic flow networks with a polytree topology, Automatica 66 (2016) 246–253.
[37] G. Como, E. Lovisari, K. Savla, Convexity and robustness of dynamic traffic assignment and freeway network control,
Transportation Research Part B: Methodological 91 (2016) 446–465.
[38] I. Karafyllis, M. Papageorgiou, Global exponential stability for discrete-time networks with applications to traffic networks,
IEEE Transactions on Control of Network Systems 2 (1) (2015) 68–77.
[39] S. Coogan, M. Arcak, A. A. Kurzhanskyi, Mixed monotonicity of partial first-in-first-out traffic flow models, available at
http://arxiv.org/abs/1511.05081.
[40] K. Savla, G. Como, M. A. Dahleh, Robust network routing under cascading failures, IEEE Transactions on Network
Science and Engineering 1 (1) (2014) 53–66.
[41] L. R. Ford, D. R. Fulkerson, Maximal flow through a network, Canadian Journal of Mathematics 8 (1956) 399–404.
[42] P. Elias, A. Feinstein, C. Shannon, A note on the maximum flow through a network, IRE Transactions on Information
Theory 2 (4) (1956) 117–119.
[43] L. Tassiulas, A. Ephremides, Stability properties of constrained queueing systems and scheduling policies for maximum
throughput in multihop radio networks, IEEE Transactions on Automatic Control 37 (12).
[44] A. Ziliaskopoulos, A linear programming model for the single destination system optimum dynamic traffic assignment
problem, Transportation Science 34 (1) (2000) 37–49.
[45] G. Gomes, R. Horowitz, Optimal freeway ramp metering using the asymmetric cell transmission model, Transportation
Research Part C 14 (4) (2006) 244–268.
[46] G. Como, K. Savla, D. Acemoglu, M. A. Dahleh, E. Frazzoli, Stability analysis of transportation networks with multiscale
driver decisions, SIAM Journal on Control and Optimization 51 (1) (2013) 230–252.
[47] G. Nilsson, G. Como, E. Lovisari, On resilience of multicommodity dynamical flow networks, in: Proceedings of 53rd
Conference on Decision and Control, CDC’14, 2014.
19
| 3cs.SY
|
GORENSTEIN INJECTIVE ENVELOPES OF ARTINIAN MODULES
MASSOUMEH NIKKHAH BABAEI AND KAMRAN DIVAANI-AAZAR
Abstract. Let R be a commutative Noetherian ring and A an Artinian R-module. We prove that if A has
arXiv:1306.4452v1 [math.AC] 19 Jun 2013
finite Gorenstein injective dimension, then A possesses a Gorenstein injective envelope which is special and
Artinian. This, in particular, yields that over a Gorenstein ring any Artinian module possesses a Gorenstein
injective envelope which is special and Artinian.
1. Introduction and Prerequisites
Throughout this paper, R is a commutative Noetherian ring with an identity. It is a standard fact that
any R-module M possesses an injective envelope ER (M ). Contrary to that it is not known yet that any
R-module has a Gorenstein injective envelope. For an R-module M , let GR (M ) denote the Gorenstein
injective envelope of M if it exists. It is known that if A is an Artinian R-module, then ER (A) is also
Artinian. There is much evidence supporting Holm’s claim “Every result in classical homological algebra
should have a counterpart in Gorenstein homological algebra”; see [H]. Hence, the following question arises:
Does any Artinian R-module A admit an Artinian Gorenstein injective envelope? In this paper, we are
concerned with this question. Belshoff and Enochs [BE, Theorem 1] have given a partial positive answer
to this question. They proved that if R is a Gorenstein local ring with the residue field k, then GR (k)
is Artinian. We show that if A is an Artinian R-module with finite Gorenstein injective dimension, then
GR (A) exists, it is Artinian and idR (GR (A)/A) < ∞. In particular, we conclude that if R is Gorenstein,
then any Artinian R-module admits an Artinian Gorenstein injective envelope. The following are two of the
remaining open questions in the theory of Gorenstein dimensions; see [CFoH, Questions 3.17 and 3.19]:
Is every direct limit of Gorenstein injective R-modules Gorenstein injective?
Is Mp Gorenstein injective over Rp for all Gorenstein injective R-modules M and all prime ideals p of R?
Assuming a positive answer to the first question, we provide an affirmative answer to the second question.
Below, we recall some definitions and notations.
An R-module M is said to be Gorenstein injective if there exists an exact sequence
I • : · · · −→ I1 −→ I0 −→ I 0 −→ I 1 −→ · · ·
of injective R-modules such that M ∼
= im(I0 −→ I 0 ) and that HomR (I, I • ) is exact for all injective Rmodules I. The notion of Gorenstein injective dimension of an R-module M , Gid M , is defined as the
infimum of the length of right resolutions of M which are consisting of Gorenstein injective R-modules.
Let X be a class of R-modules and M an R-module. An R-homomorphism φ : M −→ X where X ∈ X
is called a X -preenvelope of M if for any X ′ ∈ X , the induced R-homomorphism HomR (X, X ′ ) −→
2010 Mathematics Subject Classification. 13D05; 13E10.
Key words and phrases. Gorenstein injective dimension; Gorenstein injective envelopes; Gorenstein injective modules; Gorenstein projective covers; Gorenstein projective modules; injective covers; Henselian rings.
The second author was supported by a grant from IPM (No. 91130212).
1
2
M. NIKKHAH BABAEI AND K. DIVAANI-AAZAR
HomR (M, X ′ ) is surjective. If, moreover, any R-homomorphism f : X −→ X such that f φ = φ is an
automorphism, then φ is called a X -envelope of M . From the definition, it becomes clear that if X envelopes exist, then they are unique up to isomorphism. A X -(pre)envelope φ : M −→ X is called special
if ExtiR (Coker φ, X ′ ) = 0 for all X ′ ∈ X and all i > 0.
If X is the class of Gorenstein injective R-modules, then a X -(pre)envelope is called Gorenstein injective
(pre)envelope. By [EL, Corollary 2.7], over a commutative Noetherian ring, any module has a Gorenstein
injective preenvelope. In fact by [K, Theorem 7.12] any such a module has a special Gorenstein injective
preenvelope. But, it is not known whether any module over an arbitrary commutative Noetherian ring has
a Gorenstein injective envelope or not. Since any injective R-module is Gorenstein injective, it follows that
any Gorenstein injective preenvelope is a monomorphism. If M admits a Gorenstein injective envelope, then
without loss of generality, we may assume that M has a Gorenstein injective envelope i : M ֒→ GR (M ),
where M is a submodule of the Gorenstein injective R-module GR (M ) and i is the inclusion map. For
simplicity, sometimes we call GR (M ) the Gorenstein injective envelope of M . If φ : M −→ E is a Gorenstein
injective (pre)envelope such that Coker φ has finite injective dimension, then using [CFrH, Lemma 2.2], we
can deduce that φ : M −→ E is a special Gorenstein injective (pre)envelope of M .
An R-module M is said to be Gorenstein projective if there exists an exact sequence
P• : · · · −→ P1 −→ P0 −→ P 0 −→ P 1 −→ · · ·
of projective R-modules such that M ∼
= im(P0 −→ P 0 ) and that HomR (P• , P ) is exact for all projective Rmodules P . The Gorenstein projective dimension of an R-module M , GpdR M , is defined as the infimum of
the length of left resolutions of M which are consisting of Gorenstein projective R-modules.
Let X be a class of R-modules and M an R-module. An R-homomorphism φ : X −→ M where
X ∈ X is called a X -precover of M if for any X ′ ∈ X , the induced R-homomorphism HomR (X ′ , X) −→
HomR (X ′ , M ) is surjective. If, moreover, any R-homomorphism f : X −→ X such that φf = φ is an
automorphism, then φ is called a X -cover of M . Obviously, if X -covers exist, then they are unique up to
isomorphism. A X -(pre)cover φ : X −→ M is called special if ExtiR (X ′ , Ker φ) = 0 for all X ′ ∈ X and
all i > 0. If X is the class of Gorenstein projective R-modules, then a X -(pre)cover is called Gorenstein
projective (pre)cover. As any projective R-module is Gorenstein projective, it follows that any Gorenstein
projective precover is an epimorphism. If X is the class of injective R-modules, then a X -(pre)cover is
called injective (pre)cover. By [EJ, Theorem 5.4.1 and Corollary 7.2.3], over a commutative Noetherian ring,
every R-module admits a special injective cover.
2. The Results
Theorem 2.7 is our main result. To prove it, we need to the following five lemmas. We start by the
following special case of [T2, Corollary 2.5] that can be easily deduced from [H, Corollary 2.13] and [T1,
Remark 2.6].
Lemma 2.1. Let (R, m) be a Henselian local ring and N a finitely generated R-module of finite Gorenstein
projective dimension. Then N possesses a special Gorenstein projective cover φ : V −→ N such that V is
finitely generated and Ker φ has finite projective dimension.
In what follows, the following will serve us as a technical tool.
GORENSTEIN INJECTIVE ENVELOPES ...
3
Lemma 2.2. Let f : R −→ T be a homomorphism of Noetherian commutative rings and M a T -module.
Assume that any Gorenstein injective T -module is also a Gorenstein injective R-module via f .
i) GidR M ≤ GidT M and equality holds if any T -module which is Gorenstein injective as an R-module
is also Gorenstein injective as a T -module.
ii) Assume f is flat. Let φ : M −→ E be a Gorenstein injective preenvelope of M as a T -module such
that idT (Coker φ) < ∞. Then φ : M −→ E is also a Gorenstein injective preenvelope of M as an
R-module and idR (Coker φ) < ∞.
Proof. i) Our assumption implies that any Gorenstein injective resolution of M as a T -module is also a
Gorenstein injective resolution of M as an R-module. Hence, GidR M ≤ GidT M .
Now, assume that any T -module which is Gorenstein injective as an R-module is also Gorenstein injective
as a T -module. We show that GidT M ≤ GidR M . We may assume that n := GidR M is finite. Let
E • : 0 −→ E 0 −→ E 1 −→ · · · −→ E i −→ · · ·
be a Gorenstein injective resolution of M as a T -module. Since E • is also a Gorenstein injective resolution of
M as an R-module, [CFrH, Theorem 3.3] yields that Ker(E n −→ E n+1 ) is a Gorenstein injective R-module.
Then, by the assumption Ker(E n −→ E n+1 ) is also Gorenstein injective as a T -module. Thus, we deduce
that
0 −→ E 0 −→ E 1 −→ · · · −→ E n−1 −→ Ker(E n −→ E n+1 ) −→ 0
is a Gorenstein injective resolution of M as a T -module, and so GidT M ≤ GidR M .
ii) By the assumption E is also a Gorenstein injective R-module. As any injective T -module is also
injective as an R-module, we deduce that idR (Coker φ) < ∞. Hence, [CFrH, Lemma 2.2] implies that
ExtiR (Coker φ, G) = 0 for all Gorenstein injective R-modules G and all i > 0. Therefore, for any Gorenstein
injective R-module G, from the exact sequence
φ
0 −→ M −→ E −→ Coker φ −→ 0,
we obtain the exact sequence
HomR (E, G) −→ HomR (M, G) −→ Ext1R (Coker φ, G) = 0.
This completes the argument.
Let (R, m) be a local ring. To be able to apply Lemma 2.2 ii) to the natural flat ring homomorphism
θ : R −→ R̂, we have to prove the following:
Lemma 2.3. Let f : (R, m) −→ (T, n) be a faithfully flat ring homomorphism of local rings. Then every
Gorenstein injective T -module is also a Gorenstein injective R-module via f .
Proof. Let M be a Gorenstein injective T -module. As pdT (T ⊗R T ) < ∞, [CFrH, Lemma 2.2] yields
that ExtiT (T ⊗R T, M ) = 0 for all i > 0. On the other hand, using adjointness, one can check that
ExtiR (T, M ) ∼
= ExtiT (T ⊗R T, M ) for all i ≥ 0. So, RHomR (T, M ) ≃ HomR (T, M ). Since, by [CH, Ascent
table I, (d)], HomR (T, M ) is a Gorenstein injective T -module, we deduce that GidT (RHomR (T, M )) = 0.
Now, [CW, Theorem 1.7] implies that GidR M = GidT (RHomR (T, M )), and so M is a Gorenstein injective
R-module.
4
M. NIKKHAH BABAEI AND K. DIVAANI-AAZAR
Let A be an Artinian module over a local ring (R, m). Sharp [Sh] has shown that A has a natural structure
as a module over R̂. Let θ : R −→ R̂ denote the natural ring homomorphism. The R̂-module structure of A
is such that for any element r ∈ R the multiplication by r on A has the same effect as the multiplication of
θ(r) ∈ R̂. Furthermore, a subset of A is an R-submodule of A if and only if it is a R̂-submodule of A.
Next, we prove our main result in the special case that R is local.
Lemma 2.4. Let (R, m) be a local ring and A an Artinian R-module of finite Gorenstein injective dimension.
Then A admits a Gorenstein injective envelope G(A) which is Artinian and idR (G(A)/A) < ∞.
Proof. First of all note that [Sa, Lemma 3.6] implies that GidR̂ A < ∞. On the other hand, for any Artinian
R̂-module B, it follows that B is also Artinian as an R-module and HomR (B, B) = HomR̂ (B, B). Hence, in
view of Lemma 2.3 and Lemma 2.2 ii), we may and do assume that R is complete.
Denote the exact functor HomR (−, ER (R/m)) by (−)∨ . Then N := A∨ is a finitely generated R-module
∼ A. By [CFoH, Theorem 4.16] and [CFrH, Proposition 3.8 (a)], one sees that Gpd N = GidR A <
and N ∨ =
R
∞. Hence, by Lemma 2.1, N possesses a Gorenstein projective cover φ : V −→ N such that V is finitely
generated and K := Ker φ has finite projective dimension. From the exact sequence
φ
0 −→ K ֒→ V −→ N −→ 0,
we get the exact sequence
λ
0 −→ A −→ V ∨ −→ K ∨ −→ 0.
(∗)
Applying [CFoH, Theorem 4.16] and [CFrH, Proposition 3.8 (a)] again implies that V ∨ is an Artinian
Gorenstein injective R-module. We show that λ is a Gorenstein injective envelope of A. As pdR K < ∞,
one obtains
idR (Coker λ) = idR K ∨ < ∞.
For any Gorenstein injective R-module G, [CFrH, Lemma 2.2] yields that ExtiR (K ∨ , G) = 0 for all i > 0,
and hence from (∗), we can deduce the exact sequence
HomR (V ∨ , G) −→ HomR (A, G) −→ Ext1R (K ∨ , G) = 0.
Thus λ : A −→ V ∨ is a Gorenstein injective preenvelope of A and idR (Coker λ) < ∞.
Let τ : V −→ (V ∨ )∨ denote the natural R-isomorphism. Let f : V ∨ −→ V ∨ be an R-homomorphism
such that f λ = λ. Then, we have φ(τ −1 f ∨ τ ) = φ, which yields that τ −1 f ∨ τ is an automorphism. Therefore,
f is an automorphism, and so V ∨ is a Gorenstein injective envelope of A.
The next is our last preliminary lemma for proving our main result.
Lemma 2.5. Let S be a multiplicatively closed subset of R and L a S −1 R-module. Then L is Gorenstein
injective as an R-module if and only if it is Gorenstein injective as a S −1 R-module. More generally, for any
S −1 R-module M , we have GidR M = GidS −1 R M .
Proof. First, assume that L is Gorenstein injective as an R-module. Let N be a S −1 R- modules and E • an
injective resolution of N . For any R-module M , one has the following natural isomorphisms of complexes:
HomR (M, E • ) ∼
= HomR (M, HomS −1 R (S −1 R, E • ))
∼
= HomS −1 R (M ⊗R S −1 R, E • )
∼
= HomS −1 R (S −1 M, E • ).
GORENSTEIN INJECTIVE ENVELOPES ...
5
By bringing to mind that any injective S −1 R-module is also injective as an R-module, it follows that E • is
also an injective resolution of N as an R-module, and so Exti (M, N ) ∼
= Exti −1 (S −1 M, N ) for all i ≥ 0.
R
S
R
In particular, if M is a S −1 R-module, then we have ExtiR (M, N ) ∼
= ExtiS −1 R (M, N ) for all i ≥ 0. In what
follows, we will use these isomorphisms without any further comment. Let
d−1
d0
0 −→ L −→ I 0 −→ I 1 −→ · · ·
be an augmented injective resolution of L as a S −1 R-module and I be an injective S −1 R-module. Let k ≥ 0
and j > 0 be two integers. From the exact sequence 0 −→ im dk−1 −→ I k −→ im dk −→ 0, we deduce the
j+1
S −1 R-isomorphisms Extj −1 (I, im dk ) ∼
= Ext −1 (I, im dk−1 ). Applying these isomorphisms successively for
S
R
S
R
k = i, i − 1, . . . , 0 imply that
j+i+1
ExtjS −1 R (I, im di ) ∼
= ExtS −1 R (I, L)
∼ Extj+i+1 (I, L)
=
R
= 0.
Recall by [CFrH, Theorem 3.3], a module E over a commutative Noetherian ring T is Gorenstein injective
b E) = 0 for all injective T -modules Ib and all i > 0. By [EJ, Theorem
if and only if GidT E < ∞ and ExtiT (I,
5.4.1 and Corollary 7.2.3], the S −1 R-module L admits a special injective cover φ0 : I0 −→ L. Since L
is a Gorenstein injective R-module, there are an injective R-module I˜ and an R-epimorphism I˜ −→ L.
Localizing at S yields a S −1 R-epimorphism S −1 I˜ −→ L. Hence, φ0 is an epimorphism. From the exact
φ0
sequence 0 −→ Ker φ0 ֒→ I0 −→ L −→ 0, we deduce that GidR (Ker φ0 ) < ∞. On the other hand, as φ0 is
special, one has
ExtiR (I ′ , Ker φ0 ) ∼
= ExtiS −1 R (S −1 I ′ , Ker φ0 ) = 0
for all injective R-modules I ′ and all i > 0. Thus, Ker φ0 is a Gorenstein injective R-module. By repeating
the above argument for Ker φ0 instead of L, we get a S −1 R-epimorphism φ1 : I1 −→ Ker φ0 where I1 is
an injective S −1 R-module and and Ker φ1 is a Gorenstein injective R-module. Denote the composition
φ1
I1 −→ Ker φ0 ֒→ I0 by d1 . Continuing in this way, we find injective S −1 R-modules I0 , I1 , I2 , . . . and S −1 Rd
n
In−1 ; n ≥ 1. Now, denote the sequence
homomorphisms In −→
d
d−1 φ0
d0
n
In−1 −→ · · · −→ I0 −→ I 0 −→ I 1 −→ · · ·
· · · −→ In −→
by I • . Then I • is an exact sequence of injective S −1 R-modules, L ∼
= im(I0 −→ I 0 ) and HomS −1 R (I, I • ) is
exact for all injective S −1 R-modules I. Therefore, L is a Gorenstein injective S −1 R-module.
Next, assume that L is Gorenstein injective as an S −1 R-module. Then, there is an exact sequence
I • : · · · −→ I1 −→ I0 −→ I 0 −→ I 1 −→ · · ·
of injective S −1 R-modules such that L ∼
= im(I0 −→ I 0 ) and that HomS −1 R (I, I • ) is exact for all injective
S −1 R-modules I. Then I • is also an exact sequence of injective R-modules. Let J be an injective R-module.
We have the following natural isomorphisms of complexes:
HomR (J, I • ) ∼
= HomR (J, HomS −1 R (S −1 R, I • )) ∼
= HomS −1 R (J ⊗R S −1 R, I • ).
But, J ⊗R S −1 R is an injective S −1 R-module, and so the complex HomR (J, I • ) is exact. This shows that L
is a Gorenstein injective R-module.
The last assertion follows immediately by Lemma 2.2 i).
6
M. NIKKHAH BABAEI AND K. DIVAANI-AAZAR
We can record the following immediate corollary.
Corollary 2.6. Assume that every direct limit of Gorenstein injective R-modules is Gorenstein injective.
Let S be a multiplicatively closed subset of R and M a Gorenstein injective R-module. Then S −1 M is a
Gorenstein injective S −1 R-module.
Proof. As S −1 R is a flat R-module, the Govorov-Lazard Theorem implies that S −1 R is the direct limit
of a direct system {Fθ }θ∈Θ of finitely generated free R-modules. One can easily check that Fθ ⊗R M is
(F ⊗ M ) is a
a Gorenstein injective R-module for all θ ∈ Θ. Thus, by our assumption, S −1 M ∼
= lim
−→ θ R
θ
Gorenstein injective R-module. Next, by Lemma 2.5, we have
GidS −1 R S −1 M = GidR S −1 M = 0,
and so S −1 M is is a Gorenstein injective S −1 R-module, as required.
Now, finally we are ready to prove our theorem.
Theorem 2.7. Let A be an Artinian R-module of finite Gorenstein injective dimension. Then A admits a
Gorenstein injective envelope GR (A) which is Artinian and idR (GR (A)/A) < ∞.
Proof. Let m ∈ SuppR A. Since A is an Artinian R-module, there exists an exact sequence
0 −→ A −→ E 0 −→ E 1 −→ · · · −→ E i −→ · · ·
in which each E i is an Artinian injective R-module. Then, by [CFrH, Theorem 3.3], it turns out that
Ker(E i −→ E i+1 ) is a Gorenstein injective R-module for all i ≥ GidR A. Now,
0
1
i
0 −→ Am −→ Em
−→ Em
−→ · · · −→ Em
−→ · · ·
i+1
i
)
−→ Em
is an injective resolution of the R-module Am . By [D, Lemma 2.1 (ii)], for each i ≥ 0, Ker(Em
i
i+1
is a direct summand of Ker(E i −→ E i+1 ). Hence, Ker(Em
−→ Em
) is a Gorenstein injective R-module
for all i ≥ GidR A. Thus, GidR Am ≤ GidR A, and so GidRm Am is finite by Lemma 2.5. Now, by Lemma
2.4, GRm (Am ) exists and it is an Artinian Rm -module with idRm (GRm (Am )/Am ) < ∞. It is easy to check
that any Artinian Rm -module B is also Artinian as an R-module and HomR (B, B) = HomRm (B, B). Thus,
GRm (Am ) is an Artinian R-module and Lemma 2.5 and Lemma 2.2 ii) yield that GR (Am ) = GRm (Am ) and
idR (GR (Am )/Am ) < ∞.
Suppose that SuppR A = {m1 , . . . , mn }. By [D, Lemma 2.1 (ii)], one has a natural R-isomorphism
A∼
= Am1 ⊕ · · · ⊕ Amn .
By the preceding paragraph for each i, the R-module Ami possesses an Artinian Gorenstein injective envelope
n
φi : Ami −→ GR (Ami ) such that idR (GR (Ami )/Ami ) < ∞. Now, ⊕ GR (Ami ) is an Artinian Gorenstein
i=1
injective R-module and
n
n
i=1
i=1
⊕φi : ⊕ Ami −→ ⊕ GR (Ami )
i
n
is an R-monomorphism with Coker(⊕φi ) ∼
= ⊕ GR (Ami )/Ami . Hence, ⊕φi is a Gorenstein injective preenvei
i=1
i
lope and idR (Coker(⊕φi )) < ∞. Let m and n be two distinct maximal ideals of R and B and C two R-modules
i
GORENSTEIN INJECTIVE ENVELOPES ...
7
such that SuppR B ⊆ {m} and SuppR C ⊆ {n}. Then, one can easily check that HomR (B, C) = 0. Thus, we
have
n
n
n
i=1
i=1
i=1
HomR ( ⊕ GR (Ami ), ⊕ GR (Ami )) ∼
= ⊕ HomR (GR (Ami ), GR (Ami )),
which yields that ⊕φi is our desired Gorenstein injective envelope.
i
Corollary 2.8. Let R be a Gorenstein ring and A an Artinian R-module. Then A admits a Gorenstein
injective envelope GR (A) which is Artinian and idR (GR (A)/A) < ∞.
Proof. Suppose that SuppR A = {m1 , . . . , mn }. By [D, Lemma 2.1 (ii)], one has a natural R-isomorphism
A∼
= Am1 ⊕ · · · ⊕ Amn .
Since over a Gorenstein local ring, any module has finite Gorenstein injective dimension, the proof of Theorem
2.7 yields that
GidR A = max{GidR Ami |i = 1, . . . , n} < ∞.
Hence, the claim follows by Theorem 2.7
Acknowledgement . The authors wish to thank Professor Edgar E. Enochs for his interest and comments on
this work.
References
[BE] R. Belshoff and E. Enochs, The Gorenstein injective envelope of the residue field of a local ring, Comm. Algebra, 32(2),
(2004), 599-607.
[CFoH] L.W. Christensen, H-B. Foxby and H. Holm, Beyond totally reflexive modules and back: a survay on Gorenstein
dimensions, Commutative algebra-Noetherian and non-Noetherian perspectives, 101-143, Springer, New York, 2011.
[CFrH] L.W. Christensen, A. Frankild and H. Holm, On Gorenstein projective, injective and flat dimensions—a functorial
description with applications, J. Algebra, 302(1), (2006), 231-279.
[CH] L.W. Christensen and H. Holm, Ascent properties of Auslander categories, Canad. J. Math., 61(1), (2009), 76-108.
[CW] L.W. Christensen and S. Sather-Wagstaff, Transfer of Gorenstein dimensions along ring homomorphisms, J. Pure Appl.
Algebra, 214(6), (2010), 982-989.
[D] K. Divaani-Aazar, Vanishing of the top local cohomology modules over Noetherian rings, Proc. Indian Acad. Sci. Math.
Sci., 119(1), (2009), 23-35.
[EJ] E.E. Enochs and O.M.G. Jenda, Relative Homological Algebra, de Gruyter Expositions in Mathematics, 30, Walter de
Gruyter & Co., Berlin, 2000.
[EL] E.E. Enochs and J.A. López-Ramos, Kaplansky classes, Rend. Sem. Mat. Univ. Padova, 107, (2002), 67-79.
[H] H. Holm, Gorenstein homological dimensions, J. Pure Appl. Algebra, 189(1-3), (2004), 167-193.
[K] H. Krause, The stable derived category of a noetherian scheme, Compos. Math., 141(5), (2005), 1128-1162.
[Sa] R. Sazeedeh, Gorenstein injectivity of the section functor, Forum Math., 22(6), (2010), 1117-1127.
[Sh] R.Y. Sharp, Artinian modules over commutative rings, Math. Proc. Cambridge Philos. Soc., 111(1), (1992), 25-33.
[T1] R. Takahashi, On the category of modules of Gorenstein dimension zero II, J. Algebra, 278(1), (2004), 402-410.
[T2] R. Takahashi, On the category of modules of Gorenstein dimension zero, Math. Z., 251(2), (2005), 249-256.
M. Nikkhah Babaei, Department of Mathematics, Alzahra University, Vanak, Post Code 19834, Tehran, Iran.
E-mail address: [email protected]
K. Divaani-Aazar, Department of Mathematics, Alzahra University, Vanak, Post Code 19834, Tehran, Iranand-School of Mathematics, Institute for Research in Fundamental Sciences (IPM), P.O. Box 19395-5746, Tehran,
Iran.
E-mail address: [email protected]
| 0math.AC
|
Improved Deep Speaker Feature Learning for
Text-Dependent Speaker Recognition
Lantian Li, Yiye Lin, Zhiyong Zhang, Dong Wang∗
arXiv:1506.08349v1 [cs.CL] 28 Jun 2015
Center for Speech and Language Technologies, Division of Technical Innovation and Development
Tsinghua National Laboratory for Information Science and Technology
Center for Speech and Language Technologies, Research Institute of Information Technology
Department of Computer Science and Technology, Tsinghua University, Beijing, China
{lilt,lyy,zhangzy}@cslt.riit.tsinghua.edu.cn;
[email protected]
Abstract—A deep learning approach has been proposed recently to derive speaker identifies (d-vector) by a deep neural network (DNN). This approach has been applied to text-dependent
speaker recognition tasks and shows reasonable performance
gains when combined with the conventional i-vector approach.
Although promising, the existing d-vector implementation still
can not compete with the i-vector baseline. This paper presents
two improvements for the deep learning approach: a phonedependent DNN structure to normalize phone variation, and
a new scoring approach based on dynamic time warping
(DTW). Experiments on a text-dependent speaker recognition
task demonstrated that the proposed methods can provide
considerable performance improvement over the existing d-vector
implementation.
Index Terms: d-vector, time dynamic warping, speaker recognition
I. I NTRODUCTION
Most modern speaker recognition systems are based on
human-crafted acoustic features, for example the Mel frequency cepstral coefficients (MFCC). A problem of the MFCC
(and other primary features) is that it involves plethora information besides the speaker identity, such as phone content,
channels, noises, etc. These heterogeneous and noisy information convolve together, making it difficult to be used for
either speech recognition or speaker recognition. All modern
speaker recognition systems rely on a statical model to ‘purify’
the desired speaker information. For example, in the famous
Gaussian mixture model-universal background model (GMMUBM) framework [1], the acoustic space is divided into
subspaces in the form of Gaussian components, and each
subspace roughly represents a phone. By formulating the
speaker recognition task as subtasks on the phone subspaces,
the GMM-UBM model can largely eliminate the impact of
phone content and other acoustic factors. This idea is shared
by many advanced techniques derived from GMM-UBM,
including the joint factor analysis (JFA) [2] and the i-vector
model [3].
In spite of the great success, the GMM-UBM approach
and the related methods are still limited by the lack of
discriminative capability of the acoustic features. Some researchers proposed solutions based on discriminative models.
For example, the SVM approach for GMM-UBMs [4] and
the PLDA approach for i-vectors [5]. All these discriminative
methods achieved remarkable success. Another direction is to
look for more task-oriented features, i.e., features that are more
discriminative for speaker recognition [6]. Although it seems
to be straightforward, this ‘feature engineering’ turns out to be
highly difficult. A major reason, in our mind, is that most of
the proposed delicate features are human-crafted and therefore
tend to be fragile in practical usage.
Recent research on deep learning offers a new idea of
‘feature learning’. It has been shown that with a deep neural
network (DNN), task-oriented features can be learned layer
by layer from very raw input. For example in automatic
speech recognition (ASR), phone-discriminative features can
be learned from spectra or filter bank energies (Fbanks). This
learned features are very powerful and have defeated the
MFCC that has dominated in ASR for several decades [7]. This
capability of DNNs in learning task-oriented features can be
utilized to learn speaker-discriminative features as well. A recent study shows that this is possible at least on text-dependent
tasks [8]. The authors reported that reasonable performance
can be achieved with the DNN-learned feature, and additional
performance gains can be obtained by combining the DNNbased approach and the i-vector approach.
Although the DNN-based feature learning shows great potential, the existing implementation still can not compete with
the i-vector baseline. There are at least two drawbacks with
the current implementation: First, the DNN model does not
use any information about the phone content, which leads
to difficulty when inferring speaker-discriminative features;
second, the evaluation (speaker scoring) is based on speaker
vectors (so called ‘d-vectors’), which are derived by averaging
the frame-wise DNN features. This simple average ignores the
temporal constraint that is highly important for text-dependent
tasks. Note that for tasks with a fixed test phrase, the two
drawbacks are closely linked to each other.
This paper follows the work in [8] and provides two enhancements for the DNN-based feature learning: First, phone
posteriors are involved in the DNN input so that speakerdiscriminative features can be learned easier by alleviating the
impact of phone variation; second, two scoring methods that
consider the temporal constraint are proposed: segmentation
pooling and dynamic time warping (DTW) [9].
The rest of the paper is organized as follows. Section II
describes some related work, and Section III presents the
DNN-based feature learning. The new methods are proposed
in Section IV and the experiments are presented in Section V.
Finally, Section VI concludes this paper and discusses some
future work.
II. R ELATED WORK
This paper follows the work in [8] and provides several extensions. Particularly, the speaker identity in [8] is represented
by a d-vector derived by average pooling, which is quite neat
and efficient, but loses much information of the test signal,
such as the distributional property and the temporal constraint.
One of the main contribution of this paper is to investigate how
to utilize the temporal constraint in the DNN-based approach.
The DNN model has been studied in speaker recognition
in several ways. For example, in [10], DNNs trained for
ASR were used to replace the UBM model to derive the
acoustic statistics for i-vector models. In [11], a DNN was
used to replace PLDA to improve discriminative capability of
i-vectors. All these methods rely on the generative framework,
i.e., the i-vector model. The DNN-based feature learning
presented in this paper is purely discriminative, without any
generative model involved.
III. DNN- BASED FEATURE LEARNING
It is well-known that DNNs can learn task-oriented features from raw input layer by layer. This property has been
employed in ASR where phone-discriminative features are
learned from very low-level features such as Fbanks or even
spectra [7]. It has been shown that with a well-trained DNN,
variations irrelevant to the learning task can be gradually
eliminated when the feature propagates through the DNN
structure layer by layer. This feature learning is so powerful
that in ASR, the primary Fbank feature has defeated the MFCC
feature that was carefully designed by people and dominated
in ASR for several decades.
This property can be also employed to learn speakerdiscriminative features. Actually researchers have put much
effort in searching for features that are more discriminative
for speakers [6], but the effort is mostly vain and the MFCC
is still the most popular choice. The success of DNNs in ASR
suggests a new direction, that speaker-discriminative features
can be learned from data instead of being crafted by hand. The
learning can be easily done and the process is rather similar as
in ASR, with the only difference that in speaker recognition,
the learning goal is to discriminate different speakers.
Fig. 1 presents the DNN structure used in this work for
speaker-discriminative feature learning. Following the convention of ASR, the input layer involves a window of 20dimensional Fbanks. The window size is set to 21, which was
found to be optimal in our work. The DNN structure involves
4 hidden layers, and each consists of 200 units. The units of
the output layer correspond to the speakers in the training data,
and the number is 80 in our experiment. The 1-hot encoding
scheme is used to label the target, and the training criterion
is set to cross entropy. The learning rate is set to 0.008 at
the beginning, and is halved whenever no improvement on a
cross-validation (CV) set is found. The training process stops
when the learning rate is too small and the improvement on
the CV set is too marginal. Once the DNN has been trained
successfully, the speaker-discriminative features can be read
from the last hidden layer.
In the test phase, the features are extracted for all frames of
the given utterance. To derive utterance-based representations,
an average pooling approach was used in [8], where the framelevel features are averaged and the resultant vector is used to
represent the speaker. This vector is called ‘d-vector’ in [8],
Fbank
(40*21 dim)
d-vector is the averaged activations
from the last hidden layers
P(spk1)
P(spk2)
P(spkN)
Fully-connected maxout hidden layers.
Output layer is removed in
enrollment and evaluation.
Fig. 1. The DNN structure used for learning speaker-discriminative features.
and we adopt this name in this work. The same methods used
for i-vectors can be used for d-vectors to conduct the test, for
example by computing the cosine distance or the PLDA score.
A. Comparison between i-vectors and d-vectors
The two kinds of speaker vectors, the d-vector and the ivector, are fundamentally different. I-vectors are based on a
linear Gaussian model, for which the learning is unsupervised
and the learning criterion is maximum likelihood on acoustic
features; in contrast, d-vectors are based on neural networks,
for which the learning is supervised, and the learning criterion
is maximum discrimination for speakers. This difference leads
to several advantages with d-vectors: First, it is a ‘discriminative’ vector, which represents speakers by removing speakerirrelevant variance, and so sensitive to speakers and invariant
to other disturbance; second, it is a ‘local’ speaker description
that uses only local context, so can be inferred from very short
utterances; third, it relies on ‘universal’ data to learn the DNN
model, which makes it possible to learn from large amounts
of data that are task-independent.
IV. I MPROVED DEEP FEATURE LEARNING
There are several limitations in the implementation of the
feature learning paradigm presented in the previous section.
First, it does not involve any prior knowledge in model
training, for example phone identities. Second, the simple
average pooling does not consider the temporal information
which is particularly important for text-dependent recognition
tasks. Several approaches are proposed in this section to
address these problems.
A. Phone-dependent training
A potential problem of the DNN-based feature learning
described in the previous section is that it is a ‘blind learning’,
i.e., the features are learned from raw data without any prior
knowledge. This means that the learning purely relies on
the complex deep structure of the DNN model and a large
amount of data to discover speaker-discriminative patterns.
If the training data is abundant, this is often not a problem;
d-vector
P(spk1)
Fbanks
(40*21)
P(spk2)
P(spkN)
Phone
Posteriors
(67)
Fully-connected sigmoid hidden layers.
Output layer is removed in
enrollment and evaluation.
Fig. 2. Phone-dependent DNN structure used for learning speakerdiscriminative features.
and the segmentation does not impact quality of the piece-wise
d-vectors very much.
A more theoretical treatment is based on dynamic time
warping (DTW) [1]. The DTW algorithm is a principle way
to measure similarities between two variable-length temporal
sequences. In the most simple sense, DTW searches for an
optimal path that matches two sequences with the lowest
cost, by employing the dynamic programming (DP) method to
reduce the search complexity. In our task, the DNN-extracted
features of an utterance are treated as a temporal sequence.
In test, the sequence derived from the enrollment utterance
and the sequence derived from the test utterance are matched
by DTW, where the cosine distance is used to measure the
similarity between two frame-level DNN features. Principally,
segment pooling can be regarded as a special case of DTW,
where the two sequences are in the same length, and the
matching between the two sequences is piece-wise.
V. E XPERIMENTS
B. Segment pooling and dynamic time warping
A. Database
The experiments are performed on a database that involves
a limited set of short phrases. The entire database contains
recordings of 10 short phrases from 100 speakers (gender
balanced), and each phrase contains 2 ∼ 5 Chinese characters.
For each speaker, every phrase is recorded 15 times, amounting
to 150 utterances per speaker.
The training set involves 80 randomly selected speakers,
which results in 12000 utterances in total. To prevent overfitting, a cross-validation (CV) set containing 1000 utterances
is selected from the training data, and the remaining 11000
utterances are used for model training, including the DNN
model in the d-vector approach, and the UBM, the T matrix,
the LDA and PLDA model in the i-vector approach.
The evaluation set consists of the remaining 20 speakers.
The evaluation is performed for each particular phrase. For
each phrase, there are 44850 trails, including 2100 target trails
and 42750 non-target trials. For the sake of neat presentation,
we report the results with 5 short phrases and use ‘Pn’
to denote the n-th phrase. The conclusions obtained here
generalize well to other phrases.
Text-dependent speaker recognition is essentially a sequential pattern matching problem, but the current d-vector approach derives speaker identities as single vectors by average
pooling, and then formulates speaker recognition as vector
matching. This is certainly not ideal as the temporal constraint
is totally ignored when deriving the speaker vector. A possible solution is to segment an enrollment/test utterance into
several pieces, and derive the speaker vector for each piece.
The speaker identity of the utterance is then represented by
the sequence of the piece-wise speaker vectors, and speaker
matching is conducted by matching the corresponding vector
sequences. This paper adopts a simple sequence matching
approach: the two sequences are assumed to be identical
in length, and the matching is conducted piece by piece
independently. Finally the matching score takes the average
of scores on all the pieces.
Note that for the i-vector approach, this segmentation
method is not feasible, since i-vectors are inferred from feature
distributions and so the piece-wised solution simply degrades
the quality of the i-vector of each piece. For d-vectors, this
approach is totally fine as they are inferred from local context
B. Baseline
Two baseline systems are built, one is based on i-vectors and
the other is based on d-vectors. The acoustic features of the
i-vector system are 39-dimensional MFCCs, which consist of
13 static components (including C0) and the first- and secondorder derivatives. The number of Gaussian components of the
UBM is 128, and the dimension of the i-vector is 200. The
d-vector baseline uses the DNN structure shown in Fig. 1.
The average pooling is used to derive d-vectors. The acoustic
features are 40-dimensional Fbanks, with the left and right
10 frames concatenated together. The frame-level features are
extracted from the last hidden layer, and the dimension is 200.
Table I presents the results in terms of equal error rate
(EER). It can be seen that the i-vector system generally outperforms the d-vector system in a significant way. Particularly,
the discriminative methods (LDA and PLDA) clearly improves
the i-vector system, however for the d-vector system, no improvement was found by these methods. This is not surprising,
since the d-vectors have been discriminative by themselves.
For this reason, LDA and PLDA are not considered any more
for d-vectors in the following experiments.
however in tasks with a limited amount of data, for instance
the text-dependent task in our hand, this blind learning tends
to be difficult because there are too many speaker-irrelevant
variations involved in the raw data, particularly phone contents.
A possible solution is to supply the DNN model extra
information about which phone is spoken at each frame. This
can be simply achieved by adding a phone indicator in the
DNN input. However, it is often not easy to get the phone
alignment in practice. An alternative way is to supply a vector
of phone posterior probabilities for each frame, which is a
‘soft’ phone alignment and can be easily obtained from a
phone-discriminative model. In this work, we choose to use a
DNN model that was trained for ASR to produce the phone
posteriors. Fig. 1 illustrates how the phone posteriors are
involved in the DNN structure. The training process does not
change for the new structure.
TABLE I
P ERFORMANCE OF BASELINE SYSTEMS
i-vector
d-vector
Phrase
P1
P2
P3
P4
P5
P1
P2
P3
P4
P5
cosine
2.86
1.52
3.43
3.19
3.57
10.29
10.52
10.10
10.38
9.14
EER%
LDA
1.81
2.29
3.05
2.86
3.00
9.81
10.57
9.33
9.95
9.29
The DTW results are shown in the fourth row of Table II,
denoted by ‘DNN+PT+DTW’. It can be observed that DTW
generally outperforms segment pooling.
PLDA
1.71
1.57
3.05
2.71
2.67
12.67
12.29
10.48
11.10
11.10
TABLE III
P ERFORMANCE OF SYSTEM COMBINATION
PLDA
DNN+PT+DTW
Combination
C. phone-dependent learning
In this experiment, the phone posteriors are included in the
DNN input, as shown in Fig. 2. The phone posteriors are
produced by a DNN model that was trained for ASR with a
Chinese database consisting of 6000 hours of speech data. The
phone set consists of 66 toneless initial and finals in Chinese,
plus the silence phone. The results are shown in the second
row of Table II, denoted by ‘DNN+PT’. It can be seen that
the phone-dependent training leads to marginal but consistent
performance improvement for the d-vector system.
TABLE II
P ERFORMANCE WITH IMPROVED DNN FEATURE LEARNING
DNN baseline
DNN+PT
DNN+PT+SEG
DNN+PT+DTW
P1
10.29
10.24
9.95
9.14
P2
10.52
10.05
8.90
8.38
EER%
P3
10.10
9.81
8.95
8.52
P4
10.38
9.48
9.76
8.86
P5
9.14
8.71
8.67
8.14
D. Segment pooling and DTW
As presented in Section IV, the segment pooling approach
segments an input utterance into n pieces, and derives a dvector for each piece. The scoring is conducted on the piecewise d-vectors independently, and the average of scores on
these pieces is taken as the utterance-level score. The results
are shown in Fig. 3, where seg-n means that each utterance is
segmented into n pieces. It can be seen that segment pooling
offers clear performance improvement.
For a more clear comparison, the EER results with 3 segments are shown in Table II, denoted by ‘DNN+PT+SEG’. It is
clear to see that by the segmentation, significant performance
is obtained.
Fig. 3. EER results with segment pooling.
P1
1.71
9.14
1.52
P2
1.57
8.38
1.38
EER%
P3
3.05
8.52
2.33
P4
2.71
8.86
2.33
P5
2.67
8.14
2.38
E. System combination
Following [8], we combine the best i-vector system (PLDA)
and the best d-vector system (DNN+PT+DTW). The combination is simply done by interpolating the scores obtained from
the two systems: αsiv + (1 − α)sdv , where siv and sdv are
scores from the i-vector and d-vector systems respectively, and
α is the interpolation factor. The EER results with the optimal
α are shown in Table III. It can be seen that the combination
leads to the best performance we can obtain so far.
VI. C ONCLUSIONS
This paper presented several enhancements for the DNNbased feature learning approach in speaker recognition. We
presented a phone-dependent DNN model to supply phonetic
information when learning speaker features, and proposed
two scoring methods based on a segment pooling and DTW
respectively to leverage temporal constraints. These extensions
significantly improved performance of the d-vector system.
Future work involves investigating more complicated statistical
models for d-vectors, and use large amounts of data to learn
more powerful speaker-discriminative features.
R EFERENCES
[1] D. Reynolds, T. Quatieri, and R. Dunn, “Speaker verification using
adapted gaussian mixture models,” Digital Signal Processing, vol. 10,
no. 1, pp. 19–41, 2000.
[2] P. Kenny, G. Boulianne, P. Ouellet, and P. Dumouchel, “Joint factor analysis versus eigenchannels in speaker recognition,” IEEE Transactions on
Audio, Speech, and Language Processing, vol. 15, pp. 1435–1447, 2007.
[3] ——, “Speaker and session variability in gmm-based speaker verification,” IEEE Transactions on Audio, Speech, and Language Processing,
vol. 15, pp. 1448–1460, 2007.
[4] W. Campbell, D. Sturim, and D. Reynolds, “Support vector machines
using gmm supervectors for speaker verification,” Signal Processing
Letters, IEEE, vol. 13, no. 5, pp. 308–311, 2006.
[5] S. Ioffe, “Probabilistic linear discriminant analysis,” Computer Vision
ECCV 2006, Springer Berlin Heidelberg, pp. 531–542, 2006.
[6] T. Kinnunen and H. Li, “An overview of text-independent speaker
recognition: From features to supervectors,” Speech communication,
vol. 52, no. 1, pp. 12–40, 2010.
[7] J. Li, D. Yu, J. Huang, and Y. Gong, “Improving wideband speech
recognition using mixed-bandwidth training data in cd-dnn-hmm,” SLT,
pp. 131–136, 2012.
[8] V. Ehsan, L. Xin, M. Erik, L. M. Ignacio, and G.-D. Javier, “Deep neural
networks for small footprint text-dependent speaker verification,” IEEE
International Conference on Acoustic, Speech and Signal Processing
(ICASSP), vol. 28, no. 4, pp. 357–366, 2014.
[9] D. Berndt and J. Clifford, “Using dynamic time warping to find patterns
in time series,” KDD workshop, vol. 10, no. 16, pp. 359–370, 1994.
[10] P. Kenny, V. Gupta, T. Stafylakis, P. Ouellet, and J. Alam, “Deep neural
networks for extracting baum-welch statistics for speaker recognition,”
Odyssey, 2014.
[11] J. Wang, D. Wang, Z.-W. Zhu, T. Zheng, and F. Song, “Discriminative
scoring for speaker recognition based on i-vectors,” APSIPA, 2014.
| 9cs.NE
|
KOSZUL COMPLEX OVER SKEW POLYNOMIAL RINGS
JOSEP ÀLVAREZ MONTANER, ALBERTO F. BOIX, AND SANTIAGO ZARZUELA
arXiv:1712.07836v1 [math.AC] 21 Dec 2017
Dedicated to Professor Gennady Lyubeznik on the occasion of his 60th birthday
Abstract. We construct a Koszul complex in the category of left skew polynomial rings associated
to a flat endomorphism that provides a finite free resolution of an ideal generated by a Koszul regular
sequence.
1. Introduction
Let A be a commutative Noetherian ring of characteristic p > 0 and F : A −→ A the associated
Frobenius map for which F (a) = ap for all a ∈ A. The study of A-modules with an action of
the Frobenius map has received a lot of attention over the last decades and is at the core of the
celebrated theory of tight closure developed by M. Hochster and C. Huneke in [HH90] and the
theory of F -modules introduced by G. Lyubeznik [Lyu97].
To provide an action of the Frobenius on an A-module M is equivalent to give a left A[Θ; F ]module structure on M . Here A[Θ; F ] stands for the left skew polynomial ring associated to F ,
which is an associative, N-graded, not necessarily commutative ring extension of A. More generally,
we may also consider the skew polynomial
rings A[Θ; F e ] associated to the e-th iterated Frobenius
L
M
M
map and the graded ring F = e≥0 Fe introduced by G. Lyubeznik and K. E. Smith in [LS01]
that collects all the A[Θ; F e ]-module structures on M or equivalently, all possible actions of F e on
M . In the case that F M is principally generated then it is isomorphic to A[Θ; F ]. We want also
to single out here that, under the terminology skew polynomial ring of Frobenius type, Y. Yoshino
[Yos94] studied A[Θ; F ] and left modules over it, and used this study to obtain some new results
about tight closure theory; following the same spirit, R. Y. Sharp [Sha09] also studied the ideal
theory of A[Θ; F ]. The case when A is a field was studied in detail by F. Enescu [Ene12].
One may also develop a dual notion of right skew polynomial ring associated to F that we denote
A[ε; F ]. A left A[ε; F ]-module structure on M is then equivalent to provide an action of a Cartier
operator as considered by M. Blickle and G. Böckle in [BB11].
L The corresponding ring collecting all
the A[ε; F e ]-module structures on M , that we denote C M = e≥0 CeM , was developed by K. Schwede
[Sch11] and M. Blickle [Bli13] (see [BS13] for a nice survey) and has been a hot topic in recent years
because of its role in the theory of test ideals.
Although the theory of skew polynomial rings is classical (see [GW04, Chapter 2] and [MR01,
Chapter 1, Sections 2 and 6]), the aim of this work is to go back to its basics and develop new tools
that should be potentially useful in the study of modules with a Frobenius or Cartier action. Our
approach will be in a slightly more general setting as we will consider a Noetherian commutative
ring A (not necessarily of positive characteristic) and a ring homomorphism ϕ : A −→ A. Although
2010 Mathematics Subject Classification. Primary 13A35; Secondary 13B10, 14B25, 16S36.
Key words and phrases. Cartier algebras, Frobenius algebras, skew polynomial rings.
The first author is partially supported by Generalitat de Catalunya 2014SGR-634 project and Spanish Ministerio
de Economı́a y Competitividad MTM2015-69135-P. The second and third authors are supported by Spanish Ministerio
de Economı́a y Competitividad MTM2016-7881-P. The second author is partially supported by the Israel Science
Foundation (grant No. 844/14).
1
2
J. ÀLVAREZ MONTANER, A. F. BOIX, AND S. ZARZUELA
most of the results are mild generalizations of the case of the Frobenius morphism we point out
that this general approach has already been fruitful (see [SW07]).
Now, we overview the contents of this paper for the convenience of the reader. In Section 2 we
introduce all the basics on left and right skew polynomial rings associated to a morphism ϕ that
will be denoted by A[Θ; ϕ] and A[ε; ϕ]. More generally, given an A-module M , we will consider the
corresponding rings F M,ϕ and C M,ϕ . The main result of this Section is Theorem 2.11 which states
that, whenever ϕ is a finite morphism, we have an isomorphism between the graded pieces FeER ,ϕ
and CeR,ϕ induced by Matlis duality. Here, ER is the injective hull of the residue field of R = A/I
with I ⊆ A := K[[x1 , . . . , xn ]] being an ideal.
In Section 3 we present the main result of this work. Here we consider a flat morphism ϕ satisfying
some extra condition (which is naturally satisfied in the case of the Frobenius morphism). First
we introduce the ϕ-Koszul complex which is a Koszul-type complex associated to x1 , . . . , xn ∈ A
in the category of left A[Θ; ϕ]-modules. In Theorem 3.14 we prove that, whenever x1 , . . . , xn is an
A-Koszul regular sequence (see Definition 3.13), the ϕ-Koszul complex provides a free resolution
of A/In in the category of left A[Θ; ϕ]-modules, where In is the ideal generated by x1 , . . . , xn . In
the case where A is a regular ring of positive characteristic p > 0 and ϕ = F e is the e-th iteration
of the Frobenius, we obtain a free resolution, in the category of left A[Θ; F e ]-modules, of the ideal
In (see Corollary 3.17).
To the best of our knowledge this is one of the first explicit examples of free resolutions of Amodules as modules over a skew polynomial ring. We hope that a development of this theory of
free resolutions would provide, in the case of the Frobenius morphism, more insight in the study of
modules with a Frobenius action.
2. Preliminaries
Let A be a commutative Noetherian ring and ϕ : A −→ A a ring homomorphism. Associated
to ϕ we may consider non necessarily commutative algebra extensions of A that are useful in the
sense that, some non finitely generated A-modules become finitely generated when viewed over
these extensions. In this section we will collect the basic facts on this theory keeping always an eye
on the case of the Frobenius morphism.
First we recall that ϕ allows to describe a covariant functor, the pushforward or restriction of
scalars functor ϕ∗ , from the category of left A-modules to the category of left A-modules. Namely,
given an A-module M , ϕ∗ M is the (A, A)-bimodule having the usual A-module structure on the
right but the left structure is twisted by ϕ; that is, if we denote by ϕ∗ m (m ∈ M ) an arbitrary
element of ϕ∗ M then, for any a ∈ A
a · ϕ∗ m := ϕ∗ (ϕ(a)m).
Moreover, given a map g : M −→ N of left A-modules we can define a map ϕ∗ g : ϕ∗ M −→ ϕ∗ N of
left A-modules by setting, for any m ∈ M ,
ϕ∗ g(ϕ∗ m) := ϕ∗ g(m).
2.1. Left and right skew-polynomial rings. The most basic ring extensions given by a ring
homomorphism are the so-called skew-polynomial rings or Ore extensions. This is a classical object
of study, and we refer either to [GW04, Chapter 2] or [MR01, Chapter 1, Sections 2 and 6] for
further details.
Definition 2.1. Let A be a commutative Noetherian ring and ϕ : A −→ A a ring homomorphism.
KOSZUL COMPLEX OVER SKEW POLYNOMIAL RINGS
3
• Left skew polynomial ring of A determined by ϕ:
A[Θ; ϕ] :=
AhΘi
hΘa − ϕ(a)Θ |
a ∈ Ai
,
that is, A[Θ; ϕ] is a free left A-module with basis {θ i }i∈N0 and θ · a = ϕ(a)θ for all a ∈ A.
• Right skew polynomial ring of A determined by ϕ:
A[ε; ϕ] :=
Ahεi
haε − εϕ(a) |
a ∈ Ai
,
that is, A[ε; ϕ] is a free right A-module with basis {εi }i∈N0 and a · ε = εϕ(a) for all a ∈ A.
Remark 2.2. The reader will easily note that A[Θ; ϕ]op ∼
= A[Θ; ϕ], where
= A[ε; ϕ] and A[ε; ϕ]op ∼
op
(−) denotes the opposite ring.
Both ring extensions are determined by the following universal property.
Proposition 2.3. Let B = A[Θ; ϕ] be the left (resp. B = A[ε; ϕ] be the right) skew polynomial
ring of A determined by ϕ. Suppose that we have a ring T , a ring homomorphism φ : A −→ T and
an element y ∈ T such that, for each a ∈ A,
yφ(a) = φ(ϕ(a))y
(resp. φ(a)y = yφ(ϕ(a))).
Then, there is a unique ring homomorphism B
makes the triangle
A ❄
❄❄ φ
❄❄
❄❄
❄
ψ
T
/ T such that ψ(Θ) = y (resp. ψ(ε) = y) which
⑦
⑦
⑦
/B
⑦
ψ
commutative.
This universal property allow us to provide more general examples of skew polynomial rings.
Example 2.4. Let u ∈ A. As uΘ is an element of A[Θ; ϕ] such that, for any a ∈ A,
(uΘ)a = uΘa = uϕ(a)Θ = ϕ(a)(uΘ),
it follows from the universal property for left skew polynomial rings that there is a unique A–algebra
/ A[Θ; ϕ] sending Θ′ to uΘ; we shall denote the image of this map by
homomorphism A[Θ′ ; ϕ]
A[uΘ; ϕ]. The previous argument shows that A[uΘ; ϕ] can be also regarded as a left skew polynomial
ring. Analogously, A[εu; ϕ] can be also regarded as right skew polynomial ring since we have, for
any a ∈ A,
a(εu) = εϕ(a)u = (εu)ϕ(a).
2.1.1. The case of Frobenius homomorphism. We single out the case where A is a ring of positive
characteristic p > 0 and ϕ = F is the Frobenius homomorphism. The specialization of Definition
2.1 to this case is the following:
• The Frobenius skew polynomial ring of A is the non-commutative graded ring A[Θ; F ]; that is,
the free left A-module with basis {Θe }e∈N0 and right multiplication given by
Θ · a = ap Θ.
• The Cartier skew polynomial ring of A is the non-commutative graded ring A[ε; F ]; that is,
the free right A-module with basis {εe }e∈N0 and left multiplication given by
a · ε := εap .
4
J. ÀLVAREZ MONTANER, A. F. BOIX, AND S. ZARZUELA
It turns out that A[Θ; F ] is rarely left or right Noetherian, as it is explicitly described by
Y. Yoshino in [Yos94, Theorem (1.3)]. He also provided effective descriptions of basic examples
of left modules over A[Θ; F ]. Namely, the ring A, the localizatons Aa at elements a ∈ A and
local cohomology modules HIi (A), where I is any ideal of A, have an abstract structure as finitely
generated left A[Θ; F ]-modules that we collect in the following result. We refer to [Yos94, pp.
2490–2491] for further details.
Proposition 2.5. Let A be a commutative Noetherian ring of characteristic p, let a ∈ A be any
element and let J(a) denote the left ideal of A[Θ; F ] generated by the infinite set
{ap
e −1
Θe − 1 |
e ∈ N}.
Then, the following statements hold.
(i) A has a natural structure as left A[Θ; F ]-module given by
A∼
= A[Θ; F ]/A[Θ; F ]hΘ − 1i ∼
= A[Θ; F ]/J(1).
Moreover, if A is a local ring:
(ii) The localization Aa has a natural structure as left A[Θ; F ]-module given by
Aa ∼
= A[Θ; F ]/J(a).
(iii) The Čech complex of A with respect to a1 , . . . , at
0 −→ A −→
t
M
Aai −→
i=1
M
Aai aj −→ . . . −→
t
M
Aa1 ···abi ···at −→ Aa1 ···at −→ 0
M
A[Θ; F ]/J(ai aj ) −→ . . .
i=1
1≤i<j≤t
is a complex of left A[Θ; F ]-modules which is isomorphic to the following complex:
0 −→ A[Θ; F ]/J(1) −→
t
M
A[Θ; F ]/J(ai ) −→
i=1
. . . −→
t
M
i=1
1≤i<j≤t
A[Θ; F ]/J(a1 · · · abi · · · at ) −→ A[Θ; F ]/J(a1 · · · at ) −→ 0.
(iv) Any local cohomology module HIi (A) has an abstract structure as a finitely generated left
A[Θ; F ]-module. For example, if (A, m, K) is a local ring of characteristic p of dimension
d ≥ 1, and a1 , . . . , ad is a system of parameters for A, then
H d (A) ∼
= A[Θ; F ]/ (J(a1 · · · ad ) + A[Θ; F ]ha1 , . . . , ad i) .
m
In this section we are going to give a natural generalization of skew polynomial rings associated
to a ring homomorphism ϕ. Before doing so, we briefly recall how to give a module structure over
a skew polynomial ring by means of the so-called ϕ and ϕ−1 -linear maps.
Definition 2.6. Let A be a commutative Noetherian ring, ϕ : A −→ A a ring homomorphism and
M be an A-module. Given ψ, φ ∈ EndA (M ) :
(i) We say that ψ is ϕ-linear provided ψ(am) = ϕ(a)ψ(m) for any (a, m) ∈ A × M .
(ii) We say that φ is ϕ−1 -linear provided φ(ϕ(a)m) = aφ(m) for any (a, m) ∈ A × M .
We denote by Endϕ (M ) and Endϕ−1 (M ) the A-endomorphisms of M which are ϕ-linear and
respectively. These endomorphisms can be interpreted in terms of the pushforward
functor since we have the following bijections:
ϕ−1 -linear
Endϕ (M )
/ HomA (M, ϕ∗ M ) ,
ψ 7−→ [m 7−→ ϕ∗ (ψ(m))]
Endϕ−1 (M ) −→ HomA (ϕ∗ M, M )
ψ 7−→ [ϕ∗ m 7−→ ψ(m)]
KOSZUL COMPLEX OVER SKEW POLYNOMIAL RINGS
5
An A[Θ; ϕ]-module is simply an A-module M together with a suitable action of Θ on M . Actually,
one only needs to consider a ϕ-linear map ψ : M −→ M and define Θ · m := ψ(m) for all m ∈ M .
Analogously, an A[ε; ϕ]-module is an A-module M together with an action of ε given by a ϕ−1 -linear
map φ : M −→ M ; we record all these simple remarks into the following:
Proposition 2.7. Let M be an A-module.
(i) There is a bijective correspondence between Endϕ (M ) and the left A[Θ; ϕ]-module structures
which can be attached to M .
(ii) There is a bijective correspondence between Endϕ−1 (M ) and the left A[ε; ϕ]-module structures
which can be attached to M .
More generally, we may consider the e-th powers ϕe of the ring homomorphism ϕ and define the
corresponding notion of ϕe and ϕ−e -linear maps. We may collect all these morphisms in a suitable
algebra that would provide a generalization of the Frobenius algebra introduced by G. Lyubeznik
and K. E. Smith in [LS01, Definition 3.5] and the Cartier algebra considered by K. Schwede [Sch11]
and generalized by M. Blickle [Bli13] (see also [BS13]).
Definition 2.8. Let A be a commutative Noetherian ring and ϕ : A −→ A a ring homomorphism.
Let M be an A-module. Then we define:
• Ring of ϕ-linear operators on M : Is the associative, N-graded, not necessarily commutative
ring
M
F M,ϕ :=
FeM,ϕ ,
e≥0
where FeM,ϕ := HomA (M, ϕe∗ M ) = Endϕe (M ). A product is defined as
M,ϕ
ψe′ · ψe := ϕe∗ (ψe′ ) ◦ ψe ∈ Fe+e
′.
.
for any given ψe ∈ FeM,ϕ and ψe′ ∈ FeM,ϕ
′
• Ring of ϕ−1 -linear operators on M : Is the associative, N-graded, not necessarily commutative
ring
M
C M,ϕ :=
CeM,ϕ ,
e≥0
where CeM,ϕ := HomA (ϕe∗ M, M ) = Endϕ−e (M ).
A product is defined as
M,ϕ
φe′ · φe := φe′ ◦ ϕe∗ (φe ) ∈ Ce+e
′
′
.
for any given φe ∈ CeM,ϕ and φe′ ∈ CeM,ϕ
′
Actually, the generalization of Cartier algebra given by M. Blickle in [Bli13] would be interpreted
in our context as follows:
Definition 2.9. An A-Cartier algebra with respect to ϕ is an N-graded A-algebra
M
C ϕ :=
Ceϕ
e≥0
such that, for any (a, φe ) ∈ A × Ceϕ , we have that a · φe = φe · ϕe (a). The A-algebra structure of C ϕ
/ C ϕ is
is given by the natural map from A to C0ϕ . We also assume that the structural map A
0
surjective.
6
J. ÀLVAREZ MONTANER, A. F. BOIX, AND S. ZARZUELA
We point out that C M,ϕ is generally NOT an A-Cartier algebra with respect to ϕ, since C0M,ϕ =
/ EndA (M ) is, in general, not surjective. NevertheEndA (M ) and therefore the natural map A
less, if M = A/I (where I is any ideal of A) then EndA (M ) = A/I and therefore it follows that
C A/I,ϕ is an A-Cartier algebra.
Whenever the ring of ϕ-linear (resp. ϕ−1 -linear) operators is principally generated it is isomorphic to a left (resp. right) skew polynomial ring. This is the case of the ring of ϕ-linear operators
of the ring A; notice that, when A is of prime characteristic and ϕ = F is the Frobenius map, this
was already observed by G. Lyubeznik and K. E. Smith [LS01, Example 3.6].
A,ϕ
Example 2.10. We have that F A,ϕ ∼
= A[Θ; ϕ]. Indeed, fix e ∈ N and let ψe ∈ Fe . We point out
that, for any a ∈ A,
ψe (a) = ψe (a · 1) = ϕe (a)ψe (1) = ψe (1)ϕe (a).
In this way, set
FeA,ϕ
be
/ AΘe
ψe 7−→ ψe (1)Θe .
The previous straightforward calculation shows the injectivity of this map. In fact, it is a bijective
map with inverse
/ FeA,ϕ
AΘe
aΘe 7−→ aϕe .
b
/ A[Θ; ϕ] as the unique map of rings given in degree e by be it follows
In this way, setting F A,ϕ
that b is an isomorphism of graded algebras.
2.1.2. Duality between Cartier algebras and Frobenius algebras. Let A = K[[x1 , . . . , xd ]] be a formal
power series ring with d indeterminates over a field K, I ⊆ A an ideal and R := A/I. Let
ϕ : A −→ A be a ring homomorphism such that ϕe∗ A is finitely generated as A-module. The aim
of this subsection is to establish an explicit correspondence, at the level of graded pieces, between
C R,ϕ and F ER ,ϕ given by Matlis duality. This would extend the correspondence given in the case
of the Frobenius map (cf. [BB11, Proposition 5.2] and [SY11, Theorem 1.20 and Corollary 1.21]).
Theorem 2.11. Let ϕ : A −→ A be a ring homomorphism such that ϕe∗ A is finitely generated as
A-module. Then we have that
HomA (ϕe∗ R, R)∨ ∼
= HomA (ER , ϕe∗ ER )
and
HomA (ER , ϕe∗ ER )∨ ∼
= HomA (ϕe∗ R, R).
Before proving this theorem we have to show a previous statement which we shall need during
its proof; albeit the below result was obtained by F. Enescu and M. Hochster in [EH08, Discussion
(3.4)] (see also [Yos94, Lemma (3.6)]), we review here their proof for the convenience of the reader.
/ (B, n, L) be a local homomorphism of local rings, and suppose that
Lemma 2.12. Let (A, m, K)
mB is n-primary and that L is finite algebraic over K (both these conditions hold if B is modulefinite over A). Let E := EA (K) and EB (L) denote choices of injective hulls for K over A and for L
over B, respectively. Then, the functor HomA (−, E), on B-modules, is isomorphic with the functor
HomB (−, EB (L)).
Proof. First of all, we underline that HomA (−, E), on B-modules, can be identified via adjunction
with
HomA ((−) ⊗A B, E) ∼
= HomB (−, HomA (B, E))
KOSZUL COMPLEX OVER SKEW POLYNOMIAL RINGS
7
and therefore HomA (B, E) is injective as B-module. Moreover, as mB is n-primary any element of
HomA (B, E) is killed by a power of n and therefore
HomA (B, E) ∼
= EB (L)⊕l .
In this way, it only remains to check that l = 1. Indeed, we note that
HomA (L, E) ∼
= HomA (L, K).
However, as A-module, HomA (L, K) is abstractly isomorphic to L (here we are using the assumption
that L is finite algebraic over K). Thus, all these foregoing facts imply that
EB (L) ∼
= HomA (B, E),
hence HomA ((−) ⊗A B, E) ∼
= HomB (−, EB (L)) and we get the desired conclusion.
Proof of Theorem 2.11. First of all, we underline that
HomA (ϕe R, R)∨ ∼
= HomA (ER , ϕe (R)∨ ).
∗
∗
e
Now, let E∗ be the injective hull of the residue field of ϕ∗ R. In this way, from Lemma 2.12
deduce that HomA (−, E) ∼
= Homϕe∗ A (−, E∗ ) as functors of ϕe∗ A-modules. Therefore, combining
these facts joint with the exactness of ϕe∗ it follows that
ϕe∗ (R)∨ ∼
= HomA (ϕe∗ R, E) ∼
= Homϕe∗ A (ϕe∗ R, E∗ ) ∼
= ϕe∗ HomA (R, E) ∼
= ϕe∗ ER .
we
all
Thus, taking into account this last chain of isomorphisms one obtains the first desired conclusion.
On the other hand, using once more Lemma 2.12 it turns out that
ϕe (ER )∨ ∼
= HomA (ϕe ER , E) ∼
= Homϕe A (ϕe ER , ϕe E) ∼
= ϕe (E ∨ ) ∼
= ϕe R.
∗
∗
∗
∗
∗
R
∗
∗
Thus, bearing in mind this last chain of isomorphisms it follows that
HomA (ER , ϕe ER )∨ ∼
= HomA (ϕe (ER )∨ , E ∨ ) ∼
= HomA (ϕe R, R),
∗
∗
R
∗
just what we finally wanted to show.
2.1.3. The case of Frobenius homomorphism. Once again we single out the case where A is a ring
of positive characteristic p > 0 and ϕ = F is the Frobenius homomorphism. In this case we adopt
the terminology of pe and p−e -linear maps or, following [And00], Frobenius and Cartier linear
maps.
Definition 2.13. Let M be an A-module and ψ, φ ∈ EndA (M ).
e
(i) We say that ψ is pe -linear provided ψ(am) = ap ψ(m) for any (a, m) ∈ A × M . Equivalently,
ψ ∈ HomA (M, F∗e M ).
e
(ii) We say that φ is p−e -linear provided φ(ap m) = aφ(m) for any (a, m) ∈ A × M . Equivalently,
φ ∈ HomA (F∗e M, M ).
The corresponding rings of pe and p−e -linear maps are defined as follows:
Definition 2.14. Let A be a commutative Noetherian ring of prime characteristic p and let M be
an A-module.
(i) The Frobenius algebra attached to M is the associative, N-graded, not necessarily commutative
ring
M
F M :=
HomA (M, F∗e M ) .
e≥0
(ii) The Cartier algebra attached to M is the associative, N-graded, not necessarily commutative
ring
M
C M :=
HomA (F∗e M, M ) .
e≥0
8
J. ÀLVAREZ MONTANER, A. F. BOIX, AND S. ZARZUELA
In this work we are mainly interested in the case where the Frobenius (resp. Cartier) algebra
of a module is principally generated and thus isomorphic to the left (resp. right) skew polynomial
ring. G. Lyubeznik and K. E. Smith already carried out such an example in [LS01, Example 3.7].
Example 2.15. Let (A, m, K) be a local ring of characteristic p. Then
dim(A)
F Hm
(A)
∼
= S[Θ; F ],
b
where S denotes the S2 -ification of the completion A.
Another source of examples is given by the following result.
Proposition 2.16. Let (A, m, K) be a complete F -finite local ring of characteristic p and EA will
stand for a choice of injective hull of K over A. Then, the following statements hold.
(i) If A is quasi Gorenstein then F EA is principal.
(ii) If A is normal then F EA is principal if and only if A is Gorenstein.
(iii) If A is a Q-Gorenstein normal domain then F EA is a finitely generated A-algebra if and only
if p is relatively prime with the index of A.
(iv) If A is a Q-Gorenstein normal domain then F EA is principal if and only if the index of A
divides p − 1.
dim(A)
Proof. If A is quasi Gorenstein then EA ∼
(A). But we have seen in Example 2.15 that,
= Hm
dim(A)
H
(A)
∼
m
under our assumptions, F
= A[Θ; F ]; indeed, A is complete and any quasi Gorenstein
ring is, in particular, S2 . The second part is proved in [Bli13, Example 2.7]. On the other hand,
part (iii) follows combining [KSSZ14, Proposition 4.3] and [EY16, Theorem 4.5]; finally, part (iv)
also follows from [KSSZ14, Proposition 4.3].
We point out that an explicit description of F EA can be obtained using the following result due
to R. Fedder (cf. [Fed83, pp. 465]).
Theorem 2.17. Let A = K[[x1 , . . . , xd ]] be a formal power series ring over a field K of prime
characteristic p. Let I be an arbitrary ideal of A. Then, one has that
o
Mn
e
e
(I [p ] :A I)/I [p ] Θe ,
F ER ∼
=
e≥0
where E denotes a choice of injective hull of K over A, R := A/I, Θ is the standard Frobenius
action on E and ER := (0 :E I).
In [ÀMBZ12] we used this result to study Frobenius algebras associated to Stanley-Reisner rings.
It turns out that, whenever they are principally generated, they are isomorphic to A[uΘ; F ] with
u = xp−1
· · · xp−1
1
d .
3. The ϕ-Koszul chain complex
Let S = K[x1 , . . . , xn ] be the polynomial ring with coefficients on a commutative ring K, and let
ϕ : S −→ S be a flat map of K–algebras satisfying the extra condition that for any 1 ≤ i ≤ n,
ϕ(xi ) ∈ hxi i. Thus, there are non–zero elements s1 , . . . , sn of S such that ϕ(xi ) = si xi , for each
1 ≤ i ≤ n.
Remark 3.1. If K is a field of positive characteristic p > 0 and ϕ = F e is the iterated Frobenius
e
e
morphism, then this extra condition is naturally satisfied. Indeed, F e (xi ) = xpi so si = xip −1 .
More generally [SW07, Example 2.2], if K is any field and t ≥ 1 is an integer, then the K–linear
map ϕ on S sending each xi to xti also satisfies this condition; indeed, in this case, si = xt−1
i .
KOSZUL COMPLEX OVER SKEW POLYNOMIAL RINGS
9
Our aim is to construct a Koszul complex in the category of S[Θ; ϕ] -modules that we will denote
as ϕ-Koszul complex. To begin with, let us fix some notations. Let
/ Kn
K(x1 , . . . , xn ) := 0
dn
/ Kn−1
/ ...
/ K2
d2
d1
/ K1
/ K0
/0
be the Koszul chain complex of S with respect to x1 , . . . , xn (regarded as a chain complex in the
category of left S-modules) and suppose that each differential dl is represented by right multiplica[ϕ]
tion by matrix Ml . Moreover, for each l ≥ 0 Ml denotes the matrix obtained by applying to each
entry of Ml the map ϕ. In particular, for each entry m, the sign of m is equal to the sign of ϕ(m).
Definition 3.2. We define the ϕ-Koszul chain complex with respect to x1 , . . . , xn as the chain
complex
FK• (x1 , . . . , xn ) := 0
/ FKn+1
∂n+1
Here, for each 0 ≤ l ≤ n + 1,
M
FKl :=
S[Θ; ϕ](ei1 ∧ . . . ∧ eil ) ⊕
1≤i1 <...<il ≤n
∂n
/ FKn
M
∂1
/ ...
/ FK0
/ 0.
S[Θ; ϕ](ej1 ∧ . . . ∧ ejl−1 ∧ u),
1≤j1 <...<jl−1 ≤n
where e1 , . . . , en corresponds respectively to x1 , . . . , xn and u corresponds to Θ − 1. Here, we are
adopting the convention that FK0 := S[Θ; ϕ].
∂l
/ FKl−1 as the unique homomorphism of left S[Θ; ϕ]-modules
Moreover, one defines FKl
which, on basic elements, acts in the following manner:
· Given 1 ≤ i1 < . . . < il ≤ n, set
∂l (ei1 ∧ . . . ∧ eil ) :=
l
X
(−1)r−1 xir ei1 ∧ . . . ∧ eir−1 ∧ eir+1 ∧ eir+2 ∧ . . . ∧ eil .
r=1
· Given 1 ≤ j1 < . . . < jl−1 ≤ n, set
∂l ej1 ∧ . . . ∧ ejl−1 ∧ u := (−1)l−1 Θ − (sj1 · · · sjl−1 ) ej1 ∧ . . . ∧ ejl−1
+
l−1
X
r=1
(−1)r−1 ϕ(xjr ) ej1 ∧ . . . ∧ ejr−1 ∧ ejr+1 ∧ ejr+2 ∧ . . . ∧ ejl−1 ∧ u .
Before going on, we make the following useful:
Discussion 3.3. Given a free, finitely generated left S-module M (whence M is abstractly isomorλM
/ S[Θ; ϕ]⊕r the natural isomorphism
phic to S ⊕r for some r ∈ N), we denote by S[Θ; ϕ] ⊗S S ⊕r
of left S[Θ; ϕ]-modules given by the assignment s ⊗ m 7−→ sm; the reader will easily note that, for
each 0 ≤ l ≤ n, we have the following commutative diagram:
S[Θ; ϕ] ⊗S Kl+1
1S[Θ;ϕ]⊗dl+1
/ S[Θ; ϕ] ⊗S Kl
λKl+1
L
λKl
1≤i1 <...<il+1 ≤n S[Θ; ϕ](ei1
∧ . . . ∧ eil+1 )
d′l+1
/
L
1≤i1 <...<il ≤n S[Θ; ϕ](ei1
Here, d′l+1 denotes the map ∂l+1 restricted to the direct summand
M
S[Θ; ϕ](ei1 ∧ . . . ∧ eil+1 ).
1≤i1 <...<il+1 ≤n
As we shall see quickly, this fact turns out to be very useful in what follows.
∧ . . . ∧ eil ).
10
J. ÀLVAREZ MONTANER, A. F. BOIX, AND S. ZARZUELA
The first thing we have to check out is that FK• (x1 , . . . , xn ) defines a chain complex in the
category of left S[Θ; ϕ]-modules. This fact follows from the next:
Proposition 3.4. For any 0 ≤ l ≤ n, one has that ∂l ∂l+1 = 0.
Proof. Regarding the very definition of the ∂’s, we only have to distinguish two cases.
· Given 1 ≤ i1 < . . . < il+1 ≤ n one has, keeping in mind Discussion 3.3, that
∂l ∂l+1 ei1 ∧ . . . ∧ eil+1 = d′l d′l+1 ei1 ∧ . . . ∧ eil+1
−1
= λKl−1 ◦ 1S[Θ;ϕ] ⊗ dl−1 ◦ λ−1
◦
λ
◦
◦
λ
⊗
d
ei1 ∧ . . . ∧ eil+1
1
Kl
l
S[Θ;ϕ]
Kl
Kl+1
ei1 ∧ . . . ∧ eil+1
= λKl−1 ◦ 1S[Θ;ϕ] ⊗ (dl ◦ dl+1 ) ◦ λ−1
Kl+1
= λKl−1 ◦ 1S[Θ;ϕ] ⊗ 0 ◦ λ−1
ei1 ∧ . . . ∧ eil+1 = 0;
Kl+1
indeed, notice that dl dl+1 = 0 because they are the usual chain differentials in the Koszul chain
complex of S with respect to x1 , . . . , xn .
· Given 1 ≤ j1 < . . . < jl ≤ n, one has that
∂l (∂l+1 (ej1 ∧ . . . ∧ ejl ∧ u)) = ∂l (−1)l (Θ − (sj1 · · · sjl )) (ej1 ∧ . . . ∧ ejl ) +
!
l
X
(−1)r−1 ϕ(xjr ) ej1 ∧ . . . ∧ ejr−1 ∧ ejr+1 ∧ ejr+2 ∧ . . . ∧ ejl ∧ u
=
r=1
l
X
(−1)r+l−1 (ϕ(xjr )Θ − (sj1 · · · sjl )xjr ) ej1 ∧ . . . ∧ ejr−1 ∧ ejr+1 ∧ ejr+2 ∧ . . . ∧ ejl +
r=1
l X
r−1
X
(−1)r+k−2 ϕ(xjk xjr ) ej1 ∧ . . . ∧ ejk−1 ∧ ejk+1 ∧ ejk+2 ∧ . . . ∧ ejr−1 ∧ ejr+1 ∧ ejr+2 ∧ . . . ∧ ejl ∧ u +
r=1 k=1
l X
l
X
(−1)r+k−2 ϕ(xjk xjr ) ej1 ∧ . . . ∧ ejr−1 ∧ ejr+1 ∧ ejr+2 ∧ . . . ∧ ejk−1 ∧ ejk+1 ∧ ejk+2 ∧ . . . ∧ ejl ∧ u +
r=1 k=r
l
X
(−1)l+r−2 ϕ(xjr )Θ − (sj1 · · · sjr−1 sjr+1 sjr+2 · · · sjl )ϕ(xjr ) ej1 ∧ . . . ∧ ejr−1 ∧ ejr+1 ∧ ejr+2 ∧ . . . ∧ ejl .
r=1
Starting from the top, the first summand (respectively, the second summand) cancels out the
fourth summand (respectively, the third summand) and therefore the whole expression vanishes,
just what we finally wanted to check.
Remark 3.5. The differentials of the ϕ-Koszul complex
FK• (x1 , . . . , xn ) := 0
/ FKn+1
∂n+1
/ FKn
∂n
/ ...
∂1
/ FK0
/ 0.
are described as follows:
(1) ∂n+1 is represented by right multiplication by matrix (−1)n (Θ − (s1 · · · sn )) Mn[ϕ] .
(2) For each 1 ≤ l ≤ n − 1, ∂l+1 is represented by right multiplication by matrix
Ml+1
0
,
[ϕ]
(−1)l Dl Ml
where Dl is a diagonal matrix with non-zero entries Θ − (si1 · · · sil ), 1 ≤ i1 < . . . < il ≤ n.
T
(3) ∂1 is represented by right multiplication by matrix x1 . . . xn Θ − 1 .
KOSZUL COMPLEX OVER SKEW POLYNOMIAL RINGS
11
For example, when n = 2, FK• (x1 , x2 ) boils down to the chain complex
0
/ S[Θ; ϕ]
∂3
/ S[Θ; ϕ]⊕3
∂2
/ S[Θ; ϕ]⊕3
∂1
/ S[Θ; ϕ]
where ∂3 , ∂2 and ∂1 are given by right multiplication by matrices
0
−x2
x1
s1 − Θ
Θ − (s1 s2 ) −ϕ(x2 ) ϕ(x1 )
0
ϕ(x1 )
0
s2 − Θ ϕ(x2 )
/ 0,
x1
x2
Θ−1
Now, we want to single out the following technical fact because it will play a key role during the
proof of the first main result of this paper (see Theorem 3.10).
Lemma 3.6. Preserving the notations introduced in Remark 3.5, one has, for any 0 ≤ l ≤ n, that
[ϕ]
Ml Dl−1 = Dl Ml .
Proof. Fix 0 ≤ l ≤ n. Proposition 3.4 implies that ∂l ∂l+1 = 0. Then, according to Remark 3.5, this
equality corresponds to
!
Ml
0
0
Ml+1
= 0.
[ϕ]
[ϕ]
(−1)l−1 Dl−1 Ml−1
(−1)l Dl Ml
[ϕ]
In particular, we must have (−1)l Dl Ml + (−1)l−1 Ml Dl−1 = 0, which is equivalent to say that
[ϕ]
(−1)l Dl Ml − Ml Dl−1 = 0.
[ϕ]
Whence Ml Dl−1 = Dl Ml , just what we finally wanted to show.
Before showing our main result, we want to establish a certain technical fact, which is interesting
in its own right; namely:
Proposition 3.7. S[Θ; ϕ] is a flat right S-module.
Proof. By the very definition of left skew polynomial rings,
M
S[Θ; ϕ] =
SΘe .
e≥0
Since a direct sum of right S–modules is flat if and only if it is so all its direct summands [Rot09,
Proposition 3.46 (ii)], it is enough to check that, for any e ≥ 0, SΘe is a flat right S-module.
Fix e ≥ 0. Firstly, albeit the notation SΘe might suggest that it is just a left S-module, this is
not the case because SΘe can be identified with Θe ϕe∗ S; from this point of view, it is clear that SΘe
may be also regarded as a right S-module. Therefore, keeping in mind the previous identification
/ ϕe S given by the assignment Θe ϕe s 7−→ ϕe s defines an abstract
one has that the map Θe ϕe∗ S
∗
∗
∗
isomorphism of right S-modules, whence SΘe is (abstractly) isomorphic to ϕe∗ S in the category of
right S-modules and then the result follows from the fact that ϕe∗ S is a flat right S-module because
of the flatness of ϕ; the proof is therefore completed.
Remark 3.8. When S is a commutative Noetherian regular local ring of prime characteristic p and
F is the Frobenius map on S, the fact that S[Θ; F ] is a flat right S–module was already observed
by Y. Yoshino [Yos94, Proof of Example (9.2)].
Next result provides some useful properties of FK• (x1 , . . . , xn ).
Proposition 3.9. H0 (FK• (x1 , . . . , xn )) ∼
= S/In as left S[Θ; ϕ]-modules, where In = hx1 , . . . , xn i;
moreover, Hn+1 (FK• (x1 , . . . , xn )) = 0.
12
J. ÀLVAREZ MONTANER, A. F. BOIX, AND S. ZARZUELA
Proof. First we notice that, as in the case of the Frobenius morphism studied by Y. Yoshino (see
S[Θ;ϕ]
∼
Proposition 2.5), we have S[Θ;ϕ](Θ−1)
= S. Then, using Remark 3.5 it follows that
H0 (FK• (x1 , . . . , xn )) =
S[Θ; ϕ] ∼
S[Θ; ϕ]
∼
=
= S/In .
Im(∂1 )
S[Θ; ϕ]In + S[Θ; ϕ](Θ − 1)
Now, consider the composition
S[Θ; ϕ]
∂n+1
π
/ S[Θ; ϕ] ⊕ S[Θ; ϕ]⊕n
/ S[Θ; ϕ]⊕n ,
where π denotes the corresponding projection. In this way, we have that π∂n+1 turns out to be,
[ϕ]
up to isomorphisms, 1S[Θ;ϕ] ⊗ dn (indeed, this fact follows directly from the commutative square
[ϕ]
established in Discussion 3.3); regardless, since dn is an injective homomorphism between free left
[ϕ]
S-modules, and S[Θ; ϕ] is a flat right S-module (cf. Proposition 3.7), one has that 1S[Θ;ϕ] ⊗ dn is
an injective homomorphism between free left S[Θ; ϕ]-modules. Therefore, π∂n+1 is also an injective
homomorphism, whence ∂n+1 is so. This fact concludes the proof.
Now, we state and prove the first main result of this paper, which is the following:
Theorem 3.10. The ϕ-Koszul complex FK• (x1 , . . . , xn ) provides a free resolution of S/In in the
category of left S[Θ; ϕ]-modules.
Proof. By Proposition 3.9, it is enough to check, for any 1 ≤ l ≤ n, that Hl (FK• (x1 , . . . , xn )) = 0.
So, fix 1 ≤ l ≤ n. Our goal is to show that ker(∂l ) ⊆ Im(∂l+1 ); in other words, we have to prove
∂l+1
that the chain complex FKl+1
FKl = Kl′ ⊕ Kl′′ , where
/ FKl
M
Kl′ :=
∂l
/ FKl−1 is midterm exact. First of all, remember that
S[Θ; ϕ](ei1 ∧ . . . ∧ eil ), and
1≤i1 <...<il ≤n
M
Kl′′ :=
S[Θ; ϕ](ej1 ∧ . . . ∧ ejl−1 ∧ u).
1≤j1 <...<jl−1 ≤n
Furthermore, Discussion 3.3 implies that the chain complexes
K•′ : 0
/ K′
n
d′n
/ K′
n−1
/ ...
/ K′
2
d′2
/ K′
1
d′1
/ K′
0
/0
and
(K•′ )[ϕ] : 0
/ K′
n
(d′n )[ϕ]
/ K′
n−1
/ ...
/ K′
2
(d′2 )[ϕ]
/ K′
1
(d′1 )[ϕ]
/ K′
0
/0
[ϕ]
are respectively canonically isomorphic to S[Θ; ϕ] ⊗S K• and S[Θ; ϕ] ⊗S K• ; in particular, since
S[Θ; ϕ] is a flat right S-module (cf. Proposition 3.7), K•′ and (K•′ )[ϕ] are both acyclic chain complexes
in the category of left S[Θ; ϕ]-modules. On the other hand, we also have to keep in mind that d′l and
[ϕ]
(d′l )[ϕ] are respectively represented by right multiplication by matrix Ml and Ml (cf. Proposition
3.9 and its corresponding notation).
Now, let P ∈ ker(∂l ) ⊆ FKl . Since FKl = Kl′ ⊕ Kl′′ , we may write P = (P ′ , P ′′ ) for certain
P ′ ∈ Kl′ and P ′′ ∈ Kl′′ ; in this way, as P ∈ ker(∂l ) it follows that
!
M
0
l
′ M + P ′′ (−1)l−1 D
′′ M [ϕ] ,
0 0 = P ′ P ′′
=
P
P
[ϕ]
l
l−1
l−1
(−1)l−1 Dl−1 Ml−1
which leads to the following system of equations:
[ϕ]
P ′ Ml + P ′′ (−1)l−1 Dl−1 = 0, P ′′ Ml−1 = 0.
KOSZUL COMPLEX OVER SKEW POLYNOMIAL RINGS
13
[ϕ]
In particular, since P ′′ Ml−1 = 0 one has that P ′′ ∈ ker((d′l−1 )[ϕ] ) = Im((d′l )[ϕ] ); therefore, there is
[ϕ]
Q′′ ∈ Kl′′ such that Q′′ Ml
= P ′′ . Using this fact, it follows that
[ϕ]
P ′ Ml + Q′′ (−1)l−1 Ml Dl−1 = 0.
[ϕ]
Regardless, Lemma 3.6 tells us that Ml Dl−1 = Dl Ml , whence
P ′ Ml + Q′′ (−1)l−1 Dl Ml = 0,
which is equivalent to say that P ′ + Q′′ (−1)l−1 Dl Ml = 0. In this way, one has that
P ′ + Q′′ (−1)l−1 Dl ∈ ker(d′l ) = Im(d′l+1 )
′
such that Q′ Ml+1 = P ′ + Q′′ (−1)l−1 Dl .
and therefore there exists Q′ ∈ Kl+1
′
′′
Summing up, setting Q := (Q , Q ) ∈ FKl+1 , it follows that
Ml+1
0
′
′′
′M
′′ (−1)l D
′′ M [ϕ]
Q Q
=
Q
+
Q
Q
[ϕ]
l+1
l
l
(−1)l Dl Ml
= P ′ + Q′′ (−1)l−1 Dl + Q′′ (−1)l Dl P ′′ = P ′ P ′′
and therefore we can conclude that P ∈ Im(∂l+1 ), which is exactly what we wanted to show.
The reader will easily note that, during the proof, we have obtained the below result; we specially
thank Rishi Vyas to single out to us this fact.
Proposition 3.11. There is a short exact sequence of chain complexes
0 → S[Θ; ϕ]⊗S K• (S; x1 , . . . , xn ) → FK• (x1 , . . . , xn ) → (S[Θ; ϕ] ⊗S K• (S; ϕ(x1 ), . . . , ϕ(xn ))) [1] → 0
in the category of left S[Θ; ϕ]–modules.
3.1. The ϕ-Koszul chain complex in full generality. Our next aim is to define the ϕ-Koszul
chain complex over a more general setting, and explore some specific situations on which we can
ensure that defines a finite free resolution. Let A be a commutative Noetherian ring containing a
commutative subring K, and y1 , . . . , yn denote arbitrary elements of A; in addition, assume that
we fix a flat endomorphism of K–algebras S := K[x1 , . . . , xn ]
ϕ
/ S satisfying ϕ(xi ) ∈ hxi i for any
ψ
ψ
/ A, where S
/ A is the
1 ≤ i ≤ n. In this section, we regard A as an S-algebra under S
natural homomorphism of K-algebras which sends each xi to yi . Finally, we suppose that there
exists a K–algebra homomorphism A
Φ
/ A making the square
S
ψ
/A
ϕ
Φ
S
ψ
/A
commutative.
Definition 3.12. We define the ϕ-Koszul chain complex of A with respect to y1 , . . . , yn as the
chain complex FK• (y1 , . . . , yn ; A) := A ⊗S FK• (x1 , . . . , xn ).
Under slightly different assumptions, we want to show that FK• (y1 , . . . , yn ; A) still defines a finite
free resolution; with this purpose in mind, first of all we review the following notion, which was
introduced independently in [Bou07, Definition 2 of page 157] (using a different terminology) and
by T. Kabele in [Kab71, Definition 2].
14
J. ÀLVAREZ MONTANER, A. F. BOIX, AND S. ZARZUELA
Definition 3.13. Let R be a commutative ring, let s ∈ N, and let f1 , . . . , fn be a sequence of
elements in R. It is said that f1 , . . . , fn is a Koszul regular sequence provided the Koszul chain
complex K• (f1 , . . . , fn ; R) provides a free resolution of R/In , where In = hf1 , . . . , fn i.
Next statement may be regarded as a generalization of Theorem 3.10; this is the main result of
this section.
Theorem 3.14. Let A be a commutative Noetherian ring containing a commutative subring K, and
let y1 , . . . , yn be a sequence of elements in A. Moreover, we assume that Φ is flat and y1 , . . . , yn is
an A-Koszul regular sequence. Then, FK• (y1 , . . . , yn ; A) defines a finite free resolution of A/In in
the category of left A[Θ; Φ]-modules, where In = hy1 , . . . , yn i.
Proof. The proof of this result is, mutatis mutandi, the same as the one of Theorem 3.10 replacing S
by A and x1 , . . . , xn by y1 , . . . , yn . Indeed, a simple inspection of the proof of Theorem 3.10 reveals
that we only used there the flatness of ϕ and the fact that the Koszul chain complex K• (x1 , . . . , xn )
defines a finite free resolution of K; the proof is therefore completed.
Remark 3.15. The global homological dimension of right skew polynomial rings was studied by
K. L. Fields in [Fie69, Fie70]. An upper bound for this global dimension is n + 1 when ϕ is injective
and it is exactly n + 1 in the case that ϕ is an automorphism. Passing to the opposite ring (see
Remark 2.2) we would get analogous results for left skew polynomial rings. Notice that the length
of the ϕ-Koszul complex FK• (y1 , . . . , yn ; A) is n + 1.
3.2. The case of the Frobenius homomorphism. For the convenience of the reader, we will
specialize the construction of the ϕ-Koszul complex to the case where A is a commutative Noetherian regular ring of positive characteristic p > 0, and ϕ = F e is an e-th iteration of the Frobenius
morphism. In this case, we will denote this complex simply as Frobenius-Koszul complex. Namely
we have:
FK• (x1 , . . . , xn ) := 0
/ FKn+1
∂n+1
where, for each 0 ≤ l ≤ n + 1,
M
FKl :=
A[Θ; F e ](ei1 ∧ . . . ∧ eil ) ⊕
1≤i1 <...<il ≤n
and the differentials FKl
∂l
/ FKn
∂n
M
/ ...
∂1
/ FK0
/ 0.
A[Θ; ϕ](ej1 ∧ . . . ∧ ejl−1 ∧ u),
1≤j1 <...<jl−1 ≤n
/ FKl−1 are given by:
· For 1 ≤ i1 < . . . < il ≤ n, set
∂l (ei1 ∧ . . . ∧ eil ) :=
l
X
(−1)r−1 xir ei1 ∧ . . . ∧ eir−1 ∧ eir+1 ∧ eir+2 ∧ . . . ∧ eil .
r=1
· For 1 ≤ j1 < . . . < jl−1 ≤ n, set
e
e −1
) ej1 ∧ . . . ∧ ejl−1
∂l ej1 ∧ . . . ∧ ejl−1 ∧ u := (−1)l−1 Θ − (xjp1 −1 · · · xjpl−1
+
l−1
X
e
(−1)r−1 xpjr ej1 ∧ . . . ∧ ejr−1 ∧ ejr+1 ∧ ejr+2 ∧ . . . ∧ ejl−1 ∧ u .
r=1
Remark 3.16. For n = 2, FK• (x1 , x2 ) is just
0
/ A[Θ; F e ]
∂3
/ A[Θ; F e ]⊕3
∂2
/ A[Θ; F e ]⊕3
∂1
/ A[Θ; F e ]
/ 0,
KOSZUL COMPLEX OVER SKEW POLYNOMIAL RINGS
where ∂3 , ∂2 and ∂1 are given by right multiplication by matrices
0
−x2
x1
e
e
e
e
e
e
x1p −1 − Θ
0
xp1
Θ − (x1p −1 x2p −1 ) −xp2 xp1
e
e
0
x2p −1 − Θ xp2
15
x1
x2
Θ−1
As a direct consequence of Theorem 3.14 we obtain the below:
Corollary 3.17. Let A be a commutative Noetherian regular ring of prime characteristic p, and let
y1 , . . . , yn be an A–Koszul regular sequence. Then, FK• (y1 , . . . , yn ; A) defines a finite free resolution
of A/In in the category of left A[Θ; F e ]-modules, where In = hy1 , . . . , yn i.
Acknowledgements. The second author thanks Rishi Vyas for fruitful discussions about the
material presented here.
References
[ÀMBZ12] J. Àlvarez Montaner, A. F. Boix, and S. Zarzuela. Frobenius and Cartier algebras of Stanley-Reisner
rings. J. Algebra, 358:162–177, 2012. 8
[And00]
G. W. Anderson. An elementary approach to L-functions mod p. J. Number Theory, 80(2):291–303, 2000.
7
[BB11]
M. Blickle and G. Böckle. Cartier modules: finiteness results. J. Reine Angew. Math., 661:85–123, 2011.
1, 6
[Bli13]
M. Blickle. Test ideals via algebras of p−e -linear maps. J. Algebraic Geom., 22(1):49–83, 2013. 1, 5, 8
[Bou07]
N. Bourbaki. Éléments de mathématique. Algèbre. Chapitre 10. Algèbre homologique. Springer-Verlag,
Berlin, 2007. Reprint of the 1980 original. 13
[BS13]
M. Blickle and K. Schwede. p−1 -linear maps in algebra and geometry. In Commutative algebra, pages
123–205. Springer, New York, 2013. 1, 5
[EH08]
F. Enescu and M. Hochster. The Frobenius structure of local cohomology. Algebra Number Theory,
2(7):721–754, 2008. 6
[Ene12]
F. Enescu. Finite-dimensional vector spaces with Frobenius action. In Progress in commutative algebra 2,
pages 101–128. Walter de Gruyter, Berlin, 2012. 1
[EY16]
F. Enescu and Y. Yao. The Frobenius complexity of a local ring of prime characteristic. J. Algebra,
459:133–156, 2016. 8
[Fed83]
R. Fedder. F -purity and rational singularity. Trans. Amer. Math. Soc., 278(2):461–480, 1983. 8
[Fie69]
K. L. Fields. On the global dimension of skew polynomial rings. J. Algebra, 13:1–4, 1969. 14
[Fie70]
K. L. Fields. On the global dimension of skew polynomial rings—An addendum. J. Algebra, 14:528–530,
1970. 14
[GW04]
K. R. Goodearl and R. B. Warfield, Jr. An introduction to noncommutative Noetherian rings, volume 61
of London Mathematical Society Student Texts. Cambridge University Press, Cambridge, second edition,
2004. 1, 2
[HH90]
M. Hochster and C. Huneke. Tight closure, invariant theory, and the Briançon-Skoda theorem. J. Amer.
Math. Soc., 3(1):31–116, 1990. 1
[Kab71]
T. Kabele. Regularity conditions in nonnoetherian rings. Trans. Amer. Math. Soc., 155:363–374, 1971. 13
[KSSZ14] M. Katzman, K. Schwede, A. K. Singh, and W. Zhang. Rings of Frobenius operators. Math. Proc. Cambridge Philos. Soc., 157(1):151–167, 2014. 8
[LS01]
G. Lyubeznik and K. E. Smith. On the commutation of the test ideal with localization and completion.
Trans. Amer. Math. Soc., 353(8):3149–3180 (electronic), 2001. 1, 5, 6, 8
[Lyu97]
G. Lyubeznik. F -modules: applications to local cohomology and D-modules in characteristic p > 0. J.
Reine Angew. Math., 491:65–130, 1997. 1
[MR01]
J. C. McConnell and J. C. Robson. Noncommutative Noetherian rings, volume 30 of Graduate Studies in
Mathematics. American Mathematical Society, Providence, RI, revised edition, 2001. With the cooperation
of L. W. Small. 1, 2
[Rot09]
J. J. Rotman. An introduction to homological algebra. Universitext. Springer, New York, second edition,
2009. 11
[Sch11]
K. Schwede. Test ideals in non-Q-Gorenstein rings. Trans. Amer. Math. Soc., 363(11):5925–5941, 2011. 1,
5
[Sha09]
R. Y. Sharp. Graded annihilators and tight closure test ideals. J. Algebra, 322(9):3410–3426, 2009. 1
16
[SW07]
[SY11]
[Yos94]
J. ÀLVAREZ MONTANER, A. F. BOIX, AND S. ZARZUELA
A. K. Singh and U. Walther. Local cohomology and pure morphisms. Illinois J. Math., 51(1):287–298,
2007. 2, 8
R. Y. Sharp and Y. Yoshino. Right and left modules over the Frobenius skew polynomial ring in the
F-finite case. Math. Proc. Cambridge Philos. Soc., 150(3):419–438, 2011. 6
Y. Yoshino. Skew-polynomial rings of Frobenius type and the theory of tight closure. Comm. Algebra,
22(7):2473–2502, 1994. 1, 4, 6, 11
Departament de Matemàtiques, Universitat Politècnica de Catalunya, Avinguda Diagonal 647,
Barcelona 08028, SPAIN
E-mail address: [email protected]
Department of Mathematics, Ben-Gurion University of the Negev, P.O.B. 653 Beer-Sheva 84105,
Israel.
E-mail address: [email protected]
Departament d’Àlgebra i Geometria, Universitat de Barcelona, Gran Via de les Corts Catalanes
585, Barcelona 08007, SPAIN
E-mail address: [email protected]
| 0math.AC
|
arXiv:cs/0601037v1 [cs.CL] 10 Jan 2006
Under consideration for publication in Theory and Practice of Logic Programming
1
Constraint-based Automatic Verification of
Abstract Models of Multithreaded Programs
GIORGIO DELZANNO
Dipartimento di Informatica e Scienze dell’Informazione, Università di Genova
via Dodecaneso 35, 16146 Genova - Italy
(e-mail: [email protected])
submitted 17 December 2003; revised 13 April 2005; accepted 15 January 2006
Abstract
We present a technique for the automated verification of abstract models of multithreaded
programs providing fresh name generation, name mobility, and unbounded control.
As high level specification language we adopt here an extension of communication finitestate machines with local variables ranging over an infinite name domain, called TDL
programs. Communication machines have been proved very effective for representing communication protocols as well as for representing abstractions of multithreaded software.
The verification method that we propose is based on the encoding of TDL programs
into a low level language based on multiset rewriting and constraints that can be viewed as
an extension of Petri Nets. By means of this encoding, the symbolic verification procedure
developed for the low level language in our previous work can now be applied to TDL
programs. Furthermore, the encoding allows us to isolate a decidable class of verification
problems for TDL programs that still provide fresh name generation, name mobility, and
unbounded control. Our syntactic restrictions are in fact defined on the internal structure
of threads: In order to obtain a complete and terminating method, threads are only allowed
to have at most one local variable (ranging over an infinite domain of names).
KEYWORDS: Constraints, Multithreaded Programs, Verification.
1 Introduction
Andrew Gordon (Gordon 2001) defines a nominal calculus to be a computational
formalism that includes a set of pure names and allows the dynamic generation of
fresh, unguessable names. A name is pure whenever it is only useful for comparing
for identity with other names. The use of pure names is ubiquitous in programming
languages. Some important examples are memory pointers in imperative languages,
identifiers in concurrent programming languages, and nonces in security protocols.
In addition to pure names, a nominal process calculus should provide mechanisms
for concurrency and inter-process communication. A computational model that provides all these features is an adequate abstract formalism for the analysis of multithreaded and distributed software.
2
Giorgio Delzanno
The Problem Automated verification of specifications in a nominal process calculus
becomes particularly challenging in presence of the following three features: the possibility of generating fresh names (name generation); the possibility of transmitting
names (name mobility); the possibility of dynamically adding new threads of control
(unbounded control). In fact, a calculus that provides all the previous features can
be used to specify systems with a state-space infinite in several dimensions. This
feature makes difficult (if not impossible) the application of finite-state verification
techniques or techniques based on abstractions of process specifications into Petri
Nets or CCS-like models. In recent years there have been several attempts of extending automated verification methods from finite-state to infinite-state systems
(Abdulla and Nylén 2000; Kesten et al. 2001). In this paper we are interested in investigating the possible application of the methods we proposed in (Delzanno 2001)
to verification problems of interest for nominal process calculi.
Constraint-based Symbolic Model Checking In (Delzanno 2001) we introduced a
specification language, called MSR(C), for the analysis of communication protocols
whose specifications are parametric in several dimensions (e.g. number of servers,
clients, and tickets as in the model of the ticket mutual exclusion algorithm shown
in (Bozzano and Delzanno 2002)). MSR(C) combines multiset rewriting over first
order atomic formulas (Cervesato et al. 1999) with constraints programming. More
specifically, multiset rewriting is used to specify the control part of a concurrent
system, whereas constraints are used to symbolically specify the relations over local data. The verification method proposed in (Delzanno 2005) allows us to symbolically reason on the behavior of MSR(C) specifications. To this aim, following
(Abdulla et al. 1996; Abdulla and Nylén 2000) we introduced a symbolic representation of infinite collections of global configurations based on the combination of
multisets of atomic formulas and constraints, called constrained configurations.1
The verification procedure performs a symbolic backward reachability analysis by
means of a symbolic pre-image operator that works over constrained configurations
(Delzanno 2005). The main feature of this method is the possibility of automatically
handling systems with an arbitrary number of components. Furthermore, since we
use a symbolic and finite representation of possibly infinite sets of configurations,
the analysis is carried out without loss of precision.
A natural question for our research is whether and how these techniques can be
used for verification of abstract models of multithreaded programs.
Our Contribution In this paper we propose a sound, and fully automatic verification
method for abstract models of multithreaded programs that provide name generation, name mobility, and unbounded control. As a high level specification language
we adopt here an extension with value-passing of the formalism of (Ball et al. 2001)
1
Notice that in (Abdulla et al. 1996; Abdulla and Nylén 2000) a constraint denotes a symbolic
state whereas we use the word constraint to denote a symbolic representation of the relation of
data variables (e.g. a linear arithmetic formula) used as part of the symbolic representation of
sets of states (a constrained configuration).
Constraint-based Verification of Abstract Multithreaded Programs
3
based on families of state machines used to specify abstractions of multithreaded
software libraries. The resulting language is called Thread Definition Language
(TDL). This formalism allows us to keep separate the finite control component of a
thread definition from the management of local variables (that in our setting range
over a infinite set of names), and to treat in isolation the operations to generate
fresh names, to transmit names, and to create new threads. In the present paper we
will show that the extension of the model of (Ball et al. 2001) with value-passing
makes the model Turing equivalent.
The verification methodology is based on the encoding of TDL programs into a
specification in the instance MSRN C of the language scheme MSR(C) of(Delzanno 2001).
MSRN C is obtained by taking as constraint system a subclass of linear arithmetics
with only = and > relations between variables, called name constraints (N C). The
low level specification language MSRN C is not just instrumental for the encoding
of TDL programs. Indeed, it has been applied to model consistency and mutual
exclusion protocols in (Bozzano and Delzanno 2002; Delzanno 2005). Via this encoding, the verification method based on symbolic backward reachability obtained
by instantiating the general method for MSR(C) to NC-constraints can now be applied to abstract models of multithreaded programs. Although termination is not
guaranteed in general, the resulting verification method can succeed on practical
examples as the Challenge-Response TDL program defined over binary predicates
we will illustrated in the present paper. Furthermore, by propagating the sufficient
conditions for termination defined in (Bozzano and Delzanno 2002; Delzanno 2005)
back to TDL programs, we obtain an interesting class of decidable problems for abstract models of multithreaded programs still providing name generation, name
mobility, and unbounded control.
Plan of the Paper In Section 2 we present the Thread Definition Language (TDL)
with examples of multithreaded programs. Furthermore, we discuss the expressiveness of TDL programs showing that they can simulate Two Counter Machines. In
Section 3, after introducing the MSRN C formalism, we show that TDL programs
can be simulated by MSRN C specifications. In Section 4 we show how to transfer
the verification methods developed for MSR(C) to TDL programs. Furthermore, we
show that safety properties can be decided for the special class of monadic TDL
programs. In Section 5 we address some conclusions and discuss related work.
2 Thread Definition Language (TDL)
In this section we will define TDL programs. This formalism is a natural extension
with value-passing of the communicating machines used by (Ball et al. 2001) to
specify abstractions of multithreaded software libraries.
Terminology Let N be a denumerable set of names equipped with the relations =
and 6= and a special element ⊥ such that n 6= ⊥ for any n ∈ N . Furthermore, let
V be a denumerable set of variables, C = {c1 , . . . , cm } a finite set of constants, and
L a finite set of internal action labels. For a fixed V ⊆ V, the set of expressions is
4
Giorgio Delzanno
defined as E = V ∪ C ∪ {⊥} (when necessary we will use E(V ) to explicit the set of
variables V upon which expressions are defined). The set of channel expressions is
defined as Ech = V ∪ C. Channel expressions will be used as synchronization labels
so as to establish communication links only at execution time.
A guard over V is a conjunction γ1 , . . . , γs , where γi is either true, x = e or x 6= e
with x ∈ V and e ∈ E for i : 1, . . . , s. An assignment α from V to W is a conjunction
like xi := ei where xi ∈ W , ei ∈ E(V ) for i : 1, . . . k and xr 6= xs for r 6= s. A
message template m over V is a tuple m = hx1 , . . . , xu i of variables in V .
Definition 1
A TDL program is a set T = {P1 , . . . , Pt } of thread definitions (with distinct names
for local variables control locations). A thread definition P is a tuple hQ, s0 , V, Ri,
where Q is a finite set of control locations, s0 ∈ Q is the initial location, V ⊆ V is
a finite set of local variables, and R is a set of rules. Given s, s′ ∈ Q, and a ∈ L, a
rule has one of the following forms2 :
a
• Internal move: s −−→ s′ [γ, α], where γ is a guard over V , and α is an assignment from V to V ;
a
• Name generation: s −−→ s′ [x := new], where x ∈ V , and the expression new
denotes a fresh name;
a
• Thread creation: s −−→ s′ [run P ′ with α], where P ′ = hQ′ , t, W, R′ i ∈ T , and
α is an assignment from V to W that specifies the initialization of the local
variables of the new thread;
e!m
• Message sending: s −−→ s′ [γ, α], where e is a channel expression, m is a message template over V that specify which names to pass, γ is a guard over V ,
and α is an assignment from V to V .
e?m
• Message reception: s −−−→ s′ [γ, α], where e is a channel expression, m is a
message template over a new set of variables V ′ (V ′ ∩ V = ∅) that specifies
the names to receive, γ is a guard over V ∪ V ′ and α is an assignment from
V ∪ V ′ to V .
Before giving an example, we will formally introduce the operational semantics of
TDL programs.
2.1 Operational Semantics
In the following we will use N to indicate the subset of used names of N . Every
constant c ∈ C is mapped to a distinct name nc 6= ⊥ ∈ N , and ⊥ is mapped to ⊥.
Let P = hQ, s, V, Ri and V = {x1 , . . . , xk }. A local configuration is a tuple p =
hs′ , n1 , . . . , nk i where s′ ∈ Q and ni ∈ N is the current value of the variable xi ∈ V
for i : 1, . . . , k.
A global configuration G = hN, p1 , . . . , pm i is such that N ⊆ N and p1 , . . . , pm are
local configurations defined over N and over the thread definitions in T . Note that
2
In this paper we keep assignments, name generation, and thread creation separate in order to
simplify the presentation of the encoding into MSR.
Constraint-based Verification of Abstract Multithreaded Programs
5
there is no relation between indexes in a global configuration in G and in T ; G is
a pool of active threads, and several active threads can be instances of the same
thread definition.
Given a local configuration p = hs′ , n1 , . . . , nk i, we define the valuation ρp as
ρp (xi ) = ni if xi ∈ V , ρp (c) = nc if c ∈ C, and ρp (⊥) = ⊥. Furthermore, we
say that ρp satisfies the guard γ if ρp (γ) ≡ true, where ρp is extended to constraints in the natural way (ρp (ϕ1 ∧ ϕ2 ) = ρp (ϕ1 ) ∧ ρp (ϕ2 ), etc.).
The execution of x := e has the effect of updating the local variable x of a thread
with the current value of e (a name taken from the set of used values N ). On the contrary, the execution of x := new associates a fresh unused name to x. The formula
run P with α has the effect of adding a new thread (in its initial control location)
to the current global configuration. The initial values of the local variables of the
generated thread are determined by the execution of α whose source variables are
the local variables of the parent thread. The channel names used in a rendez-vous
are determined by evaluating the channel expressions tagging sender and receiver
rules. Value passing is achieved by extending the evaluation associated to the current configuration of the receiver so as to associate the output message of the sender
to the variables in the input message template. The operational semantics is given
via a binary relation ⇒ defined as follows.
Definition 2
Let G = hN, . . . , p, . . .i, and p = hs, n1 , . . . , nk i be a local configuration for P =
hQ, s, V, Ri, V = {x1 , . . . , xk }, then:
a
• If there exists a rule s −−→ s′ [γ, α] in R such that ρp satisfies γ, then G ⇒
hN, . . . , p′ , . . .i (meaning that only p changes) where p′ = hs′ , n′1 , . . . , n′k i,
n′i = ρp (ei ) if xi := ei is in α, n′i = ni otherwise, for i : 1, . . . , k.
a
• If there exists a rule s −−→ s′ [xi := new] in R, then G ⇒ hN ′ , . . . , p′ , . . .i
′
′
′
′
where p = hs , n1 , . . . , nk i, ni is an unused name, i.e., n′i ∈ N \ N , n′j = nj
for every j 6= i, and N ′ = N ∪ {n′i };
a
• If there exists a rule s −−→ s′ [run P ′ with α] in R with P ′ = hQ′ , t0 , W, R′ i,
W = {y1 , . . . , yu }, and α is defined as y1 := e1 , . . . , yu := eu then G ⇒
hN, . . . , p′ , . . . , qi (we add a new thread whose initial local configuration is q)
where p′ = hs′ , n1 , . . . , nk i, and q = ht0 , ρp (e1 ), . . . , ρp (eu )i.
• Let q = ht, m1 , . . . , mr i (distinct from p) be a local configuration in G associated with P ′ = hQ′ , t0 , W, R′ i.
e!m
e′ ?m′
Let s −−→ s′ [γ, α] in R and t −−−−→ t′ [γ ′ , α′ ] in R′ be two rules such that
m = hx1 , . . . , xu i, m′ = hy1 , . . . , yv i and u = v (message templates match).
We define σ as the value passing evaluation σ(yi ) = ρp (xi ) for i : 1, . . . , u,
and σ(z) = ρq (z) for z ∈ W ′ .
Now if ρp (e) = ρp (e′ ) (channel names match), ρp satisfies γ, and that σ
satisfies γ ′ , then hN, . . . , p, . . . , q, . . .i ⇒ hN, . . . , p′ , . . . , q′ , . . .i where p′ =
hs′ , n′1 , . . . , n′k i, n′i = ρp (v) if xi := v is in α, n′i = ni otherwise for i : 1, . . . , k;
q′ = ht′ , m′1 , . . . , m′r i, m′i = σ(v) if ui := v is in α′ , m′i = mi otherwise for
i : 1, . . . , r.
6
Giorgio Delzanno
Definition 3
An initial global configuration G0 has an arbitrary (but finite) number of threads
with local variables all set to ⊥. A run is a sequence G0 G1 . . . such that Gi ⇒ Gi+1
for i ≥ 0. A global configuration G is reachable from G0 if there exists a run from
G0 to G.
Example 1
Let us consider a challenge and response protocol in which the goal of two agents
Alice and Bob is to exchange a pair of new names hnA , nB i, the first one created
by Alice and the second one created by Bob, so as to build a composed secret key.
We can specify the protocol by using new names to dynamically establish private
channel names between instances of the initiator and of the responder. The TDL
program in Figure 1 follows this idea. The thread Init specifies the behavior of the
initiator. He first creates a new name using the internal action f resh, and stores
it in the local variable nA . Then, he sends nA on channel c (a constant), waits for
a name y on a channel with the same name as the value of the local variable nA
(the channel is specified by variable nA ) and then stores y in the local variable
mA . The thread Resp specifies the behavior of the responder. Upon reception of
a name x on channel c, he stores it in the local variable nB , then creates a new
name stored in local variable mB and finally sends the value in mB on channel with
the same name as the value of nB . The thread M ain non-deterministically creates
new thread instances of type Init and Resp. The local variable x is used to store
new names to be used for the creation of a new thread instance. Initially, all local
variables of threads Init/Resp are set to ⊥. In order to allow process instances to
participate to several sessions (potentially with different principals), we could also
add the following rule
restart
stopA −−−−−→ initA [nA := ⊥, mA := ⊥]
In this rule we require that roles and identities do not change from session to
session.3 Starting from G0 = hN0 , hinit, ⊥ii, and running the Main thread we can
generate any number of copies of the threads Init and Resp each one with a unique
identifier. Thus, we obtain global configurations like
hN, hinitM , ⊥i,
hinitA , i1 , ⊥, ⊥i, . . . , hinitA , iK , ⊥, ⊥i,
hinitB , iK+1 , ⊥, ⊥i, . . . , hinitB , iK+L , ⊥, ⊥i i
where N = {⊥, i1, . . . , iK , iK+1 , . . . , iK+L } for K, L ≥ 0. The threads of type Init
and Resp can start parallel sessions whenever created. For K = 1 and L = 1 one
possible session is as follows.
Starting from
h{⊥, i1 , i2 }, hinitM , ⊥i, hinitA , i1 , ⊥, ⊥i, hinitB , i2 , ⊥, ⊥ii
3
By means of thread and fresh name creation it is also possible to specify a restart rule in which
a given process takes a potential different role or identity.
Constraint-based Verification of Abstract Multithreaded Programs
7
T hread Init(local idA , nA , mA );
f resh
initA −−−−→ genA
c!hnA i
genA −−−−→ waitA
n ?hyi
waitA −−A
−−−→ stopA
[nA := new]
[true]
[mA := y]
T hread Resp(local id, nB , mB );
c?hxi
initB −−−→ genB
[nB := x]
f resh
genB −−−−→ readyB
n !hm i
[mB := new]
B
readyB −−B−−−−
→ stopB
[true]
T hread M ain(local x);
id
initM −−→ create
newA
create −−−−→
initM
new
→ initM
create −−−−B
[x := new]
[run Init with idA := x, nA := ⊥, mA := ⊥, x := ⊥]
[run Resp with idB := x, nB := ⊥, mB := ⊥, x := ⊥B ]
Fig. 1. Example of thread definitions.
if we apply the first rule of thread Init to hinitA , i1 , ⊥, ⊥i we obtain
h{⊥, i1 , i2 , a1 }, hinitM , ⊥i, hgenA , i1 , a1 , ⊥i, hinitB , i2 , ⊥, ⊥ii
where a1 is the generated name (a1 is distinct from ⊥, i1 , and i2 ). Now if we apply
the second rule of thread Init and the first rule of thread Resp (synchronization
on channel c) we obtain
h{⊥, i1 , i2 , a1 }, hinitM , ⊥i, hwaitA , i1 , a1 , ⊥i, hgenB , i2 , a1 , ⊥ii
If we apply the second rule of thread Resp we obtain
h{⊥, i1 , i2 , a1 , a2 }, hinitM , ⊥i, hwaitA , i1 , a1 , ⊥i, hreadyB , i2 , a1 , a2 ii
Finally, if we apply the last rule of thread Init and Resp (synchronization on
channel a1 ) we obtain
h{⊥, i1 , i2 , a1 , a2 }, hinitM , ⊥i, hstopA , i1 , a1 , a2 i, hstopB , i2 , a1 , a2 ii
Thus, at the end of the session the thread instances i1 and i2 have both a local
copy of the fresh names a1 and a2 . Note that a copy of the main thread hinitM , ⊥i
is always active in any reachable configuration, and, at any time, it may introduce
new threads (either of type Init or Resp) with fresh identifiers. Generation of fresh
names is also used by the threads of type Init and Resp to create nonces. Furthermore, threads can restart their life cycle (without changing identifiers). Thus,
in this example the set of possible reachable configurations is infinite and contains
configurations with arbitrarily many threads and fresh names. Since names are
stored in the local variables of active threads, the local data also range over an
infinite domain.
✷
8
Giorgio Delzanno
2.2 Expressive Power of TDL
To study the expressive power of the TDL language, we will compare it with the
Turing equivalent formalism called Two Counter Machines. A Two Counters Machine configurations is a tuple hℓ, c1 = n1 , c2 = n2 i where ℓ is control location taken
from a finite set Q, and n1 and n2 are natural numbers that represent the values
of the counters c1 and c2 . Each counter can be incremented or decremented (if
greater than zero) by one. Transitions combine operations on individual counters
with changes of control locations. Specifically, the instructions for counter ci are as
follows
Inc: ℓ1 : ci := ci + 1; goto ℓ2 ;
Dec: ℓ1 : if ci > 0 then ci := ci − 1; goto ℓ2 ; else goto ℓ3 ;
A Two Counter Machine consists then of a list of instructions and of the initial
state hℓ0 , c1 = 0, c2 = 0i. The operational semantics is defined according to the intuitive semantics of the instructions. Problems like control state reachability are
undecidable for this computational model.
The following property then holds.
Theorem 1
TDL programs can simulate Two Counter Machines.
Proof
In order to define a TDL program that simulates a Two Counter Machine we
proceed as follows. Every counter is represented via a doubly linked list implemented
via a collection of threads of type Cell and with a unique thread of type Last
pointing to the head of the list. The i-th counter having value zero is represented
as the empty list Cell(i, v, v), Last(i, v, w) for some name v and w (we will explain
later the use of w). The i-th counter having value k is represented as
Cell(i, v0 , v0 ), Cell(i, v0 , v1 ), . . . , C(i, vk−1 , vk ), Last(i, vk , w)
for distinct names v0 , v1 , . . . , vk . The instructions on a counter are simulated by
sending messages to the corresponding Last thread. The messages are sent on
channel Zero (zero test), Dec (decrement), and Inc (increment). In reply to each
of these messages, the thread Last sends an acknowledgment, namely Y es/N o for
the zero test, DAck for the decrement, IAck for the increment operation. Last
interacts with the Cell threads via the messages tstC, decC, incC acknowledged
by messages z/nz, dack. iack. The interactions between a Last thread and the Cell
threads is as follows.
Zero Test Upon reception of a message hxi on channel Zero, the Last thread with
local variables id, last, aux checks that its identifier id matches x - see transition
from Idle to Busy - sends a message hid, lasti on channel tstC directed to the cell
pointed to by last (transition from Busy to W ait), and then waits for an answer. If
the answer is sent on channel nz, standing for non-zero, (resp. z standing for zero)
- see transition from W ait to AckN Z (resp. AckZ) - then it sends its identifier on
Constraint-based Verification of Abstract Multithreaded Programs
9
Thread Last(local id, last, aux);
(Zero test)
Zero?hxi
Idle −−−−−−→ Busy
[id = x]
tstC!hid,lasti
Busy −−−−−−−−−→ W ait
nz?hxi
W ait −−−−→ AckN Z
[id = x]
z?hxi
W ait −−−→ AckZ
[id = x]
Y es!hidi
AckZ −−−−−→ Idle
No!hidi
AckN Z −−−−−→ idle
(Decrement)
Dec?hxi
Idle −−−−−→ Dbusy
[id = x]
decC!hid,lasti
DBusy −−−−−−−−−→ DW ait
dack?hx,ui
DW ait −−−−−−−→ DAck
[id = x, last := u]
DAck!hidi
DAck −−−−−−−→ Idle
(Increment)
Inc?hxi
Idle −−−−−→ IN ew
[id = x]
new
[aux := new]
run
[run Cell with idc := id; prev := last; next := aux]
IN ew −−−→ IRun
IRun −−→ IAck
IAck!hidi
IAck −−−−−−→ Idle
[last := aux]
Fig. 2. The process defining the last cell of the linked list associated to a counter
channel N o (resp. Y es) as an acknowledgment to the first message - see transition
from AckN Z (resp. Z) to Idle. As shown in Fig. 3, the thread Cell with local
variables idc, prev, and next that receives the message tstC, i.e., pointed to by a
thread Last with the same identifier as idc, sends an acknowledgment on channel
z (zero) if prev = next, and on channel nz (non-zero) if prev 6= next.
Decrement Upon reception of a message hxi on channel Dec, the Last thread with
local variables id, last, aux checks that its identifier id matches x (transition from
Idle to Dbusy), sends a message hid, lasti on channel decC directed to the cell
pointed to by last (transition from Busy to W ait), and then waits for an answer.
10
Giorgio Delzanno
Thread Cell(local idc, prev, next);
(Zero test)
tstC?hx,ui
idle −−−−−−−→ ackZ [x = idc, u = next, prev = next]
tstC?hx,ui
idle −−−−−−−→ ackN Z [x = idc, u = next, prev 6= next]
z!hidci
ackZ −−−−→ idle
nz!hidci
ackN Z −−−−−→ idle
(Decrement)
dec?hx,ui
idle −−−−−−→ dec [x = idc, u = next, prev 6= next]
dack!hidc,previ
dec −−−−−−−−−−→ idle
Fig. 3. The process defining a cell of the linked list associated to a counter
If the answer is sent on channel dack (transition from DW ait to DAck) then it
updates the local variable last with the pointer u sent by the thread Cell, namely
the prev pointer of the cell pointed to by the current value of last, and then sends
its identifier on channel DAck to acknowledge the first message (transition from
DAck to Idle).
As shown in Fig. 3, a thread Cell with local variables idc, prev, and next that
receives the message decC and such that next = last sends as an acknowledgment
on channel dack the value prev.
Increment To simulate the increment operation, Last does not have to interact with
existing Cell threads. Indeed, it only has to link a new Cell thread to the head of
the list (this is way the Cell thread has no operations to handle the increment
operation). As shown in Fig. 2 this can be done by creating a new name stored in
the local variable aux (transition from IN ew to IRun) and spawning a new Cell
thread (transition from IRun to IAck) with prev pointer equal to last, and next
pointer equal to aux. Finally, it acknowledges the increment request by sending its
identifier on channel IAck and updates variable last with the current value of aux.
Two Counter Machine Instructions We are now ready to use the operations provided by the thread Last to simulate the instructions of a Two Counter Machine.
As shown in Fig. 4, we use a thread CM with two local variables id1 , id2 to represent the list of instructions of a 2CM with counters c1 , c2 . Control locations of the
Two Counter Machines are used as local states of the thread CM . The initial local
state of the CM thread is the initial control location. The increment instruction
on counter ci at control location ℓ1 is simulated by an handshaking with the Last
thread with identifier idi : we first send the message Inc!hidi i, wait for the acknowledgment on channel IAck and then move to state ℓ2 . Similarly, for the decrement
Constraint-based Verification of Abstract Multithreaded Programs
11
Thread CM (local id1 , id2 );
..
.
(Instruction : ℓ1 : ci := ci + 1; goto ℓ2 ; )
Inc!hid i
ℓ1 −−−−−−i→ waitℓ1
IAck!hxi
waitℓ1 −−−−−−→ ℓ2 [x = idi ]
..
.
(Instruction : ℓ1 : ci > 0 then ci := ci − 1; goto ℓ2 ; else goto ℓ3 ; )
Zero!hid i
ℓ1 −−−−−−−i→ waitℓ1
NZAck?hxi
waitℓ1 −
−−−−−−−
→ decℓ1 [x = idi ]
Dec!hid i
decℓ1 −−−−−−i→ wdecℓ1
DAck?hyi
wdecℓ1 −−−−−−→ ℓ2 [y = idi ]
ZAck?hxi
waitℓ1 −−−−−−→ ℓ3 [x = idi ]
..
.
Fig. 4. The thread associated to a 2CM.
Thread Init(local nid1 , p1 , nid2 , p2 );
f reshId
init −−−−−→ init1 [nid1 := new]
f reshP
init1 −−−−−→ init2 [p1 := new]
runC
init2 −
−−−→ init3 [run Cell with idc := nid1 ; prev := p1 ; next := p1 ]
runL
init3 −−−→ init4 [run Last with idc := nid1 ; last := p1 ; aux := ⊥]
f reshId
init4 −−−−−→ init5 [nid2 := new]
f reshP
init5 −−−−−→ init6 [p2 := new]
runC
init6 −
−−−→ init7 [run Cell with idc := nid2 ; prev := p2 ; next := p2 ]
runL
init7 −−−→ init8 [run Last with idc := nid2 ; last := p2 ; aux := ⊥]
runCM
init8 −−−−−→ init9 [run 2CM with id1 := nid1 ; id2 := nid2 ]
Fig. 5. The initialization thread.
instruction on counter ci at control location ℓ1 we first send the message Zero!hidi i.
If we receive an acknowledgment on channel N ZAck we send a Dec request, wait
for completion and then move to ℓ2 . If we receive an acknowledgment on channel
ZAck we directly move to ℓ3 .
12
Giorgio Delzanno
Initialization The last step of the encoding is the definition of the initial state of the
system. For this purpose, we use the thread Init of Fig. 5. The first four rules of Init
initialize the first counter: they create two new names nid1 (an identifier for counter
c1 ) and p1 , and then spawn the new threads Cell(nid1 , p1 , p1 ), Last(nid1 , p1 , ⊥).
The following four rules spawns the new threads Cell(nid2 , p2 , p2 ), Last(nid2 , p2 , ⊥).
After this stage, we create a thread of type 2CM to start the simulation of the instructions of the Two Counter Machines. The initial configuration of the whole
system is G0 = hinit, ⊥, ⊥i. By construction we have that an execution step from
hℓ1 , c1 = n1 , c2 = n2 i to hℓ2 , c1 = m1 , c2 = m2 i is simulated by an execution run going from a global configuration in which the local state of thread CM is hℓ1 , id1 , id2 i
and in which we have ni occurrences of thread Cell with the same identifier idi
for i : 1, 2, to a global configuration in which the local state of thread CM is
hℓ2 , id1 , id2 i and in which we have mi occurrences of thread Cell with the same
identifier idi for i : 1, 2. Thus, every executions of a 2CM M corresponds to an execution of the corresponding TDL program that starts from the initial configuration
G0 = hinit, ⊥, ⊥i.
As a consequence of the previous theorem, we have the following corollary.
Corollary 1
Given a TDL program, a global configurations G, and a control location ℓ, deciding
if there exists a run going from G0 to a global configuration that contains ℓ (control
state reachability) is an undecidable problem.
3 From TDL to MSRN C
As mentioned in the introduction, our verification methodology is based on a translation of TDL programs into low level specifications given in MSRN C . Our goal is
to extend the connection between CCS and Petri Nets (German and Sistla 1992)
to TDL and MSR so as to be able to apply the verification methods defined in
(Delzanno 2005) to multithreaded programs. In the next section we will summarize
the main features of the language MSRN C introduced in (Delzanno 2001).
3.1 Preliminaries on MSRN C
N C-constraints are linear arithmetic constraints in which conjuncts have one of
the following form: true, x = y, x > y, x = c, or x > c, x and y being two variables
from a denumerable set V that range over the rationals, and c being an integer.
The solutions Sol of a constraint ϕ are defined as all evaluations (from V to Q)
that satisfy ϕ. A constraint ϕ is satisfiable whenever Sol(ϕ) 6= ∅. Furthermore, ψ
entails ϕ whenever Sol(ψ) ⊆ Sol(ϕ). N C-constraints are closed under elimination
of existentially quantified variables.
Let P be a set of predicate symbols. An atomic formula p(x1 , . . . , xn ) is such that
p ∈ P, and x1 , . . . , xn are distinct variables in V. A multiset of atomic formulas is
indicated as A1 | . . . | Ak , where Ai and Aj have distinct variables (we use variable
renaming if necessary), and | is the multiset constructor.
Constraint-based Verification of Abstract Multithreaded Programs
13
In the rest of the paper we will use M, N , . . . to denote multisets of atomic formulas,
ǫ to denote the empty multiset, ⊕ to denote multiset union and ⊖ to denote multiset
difference. An MSRN C configuration is a multiset of ground atomic formulas, i.e.,
atomic formulas like p(d1 , . . . , dn ) where di is a rational for i : 1, . . . , n.
An MSRN C rule has the form M −→ M′ : ϕ, where M and M′ are two (possibly
empty) multisets of atomic formulas with distinct variables built on predicates in
P, and ϕ is an N C-constraint. The ground instances of an MSRN C rule are defined
as
Inst(M −→ M′ : ϕ) = {σ(M) −→ σ(M′ ) | σ ∈ Sol(ϕ)}
where σ is extended in the natural way to multisets, i.e., σ(M) and σ(M′ ) are
MSRN C configurations.
An MSRN C specification S is a tuple hP, I, Ri, where P is a finite set of predicate
symbols, I is finite a set of (initial) MSRN C configurations, and R is a finite set of
MSRN C rules over P.
The operational semantics describes the update from a configuration M to one of its
possible successor configurations M′ . M′ is obtained from M by rewriting (modulo
associativity and commutativity) the left-hand side of an instance of a rule into the
corresponding right-hand side. In order to be fireable, the left-hand side must be
included in M. Since instances and rules are selected in a non deterministic way, in
general a configuration can have a (possibly infinite) set of (one-step) successors.
Formally, a rule H −→ B : ϕ from R is enabled at M via the ground substitution
σ ∈ Sol(ϕ) if and only if σ(H) 4 M. Firing rule R enabled at M via σ yields the
new configuration
M′ = σ(B) ⊕ (M ⊖ σ(H))
We use M ⇒MSR M′ to denote the firing of a rule at M yielding M′ .
A run is a sequence of configurations M0 M1 . . . Mk with M0 ∈ I such that
Mi ⇒MSR Mi+1 for i ≥ 0. A configuration M is reachable if there exists M0 ∈ I
∗
∗
such that M0 ⇒MSR M, where ⇒MSR is the transitive closure of ⇒MSR . Finally, the successor and predecessor operators P ost and P re are defined on a set
of configurations S as P ost(S) = {M′ |M ⇒MSR M′ , M ∈ S} and P re(S) =
{M|M ⇒MSR M′ , M′ ∈ S}, respectively. P re∗ and P ost∗ denote their transitive
closure.
As shown in (Delzanno 2001; Bozzano and Delzanno 2002), Petri Nets represent a
natural abstractions of MSRN C (and more in general of MSR rule with constraints)
specifications. They can be encoded, in fact, in propositional MSR specifications
(e.g. abstracting away arguments from atomic formulas).
3.2 Translation from TDL to MSRN C
The first thing to do is to find an adequate representation of names. Since all we
need is a way to distinguish old and new names, we just need an infinite domain
in which the = and 6= relation are supported. Thus, we can interpret names in N
14
Giorgio Delzanno
either as integer of as rational numbers. Since operations like variable elimination
are computationally less expensive than over integers, we choose to view names
as non-negative rationals. Thus, a local (TDL) configuration p = hs, n1 , . . . , nk i
is encoded as the atomic formula p• = s(n1 , . . . , nk ), where ni is a non-negative
rational. Furthermore, a global (TDL) configuration G = hN, p1 , . . . , pm i is encoded
as an MSRN C configuration G•
p•1 | . . . | p•m | f resh(n)
where the value n in the auxiliary atomic formula f resh(n) is an rational number
strictly greater than all values occurring in p•1 , . . . , p•m . The predicate f resh will
allow us to generate unused names every time needed.
The translation of constants C = {c1 , . . . , cm }, and variables is defined as follows:
x• = x for x ∈ V, ⊥• = 0, c•i = i for i : 1, . . . , m. We extend ·• in the natural way
on a guard γ, by decomposing every formula x 6= e into x < e• and x > e• . We will
call γ • the resulting set of N C-constraints.4
Given V = {x1 , . . . , xk }, we define V ′ as the set of new variables {x′1 , . . . , x′k }.
Now, let us consider the assignment α defined as x1 := e1 , . . . , xk := ek (we add
assignments like xi := xi if some variable does not occur as target of α). Then, α•
is the N C-constraint x′1 = e•1 , . . . , x′k = e•k .
The translation of thread definitions is defined below (where we will often refer to
Example 1).
Initial Global Configuration Given an initial global configuration consisting of the
local configurations hsi , ni1 , . . . , niki i with nij = ⊥ for i : 1, . . . , u, we define the
following MSRN C rule
init → s1 (x11 , . . . , x1k1 ) | . . . | su (xu1 , . . . , xuku ) | f resh(x) :
x > C, x11 = 0, . . . , xuku = 0
here C is the largest rational used to interpret the constants in C.
For each thread definition P = hQ, s0 , V, Ri in T with V = {x1 , . . . , xk } we translate
the rules in R as described below.
a
Internal Moves For every internal move s −−→ s′ [γ, α], and every ν ∈ γ • we define
s(x1 , . . . , xk ) → s′ (x′1 , . . . , x′k ) : ν, α•
a
Name Generation For every name generation s −−→ s′ [xi := new], we define
^
s(x1 , . . . , xk ) | f resh(x) → s′ (x′1 , . . . , x′k ) | f resh(y) : y > x′i , x′i > x,
x′j = xj
j6=i
f resh
For instance, the name generation initA −−−−→ genA [n := new] is mapped into the
MSRN C rule initA (id, x, y)| f resh(u) −−→ genA (id′ , x′ , y ′ ) | f resh(u′ ) : ϕ where ϕ
4
As an example, if γ is the constraint x = 1, x 6= z then γ • consists of the two constraints
x = 1, x > z and x = 1, z > x.
Constraint-based Verification of Abstract Multithreaded Programs
15
is the constraint u′ > x′ , x′ > u, y ′ = y, id′ = id. The constraint x′ > u represents
the fact that the new name associated to the local variable n (the second argument
of the atoms representing the thread) is fresh, whereas u′ > x′ updates the current
value of f resh to ensure that the next generated names will be picked up from
unused values.
Thread Creation Let P = hQ′ , t0 , V ′ , R′ i and V ′ = {y1 , . . . , yu }. Then, for every
a
thread creation s −−→ s′ [run P with α], we define
s(x1 , . . . , xk ) → s′ (x′1 , . . . , x′k ) | t(y1′ , . . . , yu′ ) : x′1 = x1 , . . . , x′k = xk , α• .
new
A
initM [run Init with id := x, . . .] of Example 1.
E.g., consider the rule create −−−−→
Its encoding yields the MSRN C rule create(x) −−→ initM (x′ ) | initA (id′ , n′ , m′ ) : ψ,
where ψ represents the initialization of the local variables of the new thread x′ =
x, id′ = x, n′ = 0, m′ = 0.
Rendez-vous The encoding of rendez-vous communication is based on the use of
constraint operations like variable elimination. Let P and P ′ be a pair of thread
definitions, with local variables V = {x1 , . . . , xk } and V ′ = {y1 , . . . , yl } with V ∩
e′ ?m′
e!m
V ′ = ∅. We first select all rules s −−→ s′ [γ, α] in R and t −−−−→ t′ [γ ′ , α′ ] in R′ ,
such that m = hw1 , . . . , wu i, m′ = hw1′ , . . . , wv′ i and u = v. Then, we define the new
MSRN C rule
s(x1 , . . . , xk ) | t(y1 , . . . , yl ) → s′ (x′1 , . . . , x′k ) | t′ (y1′ , . . . , yl′ ) : ϕ
for every ν ∈ γ • and ν ′ ∈ γ ′• such that the NC-constraint ϕ obtained by eliminating w1′ , . . . , wv′ from the constraint ν ∧ ν ′ ∧ α• ∧ α′• ∧ w1 = w1′ ∧ . . . ∧ wv = wv′
nA ?hyi
is satisfiable. For instance, consider the rules waitA −−−−−→ stopA [mA := y] and
nB !hmB i
readyB −−−−−−→ stopB [true]. We first build up a new constraint by conjoining the
NC-constraints y = mB (matching of message templates), and nA = nB , m′A =
y, n′A = nA , m′B = mB , n′B = nB , id′1 = id1 , id′2 = id2 (guards and actions of
sender and receiver). After eliminating y we obtain the constraint ϕ defined as
nB = nA , m′A = mB , n′A = nA , m′B = mB , n′B = nB , id′1 = id1 , id′2 = id2 defined
over the variables of the two considered threads. This step allows us to symbolically
represent the passing of names. After this step, we can represent the synchronization of the two threads by using a rule that simultaneously rewrite all instances
that satisfy the constraints on the local data expressed by ϕ, i.e., we obtain the
rule
waitA (id1 , nA , mA )| readyB (id2 , nB , mB ) −→
stopA (id′1 , n′A , m′A ) | stopB (id′2 , n′B , m′B ) : ϕ
The complete translation of Example 1 is shown in Fig. 6 (for simplicity we have
applied a renaming of variables in the resulting rules). An example of run in the
resulting MSRN C specification is shown in Figure 7. Note that, a fresh name is
selected between all values strictly greater than the current value of f resh (e.g. in
the second step 6 > 4), and then f resh is updated to a value strictly greater than
all newly generated names (e.g. 8 > 6 > 4).
16
Giorgio Delzanno
init −−→ f resh(x) | initM (y) : x > 0, y = 0.
f resh(x) | initM (y) −−→ f resh(x′ ) | create(y ′ ) : x′ > y ′ , y ′ > x.
create(x) −−→ initM (x′ ) | initA (id′ , n′ , m′ ) : x′ = x, id′ = x, n′ = 0, m′ = 0.
create(x) −−→ initM (x′ ) | initB (id′ , n′ , m′ ) : x′ = x, id′ = x, n′ = 0, m′ = 0.
initA (id, n, m)| f resh(u) −−→ genA (id, n′ , m) | f resh(u′ ) : u′ > n′ , n′ > u.
genA (id1 , n, m)| initB (id2 , u, v) −−→ waitA (id1 , n, m) | genB (id′2 , u′ , v ′ ) : u′ = n, v ′ = v
genB (id, n, m)| f resh(u) −−→ readyB (id, n, m′ ) | f resh(u′ ) : u′ > m′ , m′ > u.
waitA (id1 , n, m)| readyB (id2 , u, v) −−→ stopA (id1 , n, m′ ) | stopB (id2 , u, v) : n = u, m′ = v.
stopA (id, n, m) −−→ initA (id′ , n′ , m′ ) : n′ = 0, m′ = 0, id′ = id.
stopB (id, n, m) −−→ initB (id′ , n′ , m′ ) : n′ = 0, m′ = 0, id′ = id.
Fig. 6. Encoding of Example 1: for simplicity we embed constraints like x = x′ into
the MSR formulas.
init ⇒ . . . ⇒ f resh(4) | initM (0) | initA (2, 0, 0) | initB (3, 0, 0)
⇒ f resh(8) | initM (0) | genA (2, 6, 0) | initB (3, 0, 0)
⇒ f resh(8) | initM (0) | waitA (2, 6, 0) | genB (3, 6, 0)
⇒ . . . ⇒ f resh(16) | initM (0) | waitA (2, 6, 0) | genB (3, 6, 0) | initA (11, 0, 0)
Fig. 7. A run in the encoded program.
Let T = hP1 , . . . , Pt i be a collection of thread definitions and G0 be an initial
global state. Let S be the MSRN C specification that results from the translation
described in the previous section.
Let G = hN, p1 , . . . , pn i be a global configuration with pi = hsi , vi1 , . . . , viki i, and
let h : N ❀ Q+ be an injective mapping. Then, we define G• (h) as the MSRN C
configuration
s1 (h(v11 ), . . . , h(v1k1 )) | . . . | sn (h(vn1 ), . . . , h(vnkn )) | f resh(v)
where v is a the first value strictly greater than all values in the range of h. Given
an MSRN C configuration M defined as s1 (v11 , . . . , v1k1 ) | . . . | sn (vn1 , . . . , vnkn )
with sij ∈ Q+ , let V (M) ⊆ Q+ be the set of values occurring in M. Then, given a
bijective mapping f : V (M) ❀ N ⊆ N , we define M• (f ) as the global configuration hN, p1 , . . . , pn i where pi = hsi , f (vi1 ), . . . , f (viki )i.
Based on the previous definitions, the following property then holds.
Theorem 2
For every run G0 G1 . . . in T with corresponding set of names N0 N1 . . ., there exist
sets D0 D1 . . . and bijective mappings h0 h1 . . . with hi : Ni ❀ Di ⊆ Q+ for i ≥ 0,
such that init G•0 (h0 )G•1 (h1 ) . . . is a run of S. Vice versa, if init M0 M1 . . . is a
run of S, then there exist sets N0 N1 . . . in N and bijective mappings f0 f1 . . . with
fi : V (Mi ) ❀ Ni for i ≥ 0, such that M•0 (f0 )M•1 (f1 ) . . . is a run in T .
Proof
We first prove that every run in T is simulated by a run in S.
Let G0 . . . Gl be a run in T , i.e., a sequence of global states (with associated set
Constraint-based Verification of Abstract Multithreaded Programs
17
of names N0 . . . Nl ) such that Gi ⇒ Gi+1 and Ni ⊆ Ni+1 for i ≥ 0.
We prove that it can be simulated in S by induction on its length l.
Specifically, suppose that there exist sets of non negative rationals D0 . . . Dl and
bijective mappings h0 . . . hl with hi : Ni ❀ Di for 0 ≤ i ≤ l, such that
c0 (h0 ) . . . G
cl (hl )
init G
is a run of S. Furthermore, suppose Gl ⇒ Gl+1 .
We prove the thesis by a case-analysis on the type of rule applied in the last step
of the run.
Let Gl = hNl , p1 , . . . , pr i and pj = hs, n1 , . . . , nk i be a local configuration for the
thread definition P = hQ, s, V, Ri with V = {x1 , . . . , xk } and ni ∈ Nl for i : 1, . . . , k.
a
Assignment Suppose there exists a rule s −−→ s′ [γ, α] in R such that ρpj satisfies γ,
Gl = hNl , . . . , pj , . . .i ⇒ hNl+1 , . . . , p′j , . . .i = Gl+1 Nl = Nl+1 , p′j = hs′ , n′1 , . . . , n′k i,
and if xi := yi occurs in α, then n′i = ρpj (yi ), otherwise n′i = ni for i : 1, . . . , k.
The encoding of the rule returns one M SRN C rule having the form
s(x1 , . . . , xk ) → s′ (x′1 , . . . , x′k ) : γ ′ , α
b
for every γ ′ ∈ γ
b.
cl (hl ) is a multiset of atomic formulas that contains the
By inductive hypothesis, G
formula s(hl (n1 ), . . . , hl (nk )).
Now let us define hl+1 as the mapping from Nl to Dl such that hl+1 (n′i ) = hl (nj ) if
xi := xj is in α and hl+1 (n′i ) = 0 if xi := ⊥ is in α. Furthermore, let us the define
the evaluation
σ = hx1 7→ hl (n1 ), . . . , xk 7→ hl (nk ), x′1 7→ hl+1 (n′1 ), . . . , x′k 7→ hl+1 (n′k )i
Then, by construction of the set of constraints b
γ and of the constraint α
b, it follows
that σ is a solution for γ ′ , α
b for some γ ′ ∈ b
γ . As a consequence, we have that
s(n1 , . . . , nk ) → s′ (n′1 , . . . , n′k )
is a ground instance of one of the considered M SRN C rules.
cl (hl ), if we apply a rewriting step
Thus, starting from the M SRN C configuration G
we obtain a new configuration in which s(n1 , . . . , nk ) is replaced by s′ (n′1 , . . . , n′k ),
c
[
and all the other atomic formulas in G
l+1 (hl+1 ) are the same as in Gl (hl ). The
[
resulting M SRN C configuration coincides then with the definition of G
l+1 (hl+1 ).
Creation of new names Let us now consider the case of fresh name generation.
a
Suppose there exists a rule s −−→ s′ [xi := new] in R, and let n 6∈ Nl , and suppose
hNl , . . . , pj , . . .i ⇒ hNl+1 , . . . , p′j , . . .i where Nl+1 = Nl ∪ {v}, p′j = hs′ , n′1 , . . . , n′k i
where n′i = n, and n′j = nj for j 6= i.
We note than that the encoding of the previous rule returns the M SRN C rule
s(x1 , . . . , xk ) | f resh(x) → s′ (x′1 , . . . , x′k ) | f resh(x′ ) : ϕ
where ϕ consists of the constraints y > x′i , x′i > x and x′j = xj for j 6= i. By
cl (hl ) is a multiset of atomic formulas that contains the
inductive hypothesis, G
18
Giorgio Delzanno
formulas s(hl (n1 ), . . . , hl (nk )) and f resh(v) where hl is a mapping into Dl , and v
is the first non-negative rational strictly greater than all values occurring in the
formulas denoting processes.
Let v be a non negative rational strictly greater than all values in Dl . Furthermore,
let us define v ′ = v + 1 and Dl+1 = Dl ∪ {v, v ′ }.
Furthermore, we define hl+1 as follows hl+1 (n) = hl (n) for n ∈ Nl , and hl+1 (n′i ) =
hl+1 (n) = v ′ . Furthermore, we define the following evaluation
σ
= h x 7→ v, x1 7→ hl (n1 ), . . . , xk 7→ hl (nk ),
x′ 7→ v ′ , x′1 7→ hl+1 (n′1 ), . . . , x′k 7→ hl+1 (n′k ) i
Then, by construction of σ and α
b, it follows that σ is a solution for α
b. Thus,
s(n1 , . . . , nk ) | f resh(v) → s′ (n′1 , . . . , n′k ) | f resh(v ′ )
is a ground instance of the considered M SRN C rule.
cl (hl ), if we apply a rewriting step we
Starting from the M SRN C configuration G
obtain a new configuration in which s(n1 , . . . , nk ) and f resh(v) are substituted by
[
s′ (n′1 , . . . , n′k ) and f resh(v ′ ), and all the other atomic formulas in G
l+1 (hl+1 ) are
c
the same as in Gl (hl ). We conclude by noting that this formula coincides with the
[
definition of G
l+1 (hl+1 ).
For sake of brevity we omit the case of thread creation whose only difference from
the previous cases is the creation of several new atoms instead (with values obtained
by evaluating the action) of only one.
Rendez-vous Let pi = hs, n1 , . . . , nk i and pj = ht, m1 , . . . , mu i two local configurations for threads P 6= P ′ , ni ∈ Nl for i : 1, . . . , k and mi ∈ Nl for i : 1, . . . , u.
c!m
c?m′
Suppose s −−→ s′ [γ, α] and t −−−→ t′ [γ ′ , α′ ], where m = hxi1 , . . . , xiv i, and m′ =
hy1 , . . . , yv i ( all defined over distinct variables) are rules in R.
Furthermore, suppose that ρpi satisfies γ, and that ρ′ (see definition of the operational semantics) satisfies γ ′ , and suppose that Gl = hNl , . . . , pi , . . . , pj , . . .i ⇒
hNl+1 , . . . , p′i , . . . , p′j , . . .i = Gl+1 , where Nl+1 = Nl , p′i = hs′ , n′1 , . . . , n′k i, p′j =
ht′ , m′1 , . . . , m′u i, and if xi := e occurs in α, then n′i = ρpi (e), otherwise n′i = ni
for i : 1, . . . , k; if ui := e occurs in α′ , then m′i = ρ′ (e), otherwise m′i = mi for
i : 1, . . . , u.
cl (hl ) is a multiset of atomic formulas that contains the
By inductive hypothesis, G
formulas s(hl (n1 ), . . . , hl (nk )) and t(hl (m1 ), . . . , hl (mu )).
Now, let us define hl+1 as the mapping from Nl to Dl such that hl+1 (n′i ) = hl (nj )
if xi := xj is in α, hl+1 (m′i ) = hl (mj ) if ui := uj is in α′ , hl+1 (n′i ) = 0 if xi := ⊥ is
in α, hl+1 (m′i ) = 0 if ui := ⊥ is in α′ .
Now, let us define σ as the evaluation from Nl to Dl such that
σ = σ1 ∪ σ2
σ1 = hx1 7→ hl (n1 ), . . . , xk 7→ hl (nk ), u1 7→ hl (m1 ), . . . , uu 7→ hl (mu )i
σ2 = hx′1 7→ hl+1 (n′1 ), . . . , x′k 7→ hl+1 (n′k ), u′1 7→ hl+1 (m′1 ), . . . , u′u 7→ hl+1 (m′u )i.
Then, by construction of the sets of constraints γ
b, γb′ , α
b and αb′ it follows that σ is
′
′
′
a solution for the constraint ∃w1 . . . . ∃wp .θ ∧ θ ∧ α
b ∧ αb′ ∧ w1 = w1′ ∧ . . . ∧ wp = wp′
for some θ ∈ γ
b and θ′ ∈ γb′ . Note in fact that the equalities wi = wi′ express the
Constraint-based Verification of Abstract Multithreaded Programs
19
passing of values defined via the evaluation ρ′ in the operational semantics.
As a consequence,
s(n1 , . . . , nk ) | t(m1 , . . . , mu ) → s′ (n′1 , . . . , n′k ) | t′ (m′1 , . . . , m′u )
is a ground instance of one of the considered M SRN C rules.
cl (hl ), if we apply a rewriting
Thus, starting from the M SRN C configuration G
step we obtain a new configuration in which s(n1 , . . . , nk ) has been replaced by
s′ (n′1 , . . . , n′k ), and t′ (m′1 , . . . , m′k ) has been replaced by t(m′1 , . . . , m′u ), and all the
cl (hl ). This formula coincides with the definition
other atomic formulas are as in G
[
of Gl+1 (hl+1 ).
The proof of completeness is by induction on the length of an MSR run, and by
case-analysis on the application of the rules. The structure of the case analysis is
similar to the previous one and it is omitted for brevity.
4 Verification of TDL Programs
Safety and invariant properties are probably the most important class of correctness
specifications for the validation of a concurrent system. For instance, in Example
1 we could be interested in proving that every time a session terminates, two instances of thread Init and Resp have exchanged the two names generated during
the session. To prove the protocol correct independently from the number of names
and threads generated during an execution, we have to show that from the initial configuration G0 it is not possible to reach a configuration that violates the
aforementioned property. The configurations that violate the property are those in
which two instances of Init and Resp conclude the execution of the protocol exchanging only the first nonce. These configurations can be represented by looking
at only two threads and at the relationship among their local data. Thus, we can
reduce the verification problem of this safety property to the following problem:
Given an initial configuration G0 we would like to decide if a global configuration that contains at least two local configurations having the form hstopA , i, n, mi
and hstopB , i′ , n′ , m′ i with n′ = n and m 6= m′ for some i, i′ , n, n′ , m, m′ is reachable. This problem can be viewed as an extension of the control state reachability
problem defined in (Abdulla and Nylén 2000) in which we consider both control locations and local variables. Although control state reachability is undecidable (see
Corollary 1), the encoding of TDL into MSRN C can be used to define a sound and
automatic verification methods for TDL programs. For this purpose, we will exploit
a verification method introduced for MSR(C) in (Delzanno 2001; Delzanno 2005).
In the rest of this section we will briefly summarize how to adapt the main results
in (Delzanno 2001; Delzanno 2005) to the specific case of MSRN C .
Let us first reformulate the control state reachability problem of Example 1 for
the aforementioned safety property on the low level encoding into MSRN C . Given
the MSRN C initial configuration init we would like to check that no configuration
in P ost∗ ({init}) has the following form
{stopA (a1 , v1 , w1 ), stopB (a2 , v2 , w2 )} ⊕ M
20
Giorgio Delzanno
for ai , vi , wi ∈ Q i : 1, 2 and an arbitrary multiset of ground atoms M. Let us call
U the set of bad MSRN C configurations having the aforementioned shape. Notice
that U is upward closed with respect to multiset inclusion, i.e., if M ∈ U and
M 4 M′ , then M′ ∈ U . Furthermore, for if U is upward closed, so is P re(U ).
On the basis of this property, we can try to apply the methodology proposed in
(Abdulla and Nylén 2000) to develop a procedure to compute a finite representation
R of P re∗ U ). For this purpose, we need the following ingredients:
1. a symbolic representation of upward closed sets of configurations (e.g. a set
of assertions S whose denotation [[S]] is U );
2. a computable symbolic predecessor operator SP re working on sets of formulas
such that [[SP re(S)]] = P re([[S]]);
3. a (decidable) entailment relation Ent to compare the denotations of symbolic
representations, i.e., such that Ent(N, M ) implies [[N ]] ⊆ [[M ]]. If such a relation Ent exists, then it can be naturally extended to sets of formulas as
follows: EntS (S, S ′ ) if and only if for all N ∈ S there exists M ∈ S ′ such that
Ent(N, M ) holds (clearly, if Ent is an entailment, then EntS (S, S ′ ) implies
[[S]] ⊆ [[S ′ ]]).
The combination of these three ingredients can be used to define a verification
methods based on backward reasoning as explained next.
Symbolic Backward Reachability Suppose that M1 , . . . , Mn are the formulas of our
assertional language representing the infinite set U consisting of all bad configurations. The symbolic backward reachability procedure (SBR) procedure computes a
chain {Ii }i≥0 of sets of assertions such that
I0 = {M1 , . . . , Mn }
Ii+1 = Ii ∪ SP re(Ii ) for i ≥ 0
The procedure SBR stops when SP re produces only redundant information, i.e.,
EntS (Ii+1 , Ii ). Notice that EntS (Ii , Ii+1 ) always holds since Ii ⊆ Ii+1 .
Symbolic Representation In order to find an adequate represention of infinite sets of
MSRN C configurations we can resort to the notion of constrained configuration introduced in (Delzanno 2001) for the language scheme MSR(C) defined for a generic
constraint system C. We can instantiate this notion with N C constraints as follows.
A constrained configuration over P is a formula
p1 (x11 , . . . , x1k1 ) | . . . | pn (xn1 , . . . , xnkn ) : ϕ
where p1 , . . . , pn ∈ P, xi1 , . . . , xiki ∈ V for any i : 1, . . . n and ϕ is an N C-constraint.
.
The denotation a constrained configuration M = (M : ϕ) is defined by taking the
upward closure with respect to multiset inclusion of the set of ground instances,
namely
[[M ]] = {M′ | σ(M) 4 M′ , σ ∈ Sol(ϕ)}
Constraint-based Verification of Abstract Multithreaded Programs
21
This definition can be extended to sets of MSRN C constrained configurations with
disjoint variables (we use variable renaming to avoid variable name clashing) in the
natural way.
In our example the following set SU of MSRN C constrained configurations (with
distinct variables) can be used to finitely represent all possible violations U to the
considered safety property
SU = { stopA (i1 , n1 , m1 ) | stopB (i2 , n2 , m2 ) : n1 = n2 , m1 > m2
stopA (i1 , n1 , m1 ) | stopB (i2 , n2 , m2 ) : n1 = n2 , m2 > m1 }
Notice that we need two formulas to represent m1 6= m2 using a disjunction
of > constraints. The MSRN C configurations stopB (1, 2, 6) | stopA (4, 2, 5), and
stopB (1, 2, 6) | stopA (4, 2, 5) | waitA (2, 7, 3) are both contained in the denotation
of SU . Actually, we have that [[SU ]] = U . This symbolic representation allows us to
reason on infinite sets of MSRN C configurations, and thus on global configurations
of a TDL program, forgetting the actual number or threads of a given run.
To manipulate constrained configurations, we can instantiate to N C-constraints
the symbolic predecessor operator SP re defined for a generic constraint system in
(Delzanno 2005). Its definition is also given in Section Appendix A in Appendix.
From the general properties proved in (Delzanno 2005), we have that when applied
to a finite set of MSRN C constrained configurations S, SP reN C returns a finite set
of constrained configuration such that [[SP reN C (S)]] = P re([[S]]), i.e., SP reN C (S)
is a symbolic representation of the immediate predecessors of the configurations in
the denotation (an upward closed set) of S. Similarly we can instantiate the generic
entailment operator defined in (Delzanno 2005) to MSRN C constrained configurations so as to obtain an a relation Ent such that EntN C (N, M ) implies [[N ]] ⊆ [[M ]].
Based on these properties, we have the following result.
Proposition 1
Let T be a TDL program with initial global configuration G0 , Furthermore, let
S be the corresponding MSRN C encoding. and SU be the set of MSRN C constrained configurations denoting a given set of bad TDL configurations. Then,
init 6∈ SP re∗N C (SU ) if and only if there is no finite run G0 . . . Gn and mappings h0 , . . . , hn from the names occurring in G to non-negative rationals such
that init• G•0 (h0 ) . . . G•n (hn ) is a run in S and G•n (hn ) ∈ [[U ]].
Proof
Suppose init 6∈ SP re∗N C (U ). Since [[SP reN C (S)]] = pre([[S]]) for any S, it follows
that there cannot exist runs initM0 . . . Mn in S such that Mn ∈ [[U ]]. The thesis
then follows from the Theorem 2.
As discussed in (Bozzano and Delzanno 2002), we have implemented our verification procedure based on M SR and linear constraints using a CLP system with linear
arithmetics. By the translation presented in this paper, we can now reduce the verification of safety properties of multithreaded programs to a fixpoint computation
built on constraint operations. As example, we have applied our CLP-prototype
to automatically verify the specification of Fig. 6. The unsafe states are those described in Section 4. Symbolic backward reachability terminates after 18 iterations
22
Giorgio Delzanno
and returns a symbolic representation of the fixpoint with 2590 constrained configurations. The initial state init is not part of the resulting set. This proves our
original thread definitions correct with respect to the considered safety property.
4.1 An Interesting Class of TDL Programs
The proof of Theorem 1 shows that verification of safety properties is undecidable for TDL specifications in which threads have several local variables (they
are used to create linked lists). As mentioned in the introduction, we can apply the sufficient conditions for the termination of the procedure SBR given in
(Bozzano and Delzanno 2002; Delzanno 2005) to identify the following interesting
subclass of TDL programs.
Definition 4
A monadic TDL thread definition P = hQ, s, V, Ri is such that V is at most a
singleton, and every message template in R has at most one variable.
A monadic thread definition can be encoded into the monadic fragment of MSRN C
studied in (Delzanno 2005). Monadic MSRN C specifications are defined over atomic
formulas of the form p or p(x) with p is a predicate symbol and x is a variable, and
on atomic constraints of the form x = y, and x > y. To encode a monadic TDL
thread definitions into a Monadic MSRN C specification, we first need the following
observation. Since in our encoding we only use the constant 0, we first notice that
we can restrict our attention to MSRN C specifications in which constraints have no
constants at all. Specifically, to encode the generation of fresh names we only have
to add an auxiliary atomic formula zero(z), and refer to it every time we need to
express the constant 0. As an example, we could write rules like
init −−→ f resh(x) | initM (y) | zero(z) : x > z, y = z
for initialization, and
create(x) | zero(z) −−→ initM (x′ ) | initA (id′ , n′ , m′ ) | zero(z) :
x′ = x, id′ = x, n′ = z, m′ = z, z ′ = z
for all assignments involving the constant 0. By using this trick an by following the
encoding of Section 3, the translation of a collection of monadic thread definitions
directly returns a monadic MSRN C specification. By exploiting this property, we
obtain the following result.
Theorem 3
The verification of safety properties whose violations can be represented via an
upward closed set U of global configurations is decidable for a collection T of
monadic TDL definitions.
Proof
Let S be the MSRN C encoding of T and SU be the set of constrained configuration
such that SU = U . The proof is based on the following properties. First of all, the
MSRN C specification S is monadic. Furthermore, as shown in (Delzanno 2005),
Constraint-based Verification of Abstract Multithreaded Programs
23
the class of monadic MSRN C constrained configurations is closed under application
of the operator SP reN C . Finally, as shown in (Delzanno 2005), there exists an
entailment relation CEnt for monadic constrained configurations that ensures the
termination of the SBR procedure applied to a monadic MSRN C specification.
Thus, for the monadic MSRN C specification S, the chain defined as I0 = SU ,
Ii+1 = Ii ∪ SP re(Ii ) always reaches a point k ≥ 1 in which CEntS (Ik+1 , Ik ), i.e.
[[Ik ]] is a fixpoint for P re. Finally, we note that we can always check for membership
of init in the resulting set Ik .
As shown in (Schnoebelen 2002), the complexity of verification methods based on
symbolic backward reachability relying on the general results in (Abdulla and Nylén 2000;
Finkel and Schnoebelen 2001) is non primitive recursive.
5 Conclusions and Related Work
In this paper we have defined the theoretical grounds for the possible application
of constraint-based symbolic model checking for the automated analysis of abstract
models of multithreaded concurrent systems providing name generation, name mobility, and unbounded control. Our verification approach is based on an encoding
into a low level formalism based on the combination of multiset rewriting and
constraints that allows us to naturally implement name generation, value passing,
and dynamic creation of threads. Our verification method makes use of symbolic
representations of infinite set of system states and of symbolic backward reachability. For this reason, it can be viewed as a conservative extension of traditional
finite-state model checking methods. The use of symbolic state analysis is strictly
related to the analysis methods based on abstract interpretation. A deeper study
of the connections with abstract interpretation is an interesting direction for future
research.
Related Work The high level syntax we used to present the abstract models of
multithreaded programs is an extension of the communicating finite state machines
used in protocol verification (Bochmann 1978), and used for representing abstraction of multithreaded software programs (Ball et al. 2001). In our setting we enrich
the formalism with local variables, name generation and mobility, and unbounded
control. Our verification approach is inspired by the recent work of Abdulla and
Jonsson. In (Abdulla and Jonsson 2003), Abdulla and Jonsson proposed an assertional language for Timed Networks in which they use dedicated data structures
to symbolically represent configurations parametric in the number of tokens and
in the age (a real number) associated to tokens. In (Abdulla and Nylén 2000), Abdulla and Nylén formulate a symbolic algorithm using existential zones to represent the state-space of Timed Petri Nets. Our approach generalizes the ideas of
(Abdulla and Jonsson 2003; Abdulla and Nylén 2000) to systems specified via multiset rewriting and with more general classes of constraints. In (Abdulla and Jonsson 2001),
the authors apply similar ideas to (unbounded) channel systems in which messages
can vary over an infinite name domain and can be stored in a finite (and fixed a
24
Giorgio Delzanno
priori) number of data variables. However, they do not relate these results to multithreaded programs. Multiset rewriting over first order atomic formulas has been proposed for specifying security protocols by Cervesato et al. in (Cervesato et al. 1999).
The relationships between this framework and concurrent languages based on process algebra have been recently studied in (Bistarelli et al. 2005). Apart from approaches based on Petri Net-like models (as in (German and Sistla 1992; Ball et al. 2001)),
networks of finite-state processes can also be verified by means of automata theoretic techniques as in (Bouajjani et al. 2000). In this setting the set of possible
local states of individual processes are abstracted into a finite alphabet. Sets of
global states are represented then as regular languages, and transitions as relations
on languages. Differently from the automata theoretic approach, in our setting
we handle parameterized systems in which individual components have local variables that range over unbounded values. The use of constraints for the verification
of concurrent systems is related to previous works connecting Constraint Logic
Programming and verification, see e.g. (Delzanno and Podelski 1999). In this setting transition systems are encoded via CLP programs used to encode the global
state of a system and its updates. In the approach proposed in (Delzanno 2001;
Bozzano and Delzanno 2002), we refine this idea by using multiset rewriting and
constraints to locally specify updates to the global state. In (Delzanno 2001), we
defined the general framework of multiset rewriting with constraints and the corresponding symbolic analysis technique. The language proposed in (Delzanno 2001) is
given for a generic constraint system C (taking inspiration from CLP the language
is called M SR(C)). In (Bozzano and Delzanno 2002), we applied this formalism to
verify properties of mutual exclusion protocols (variations of the ticket algorithm)
for systems with an arbitrary number of processes. In the same paper we also formulated sufficient conditions for the termination of the backward analysis. The
present paper is the first attempt of relating the low level language proposed in
(Delzanno 2001) to a high level language with explicit management of names and
threads.
Acknowledgments The author would like to thank Ahmed Bouajjani, Andrew Gordon, Fabio Martinelli, Catuscia Palamidessi, Luca Paolini, and Sriram Rajamani
and the anonymous reviewers for several fruitful comments and suggestions.
References
Abdulla, P. A., Cerāns, K., Jonsson, B., and Tsay, Y.-K. 1996. General Decidability Theorems for Infinite-State Systems. In Proceedings 11th Annual International
Symposium on Logic in Computer Science (LICS’96). IEEE Computer Society Press,
New Brunswick, New Jersey, 313–321.
Abdulla, P. A. and Jonsson, B. 2001. Ensuring Completeness of Symbolic Verification
Methods for Infinite-State Systems. Theoretical Computer Science 256, 1-2, 145–167.
Abdulla, P. A. and Jonsson, B. 2003. Model checking of systems with many identical
timed processes. Theoretical Computer Science 290, 1, 241–264.
Abdulla, P. A. and Nylén, A. 2000. Better is Better than Well: On Efficient Verification
of Infinite-State Systems. In Proceedings 15th Annual International Symposium on
Constraint-based Verification of Abstract Multithreaded Programs
25
Logic in Computer Science (LICS’00). IEEE Computer Society Press, Santa Barbara,
California, 132–140.
Ball, T., Chaki, S., and Rajamani, S. K. 2001. Parameterized Verification of Multithreaded Software Libraries. In 7th International Conference on Tools and Algorithms
for Construction and Analysis of Systems (TACAS 2001), Genova, Italy, April 2-6,.
LNCS, vol. 2031. Springer-Verlag, 158–173.
Bistarelli, S., Cervesato, I., Lenzini, G., and Martinelli, F. 2005. Relating multiset
rewriting and process algebras for security protocol analysis. Journal of Computer
Security 13, 1, 3–47.
Bochmann, G. V. 1978. Finite state descriptions of communicating protocols. Computer
Networks 2, 46–57.
Bouajjani, A., Jonsson, B., Nilsson, M., and Touili, T. 2000. Regular Model Checking. In Proceedings 12th International Conference on Computer Aided Verification
(CAV’00), E. A. Emerson and A. P. Sistla, Eds. LNCS, vol. 1855. Springer-Verlag,
Chicago, Illinois, 403–418.
Bozzano, M. and Delzanno, G. 2002. Algorithmic verification of invalidation-based
protocols. In 14th International Conference on Computer Aided Verification, CAV ’02.
Lecture Notes in Computer Science, vol. 2404. Springer.
Cervesato, I., Durgin, N., Lincoln, P., Mitchell, J., and Scedrov, A. 1999. A
Meta-notation for Protocol Analysis. In 12th Computer Security Foundations Workshop
(CSFW’99). IEEE Computer Society Press, Mordano, Italy, 55–69.
Delzanno, G. 2001. An Assertional Language for Systems Parametric in Several Dimensions. In Verification of Parameterized Systems - VEPAS 2001. ENTCS, vol. 50.
Delzanno, G. 2005. Constraint Multiset Rewriting. Tech. Rep. TR-05-08, Dipartimento
Informatica e Scienze dell’Informazione, Università di Genova, Italia.
Delzanno, G. and Podelski, A. 1999. Model checking in CLP. In Proceedings 5th
International Conference on Tools and Algorithms for Construction and Analysis of
Systems (TACAS’99). Lecture Notes in Computer Science, vol. 1579. Springer-Verlag,
Amsterdam, The Netherlands, 223–239.
Finkel, A. and Schnoebelen, P. 2001. Well-Structured Transition Systems Everywhere!
Theoretical Computer Science 256, 1-2, 63–92.
German, S. M. and Sistla, A. P. 1992. Reasoning about Systems with Many Processes.
Journal of the ACM 39, 3, 675–735.
Gordon, A. D. 2001. Notes on nominal calculi for security and mobility. In Foundations
of Security Analysis and Design, Tutorial Lectures. Lecture Notes in Computer Science,
vol. 2171. Springer, 262–330.
Kesten, Y., Maler, O., Marcus, M., Pnueli, A., and Shahar, E. 2001. Symbolic
model checking with rich assertional languages. Theoretical Computer Science 256, 1,
93–112.
Schnoebelen, P. 2002. Verifying Lossy Channel Systems has Nonprimitive Recursive
Complexity. Information Processing Letters 83, 5, 251–261.
Appendix A Symbolic Predecessor Operator
Given a set of MSRN C configurations S, consider the MSRN C predecessor operator
P re(S) = {M|M ⇒MSR M′ , M′ ∈ S}. In our assertional language, we can define a
symbolic version SP reN C of P re defined on a set S containing MSRN C constrained
26
Giorgio Delzanno
multisets (with disjoint variables) as follows:
SP reN C (S) = { (A ⊕ N : ξ) |
(A −→ B : ψ) ∈ R, (M : ϕ) ∈ S,
M′ 4 M, B ′ 4 B,
(M′ : ϕ) =θ (B ′ : ψ), N = M ⊖ M′ ,
ξ ≡ (∃x1 . . . . xk .θ)
and x1 , . . . , xk are all variables not in A ⊕ N }.
where =θ is a matching relation between constrained configurations that also takes
in consideration the constraint satisfaction, namely
(A1 | . . . | An : ϕ) =θ (B1 | . . . | Bm : ψ)
provided m = n and there exists a permutation j1 , . . . , jn of 1, . . . , n such that
Vn
the constraint θ = ϕ ∧ ψ ∧ i=1 Ai = Bji is satisfiable; here p(x1 , . . . , xr ) =
q(y1 , . . . , ys ) is an abbreviation for the constraints x1 = y1 ∧ . . . ∧ xr = ys if p = q
and s = r, f alse otherwise.
As proved in (Delzanno 2005), the symbolic operator SP reN C returns a set of
MSRN C constrained configurations and it is correct and complete with respect to
P re, i.e., [[SP reN C (S)]] = P re([[S]]) for any S. It is important to note the difference
between SP reN C and a simple backward rewriting step.
For instance, given the constrained configurations M defined as p(x, z) | f (y) : z >
y and the rule s(u, m) | r(t, v) → p(u′ , m′ ) | r(t′ , v ′ ) : u = t, m′ = v, v ′ =
v, u′ = u, t′ = t (that simulates a rendez-vous (u, t are channels) and value passing
(m′ = v)), the application of SP re returns s(u, m) | r(t, v) | f (y) : u = t, v > y as
well as s(u, m) | r(t, v) | p(x, z) | f (y) : u = t, x > y (the common multiset here is
ǫ).
| 2cs.AI
|
arXiv:1801.06879v1 [physics.comp-ph] 21 Jan 2018
Bayesian Deep Convolutional Encoder-Decoder Networks
for Surrogate Modeling and Uncertainty Quantification
Yinhao Zhu, Nicholas Zabaras∗
Center for Informatics and Computational Science,
311I Cushing Hall, University of Notre Dame, Notre Dame, IN 46556, U.S.A
Abstract
We are interested in the development of surrogate models for uncertainty
quantification and propagation in problems governed by stochastic PDEs
using a deep convolutional encoder-decoder network in a similar fashion to
approaches considered in deep learning for image-to-image regression tasks.
Since normal neural networks are data intensive and cannot provide predictive uncertainty, we propose a Bayesian approach to convolutional neural
nets. A recently introduced variational gradient descent algorithm based
on Stein’s method is scaled to deep convolutional networks to perform approximate Bayesian inference on millions of uncertain network parameters.
This approach achieves state of the art performance in terms of predictive
accuracy and uncertainty quantification in comparison to other approaches
in Bayesian neural networks as well as techniques that include Gaussian
processes and ensemble methods even when the training data size is relatively small. To evaluate the performance of this approach, we consider
standard uncertainty quantification benchmark problems including flow in
heterogeneous media defined in terms of limited data-driven permeability
realizations. The performance of the surrogate model developed is very
good even though there is no underlying structure shared between the input
(permeability) and output (flow/pressure) fields as is often the case in the
image-to-image regression models used in computer vision problems. Studies are performed with an underlying stochastic input dimensionality up
to 4, 225 where most other uncertainty quantification methods fail. Uncertainty propagation tasks are considered and the predictive output Bayesian
∗
Corresponding author
Email addresses: [email protected] (Yinhao Zhu), [email protected] (Nicholas
Zabaras)
URL: https://cics.nd.edu/ (Nicholas Zabaras)
Preprint submitted to Journal of Computational Physics
January 23, 2018
statistics are compared to those obtained with Monte Carlo estimates.
Keywords: Uncertainty Quantification, Bayesian Neural Networks,
Convolutional Encoder-Decoder Networks, Deep Learning, Porous Media
Flows
1. Introduction
Uncertainty in complex systems arises from model error and model parametrization, unknown/incomplete material properties, boundary conditions or forcing terms, and other. Uncertainty propagation takes place by reformulating
the problem of interest as a system of stochastic partial differential equations
(SPDEs). Solution of such problems often needs to rely on the solution of
the deterministic problem at a finite number of realizations of the random
input using Monte Carlo sampling, or collocation methods. Considering the
computational cost of solving complex multiscale/multiphysics deterministic problems, one often relies on Bayesian surrogate models that are trained
with only a small number of deterministic solution runs while at the same
time they are capable of capturing the epistemic uncertainty introduced
from the limited training data [1].
For realistic problems in science and engineering, we only have access to
limited number (e.g. 100 or so) of deterministic simulation runs. Vanilla
Monte Carlo for uncertainty propagation is thus hopeless. A dominant solution is to train a surrogate model using the limited simulation-based training
data, and then perform prediction and uncertainty propagation tasks using
the surrogate instead of solving the actual PDEs. Unfortunately most existing surrogate models have difficulty scaling to high-dimensional problems,
such as the ones based on Gaussian processes (GP) [2, 3] or generalized polynomial chaos expansions (gPC [4]). High dimensionality often arises from
the discretization of properties with small correlation lengths (e.g. permeability in heterogeneous media flows), random distributed sources or force
input fields with multiple scales [5].
To alleviate the curse of stochastic input dimensionality, we usually assume that the given input data lie on an embedded non-linear manifold
within the higher dimensional space. This intrinsic dimensionality is captured by dimensionality reduction techniques [6], such as the KarhunenLoève expansion (KLE), t-SNE [7], auto-encoders [8], probabilistic methods
like variational auto-encoders [9], Gaussian process latent variable models
(GP-LVM) [10], and many more. Most dimensionality reduction models are
unsupervised learning problems that do not explicitly take the regression
2
task into account. Thus the classical approach to uncertainty quantification
is to first reduce the dimensionality of the input to obtain a low-dimensional
latent representation of the input field, then to built a regression model
from this latent representation to the output. This approach is certainly not
efficient is particular when the map from the latent representation to the
physical space of the input data is not available. In [2], KLE was used for dimensionality reduction of the permeability field and then GP was performed
as independent task for Bayesian regression. In [11], this approach was
taken one step further with the probabilistic mappings from input to latent
space and from latent space to output being modeled by generalized linear
models both trained simultaneously end-to-end using stochastic variational
inference instead of performing the unsupervised and supervised/regression
tasks separately.
One of the essential upcoming approaches for handling high-dimensional
data is to learn the latent input representation automatically by supervision with the output in regression tasks. This is the central idea of deep
neural networks [12], especially convolutional neural networks (CNNs) [13]
which stack (deeper) layers of linear convolutions with nonlinear activations
to automatically extract the multi-scale features or concepts from highdimensional input [14], thus alleviating the hand-craft feature engineering,
such as searching for the right set of basis functions, or relying on expert
knowledge.
However, the general perspective for using deep neural networks [15, 16]
in the context of surrogate modeling is that physical problems in uncertainty quantification (UQ) are not big data problems, thus not suitable for
addressing them with deep learning approaches. However, we argue otherwise in the sense that each simulation run generates large amount of data
which potentially reveal the essential characteristics about the underlying
system. In addition, even for a relatively small dataset, deep neural networks show unique generalization property [17, 18]. These are typically
over-parameterized models (hundreds and thousands of times more parameters than training data), but they do not overfit, i.e. the test error does not
grow as the network parameters increase. Deep learning has been explored
as a competitive methodology across fiels such as fluid mechanics [19, 20],
hydrology [21], bioinformatics [22], high energy physics [23] and other.
This unique generalization behavior makes it possible to use deep neural
networks for surrogate modeling. They are capable to capture the complex
nonlinear mapping between high-dimensional input and output due to their
expressiveness [24], while they only use small number of data from simulation runs. In addition, there has been a resurgence of interest in putting
3
deep neural network under a formal Bayesian framework. Bayesian deep
learning [25, 26, 27, 28, 29, 30, 31, 32, 33] enables the network to express its
uncertainty on its predictions when using a small number of training data.
The Bayesian neural networks can quantify the predictive uncertainty by
treating the network parameters as random variables, and perform Bayesian
inference on those uncertain parameters conditioned on limited observations.
In this work, we mainly consider surrogate modeling of physical systems
governed by stochastic partial differential equations with high-dimensional
stochastic input such as flow in random porous media [34]. The spatially
discretized stochastic input field and the corresponding output fields are
high-dimensional. We adopt an end-to-end image-to-image regression approach for this challenging surrogate modeling problem. More specifically,
a fully convolutional encoder-decoder network is designed to capture the
complex mapping directly from the high-dimensional input field to the output fields without using any explicit intermediate dimensionality reduction
method. To make the model more parameter efficient and compact, we use
DenseNet to build the feature extractor within the encoder and decoder
paths [35]. Intuitively, the encoder network extracts the multi-scale features
from the input data which are used by the decoder network to reconstruct
the output fields. In similarity with problems in computer vision, we treat
the input-output map as an image-to-image map. To account for the limited training data and endow the network with uncertainty estimates, we
further treat the convolutional encoder-decoder network to be Bayesian and
scale a recently proposed approximate inference method called Stein Variational Gradient Descent to modern deep convolutional networks. We will
show that the methodology can learn a Bayesian surrogate for a problem
with an intrinsic dimensionality of 50, achieving promising results on both
predictive accuracy and uncertainty estimates using as few as 32 training
data. More importantly we develop a surrogate for the case of 4225 dimensionality using 512 training data. We also show these uncertainty estimates
are well-calibrated using a reliability diagram. To this end, we believe that
Bayesian neural networks are strong candidates for surrogate modeling and
uncertainty propagation in high-dimensional problems with limited training
data.
The remaining of the paper is organized as follows: In Section 2, we
present the problem setup for surrogate modeling with high-dimensional input and the proposed approach in treating it as an image regression problem.
We then introduce the CNNs and the encoder-decoder network used in our
model. In Section 3, we present the Bayesian formulation of neural networks
and a non-parametric variational method for the underlying challenging ap4
proximate inference task. In Section 4, we provide implementation details
and show the performed experiments on a porous media flow problem. We
finally conclude and discuss the various unexplored research directions in
Section 5.
2. Methodology
2.1. Surrogate Modeling as Image-to-Image Regression
The physical systems considered here are modeled by stochastic PDEs
(SPDEs) with solutions y(s, x(s)), i.e. the model response y ∈ Rdy at the
spatial location s ∈ S ⊂ Rds (ds = 1, 2, 3), with one realization x(s) of the
random field {x(s, ω), s ∈ S, ω ∈ Ω}, where S is the index set and Ω is
the sample space. The formulation allows for multiple input channels (i.e.
dx > 1) even though our interest here is on one input property represented
as a vector random field. This random field appears in the coefficients of the
SPDEs, and is used to model material properties, such as the permeability or
porosity fields in geological media flows. We assume the computer simulation
for the physical systems is performed over a given set of spatial grid locations
S = {s1 , · · · , sns } (e.g. mesh nodes in finite element methods). In this case,
the random field x is discretized over the fixed grids S, thus is equivalent
to a high-dimensional random vector, denoted as x, where x ∈ X ⊂ Rdx ns .
The corresponding response y is solved over S, thus can be represented as
a vector y ∈ Y ⊂ Rdy ns .
With the discretization described above and assuming for simplicity fixed
boundary and initial conditions and source terms as appropriate, we consider
the computation simulation as a black-box mapping of the form:
η : X → Y.
(1)
In order to tackle the limitations of using the deterministic computationally expensive simulator for uncertainty propagation, a surrogate model
y = f (x, θ) is trained using limited simulation data D = {xi , yi }N
i=1 , to approximate the ‘ground-truth’ simulation-induced function y = η(x), where
θ are the model parameters, and N is the number of simulation runs (number of training simulation-based data).
Let us consider that the PDEs governing the physical system described
above are solved over 2D regular grids of H × W , where H and W denote
the number of grid points in the two axes of the spatial domain (height and
width), and ns = H · W . It is very natural to organize the simulation data
i
dx ×H×W is one input
as an image dataset D = {xi , yi }N
i=1 , where x ∈ R
field realization, and yi ∈ Rdy ×H×W is the simulated steady-state output
5
fields for xi discretized over the same grids. Here dx , dy are the number
of dimensions for the input x and the output y at one location. These are
treated herein as the number of channels in input and output images, similar
to RGB channels in natural images. It is easy to generalize to the 3D spatial
domain by adding an extra depth axis to the images, e.g. xi ∈ Rdx ×D×H×W ,
and yi ∈ Rdy ×D×H×W .
Therefore, we transform the surrogate modeling problem to an imageto-image regression problem, with the regression function as
η : Rdx ×H×W → Rdy ×H×W .
(2)
In distinction from an image classification problem which requires imagewise prediction, the image regression problem is concerned with pixel-wise
predictions, e.g. predicting the depth of each pixel in an image, or in our
physical problem, predicting the output fields at each grid point. Such
problems have been intensively studied within the computer vision community by leveraging the rapid recent progress of convolutional neural networks (CNNs), such as AlexNet [13], VGG [36], Inception [37], ResNet [38],
DenseNet [35], and many more. A common model design pattern for semantic segmentation [39] or depth regression [40] is the encoder-decoder architecture. The intuition behind regression between two high-dimensional objects
is to go through a coarse-refine process, i.e. to reduce the spatial dimension of the input image to high-level coarse features using an encoder, and
then recover the spatial dimension by refining the coarse features through a
decoder. One of the characteristics shared by those vision tasks is that the
input and output images share the underlying structure, or they are different
renderings of the same underlying structure [41]. However, for our surrogate
modeling tasks, the input and output images appear to be quite different,
due to the complex physical influences (defined by PDEs) of the random
input field, forcing terms and boundary conditions on the system response.
This was after all the reason of pursuing the training of a surrogate model
that avoids the repeated solution of the PDEs for different input realizations.
Surprisingly, as we will discuss later on in this paper, the encoder-decoder
network still works very well.
Remark 1. The training data for the surrogate model of interest here include the realizations of the random input field and the corresponding multioutput obtained from simulation. Of interest is to address problems with
limited training data sets considering the high-computational cost of each
simulation run. However, note that key UQ tasks include the ability to
predict the system response and our confidence on it using input realiza6
tions (testing dataset) consistent with the given training data but also the
computation of the statistics of the response induced by the random input.
Both of these tasks require the availability of a high-number of input data
sets (e.g. 500 data points for testing, and 104 input data points for Monte
Carlo calculation of the output statistics). The problem of generating more
input realizations using only the training dataset is the solution of a generative model problem. There has been significant progress in recent years in
the topic, such as the generative adversarial networks (GANs) [42] and its
ever exploding variants, variational auto-encoders (VAEs) [9], autoregressive
models like PixelCNN [43], PixelRNN [44], and other. However, note that in
this work our focus is on the image-to-image mapping and its performance
on uncertainty quantification tasks. We will thus assume that enough input samples are provided both for testing and output statistics calculation
even though only a small dataset will be used for training. In our examples
in Section 4, synthetic log-permeability datasets are generated by sampling
a Gaussian random field with an exponential kernel. The output for each
permeability sample is generated using a deterministic simulator.
2.2. Dense Convolutional Encoder-Decoder Networks
In this subsection, we briefly introduce a state-of-the-art CNN architecture called DenseNet [35] and fully convolutional encoder-decoder networks
developed in computer vision, and then present how to utilize these advances to build our baseline network for surrogate modeling in uncertainty
quantification.
2.2.1. Densely Connected Convolutional Networks
DenseNet [35] is a recently proposed CNN architecture which extends the
ideas of ResNet [38] and Highway Networks [45] to create dense connections
between all layers, so as to improve the information (gradient) flow through
the network for better parameter efficiency.
Let xl be the output of the lth layer. Traditionally CNNs pass the
output of one layer only to the input of the next layer, i.e. xl = hl (xl−1 ),
where hl denotes the nonlinear function of the lth hidden layer. In current
CNNs, h is commonly defined as a composition of Batch Normalization [46]
(BatchNorm), Rectified Linear Unit [47] (ReLU) and Convolution (Conv)
or transposed convolution (ConvT) [48]. ResNets [38] create an additional
identity mapping that bypasses the nonlinear layer, i.e. xl = hl (xl−1 )+xl−1 .
In this way, the nonlinear layer only needs to learn a residual function which
facilitates the training of deeper networks.
7
DenseNets [35] introduce connections from any layer to all subsequent
layers, i.e. xl = hl ([xl−1 , xl−2 , · · · , x0 ]). To put it in another way, the input
features of one layer are concatenated to the output features of this layer
and this serves as the input features to the next layer. Assume that the
input image has K0 channels, and each layer outputs K feature maps, then
the lth layer would have input with K0 + (l − 1) · K feature maps, i.e. the
number of feature maps in DenseNet grows linearly with the depth. K is
here referred to as the growth rate.
For image regression based on encoder-decoder networks, downsampling
and upsampling are required to change the size of feature maps, which makes
concatenation of feature maps unfeasible. Dense blocks and transition layers
are introduced to solve this problem and modularize the network design.
A dense block contains multiple densely connected layers whose input and
output feature maps are of the same size. It contains two design parameters,
namely the number L of layers within and the growth rate K for each layer.
An illustration of the dense block is in Fig. 1.
(a) Dense block.
(b) The second layer.
Figure 1: (a) A dense block contains L = 3 layers h1 , h2 , h3 with growth rate K = 2.
(b) The second layer h2 of the dense block, where x2 = h2 ([x1 , x0 ]) is its output feature
map. Notice that the input to the third layer is the concatenation of the output and input
features of h2 , i.e. [x2 , x1 , x0 ]. As is often the case, each layer is composed of Batch Normalization [46] (BatchNorm), Rectified Linear Unit [47] (ReLU) and Convolution (Conv).
The convolution kernel has kernel size k = 3, stride s = 1 and zero padding p = 1, which
keep the size of the feature maps the same as the input.
Transition layers are used to change the size of feature maps and reduce
their number between dense blocks. More specifically, the encoding layer
typically halfs the size of feature maps, while the decoding layer doubles the
feature map size. Both of the two layers reduce the number of feature maps.
This is illustrated in Fig. 2.
8
(a) Encoding layer.
(b) Decoding layer.
Figure 2: Both (a) encoding layer and (b) decoding layer contain two convolutions. The
first convolution reduces the number of feature maps while keeps their size the same using
a kernel with parameters k = 1, s = 1, p = 0; the second convolution changes the size
of the feature maps but not their number using a kernel k = 3, s = 2, p = 1. The main
difference between (a) and (b) is in the type of the second convolution layer, which is
Conv and ConvT respectively, for downsampling and upsampling. Note that no pooling
is used at transition layers for maintaining the location information. The colored feature
maps used here are independent from the feature maps with the same color shown in other
figures.
2.2.2. Fully Convolutional Networks
The fully convolutional networks (FCNs) [49] are extensions of CNNs for
pixel-wise prediction, e.g. semantic segmentation. FCNs replace the fully
connected layers in CNNs with convolution layers, add upsampling layers in
the end to recover the input spatial resolution, and introduce the skip connections between feature maps in downsampling and upsampling path to
recover finer information lost in the downsampling path. Most of the recent
work focuses in improving the upsampling path and increase the connectivity within and between upsampling and downsampling paths. U-nets [39]
extend the upsampling path as symmetric to the downsampling path and
add skip connections between each size of feature maps in the downsampling and upsampling paths. Within SegNets [50], the decoder uses pooling
indices computed in the max-pooling step of the corresponding encoder to
perform non-linear upsampling. Fully convolutional DenseNets [51] extend
DenseNets to FCNs, which are closest to our network design but with several differences. We keep all the feature maps of a dense block concatenated
so far before passing to the transition layers, while they only keep the output feature maps of the last convolution layer within the dense block. The
feature maps explosion problem is addressed by the first convolution layer
9
within the transition layer. Besides that we do not use skip connections
between encoding and decoding paths because of the weak correspondence
between the input and output images. We also do not use max-pooling for
encoding layers, instead we use convolution with stride 2.
2.3. Network architecture: DenseED
We follow the fully convolutional networks (FCNs) [49] for image segmentation without using any fully connected layers, and encode-decoder
architecture similar to U-net [39] and SegNet [50] but without the concatenation of feature maps between the encoder paths and decoder paths.
Furthermore, we adapt the DenseNet [35] structure into the encoder and
decoder networks. After extensive hyperparameter and architecture search,
we arrived at a baseline dense convolutional encoder-decoder network, called
DenseED, similar to the network proposed in [51] but with noticeable differences as stated above and shown in Fig. 3.
Figure 3: Network architecture: DenseED.
In the encoding path, the input field realizations are fed into the first
convolution layer with large kernel size k = 7, stride s = 2 and zero padding
p = 2. Then the extracted feature maps are passed through an alternative
cascade of dense blocks and encoding layers as introduced in Figs. 1 and 2.
The dense block after the last encoding layer outputs the high-level coarse
feature maps extracted from the input, as shown in purple at the right end
of the network in Fig. 3, which are subsequently fed into the decoder path.
The decoding network consists of an alternation of dense blocks and decoding
layers, with the last decoding layer directly leading to the prediction of the
output fields.
10
2.4. Network architecture engineering and hyperparameter search
Network architecture engineering and hyperparameter search are among
the main challenges and source of innovations in deep learning, mostly
problem-specific and empirical. The general network architecture is introduced in Section 2.2, which is based on the recent development of neural
network design for image segmentation. For our image regression problem,
the main design considerations include the following:
• Downsampling layers: convolution or pooling;
• Upsampling layers: bilinear upsampling or transposed convolution;
• Smallest spatial dimensions of feature maps: this is determined by the
number of downsampling layers;
• Add or not of skip connections between the encoding and decoding
paths;
• Kernel of convolution layers, kernel size k, stride s, zero padding p;
• Number of layers L and growth rate K within each dense block;
• Regularizations: weight decay, batch normalization, dropout, etc;
• Optimizer: stochastic gradient descent algorithms and their variants,
such as Adam, Adagrad, RMSprop, and others;
• Training hyperparameters: batch size, learning rate and its scheduler.
The details of architecture search and hyperparameter selection for the
particular problem considered are presented in Appendix A where we also
report various experiments using DenseED for surrogate modeling with limited training data. No overfitting was observed in our calculations and the
obtained results were quite good. This is an intriguing and active research
topic in the deep learning community [17].
In our non-Bayesian calculations, we have considered L2 or L1 regularized MSE training loss function. Given an input image x, and a target image
y, the prediction f (x, w), the regularized MSE loss is
n
2
1 X
L(f (x, w), y) =
(3)
fi − yi + αΩ(w),
n
i=1
1 >
where the
Ppenalty function Ω(w) = 2 w w for L2 regularization, and Ω(w) =
kwk1 = i |wi | for L1 regularization, and n = Cout ·Hout ·Wout is the number
11
of pixels in all channels of one output image. w denotes all the parameters
in the network. For our case, it includes the kernel weights in all the convolution and transposed convolution layers (no bias is used in convolutional
kernel), the scale and shift parameters in all the batch normalization layers.
See Section 4.3 for an example of the fully convolutional encoder-decoder
network used for the Darcy flow problem. Note that L2 regularization is
implemented in PyTorch optimizers by specifying weight decay, which is α
in Eq. (3).
The network architecture selected for the non-Bayesian model is the same
as that used for our Bayesian model introduced next.
Remark 2. In the encoder-decoder network, the batch normalization layer
used after each convolutional layer can also be considered as an effective regularizer1 . It is commonly adopted nowadays in deep convolutional networks2
replacing dropout3 .
3. Bayesian Neural Networks
Consider a deterministic neural net y = f (x, w) with input x, output y,
and all parameters w including the weights and biases. 4 Bayesian neural
networks (BNNs) treat the parameters w as random variables instead of
deterministic unknowns to account for epistemic uncertainty induced by
lack of training data. Besides that, usually additive noise n is introduced to
model the aleatoric uncertainty which can not be reduced by having more
observations, also to make the probabilistic model have an explicit likelihood
depending on the noise distribution, i.e.
y = f (x, w) + n,
(4)
where f (x, w) is the output of a neural network with the uncertain w, and
n is the additive noise.
3.1. Sparsity inducing prior on weights
Given the large amount of ‘un-interpretable’ parameters w in a deep
neural net, there are not many choices of priors. But the demand for com1
https://openreview.net/forum?id=BJlrSmbAZ¬eId=BJlrSmbAZ
https://github.com/pytorch/vision/tree/master/torchvision/models
3
https://www.reddit.com/r/MachineLearning/comments/5l3f1c/d_what_
happened_to_dropout/
4
http://pytorch.org/docs/master/nn.html?ht=conv2d#torch.nn.functional.
conv2d
2
12
pression [52, 33] of the neural net for lower memory and computation cost
calls for sparsity promoting priors. We assume a fully factorized Gaussian
prior with zero mean and Gamma-distributed precision α on parameters w
p(w | α) = N (w | 0, α−1 I),
p(α) = Gamma(α | a0 , b0 ).
(5)
This results in a student’s t-prior for w, which has heavy tails and more
mass close to zero.
3.2. Additive Noise Model
Additive noise can be considered of the following form:
• Output-wise: n = σ, same for all output pixels;
• Channel-wise: n = [σ1 1 , · · · , σdy dy ], same across each of the dy output channels/fields;
• Pixel-wise: n = σ
, distinct for each output pixel.
d
y
Here, σ, {σi }i=1
are scalars, σ is a field with the same dimension as the
output y and
denotes the element-wise product operator. In this work,
we have considered both Gaussian noise, ∼ N (0, I) and Laplacian noise,
∼ Laplace(0, I)5 . In the numerical results discussed in Section 4, we
concentrate in the the output-wise and channel-wise cases above. We treat
the noise precision β = 1/σ 2 as a random variable with a conjugate prior
p(β) = Gamma(β | a1 , b1 ). For the generated training data, a priori we
assume that the noise variance (also known as nugget [53]) to be very small
e.g. 10−6 . Thus the values a1 = 2, b1 = 2 · 10−6 provide a good initial guess
for the prior hyperparameters.
Remark 3. We can also model the noise varying with input (pixel-wise
noise model), resulting in a heteroscedastic noise model, i.e. n(x, w) =
σ(x, w) or n(x, w) = σ(x, w) . Again can be Gaussian or Laplacian.
The heteroscedastic noise [54, 55, 56] can be implemented as extending the
output of the neural net as:
[f (x, w), σ 2 (x, w)] or [f (x, w), σ 2 (x, w)].
(6)
The output of the system becomes y = f (x, w) + n(x, w). The pixel-wise
case σ 2 (x, w) may help capture large variations for example near discontinuous regions of the output. In practice, we apply a softplus transformation
5
http://homepages.inf.ed.ac.uk/rbf/CVonline/LOCAL_COPIES/VELDHUIZEN/
node11.html
13
to the second part of the output of the neural net to enforce the positive
variance constraint, i.e. σ 2 = log(1 + exp(·)) + eps, where eps = 10−10 for
numerical stability.
3.3. Stein Variational Gradient Descent (SVGD)
Approximate inference for Bayesian deep neural network is a daunting
task because of the large number of uncertain parameters, e.g. tens or
hundreds of millions in modern deep networks. In our surrogate problem,
the task is to find a high-dimensional posterior distribution over millions of
random variables using less than hundreds or thousands of training data.
As reviewed in Section 1, most of variational inference methods [57] restrict the approximate posterior within certain parametric variational family, while sampling-based methods are slow and difficult to converge. Here
we adopt a recently proposed non-parametric variational inference method
called stochastic variational gradient descent (SVGD) [29, 58] that is similar to standard gradient descent while maintaining the efficiency of particle
methods.
For a prescribed probabilistic model with likelihood function p(y | θ, x)
and prior p0 (θ), we are interested in Bayesian inference of the uncertain
parameters θ, i.e. to find the posterior distribution p(θ | D), where D
denote the i.i.d. observations (training data), i.e. D = {xi , yi }N
i=1 . For the
BNNs with homoescedastic Gaussian noise case, θ = {w, β}. Variational
inference aims to approximate the target posterior distribution p(θ | D) with
a variational distribution q ∗ (θ) which lies in a restricted set of distributions
Q by minimizing the KL divergence between the two, i.e.
q ∗ (θ) = arg min KL(q(θ) k p(θ | D)) = arg min Eq [log q(θ)−log p̃(θ | D)+log Z],
q∈Q
q∈Q
QN
where p̃(θ | D) = p(D | Rθ)p0 (θ) = i=1 p(yi | θ, xi )p0 (θ) is the unnormalized posterior, and Z = p̃(θ)dθ is the normalization constant or model
evidence, which is usually computationally intractable, but can be ignored
when we optimize the KL divergence.
The variational family considered here is a set of distributions obtained
by smooth transforms from an initial tractable distribution (e.g. the prior)
represented in terms of particles. The transforms applied to each particle
take the following form:
T(θ) = θ + φ(θ),
(7)
where is the step size, φ(θ) ∈ F is the perturbation direction within a
14
function space F. When is small, T transforms the initial density q(θ) to
q[T] (θ) = q(T−1 (θ))|det(∇T−1 (θ))|.
Instead of using parametric form for the variational posterior, a particle
approximation
is used, i.e. a set of particles {θ i }Si=1 with empirical measure
1 PS
µS (dθ) = S i=1 δ(θ−θ i )dθ. We would like to have µ to weakly converge to
the measure of the true posterior νp (dθ) = p(θ)dθ. We apply the transform
T to those particles, and denote the pushforward measure of µ as Tµ.
The problem is to find out the direction to maximally decrease the KL
divergence of the variational approximation and the target distribution, i.e.
to solve the following functional optimization problem:
n
o
d
max − KL(Tµ k νp )|=0 .
(8)
φ∈F
d
It turns out that [29]
d
KL(Tµ k νp )|=0 = Eµ [Tp φ],
d
where Tp is called the Stein operator associated to the distribution p,
−
Tp φ =
(9)
∇ · (pφ)
(∇p) · φ + p(∇ · φ)
=
= (∇ log p) · φ + ∇ · φ.
p
p
The expectation Eµ [Tp φ] evaluates the difference between p and µ, and its
maximum is defined as the Stein discrepancy,
S(µ, p) = max Eµ [Tp φ].
φ∈F
(10)
It has been shown [59] that when the functional space F is chosen to
be the unit ball in a product reproducing kernel Hilbert space H with the
positive kernel k(x, x0 ), the maximal direction to perturb (or the Stein discrepancy) has a closed-form solution,
0
φ∗ (θ) ∝ Eθ0 ∼µ [Tpθ k(θ, θ 0 )] = Eθ0 ∼µ [∇θ0 log p(θ 0 )k(θ, θ 0 ) + ∇θ0 k(θ, θ 0 )].
Thus we have the following algorithm to transform an initial distribution
µ0 to the target posterior νp .
This is an one-line algorithm, where the gradient φ(θ) pushes the particles towards the high posterior region by kernel smoothed gradient term
k(·, ·)∇ log p, while maintaining a degree of diversity by repulsive force term
∇k(θ, θ 0 ). When the number of particles becomes 1, then the algorithm reduces to the MAP estimate of the posterior. This algorithm is implemented
in PyTorch with GPU acceleration and scaled to our deep convolutional
encoder-decoder network DenseED.
15
Algorithm 1: Bayesian inference by Stein Variational Gradient Descent.
Input: A set of initial particles {θ0i }Si=1 , score function ∇ log p(θ),
kernel k(θ, θ 0 ), step-size scheme {t }
Result: A set of particles θ i that approximate the target posterior
for iteration t do
i
θt+1
← θti + t φ(θti ) ;
i
P h
φ(θti ) = S1 Si=1 k(θtj , θti )∇θj log p(θtj ) + ∇θj k(θtj , θti )
t
t
end
Here we use one toy example to illustrate the idea of SVGD. We start
with the 20 particles from the Normal distribution N (−10, 1), and tranport the particles iteratively to the target Gaussian mixture distribution
0.8N (−2, 1) + 0.2N (2, 1) with Algorithm 1.
0.35
0.35
0.30
0.30
0.25
0.25
0.20
0.20
0.20
0.15
0.15
0.15
0.35
0.30
0.25
0.30
0.30
0.25
0.25
0.20
0.20
0.15
0.15
0.10
0.10
0.10
0.10
0.05
0.05
0.05
0.05
0.00
15.0
12.5
10.0
7.5
5.0
2.5
0.0
2.5
(a) Iter 0
5.0
0.00
15.0
12.5
10.0
7.5
5.0
2.5
0.0
2.5
5.0
(b) Iter 50
0.00
15.0
12.5
10.0
7.5
5.0
2.5
0.0
2.5
5.0
(c) Iter 100
0.00
0.10
0.05
15.0
12.5
10.0
7.5
5.0
2.5
0.0
2.5
5.0
(d) Iter 200
0.00
15.0
12.5
10.0
7.5
5.0
2.5
0.0
2.5
5.0
(e) Iter 450
Figure 4: Example of transporting 20 particles from N (−10, 1) to Gaussian mixture
0.8N (−2, 1) + 0.2N (2, 1) with SVGD using 20 particles. The green dash represents the
target Gaussian mixture.
3.3.1. Implementation and Training
• We use S samples of θ to approximate the empirical measure of the
posterior. They are initialized and stored in 20 deterministic DenseED
neural networks.
• At each step t, for each model j, the gradient of its joint likelihood
(or unnormalized posterior) ∇θj log p(θtj ) is computed by the autot
matic differentiation
tool
in
PyTorch.
We first compute the joint likeQN
j
j
lihood log p(θt ) = i=1 p(yi | θt , xi )p(θtj ) by feeding forward the data
j
{xi , yi }N
i=1 , then back propagate to compute its gradient ∇θtj log p(θt ).
Note that this gradient is stored in PyTorch module associated to the
weights θtj in each network.
h
i
• Then we can proceed to compute the kernel matrix k(θtj , θti )
,
i,j∈{1,···,S}
16
and its gradient ∇θj k(θtj , θti ), as well as the kernel weighted gradient
t
of the joint likelihood k(θtj , θti )∇θj log p(θtj ). For this to happen, we
t
need to vectorize (extract out) the parameters θtj and the computed
gradient ∇θj log p(θtj ) from each neural network. The optimal perturt
bation direction φ is further computed by the sum of the two terms
as shown in the SVGD algorithm.
• After {φ(θtj )}Sj=1 is computed, we send φ(θtj ) back to each neural network the gradient w.r.t. its parameters θtj (overwriting the previously
computed ∇θj log p(θtj )), and updating locally in each neural network
t
using PyTorch’s optimization library such as Adam or SGD to comj
pute θt+1
.
• With one iteration complete, the algorithm repeats the above steps
until convergence in the parameters is achieved.
3.4. Uncertainty Quantification
Of interest to the classical UQ problem is the computation of the posterior predictive distribution (predict the system response for a test input)
as well as the computation of the output response averaged over the input
probability distribution. In particular, we are interested in computing the
following:
• Predictive uncertainty at x∗ : p(y∗ | x∗ , D), and in particular the moments E[y∗ | x∗ , D], Var(y∗ | x∗ , D).
• Propagated uncertainty to the system response by integrating over
p(x): p(y | θ), θ ∼ p(θ | D), and in particular E[y | θ], Var(y | θ). One
can use these moments
to compute
the
output
h
i
statistics
of conditional
statistics, e.g. Eθ E[y | θ] , Varθ E[y | θ] and Eθ Var(y | θ) ,
Varθ Var(y | θ) .
We can use Monte Carlo to approximate the moments of the predictive
distribution
Z
∗
∗
p(y | x , D) = p(y∗ | x∗ , w, β)p(w, β | D)dwdβ,
(11)
17
with mean (by the law of total expectation)
h
i
E[y∗ | x∗ , D] = Ep(w,β|D) E[y∗ | x∗ , w, β]
= Ep(w|D) [f (x∗ , w)]
≈
S
1X
f (x∗ , wi ),
S
wi ∼ p(w | D).
(12)
i=1
Note that the SVGD algorithm provides a sample representation of the
joint posterior of all parameters p(w, β | D). To obtain the samples of the
marginal posterior p(w | D) as needed above, one simply needs to use the
samples corresponding to w.
The predictive covariance can also be easily calculated using the law
of total variance. The variance and expectation below are w.r.t. to the
posterior of the parameters. We can show the following:
h
i
Cov(y∗ | x∗ , D) = Ew,β Cov(y∗ | w, β, x∗ ) + Covw,β E[y∗ | w, β, x∗ ]
= Ew,β [β −1 I] + Covw,β (f (x∗ , w))
∗
= Eβ [β −1 I] + Ew [f (x∗ , w)f > (x∗ , w)] − Ew [f (x∗ , w)] E>
w [f (x , w)]
S
1 X i −1
≈
(β ) I + f (x∗ , wi )f > (x∗ , wi )
S
i=1
−
S
1 X
S
f (x∗ , wi )
S
1 X
i=1
S
f (x∗ , wi )
>
,
i=1
(13)
where β i ∼ p(β | D), wi ∼ p(w | D). The predictive variance is the diagonal
of the predictive covariance:
Var(y∗ | x∗ , D) = diag Cov(y∗ | x∗ , D)
S
S
1 X
2
1 X i −1
2 ∗
i
(β ) 1 + f (x , w ) −
f (x∗ , wi ) (14)
,
=
S
S
i=1
i=1
where 1 is a vector of ones with the same dimension as f , and the square
(·)2 is applied element-wise to the vectors.
The above computation is the prediction at a specific input x∗ . We
would also like to compute the average prediction over the distribution of
the uncertain input. We first compute the output statistics given the realizations of the uncertain parameters θ = {w, β}, where θ ∼ p(θ | D). The
18
conditional predictive mean is
E[y | θ] = Ex E[y | x, θ] = Ex [f (x, w)] ≈
M
1 X
f (xj , w),
M
xj ∼ p(x),
j=1
(15)
and the conditional predictive covariance is
Cov(y | θ) = Ex [Cov(y | x, θ)] + Covx (E[y | x, θ])
= Ex [(β)−1 I] + Covx (f (x, w))
≈β
−1
M
M
M
1 X
1 X
>
1 X
j
> j
j
I+
f (x , w)f (x , w) −
f (x , w)
f (xj , w) .
M
M
M
j=1
j=1
j=1
(16)
Also the conditional predictive variance (at each spatial location) is
Var(y | θ) = diag Cov(y | θ) = β
−1
M
M
1 X
2
1 X 2 j
1+
f (x , w)−
f (xj , w) ,
M
M
j=1
j=1
(17)
where 1 is a vector of ones with the same dimension as y, and the square
operator is here applied element-wise to the vectors. Then we can further
compute the statistics of the above conditional
statistics
due
h
i
to the uncer
tainty in the surrogate, i.e. θ, such as Eθ E[y | θ] , Varθ E[y | θ] and
Eθ Var(y | θ) , Varθ Var(y | θ) , which are the sample means and sample variances in each output dimension of of the conditional predictive mean
and variance.
4. Numerical Implementation and Results
We study the two-dimensional, single phase, steady-state flow through
a random permeability field following the case study in Section 3.2 in [2].
Consider the random permeability field K on a unit square spatial domain
S = [0, 1]2 , the pressure field p and velocity field u of the fluid through the
porous media are governed by Darcy’s law:
u(s) = −K(s)∇p(s),
∇ · u(s) = f (s),
u(s) · n̂(s) = 0,
Z
p(s)ds = 0,
s ∈ S,
s ∈ ∂S,
S
19
s ∈ S,
(18)
where n̂ denotes the unit normal vector to the boundary and the source
term f is used to model an injection well on the left-bottom corner of S and
a production well on the right-top corner. We also enforce no-flux boundary
condition, and an integral constraint to ensure the uniqueness of the solution
as in [2]. More specifically,
if |si − 12 w|≤ 21 w, for i = 1, 2,
r,
f (s) = −r,
(19)
if |si − 1 + 12 w|≤ 21 w, for i = 1, 2,
0,
otherwise,
where r is the rate of the wells and w is their size. The input log-permeability
field is restricted in this work to be a Gaussian random field, i.e.
K(s) = exp(G(s)),
G(·) ∼ N (m, k(·, ·)),
(20)
where m is the constant mean and covariance function k is specified in the
following form using the L2 norm in the exponent instead of the L1 norm
in [2], i.e.
(21)
k(s, s0 ) = exp(− s − s0 2 /l).
4.1. Datasets
The Gaussian random field [60] with exponential kernel for the onedimensional case corresponds to the Ornstein-Uhlenbeck process which is
mean-square continuous but not mean-square differentiable. Thus when we
discretize the field over a grid, the field value jumps (varies highly) as we
move from pixel to pixel. The field does not become smoother when we
use a finer grid over a fixed spatial domain. This high variability creates a
significant challenge for data-driven models to capture, i.e. the intrinsic dimensionality of the discretized random field is the total number of pixels, e.g.
4, 225 for 65×65 grids (which will be our reference grid for our calculations).
However, a common assumption for natural images is that the underlying dimensionality is actually small (few hundreds) despite their complex appearance. To evaluate the generality and effectiveness of the methodology, we
use KLE to control the intrinsic dimensionality of the permeability dataset.
We evaluated our model using datasets produced with increasing dimensionality of 50, 500, 4225 (called KLE50, KLE500, and KLE4225, respectively).
Notice that when the number of KLE terms is 4225, the permeability field
is directly sampled from the exponential Gaussian field without any dimensionality reduction. The intrinsic dimensionality of dataset is hidden from
our model, i.e. our model do not built a map from the KLE terms to the
system output. Instead, it models an end-to-end mapping from input fields
20
to output fields. In fact, we will show one specific network architecture that
works well for all three datasets obtained from the different intrinsic input
data dimensions.
We consider solving the Darcy flow Eq. (18) over a unit squared domain
S = [0, 1]2 with fixed 65 × 65 grid, and length scale l = 0.1, kernel mean
m = 0, rate of source r = 10, size of source w = 0.125. The ratio of the
cumulative sum of eigenvalues (in decreasing order) over the total sum of
them is shown in Fig. 5. The Darcy flow equation is solved using mixed finite
Figure 5: KLE profile.
element formulation implemented in FEniCS [61] with third-order RaviartThomas elements for the velocity, and fourth-order discontinuous elements
for the pressure. The sample input permeability field and computed output
pressure and velocity fields for three datasets are shown in Fig. 6.
When the available data is limited, it is common practice to use crossvalidation to evaluate the model. Since our dataset is synthetically generated, we have access to any number of training and test data up to computing constraints to solve the Darcy flow equations. Our current data
includes four sets: the training set, validation set, test set, and uncertainty
propagation set. The training set is sampled using the simplest design of
experiment method, Latin hypercube sampling. More specifically, the KLE
for the log-permeability field is
q
X
p
G(s) = m +
λk zk φk (s),
(22)
k=1
where λk and φk (s) are the eigenvalues and eigenfunctions of the exponential
covariance function of the Gaussian field specified in Eqs. (20) and (21),
zk ’s are i.i.d. standard Normal, and q is the number of KLE coefficients
maintained in the expansion. The maximum number that can be used is
21
(a) KLE50
(b) KLE500
(c) KLE4225
Figure 6: Sample permeability K obtained from the (a) KLE50 dataset, (b) KLE500
dataset, and (c) KLE4225 dataset (no dimensionality reduction) and the corresponding
velocity components ux , uy and pressure p obtained from the simulator. All figures are
shown using pixels (imshow) to reveal the high variability of the input and output fields.
finite and equal to the number of grid points used in the discetization of the
field over the unit square. We first use Latin hypercube design to sample
ξk from the hypercube [0, 1]q , then obtain the eigenvalue by zk = Φ−1 (ξk ),
where Φ is the cumulative distribution function of the standard normal
distribution. The KLE50 case contains 32, 64, 128, and 256 training data;
KLE500 contains 64, 128, 256, and 512 training data; and KLE4225 contains
128, 256, 512, and 1024 training data.
The log-permeability fields in the other three sets are reconstructed directly with zk , which are sampled from standard normal. The validation
and test set each contains 500 input permeability fields, and the dataset
for uncertainty propagation contains 10, 000 realizations. All datasets are
organized as images as discussed in Section 2.1.
22
4.2. Evaluation Metrics
Several metrics are used to evaluate the trained models on test data
{xi , yi }Ti=1 . In particular, we consider the following:
Coefficient of determination (R2 -score):
PT
i
i 2
i=1 y − ŷ 2
2
R = 1 − PT
,
(23)
2
i
i=1 ky − ȳk2
P
where ŷi is the output mean of the Bayesian surrogate, i.e. Si=1 f (x, wi )/S
as in Eq. (12) or predictive output of the non-Bayesian surrogate, i.e. just
f (x), yi is the test target, ȳ is the mean of test target, and T is the total
number of test data. This metric enables the comparison between different
datasets since the error is normalized, with the score closer to 1 corresponding to better regression. This is the only metric used for evaluating nonBayesian surrogate, the following metrics are additional metrics for evaluating the Bayesian surrogate. Note that this metric is also used for tracking
the performance of the training process, thus it is evaluated for both the
training and test data sets.
Root Mean Squared Error (RMSE):
v
u
T
u1 X
t
kŷi − yi k22 .
T
i=1
This is a common metric for regression that is used in our experiments for
monitoring the convergence of training.
Mean Negative Log-Probability (MNLP):
T
1X
MNLP = −
log p(yi | xi , D).
T
i=1
This metric evaluates the likelihood of the observed data. It is is used to
assess the quality of the predictive model.
Predictive uncertainty and Propagated Uncertainty: These metrics were introduced in Section 3.4.
Estimated Distributions: They include histograms or kernel density estimates for the output fields at certain locations of the physical domain.
Reliability Diagram: Given a trained Bayesian surrogate and a test data set,
we can compute the p% predictive interval for each test data point based on
the Gaussian quantiles using the predictive mean and variance [62]. We then
compute the frequency of the test targets that fall within this predictive
interval. For a well-calibrated regression model, the observed frequency
23
should be close to p%. The reliability diagram is the plot of the observed
frequency with respect to p. Thus a well-calibrated model should have a
reliability diagram close to the diagonal.
4.3. Non-Bayesian Surrogate Model
The hyperparameters to search include the parameters that determine
the network architecture and the ones that specify training process, which
both affect model performance. We use Hyperband [63] algorithm to optimize those hyperparameters with a constraint that the number of model
parameters being less than 0.25 million. The details of these experiments
are given in Appendix A. The network configuration with the highest R2 score that Hyperband finds is shown in Fig. 7 with more details provided
in Table 1. This configuration is referred to as DenseED-c16. The 2nd–4th
columns of Table 1 show the number Cf of output feature maps, the spatial
resolution Hf × Wf of output feature maps, the number of parameters of
each layer in the network.
Figure 7: DenseED-c16 with blocks (3, 6, 3), growth rate 16, and 48 initial feature maps
after the first convolution layer (yellow in the figure). There are in total 19 (conv) layers
and 241, 164 parameters in the network. The number of network parameters is optimized
based on the generalization error analysis reported in Appendix B. It contains two downsampling layers, thus the smallest spatial dimension (code dimension) of feature maps
(purple in the figure) is 16 × 16, hence the network is named DenseED-c16. The first
convolution kernel is of k7s2p2. The last transposed conv kernel in the decoding layer is
of k5s2p1. The number of its output feature maps is 3 (corresponding to 3 output fields).
For the decoding layers, the output padding is set to 1. The other conv kernels in dense
blocks and encoding, decoding layers are described in Section 2.2.
The network DenseED-c16 is trained with Adam [64], a variant of stochastic gradient descent, with the loss function being L2 regularized MSE which
is implemented as weight decay in modern neural net frameworks, such as
PyTorch and TensorFlow. Other loss functions may achieve better results,
24
Table 1: DenseED-c16 architecture for the Darcy flow dataset
Layers
Input
Convolution k7s2p2
Dense Block (1) K16L3
Encoding Layer
Dense Block (2) K16L6
Decoding Layer (1)
Dense Block (3) K16L3
Decoding Layer (2)
Resolution Hf × Wf
65 × 65
32 × 32
32 × 32
16 × 16
16 × 16
32 × 32
32 × 32
65 × 65
Cf
1
48
96
48
144
72
120
3
Number of parameters
2352
28032
25632
77088
57456
38544
12060
such as smoothed L1 loss, or conditional GAN loss [41]. This requires further
investigations to be considered in future publication. The initial learning
rate is 0.015, weight decay (regularization on weights) is 0.0005, the batch
size is 16. We also use a learning rate scheduler which drops 10 times on
plateau of the rooted MSE. The model is trained 200 epochs. We train the
model with the dataset introduced in Section 4.1.
Training the deterministic neural networks with L2 regularized MSE is
equivalent to finding the maximum a posterior of the uncertain parameters
in Bayesian neural networks whose prior is independent normal. The typical
training process is shown in Fig. 8.
KLE50 train
KLE50 test
KLE500 train
KLE500 test
KLE4225 train
KLE4225 test
0.12
0.10
RMSE
0.08
0.06
0.04
0.02
0
25
50
75
100 125
Epochs
150
175
200
Figure 8: Training process of DenseED-c16 with 128 training data.
We train each network with different number of training data of KLE50,
KLE500, and KLE4225. The validation R2 -score is shown in Fig. 9, which
shows that, with the same training data, the R2 -score is closer to 1 when
the intrinsic dimensionality is smaller, and the R2 -score is higher with more
25
training data of the same dimensionality. Note that the score is more than
0.9 with reasonably small size training data set for all the three cases which
have dimensionality from 50 to 4225. This shows the effectiveness of the
network DenseED-c16 for both low-dimensional and high-dimensional problems.
0.975
0.950
Test R 2-score
0.925
0.900
0.875
0.850
0.825
KLE50
KLE500
KLE4225
0.800
0.775
25
26
28
27
Number of training data
29
210
Figure 9: Test R2 scores for the non-Bayesian surrogate.
The prediction of the output fields can be easily obtained in the test time
by feeding the test input permeability field x∗ into the trained network, i.e.
ŷ∗ = f (x∗ ). We show the prediction of the test input shown in Fig. 6
using DenseED-c16, which is trained with three datasets (KLE50, KLE500,
KLE4225) in Figs. 10, 11, and 12, respectively. The predictions are quite
good even for the KLE4225 case, where both the input and output fields
vary rapidly in certain regions of the domain.
4.4. Bayesian Surrogate Model
For all the experiments we only consider the homoscedastic noise model
for Bayesian neural networks, i.e. output-wise Gaussian noise with Gamma
prior on its precision β, and Student’s t-prior on w.
The set of all uncertain parameters is denoted as θ = {w, β}. We apply
SVGD to the Bayesian neural network with S samples {θ i }Si=1 from the
posterior p(θ | D), i.e. S set of deterministic model parameters {wi }Si=1 of
DenseED’s and noise precision {β i }Si=1 . In implementation, this corresponds
to S different initializations for the deterministic DenseED and noise precision
(a scalar). We update the parameters of S DenseED’s and the corresponding
noise precision using the SVGD algorithm as in Algorithm 1. The kernel is
chosen to be k(x, x0 ) = exp(− kx − x0 k22 /h), with median heuristic for the
choice of the kernel bandwidth h = H 2 /log S, where H is the median of the
pairwise distances between the current samples {θ i }Si=1 . We typically use
26
(a) 32 training data
(b) 128 training data
Figure 10: Prediction for the input realization shown in Fig. 6a from the KLE50 dataset
using DenseED-c16 which is trained with datasets of sizes (a) 32 and (b) 128, respectively.
In both subfigures, the first row shows the three test target fields (simulation output), i.e.
pressure p and velocity uy , ux , the second row shows the corresponding model predictions,
the third row shows the error.
(a) 64 training data
(b) 256 training data
Figure 11: Prediction for the input realization as shown in Fig. 6b from KLE500 dataset
using DenseED-c16 which is trained with datasets of sizes (a) 64 and (b) 256, respectively.
In both subfigures, the first row shows the three test target fields (simulation output), i.e.
pressure p and velocity uy , ux , the second row shows the corresponding model predictions,
the third row shows the error.
27
(a) 128 training data
(b) 512 training data
Figure 12: Prediction for the input realization as shown in Fig. 6c from KLE4225 dataset
using DenseED-c16 which is trained with datasets of sizes (a) 128 and (b) 512, respectively.
In both subfigures, the first row shows the three test target fields (simulation output), i.e.
pressure p and velocity uy , ux , the second row shows the corresponding model predictions,
the third row shows the error.
S = 20 samples of θ to approximate the empirical measure of the posterior.
For large number of training data, the unnormalized
is evaluated
Q posterior
i | θ, xi )p (θ) ≈
p(y
using mini-batches of training data, i.e. p̃(θ | D) = N
0
i=1
Q
i | θ, xi )p (θ). We observe that even for small training data
p(y
N/B B
0
i=1
such as 512, using smaller batch size (e.g. 16) helps to get lower training and
test errors, but with more time for training. We use Adam [64] to update θ
using the gradient φ, instead of the vanilla stochastic gradient descent, for
300 epochs, with learning rate 0.002 for w and 0.01 for β, and a learning
rate scheduler that decreases by 10 times when the training RMSE is on
plateau.
The algorithm is implemented in PyTorch and runs on a single NVIDIA
GeForce GTX 1080 Ti X GPU which requires about 2000 − 7000 seconds
for training 300 epochs, when the training data size varies from 32 to 512.
The training time depends heavily on the training mini-batch size, which
is 16 for all cases. Potential ways to speed up significantly the training
process include increasing the mini-batch size, or implementing the SVGD
in parallel using multi-GPUs. The python source code will become available
upon publication at https://github.com/bmmi/bayesnn.
We report next the R2 -score computed similar to the non-Bayesian case,
except the predicted output mean is used to compare with the test target.
The scores are shown in Fig. 13. We can see that the Bayesian surrogate
improves the R2 -score significantly over the non-Bayesian version.
28
Test R 2-score
0.95
0.90
0.85
KLE50 - Bayesian
KLE50 - Non-Bayesian
KLE500 - Bayesian
KLE500 - Non-Bayesian
KLE4225 - Bayesian
KLE4225 - Non-Bayesian
0.80
26
25
28
27
Number of training data
29
210
Figure 13: R2 –scores comparison for the non-Bayesian and Bayesian models.
We also report the MNLP for test data in Fig. 14, which is a proper scoring rule and usually used to access the quality of predictive uncertainty [65].
1e6
KLE50
KLE500
KLE4225
0.8
Test MNLP
0.6
0.4
0.2
0.0
25
26
28
27
Number of training data
29
210
Figure 14: MNLP.
The R2 -score gives us a general estimate of how well the regression performs. For a given input permeability field, the Bayesian neural network can
predict the mean of corresponding output fields, and also gives uncertainty
estimate represented as predictive variance at each spatial location, which is
unavailable for deterministic models, and desirable when the training data
is small. In Figs. 15, 16, and 17, we show predictions for the test input
shown in Fig. 6 with training data from KLE50, KLE500, and KLE4225,
respectively. We can see that the predictive accuracy improves as the size
of the training dataset increases, while the predictive uncertainty drops.
29
(a) 64 training data
(b) 256 training data
Figure 15: Prediction for the input realization shown in Fig. 6a from the KLE50 dataset.
The rows from top to bottom show the simulation output fields (ground truth), predictive
mean E[y∗ | x∗ , D], the error of the above two, and two standard deviation of predictive output distribution per pixel Var(y∗ | 30
x∗ , D). The three columns from left to right
correspond to pressure field p, and two velocity fields uy , ux , respectively.
(a) 64 training data
(b) 256 training data
Figure 16: Prediction for the input realization shown in Fig. 6b from the KLE500 dataset.
The rows from top to bottom show the simulation output fields (ground truth), predictive
mean E[y∗ | x∗ , D], the error of the above two, and two standard deviation of predictive output distribution per pixel Var(y∗ | 31
x∗ , D). The three columns from left to right
correspond to pressure field p, and two velocity fields uy , ux , respectively.
(a) 128 training data
(b) 512 training data
Figure 17: Prediction for the input realization shown in Fig. 6c from the KLE4225 dataset.
The rows from top to bottom show the simulation output fields (ground truth), predictive
mean E[y∗ | x∗ , D], the error of the above two, and two standard deviation of predictive output distribution per pixel Var(y∗ | 32
x∗ , D). The three columns from left to right
correspond to pressure field p, and two velocity fields uy , ux , respectively.
We also performed uncertainty propagation by feeding the trained Bayesian
surrogate with 10, 000 input realizations sampled from the Gaussian field,
and calculating the output statistics as in Section 3.4. In Figs. 18, 19
and 20, we show the uncertainty propagation results and compare with
Monte Carlo using the 10, 000 UP data for the Bayesian surrogate trained
with the datasets KLE50, KLE500, and KLE4225.
We show the estimate of pressure p, velocity components ux , uy at locations (0.85, 0.88), and (0.68, 0.05) on the unit square for 128, and 512 training
data of KLE4225 in Fig. 21, and 22, respectively. The PDF is obtained by
kernel density estimation using the predictive mean. We can see that the
density estimate is close to the Monte Carlo result even when the training
dataset is small, and becomes closer as the training dataset increases. From
Fig. 22, we observe that the predictions for the velocity fields are better
than the pressure field especially in locations away from the diagonal of the
unit square domain, and this is in general the case for our current network
architecture, where the three output fields are treated the same.
In order to access the quality of the computed uncertainty, we adopt the
reliability diagram which expresses the discrepancy between the predictive
probability and the frequency of falling in the predictive interval for the
test data. The diagram is shown in Fig. 23. Overall our models are wellcalibrated since they are quite close to the ideal diagonal, especially the
case when the training dataset size is 128 as shown in Fig. 23d. In general
the model turns to be over-confident (small predictive uncertainty) when
the training data is small, and gradually becomes prudent (larger predictive
uncertainty) when the training data increases. The main reason for this
observation is that the predictive uncertainty is dominated by the variation
seen in the training data, which is small when small data is observed. The
initial learning rates and their scheduling scheme for network parameters
w and noise precision β may also play roles here since the uncertainty is
determined by the optimum that that stochastic optimization obtained.
The training processes with different datasets are shown in Fig. 24. We
can see that the training and test RMSE converges around 50 ∼ 75 epochs of
training for KLE50, and KLE500, but the convergence for KLE4225 seems
to take longer time. The training dataset size is 256 for all three sets.
To empirically show that using 20 samples of the Bayesian neural nets
for the SVGD algorithm is sufficient in our problem, we show in Fig. 25 the
convergence of the test and training RMSE when we vary the number of
samples.
33
(a) 32 training data
(b) 128 training data
(c) 32 training data
(d) 128 training data
Figure 18: Uncertainty propagation for KLE50: In (a) and (b) from top to bottom, we
show the Monte Carlo output mean, predictive output mean Eθ [E[y | θ]], the error of the
above two, and two standard deviation of conditional predictive mean Varθ (E[y | θ]). The
three columns from left to right correspond to pressure field p, and two velocity fields uy ,
ux , respectively. In (c) and (d) from top to bottom, we show the Monte Carlo output
variance, predictive output variance Eθ [Var(y | θ)], the error of the above two, and two
standard deviation of conditional predictive variance Varθ (Var(y | θ)). The three columns
are the same as (a) and (b).
34
(a) 64 training data
(b) 256 training data
(c) 64 training data
(d) 256 training data
Figure 19: Uncertainty propagation for KLE500: (a), (b), (c), and (d) refer to Fig. 18.
35
(a) 128 training data
(b) 512 training data
(c) 128 training data
(d) 512 training data
Figure 20: Uncertainty propagation for KLE4225: (a), (b), (c), and (d) refer to Fig. 18.
36
2.00
Monte Carlo
surrogate
8
Monte Carlo
surrogate
1.75
2.0
1.50
6
1.25
1.0
0.75
0.50
2
0
1.5
1.00
4
0.5
0.25
0.6
0.4
0.2
0.0
0.00
Monte Carlo
surrogate
2.5
0.0
0.5
1.0
0.0
1.5
0.0
0.5
1.0
1.5
(a) Location (0.85, 0.88) 128 training data
2.00
Monte Carlo
surrogate
8
Monte Carlo
surrogate
1.75
1.50
6
1.5
1.25
1.00
4
1.0
0.75
0.50
2
0.5
0.25
0
Monte Carlo
surrogate
2.0
0.6
0.4
0.2
0.0
0.00
0.0
0.5
1.0
0.0
1.5
0.0
0.5
1.0
1.5
(b) Location (0.85, 0.88) 512 training data
Figure 21: Distribution estimate for the pressure, and the two velocity components
(from left to right) at location (0.85, 0.88) with Bayesian surrogate trained with KLE4225
dataset.
Monte Carlo
surrogate
40
Monte Carlo
surrogate
8
30
6
15
20
4
10
10
2
5
0
0
0.05
0.00
0.05
0.10
0.0
0.2
0.4
0.6
0.8
Monte Carlo
surrogate
20
0
0.2
0.0
0.2
0.4
(a) Location (0.68, 0.05) 128 training data
Monte Carlo
surrogate
35
30
25
Monte Carlo
surrogate
8
Monte Carlo
surrogate
20
6
15
4
10
2
5
20
15
10
5
0
0.05
0.00
0.05
0.10
0
0.0
0.2
0.4
0.6
0.8
0
0.2
0.0
0.2
0.4
(b) Location (0.68, 0.05) 512 training data
Figure 22: Distribution estimate for the pressure, and the two velocity components
(from left to right) at location (0.68, 0.05) with Bayesian surrogate trained with KLE4225
dataset.
37
1.0
0.8
num_train = 64
num_train = 128
num_train = 256
num_train = 512
Ideal
0.8
0.6
Frequency
Frequency
1.0
num_train = 32
num_train = 64
num_train = 128
num_train = 256
Ideal
0.4
0.2
0.6
0.4
0.2
0.0
0.0
0.0
0.2
0.4
0.6
Probability
0.8
1.0
0.0
(a) KLE50
1.0
1.0
0.8
1.0
0.8
1.0
KLE 50
KLE 500
KLE 4225
Ideal
0.8
Frequency
Frequency
0.4
0.6
Probability
(b) KLE500
num_train = 128
num_train = 256
num_train = 512
num_train = 1024
Ideal
0.8
0.2
0.6
0.4
0.6
0.4
0.2
0.2
0.0
0.0
0.0
0.2
0.4
0.6
Probability
0.8
0.0
1.0
0.2
0.4
0.6
Probability
(d) All cases with 128 training data
(c) KLE4225
Figure 23: Reliability diagrams. Subfigures (a), (b), (c) show the reliability diagram
for KLE50, KLE500, KLE4225, respectively. (d) shows the diagrams when the training
dataset size is 128 for all three datasets.
KLE50 train
KLE50 test
KLE500 train
KLE500 test
KLE4225 train
KLE4225 test
0.14
0.12
RMSE
0.10
0.08
0.06
0.04
0.02
0.00
0
50
100
150
200
Number of training data
250
300
Figure 24: Test and training RMSE for training of SVGD using 128 training data from
KLE50, KLE500, and KLE4225.
38
train
test
0.020
0.018
RMSE
0.016
0.014
0.012
0.010
0.008
0
10
20
30
40
50
Number of particles
60
70
80
Figure 25: Test and training error using SVGD to train the Bayesian NN with 128 training
data of KLE500. As we increase the number of posterior samples for θ, i.e. the number of deterministic neural networks (and noise precision), both training and test errors
drop and become steady. This plot empirically supports the reason we choose 20 model
instances/samples for SVGD.
5. Conclusions
In this work, we explore to use an alternative Bayesian surrogate, i.e. a
Bayesian neural network, to predict the output fields of a systems governed
by stochastic partial differential equations with high-dimensional stochastic input. The approach is built on the highly expressive deep convolutional
encoder-decoder networks. The end-to-end image-to-image regression avoids
the usual linear dimensionality reduction step, and achieves promising results in terms of predictive performance and uncertainty modeling, even
with limited training data.
We show in experiments that one network (DenseED-c16) works well
across problems with different intrinsic dimensionality. We believe the performance at the small dataset domain of deep neural networks is due to
its unique generalization property [66, 67] which effectively says the overparameterized deep neural nets lack over-fitting.
The Bayesian model provides uncertainty estimates for our predictions
thus accounting for epistemic uncertainty when training with small datasets.
We show that Bayesian inference based on SVGD works well for training the
Bayesian neural network, and presented results on uncertainty propagation
tasks. The uncertainty of the trained model is well-calibrated by investigating the reliability diagrams.
39
There are three potential directions to improve the regression performance of non-Bayesian surrogate:
• Add skip connections between the encoder and decoder path in each
level of feature map size. The reason is that the input permiability
field directly affects the output velocity field pixel-wise through the
equation u(s) = −K(s)∇p(s). Concatenate the features extracted in
the encoder path to the decoder path may help to recover the discontinuity of the output velocity fields which are caused by the discontinuity
in the input.
• Conditional GANs loss. The idea is that we should use a stronger loss
function such as an adversarial classifier to enforce the model output to
obey the discontinuity in the data. Similar situation has been explored
in jet maps in high-energy particle physics [68].
• Using group convolution in the last decoding layer to separate the
influences between the pressure field and the velocity fields in the end.
This is to address the observation that the prediction performance for
the velocity fields is better than for the pressure field using our current
network architecture DenseED-c16.
For the Bayesian neural network, the inference task is genuinely difficult
since it is asked to find the posterior of millions of random variables based on
hundreds of training data. Exploring other priors for the network weights,
and using recent advances in variational inference for Bayesian neural networks are definitely valuable directions to pursue.
The image-to-image regression approach can be used to handle the prediction of systems with different source terms, by adding the source field as
another channel in the input besides the material property field. For implementation, one only needs to increase the number of input channels by 1,
and leave everything else unchanged.
Our current network does not utilize any information from physics, such
as the governing equations or constraints between the three output fields.
Incorporating physics information into deep neural networks is a more principled ways to improve the model performance.
Acknowledgments
This work was supported from the University of Notre Dame, the Center for Research Computing (CRC) and the Center for Informatics and
40
Computational Science (CICS). Early developments were supported by the
Computer Science and Mathematics Division of ORNL under the DARPA
EQUiPS program. N.Z. thanks the Technische Universität München, Institute for Advanced Study for support through a Hans Fisher Senior Fellowship, funded by the German Excellence Initiative and the European Union
Seventh Framework Programme under Grant Agreement No. 291763. The
authors acknowledge Dr. Steven Atkinson from the CICS for generating and
providing us the Darcy flow datasets.
Appendix A. Experiments on Darcy Flow Problem
We optimize the network hyperparameters following the general architecture
as shown in Fig. 3. The encoding layer in Fig. 2(a) down-samples its input
feature maps by 2 using convolution of stride 2 as in the first convolution
layer. The decoding layer in Fig. 2(b) up-samples the feature maps by 2
using transposed convolution instead of pooling layers (as commonly seen
in image classification networks) since the location information is critical
for regression. The number of down-sampling and up-sampling layers are
the same. The datasets are described in Section 4.1. We aim to find one
general network architecture that works well across different datasets from
both Bayesian and non-Bayesian network models.
Experiment 1 - Spatial dimension of feature maps at the coarsest scale: This
is determined by how many down-sampling layers are used. Shrinking the
spatial dimension of feature maps can extract high-level or coarse information of the input permeability field, which is subsequently used to predict the
output pressure and velocity fields. This design choice is related to a central
concept in CNNs called receptive field [69] of an unit within a certain layer,
which is the region in the input image that affects this unit feedforward, or
the region in the output image that affects this unit when back-propagating
gradients. For dense prediction tasks such as our surrogate modeling problems, it is important for each pixel in the code feature maps to have the
suitable receptive field in the input and output images. We observe that
for the dataset generated with fewer KLE terms, both the input and output
velocity fields are smoother, or the output has stronger correlation across
pixels, such as KLE50 in Fig. 6a. But for KLE500 or KLE4225 in Figs. 6b
and 6c, the fields vary more rapidly from pixel to pixel, but still there is
weak long-range correlation between pixel values.
41
In our experimental setup we have used no skip connections, the convolution
kernels were fixed to be k7s2p2 for the first Conv layer, k3s1p1 in the layers
within dense block, k1s1p0 and k3s2p1 in the encoding and decoding layers,
and k5s2p2 in the last ConvtT layer. No dropout was used and the growth
rate of dense blocks was taken as 16.
The loss function is taken as the regularized MSE in Eq. (3), and the validation metric as the R2 -score in Eq. (23). Adam optimizer is used for training
200 epochs, with learning rate 0.001 and a plateau scheduler on the test
RMSE. Batch size is always smaller than the number of training data (e.g.
16, 32, 64 for dataset with 32, 64, 128 data, respectively). Weight decay is
set to 0.0005.
We vary the dense block configurations, i.e. a list of integers specifying
the number of layers within each dense block. This list contains odd number
of integers because of the symmetry between encoding and decoding paths.
For example, blocks (3, 6, 3) specify the network architecture in Fig. A.26a.
K16L6 in the second dense block means there are 6 layers within the dense
blocks, and the growth rate for each layer is 16. In this case, the code
dimension (purple feature maps) is 16×16 since there are two down-sampling
layers in the encoding path.
The second block configuration is (12, ), i.e. only one dense block sits
after the first Conv layer, as shown in Fig. A.26b or (2, 2, 4, 2, 2) as shown
in Fig. A.27.
The validation R2 -scores of the above three networks DenseED-c16, DenseED-c8,
DenseED-c32 on test set (500 Monte Carlo samples from each KLE case)
are summarized in Tables A.2 , A.3, A.4. From these tables, it can be seen
that the DenseED − c16 network works well for all three datasets. The
network DenseED-c8 does not work well for the KLE4225 dataset, since its
code dimension is too small. This enforces the output field to be too smooth,
which is still favorable to the KLE50 dataset. We can also clearly note that
the network DenseED-c32 does not work well for KLE50, even though it
performs well for the KLE4225 dataset.
Predictions and learning curve for the selected network design DenseED-c16
are presented in Section 4.3.
Experiment 2 - Hyperparameter optimization with Hyperband: The hyperparameters to select include the ones specifying the network architecture
and the training process. We separately optimize those two sets of hyperparameters with one of them fixed. Hyperband [63] is a bandit random search
algorithm which has several rounds of successive halving. There are two
input for this algorithm, i.e. R = 243, the maximum iterations for each hy42
(a) DenseED-c16 with blocks (3, 6, 3)
(b) DenseED-c32 with blocks
(12, )
Figure A.26: Two configurations DenseED-c16 and DenseED-c32 of DenseED.
Figure A.27: Configuration DenseED-c8 with blocks (2, 2, 4, 2, 2).
Table A.2: Test R2 -score for DenseED-c16.
32
64
128
256
512
KLE50
0.718
0.883
0.947
0.970
-
KLE500
0.551
0.817
0.913
0.954
0.976
43
KLE4225
0.280
0.662
0.829
0.927
0.963
Table A.3: Test R2 -score for DenseED-c8.
32
64
128
256
512
KLE50
0.661
0.883
0.931
0.963
-
KLE500
0.344
0.687
0.842
0.934
0.971
KLE4225
0.052
0.411
0.607
0.753
0.829
Table A.4: Test R2 -score for DenseED-c32.
32
64
128
256
512
KLE50
0.563
0.811
0.893
0.931
-
KLE500
0.558
0.807
0.899
0.937
0.954
KLE4225
0.413
0.704
0.863
0.909
0.939
perparameter configuration (here each iteration corresponding to one epoch
of training) and η = 3 which specifies to keep 1/3 of the best configurations
after running through all the candidate configurations.
The search space for network architecture with the search space specified
in Table A.5 with the constraint that the number of parameters being less
than 0.25 million, gives the network architecture presented in Table 1.
Table A.5: Network architecture hyperparameters and associated ranges for the
DenseED-c16 network.
Hyperparameters
Encoding dense block layers
Bottom dense block layers
Decoding dense block layers
Growth rate
Features after 1st Conv
Type
Q-uniform
Q-uniform
Q-uniform
Categorical
Categorical
Values
[1, 8]
[3, 8]
[1, 8]
16, 32, 48
32, 48
The search for the training process hyperparameters as in Table A.6 gives
approximately the following hyperparameters: initial learning rate 0.002,
weight decay 0.0005, batch size 16, and Adam optimizer.
Note that the hyperparameters found by Hyperband are only sub-optimal,
but indicate the potential range. The actual hyperparameters used for training are presented in Section 4.3.
44
Table A.6: Training Hyperparameters and associated ranges for the DenseED-c16 network.
Hyperparameters
Initial learning rate
Weight decay
Batch size
Optimizer
Type
Uniform
Uniform
Categorical
Categorical
Values
[10−4 , 5 × 10−3 ]
[5 × 10−5 , 10−2 ]
16, 24
Adam, RMSprop
Appendix B. Generalization behavior
Several numerical experiments were conducted to show the generalization behavior of the network DenseED-c16. The network was trained with
different block configurations while keeping the number of training data to
be 256. The number of parameters ranged from 37, 892 to 805, 204. The
training and test RMSE are shown in Fig. B.28. Clearly, the model is overparameterized but still shows no overfitting behavior, i.e. the training error
does not become smaller and the test (generalization) error does not go
higher as we increase the number of parameters. When the number of parameters is less than 200, 000 (still in the over-parameterized regime), there
is plenty of room to improve in both training loss and test loss. We configure the number of parameters of the baseline network DenseED-c16 as
in Fig. 7 to have 241, 164 parameters, which is favorable according to the
generalization curve in Fig. B.28.
0.08
test
train
0.07
RMSE
0.06
0.05
0.04
0.03
0.02
0.01
0
100000 200000 300000 400000 500000 600000 700000 800000
Number of parameters
Figure B.28: Non-overfitting of over-parameterized neural network DenseED-c16.
45
References
[1] I. Bilionis, N. Zabaras, Bayesian uncertainty propagation using Gaussian processes, Handbook of Uncertainty Quantification (2016) 1–45.
[2] I. Bilionis, N. Zabaras, B. A. Konomi, G. Lin, Multi-output separable
Gaussian process: towards an efficient, fully Bayesian paradigm for uncertainty quantification, Journal of Computational Physics 241 (2013)
212–239.
[3] I. Bilionis, N. Zabaras, Multi-output local Gaussian process regression:
Applications to uncertainty quantification, Journal of Computational
Physics 231 (17) (2012) 5718–5746.
[4] D. Xiu, G. E. Karniadakis, The Wiener–Askey polynomial chaos for
stochastic differential equations, SIAM journal on scientific computing
24 (2) (2002) 619–644.
[5] S. Torquato, Random heterogeneous materials: microstructure and
macroscopic properties, Vol. 16, Springer Science & Business Media,
2013.
[6] L. J. P. van der Maaten, E. O. Postma, H. J. van den Herik, Dimensionality Reduction: A Comparative Review, Tech. rep., Tilburg
University (2009).
URL http://homepage.tudelft.nl/19j49/Matlab_Toolbox_for_
Dimensionality_Reduction_files/TR_Dimensiereductie.pdf
[7] L. v. d. Maaten, G. Hinton, Visualizing data using t-sne, Journal of
Machine Learning Research 9 (Nov) (2008) 2579–2605.
[8] G. E. Hinton, R. R. Salakhutdinov, Reducing the dimensionality of data
with neural networks, science 313 (5786) (2006) 504–507.
[9] D. P. Kingma, M. Welling, Auto-encoding variational bayes, arXiv
preprint arXiv:1312.6114.
[10] N. D. Lawrence, Gaussian process latent variable models for visualisation of high dimensional data, in: Advances in neural information
processing systems, 2004, pp. 329–336.
[11] C. Grigo, P.-S. Koutsourelakis, Bayesian model and dimension reduction for uncertainty propagation: applications in random media, arXiv
preprint arXiv:1711.02475.
46
[12] Y. Bengio, Learning Deep Architectures for AI, Found. Trends Mach.
Learn. 2 (1) (2009) 1–127. doi:10.1561/2200000006.
URL http://dx.doi.org/10.1561/2200000006
[13] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with
deep convolutional neural networks, in: Advances in neural information
processing systems, 2012, pp. 1097–1105.
[14] M. D. Zeiler, R. Fergus, Visualizing and understanding convolutional
networks, in: European conference on computer vision, Springer, 2014,
pp. 818–833.
[15] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553)
(2015) 436–444.
[16] G. E. Hinton, S. Osindero, Y.-W. Teh, A fast learning algorithm for
deep belief nets, Neural computation 18 (7) (2006) 1527–1554.
[17] C. Zhang, S. Bengio, M. Hardt, B. Recht, O. Vinyals, Understanding deep learning requires rethinking generalization, arXiv preprint
arXiv:1611.03530.
[18] G. K. Dziugaite, D. M. Roy, Computing nonvacuous generalization
bounds for deep (stochastic) neural networks with many more parameters than training data, arXiv preprint arXiv:1703.11008.
[19] J. N. Kutz, Deep learning in fluid dynamics, Journal of Fluid Mechanics
814 (2017) 1–4.
[20] S. Chan, A. H. Elsheikh, A machine learning approach for efficient uncertainty quantification using multiscale methods, CoRR
abs/1711.04315. arXiv:1711.04315.
URL http://arxiv.org/abs/1711.04315
[21] J. Marçais, J.-R. de Dreuzy, Prospective interest of deep learning for
hydrological inference, Groundwater 55 (5) (2017) 688–692.
[22] S. Min, B. Lee, S. Yoon, Deep learning in bioinformatics, Briefings in
bioinformatics 18 (5) (2017) 851–869.
[23] P. Baldi, P. Sadowski, D. Whiteson, Searching for exotic particles
in high-energy physics with deep learning, Nature communications 5
(2014) 4308.
47
[24] M. Raghu, B. Poole, J. Kleinberg, S. Ganguli, J. Sohl-Dickstein,
On the expressive power of deep neural networks, arXiv preprint
arXiv:1606.05336.
[25] D. J. MacKay, A practical Bayesian framework for backpropagation
networks, Neural computation 4 (3) (1992) 448–472.
[26] R. M. Neal, Bayesian learning for neural networks, Vol. 118, Springer
Science & Business Media, 2012.
[27] Y. Gal, Z. Ghahramani, Dropout as a Bayesian approximation: Representing model uncertainty in deep learning, in: international conference
on machine learning, 2016, pp. 1050–1059.
[28] C. Blundell, J. Cornebise, K. Kavukcuoglu, D. Wierstra, Weight uncertainty in neural networks, arXiv preprint arXiv:1505.05424.
[29] Q. Liu, D. Wang, Stein variational gradient descent: A general purpose Bayesian inference algorithm, in: Advances In Neural Information
Processing Systems, 2016, pp. 2378–2386.
[30] D. P. Kingma, T. Salimans, M. Welling, Variational dropout and the
local reparameterization trick, in: Advances in Neural Information Processing Systems, 2015, pp. 2575–2583.
[31] J. M. Hernández-Lobato, R. Adams, Probabilistic backpropagation for
scalable learning of bayesian neural networks, in: International Conference on Machine Learning, 2015, pp. 1861–1869.
[32] C. Louizos, M. Welling, Multiplicative normalizing flows for variational
bayesian neural networks, arXiv preprint arXiv:1703.01961.
[33] C. Louizos, K. Ullrich, M. Welling, Bayesian Compression for Deep
Learning, arXiv preprint arXiv:1705.08665.
[34] S. Chan, A. H. Elsheikh, A machine learning approach for efficient uncertainty quantification using multiscale methods, Journal
of Computational Physics 354 (Supplement C) (2018) 493 – 511.
doi:https://doi.org/10.1016/j.jcp.2017.10.034.
URL
http://www.sciencedirect.com/science/article/pii/
S0021999117307933
[35] G. Huang, Z. Liu, K. Q. Weinberger, L. van der Maaten, Densely connected convolutional networks, arXiv preprint arXiv:1608.06993.
48
[36] K. Simonyan, A. Zisserman, Very deep convolutional networks for largescale image recognition, arXiv preprint arXiv:1409.1556.
[37] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking
the inception architecture for computer vision, in: Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, 2016,
pp. 2818–2826.
[38] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image
recognition, in: Proceedings of the IEEE conference on computer vision
and pattern recognition, 2016, pp. 770–778.
[39] O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for
biomedical image segmentation, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer,
2015, pp. 234–241.
[40] D. Eigen, C. Puhrsch, R. Fergus, Depth map prediction from a single
image using a multi-scale deep network, in: Advances in neural information processing systems, 2014, pp. 2366–2374.
[41] P. Isola, J. Zhu, T. Zhou, A. A. Efros, Image-to-Image Translation
with Conditional Adversarial Networks, CoRR abs/1611.07004. arXiv:
1611.07004.
URL http://arxiv.org/abs/1611.07004
[42] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,
S. Ozair, A. Courville, Y. Bengio, Generative adversarial nets, in: Advances in neural information processing systems, 2014, pp. 2672–2680.
[43] A. van den Oord, N. Kalchbrenner, L. Espeholt, O. Vinyals, A. Graves,
et al., Conditional image generation with pixelcnn decoders, in: Advances in Neural Information Processing Systems, 2016, pp. 4790–4798.
[44] A. v. d. Oord, N. Kalchbrenner, K. Kavukcuoglu, Pixel recurrent neural
networks, arXiv preprint arXiv:1601.06759.
[45] R. K. Srivastava, K. Greff, J. Schmidhuber, Training very deep networks, in: Advances in neural information processing systems, 2015,
pp. 2377–2385.
[46] S. Ioffe, C. Szegedy, Batch normalization: Accelerating deep network
training by reducing internal covariate shift, in: International Conference on Machine Learning, 2015, pp. 448–456.
49
[47] X. Glorot, A. Bordes, Y. Bengio, Deep sparse rectifier neural networks,
in: Proceedings of the Fourteenth International Conference on Artificial
Intelligence and Statistics, 2011, pp. 315–323.
[48] Theano Development Team, Theano: A Python framework for
fast computation of mathematical expressions, arXiv e-prints
abs/1605.02688.
URL http://arxiv.org/abs/1605.02688
[49] J. Long, E. Shelhamer, T. Darrell, Fully convolutional networks for
semantic segmentation, in: Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, 2015, pp. 3431–3440.
[50] V. Badrinarayanan, A. Kendall, R. Cipolla, Segnet: A deep convolutional encoder-decoder architecture for image segmentation, arXiv
preprint arXiv:1511.00561.
[51] S. Jégou, M. Drozdzal, D. Vazquez, A. Romero, Y. Bengio, The one
hundred layers tiramisu: Fully convolutional densenets for semantic segmentation, in: Computer Vision and Pattern Recognition Workshops
(CVPRW), 2017 IEEE Conference on, IEEE, 2017, pp. 1175–1183.
[52] S. Han, H. Mao, W. J. Dally, Deep compression: Compressing deep
neural networks with pruning, trained quantization and huffman coding, arXiv preprint arXiv:1510.00149.
[53] R. B. Gramacy, H. K. Lee, Cases for the nugget in modeling computer
experiments, Statistics and Computing 22 (3) (2012) 713–722.
[54] D. A. Nix, A. S. Weigend, Estimating the mean and variance of the
target probability distribution, in: Neural Networks, 1994. IEEE World
Congress on Computational Intelligence., 1994 IEEE International Conference On, Vol. 1, IEEE, 1994, pp. 55–60.
[55] Q. V. Le, A. J. Smola, S. Canu, Heteroscedastic Gaussian process regression, in: Proceedings of the 22nd international conference on Machine learning, ACM, 2005, pp. 489–496.
[56] A. Kendall, Y. Gal, What Uncertainties Do We Need in Bayesian Deep
Learning for Computer Vision?, arXiv preprint arXiv:1703.04977.
[57] D. M. Blei, A. Kucukelbir, J. D. McAuliffe, Variational inference: A
review for statisticians, Journal of the American Statistical Association 112 (518) (2017) 859–877. arXiv:https://doi.org/10.1080/
50
01621459.2017.1285773, doi:10.1080/01621459.2017.1285773.
URL https://doi.org/10.1080/01621459.2017.1285773
[58] Q. Liu, Stein variational gradient descent as gradient flow, in: Advances
in Neural Information Processing Systems, 2017, pp. 3117–3125.
[59] Q. Liu, J. Lee, M. Jordan, A kernelized Stein discrepancy for goodnessof-fit tests, in: International Conference on Machine Learning, 2016,
pp. 276–284.
[60] C. E. Rasmussen, C. K. Williams, Gaussian processes for machine learning, Vol. 1, MIT press Cambridge, 2006.
[61] M. S. Alnæs, J. Blechta, J. Hake, A. Johansson, B. Kehlet, A. Logg,
C. Richardson, J. Ring, M. E. Rognes, G. N. Wells, The FEniCS Project
Version 1.5, Archive of Numerical Software 3 (100). doi:10.11588/
ans.2015.100.20553.
[62] B. Lakshminarayanan, A. Pritzel, C. Blundell, Simple and scalable predictive uncertainty estimation using deep ensembles, arXiv preprint
arXiv:1612.01474.
[63] L. Li, K. Jamieson, G. DeSalvo, A. Rostamizadeh, A. Talwalkar, Hyperband: A novel bandit-based approach to hyperparameter optimization,
arXiv preprint arXiv:1603.06560.
[64] D. Kingma, J. Ba, Adam: A method for stochastic optimization, arXiv
preprint arXiv:1412.6980.
[65] J. Quinonero Candela, C. Rasmussen, F. Sinz, O. Bousquet,
B. Schölkopf, Evaluating predictive uncertainty challenge, in: Machine
Learning Challenges: Evaluating Predictive Uncertainty, Visual Object Classification, and Recognising Tectual Entailment, Max-PlanckGesellschaft, Springer, Berlin, Germany, 2006, pp. 1–27.
[66] C. Zhang, S. Bengio, M. Hardt, B. Recht, O. Vinyals, Understanding
deep learning requires rethinking generalization, CoRR abs/1611.03530.
arXiv:1611.03530.
URL http://arxiv.org/abs/1611.03530
[67] T. Poggio, K. Kawaguchi, Q. Liao, B. Miranda, L. Rosasco, X. Boix,
J. Hidary, H. Mhaskar, Theory of Deep Learning III: explaining the
non-overfitting puzzle, ArXiv e-printsarXiv:1801.00173.
51
[68] L. de Oliveira, M. Paganini, B. Nachman, Learning particle physics by
example: Location-aware generative adversarial networks for physics
synthesis, arXiv preprint arXiv:1701.05927.
[69] W. Luo, Y. Li, R. Urtasun, R. Zemel, Understanding the effective receptive field in deep convolutional neural networks, in: Advances in
Neural Information Processing Systems, 2016, pp. 4898–4906.
52
| 1cs.CV
|
Parallel Vertex Approximate Gradient discretization of
hybrid dimensional Darcy flow and transport in discrete
fracture networks
arXiv:1602.03641v2 [math.NA] 17 Nov 2016
F. Xing∗, R. Masson†, S. Lopez‡
November 18, 2016
Abstract
This paper proposes a parallel numerical algorithm to simulate the flow and the transport in a discrete fracture network taking into account the mass exchanges with the
surrounding matrix. The discretization of the Darcy fluxes is based on the Vertex Approximate Gradient finite volume scheme adapted to polyhedral meshes and to heterogeneous anisotropic media, and the transport equation is discretized by a first order upwind
scheme combined with an Euler explicit integration in time. The parallelization is based
on the SPMD (Single Program, Multiple Data) paradigm and relies on a distribution of
the mesh on the processes with one layer of ghost cells in order to allow for a local assembly of the discrete systems. The linear system for the Darcy flow is solved using different
linear solvers and preconditioners implemented in the PETSc and Trilinos libraries. The
convergence of the scheme is validated on two original analytical solutions with one and
four intersecting fractures. Then, the parallel efficiency of the algorithm is assessed on
up to 512 processes with different types of meshes, different matrix fracture permeability
ratios, and different levels of complexity of the fracture network.
1
Introduction
1.1
Hybrid dimensional flow and transport models
This article deals with the simulation of the Darcy flow and transport in fractured porous media
for which the fractures are modeled as interfaces of codimension one. In this framework, the
d − 1 dimensional flow and transport in the fractures is coupled with the d dimensional flow and
transport in the matrix leading to the so called hybrid dimensional Darcy flow and transport
model.
For the Darcy flow model, we focus on the particular case where the pressure is continuous
at the interfaces between the fractures and the matrix domain. This type of Darcy flow model
introduced in [3], [4] corresponds physically to pervious fractures for which the ratio of the
∗
Laboratoire de Mathématiques J.A. Dieudonné, UMR 7351 CNRS, University Nice Sophia Antipolis, team
COFFEE, INRIA Sophia Antipolis Méditerranée, Parc Valrose 06108 Nice Cedex 02, France, and BRGM
Orléans France, [email protected]
†
Laboratoire de Mathématiques J.A. Dieudonné, UMR 7351 CNRS, University Nice Sophia Antipolis,
and team COFFEE, INRIA Sophia Antipolis Méditerranée, Parc Valrose 06108 Nice Cedex 02, France,
[email protected]
‡
BRGM, scientific and Technical Center, 3 avenue Claude Guillemin, BP 36009, 45060 Orléans Cedex 2
France, [email protected]
1
normal permeability of the fracture to the width of the fracture is large compared with the
ratio of the permeability of the matrix to the size of the domain. Note that it does not cover
the case of fractures acting as barriers for which the pressure is discontinuous at the matrix
fracture interfaces (see [22], [33], [5] for discontinuous pressure models). It is also assumed
in our model that the pressure is continuous at the fracture intersections. It corresponds to
the assumption that the ratio between the permeability at the fracture intersections and the
width of the fracture is large compared to the ratio between the tangential permeability of each
fracture and its length. We refer to [24] and [40] for more general reduced models taking into
account discontinuous pressures at fracture intersections in dimension d = 2.
The hybrid dimensional transport model is derived in [3] in the case of a convection diffusion
flux for the matrix and fracture concentration. In this work, a purely advective model is
considered. It requires the specification of the transmission conditions at the matrix fracture
interfaces and at fracture intersections which, to our knowledge, have not been done so far at
the continuous level.
The discretization of the hybrid dimensional Darcy flow model with continuous pressures has
been the object of several works. In [31] a cell-centred Finite Volume scheme using a Two Point
Flux Approximation (TPFA) is proposed assuming the orthogonality of the mesh and isotropic
permeability fields. Cell-centred Finite Volume schemes can be extended to general meshes
and anisotropic permeability fields using MultiPoint Flux Approximations (MPFA) following
the ideas of [42], [39],[1] and [2]. In [3] and [29] a Mixed Finite Element (MFE) method is
proposed, and Control Volume Finite Element Methods (CVFE) using nodal unknowns have
been introduced for such models in [9], [38], [35], [34], [28]. The Hybrid Finite Volume and
Mimetic finite difference schemes, belonging to the family of Hybrid Mimetic Mixed Methods
[17], have been extended to hybrid dimensional models in [23], [6] as well as in [12], [13] in
the more general Gradient Discretization framework [18]. Non-matching discretizations of the
fracture and matrix meshes are studied in [16], [25], [8] and [40].
Regarding the hybrid dimensional advective transport model, an explicit first order upwind
scheme combined with the MPFA Darcy fluxes is used in [1], [2], and [39]. At fracture intersections, the authors neglect the accumulation term and the concentration unknown is eliminated
using the flux conservation equation in order to avoid severe restrictions on the time step caused
by the small volumes. A CVFE method is used in [38] with a first order upwind approximation
and a fully implicit time integration of the two phase flow model to avoid small time steps.
Higher order methods have also been developed in the CVFE method of [34] using a MUSCL
type second order scheme for the saturation equation and also in [29] where a Discontinuous
Galerkin method is used for the transport saturation equation with an Euler implicit time integration in the fracture network and an explicit time integration in the matrix domain. In [26],
a streamline method is developed in 2D based on the hybrid dimensional Darcy flow velocity
field. The solution is very accurate for purely advective transport but this method requires that
the fractures be expanded and seems difficult to extend to the case of a complex 3D network
in practice.
1.2
Content and objectives of this work
In this work, we focus on the Vertex Approximate Gradient (VAG) scheme introduced in [19]
for diffusion problems and extended in [11], [12], [13] to hybrid dimensional Darcy flow models.
The VAG scheme uses nodal and fracture-face unknowns in addition to the cell unknowns which
can be eliminated without any fill-in. Thanks to its essentially nodal feature, it leads to a sparse
discretization on tetrahedral or mainly tetrahedral meshes. The VAG scheme has the major
advantage, compared with the CVFE methods of [9], [38], [35] or [34], that it avoids the mixing
2
of the control volumes at the fracture matrix interfaces, which is a key feature for its coupling
with the transport model. As shown in [11] for two phase flow problems, the VAG scheme
allows for a coarser mesh size at the matrix fracture interface for a given accuracy. For the
discretization of the transport hybrid dimensional model, we will use in this work a simple first
order upwind scheme with explicit time integration. The extension to second order MUSCL
type discretization will be considered in a future work. Our main objective in this paper is to
develop a parallel algorithm for the VAG discretization of hybrid dimensional Darcy flow and
transport models, and to assess the parallel scalability of the algorithm.
Starting from the hybrid dimensional Darcy flow model of [11] and [12], we first derive the
hybrid dimensional transport model for a general fracture network taking into account fracture
intersections and the coupling with the matrix domain. Then, the VAG discretization of the
Darcy flow model is recalled and the VAG Darcy fluxes are used to discretize the transport
model with an upwind first order discretization in space and an Euler explicit time integration.
A key feature of this discretization is the definition of the control volumes which is adapted to
the heterogeneities of the porous medium. This can be achieved thanks to the fact that, on the
one hand, the VAG scheme keeps the cell unknowns and, on the other hand, the VAG Darcy
fluxes are constructed independently of the definition of the control volumes. In particular, the
control volumes are constructed in such a way that, at matrix fracture interfaces, the volume
is taken only in the fracture. Otherwise, the fracture will be enlarged artificially and the
front velocity will not be accurately approximated in the fractures as it it the case for usual
CVFE methods. Note also that we do not eliminate the concentration unknowns at fracture
intersections as was done in [39], [1] and [2] for cell centred discretizations. In the case of a
nodal discretization like the VAG scheme, this elimination is not possible since these unknowns
are connected to the matrix and it is not needed since the size of the control volumes at fracture
intersections is roughly the same as the size of any control volume located at the matrix fracture
interface.
Our parallelization of the hybrid dimensional flow and transport numerical model is based
on the SPMD (Single Program, Multiple Data) paradigm. It relies on a distribution of the
mesh on the processes with one layer of ghost cells in order to allow for a local assembly of the
discrete systems. The linear system for the Darcy flow is solved using different linear solvers
and preconditioners implemented in the PETSc and Trilinos libraries.
In order to validate the convergence of the scheme, two analytical solutions are constructed
for the hybrid dimensional flow and transport model. We consider the case of a single nonimmersed fracture as well as the case of four intersecting fractures. The analytical solutions for
the transport model are obtained by integration of the matrix and fracture equations along the
characteristics of the velocity field taking into account source terms coming from the matrix
fracture transmission conditions. Then, we study the parallel scalability of the Darcy flow and
transport solvers on up to 512 processes. Our numerical investigation includes different levels
of complexity of the fracture network with a number of fractures ranging from a few to a few
hundreds. It covers different types of meshes namely hexahedral, tetrahedral and prismatic
meshes as well as a large range of permeability ratios between the fracture network and the
matrix domain. In addition, the influence of the choices of the linear solver and of the preconditioner is also studied for the solution of the Darcy flow equation.
The paper is organized as follows. Section 2 recalls the geometrical and functional setting
introduced in [12] for a general 2D fracture network immersed in a surrounding 3D matrix.
Then, the hybrid dimensional Darcy flow and transport models are introduced. In Section
3, the VAG discretization is recalled for the Darcy flow model and extended to the transport
model. The parallel implementation of the scheme is detailed in section 4. Section 5 is devoted
3
to the description of the test cases including the analytical solutions and to the numerical
investigation of the parallel scalability of the algorithm.
2
2.1
Hybrid dimensional Darcy Flow and Transport Model
in Fractured Porous Media
Discrete Fracture Network and functional setting
Let Ω denote a bounded domain of Rd , d = 2, 3 assumed to be polyhedral for d = 3 and
polygonal for d = 2. To fix ideas the dimension will be fixed to d = 3 when it needs to be
specified, for instance in the naming of the geometrical objects or for the space discretization
in the next section. The adaptations to the case d = 2 are straightforward.
We consider the asymptotic model introduced in [3] where
S fractures are represented as
interfaces of codimension 1. Let I be a finite set and let Γ = i∈I Γi and its interior Γ = Γ \ ∂Γ
denote the network of fractures Γi ⊂ Ω, i ∈ I, such that each Γi is a planar, polygonal, simply
connected, open domain included in an oriented plane Pi of Rd . It is assumed that the angles
For all i ∈ I, let us set
of Γi are strictly smaller than 2π and that Γi ∩ Γj = ∅ for all i 6= j. S
Σi = ∂Γi , Σi,j = Σi ∩ Σj , j ∈ I \ {i}, Σi,0 = Σi ∩ ∂Ω, Σi,N = Σi \ ( j∈I\{i} Σi,j ∪ Σi,0 ), and
S
Σ = (i,j)∈I×I,i6=j Σi,j . It is assumed that Σi,0 = Γi ∩ ∂Ω. We will denote by dτ (x) the d − 1
Figure 1: Example of a 2D domain with 3 intersecting fractures Γ1 , Γ2 , Γ3 and 2 connected
components Ω1 , Ω2 .
dimensional Lebesgue measure on Γ. On the fracture network Γ, we define theP
function space
2
2
2
L (Γ) = {v = (vi )i∈I , vi ∈ L (Γi ), i ∈ I}, endowed with the norm kvkL2 (Γ) = i∈I kvi k2L2 (Γi ) .
Its subspace H 1 (Γ) is defined as the space of functions v = (vi )i∈I such that vi ∈ H 1 (Γi ), i ∈ I
with continuous
at the fracture intersections. The space H 1 (Γ) is endowed
P traces
S with the norm
2
2
kvkH 1 (Γ) = i∈I kvi kH 1 (Γi ) and its subspace with vanishing traces on Σ0 = i∈I Σi,0 is denoted
by HΣ1 0 (Γ).
Let us also consider the trace operator γi from H 1 (Ω) to L2 (Γi ) as well as the trace operator
γ from H 1 (Ω) to L2 (Γ) such that (γv)i = γi (v) for all i ∈ I.
On Ω, the gradient operator from H 1 (Ω) to L2 (Ω)d is denoted by ∇. On the fracture
network Γ, the tangential gradient ∇τ acting from H 1 (Γ) to L2 (Γ)d−1 is defined by
∇τ v = (∇τi vi )i∈I ,
where, for each i ∈ I, the tangential gradient ∇τi is defined from H 1 (Γi ) to L2 (Γi )d−1 by fixing
a reference Cartesian coordinate system of the plane Pi containing Γi . We also denote by divτi
4
the divergence operator from Hdiv (Γi ) to L2 (Γi ).
The function spaces arising in the variational formulation of the hybrid dimensional Darcy
flow model are
V = {v ∈ H 1 (Ω) such that γv ∈ H 1 (Γ)},
and its subspace
V 0 = {v ∈ H01 (Ω) such that γv ∈ HΣ1 0 (Γ)}.
The space V 0 is endowed with the following Hilbertian norm
1/2
.
kvkV 0 = k∇vk2L2 (Ω)d + k∇τ γvk2L2 (Γ)d−1
Let Ωα , α ∈ A denote the connected components of Ω \ Γ, with A being the set of connected
components of Ω \ Γ. Let us define the space Hdiv (Ω \ Γ) = {qm = (qm,α )α∈A | qm,α ∈
Hdiv (Ωα )}. Using the orientation of Pi we can define the two sides ± of the fracture Γi , for all
±
i ∈ I. For all qm ∈ Hdiv (Ω \ Γ), let γn,i
qm denote the normal trace of qm on the side ± of Γi
with the normal oriented outward from the side ±. Let us define the Hilbert function space
H(Ω, Γ) = { qm = (qm,α )α∈A , qf = (qf,i )i∈I | qm ∈ Hdiv (Ω \ Γ),
+
−
qf,i ∈ L2 (Γi )d−1 , divτi (qf,i ) − γn,i
qm − γn,i
qm ∈ L2 (Γi ), i ∈ I},
and its closed Hilbert subspace
XZ
W (Ω, Γ) = {(qm , qf ) ∈ H(Ω, Γ) |
(qm,α · ∇v + div(qm,α )v)dx
Ωα
α∈A
XZ
+
−
+
(qf,i · ∇τi γi v + (divτi (qf,i ) − γn,i
qm − γn,i
qm )γi v)dτ (x) = 0 ∀ v ∈ V 0 }.
i∈I
(1)
Γi
P
The last definition corresponds to imposing in a weak sense the conditions i∈I γn,Σi qf,i = 0 on
Σ \ Σ0 and γn,Σi qf,i = 0 on Σi,N , i ∈ I, where γn,Σi is the normal trace operator on Σi (tangent
to Γi ) with the normal oriented outward from Γi , and using the extension of γn,Σi qf,i by zero
on Σ \ Σi .
2.2
Hybrid dimensional Darcy Flow Model
In the matrix domain Ω \ Γ (resp. in the fracture network Γ), let us denote by Λm ∈ L∞ (Ω)d×d
(resp. Λf ∈ L∞ (Γ)(d−1)×(d−1) ) the permeability tensor such that there exist λm ≥ λm > 0 (resp.
λf ≥ λf > 0) with
λm |ξ|2 ≤ (Λm (x)ξ, ξ) ≤ λm |ξ|2 for all ξ ∈ Rd , x ∈ Ω,
(resp. λf |ξ|2 ≤ (Λf (x)ξ, ξ) ≤ λf |ξ|2 for all ξ ∈ Rd−1 , x ∈ Γ).
We also denote by µ the fluid viscosity which is assumed constant and by df ∈ L∞ (Γ) the
width of the fractures assumed to be such that there exist df ≥ df > 0 with df ≤ df (x) ≤ df
for all x ∈ Γ.
Given ū ∈ V , the strong formulation of the hybrid dimensional Darcy flow model amounts
to: find u ∈ V and (qm , qf ) ∈ W (Ω, Γ) such that u − ū ∈ V 0 and
div(qm,α ) = 0
on Ωα , α ∈ A,
Λm
qm,α = − µ ∇u
on Ωα , α ∈ A,
(2)
+
−
divτi (qf,i ) − γn,i qm − γn,i qm = 0
on Γi , i ∈ I,
Λ
qf,i = −df µf ∇τi γi u on Γi , i ∈ I.
5
The weak formulation of (2) amounts to: find u ∈ V such that u − ū ∈ V 0 and the following
variational equation is satisfied for all v ∈ V 0 :
Z
Z
Λm (x)
Λf (x)
∇u(x) · ∇v(x)dx + df (x)
∇τ γu(x) · ∇τ γv(x)dτ (x) = 0.
(3)
µ
µ
Ω
Γ
The existence and uniqueness of the solution to (3) derives from the Lax Milgram theorem and
a Poincaré inequality stated in [12].
2.3
Hybrid dimensional transport model
Let γn be the normal trace operator on ∂Ω with the normal oriented outward from Ω. Let us
define ∂Ω− = {x ∈ ∂Ω | γn qm (x) < 0}, Σ−
i,0 = {x ∈ Σi,0 | γn,Σi qf,i (x) < 0}, i ∈ I, as well as the
following subset of Σ \ Σ0 :
X
Σ− = {x ∈ Σ \ Σ0 |
|γn,Σi qf,i (x)| =
6 0}.
i∈I
We consider a linear, purely advective model with velocity qm in the matrix domain and qf
in the fracture network. The matrix concentration is denoted by cm (cm,α in each connected
component Ωα , α ∈ A) and the fracture concentration, representing the average concentration
in the fracture width, is denoted by cf (cf,i in each fracture Γi , i ∈ I). The 2D equation in the
fracture network is as usual obtained by integration of the 3D advection equation in the width
of the fractures. For a purely advective equation, the transmission condition at the matrix
fracture interfaces states that the input normal flux in the matrix is obtained using the upwind
fracture concentration cf . At the fracture intersection Σ− , an additional unknown cf,Σ− must
be introduced and the transmission conditions state that the normal fluxes sum to zero and
that the input normal fluxes are obtained using the upwind concentration cf,Σ− .
Let be given the input boundary conditions c̄m ∈ L∞ (∂Ω− ), c̄f,i ∈ L∞ (Σ−
i,0 ), i ∈ I, and the
0
∞
0
∞
initial conditions cm ∈ L (Ω\Γ), cf ∈ L (Γ). Let us denote by φm (x) the porosity in the matrix
and by φf (x) the porosity in the fracture network. The transport
hybrid dimensional
model
amounts to: find cm ∈ L∞ (Ω\Γ)×(0, T ) , cf ∈ L∞ Γ×(0, T ) , and cf,Σ− ∈ L∞ Σ− ×(0, T ) ,
such that:
φm ∂t cm,α + div(cm,α qm,α ) = 0
on Ωα × (0, T ), α ∈ A
+
−
φ
d
∂
c
+
div
(c
q
)
=
γ
c
q
+
γ
c
q
on Γi × (0, T ), i ∈ I,
f f t f,i
τi f,i f,i
n,i m m
n,i m m
±
±
−
−
(γn,i cm qm ) = cf (γn,i qm )
on Γi × (0, T ), i ∈ I,
−
−
(γn,Σi cf,i qf,i ) = cf,Σ− (γn,Σi qf,i )
on (Σi \ Σi,0 ) × (0, T ), i ∈ I,
X
γn,Σj cf,j qf,j = 0
on (Σ \ Σ0 ) × (0, T ),
(4)
j∈I
(γn cm qm )− = c̄m (γn qm )−
on ∂Ω × (0, T ),
−
−
(γ
c
q
)
=
c̄
(γ
q
)
on Σi,0 × (0, T ), i ∈ I,
n,Σi f,i f,i
f,i n,Σi f,i
0
cm = cm
on (Ω \ Γ) × {t = 0},
cf = c0f
on Γ × {t = 0},
where the notations a+ = max(a, 0) and a− = min(a, 0) are used for all a ∈ R.
6
3
3.1
Vertex Approximate Gradient Discretization (VAG)
VAG discretization of the Darcy flow model
In the spirit of [19], we consider generalized polyhedral meshes of Ω in the sense that the edges
of the mesh are linear but its faces are not necessarily planar. Roughly speaking, each face is
assumed to be defined by the union of the triangles joining each edge of the face to a so called
face centre. This definition has the advantage to include in particular hexahedral cells with
non planar faces.
S Let M be the set of cells which are disjoint open polyhedral subsets of Ω such that
K∈M K = Ω. For each K ∈ M, it is assumed that there exists xK ∈ K \ ∂K, the socalled “centre” of the cell K, such that K is star-shaped with respect to xK . We then denote
by FK the set of interfaces of non zero d − 1 dimensional measure among the interior faces
K ∩ L, L ∈ M \ {K}, and the boundary interface K ∩ ∂Ω, which possibly splits in several
boundary faces. Let us denote by
[
F=
FK
K∈M
the set of all faces of the mesh. Note that the faces are not assumed to be planar, hence the
term “generalized polyhedral mesh”. For σ ∈ F, let Eσ be the set of interfaces of non zero,
d − 2 dimensional measure among the interfaces σ ∩ σ 0 , σ 0 ∈ F \ {σ}. Then, we denote by
[
Eσ
E=
σ∈F
the set of all edges of the mesh. Let Vσ = (e,e0 )∈Eσ2 ,e6=e0 e ∩ e0 be the set of nodes of σ. For
S
each K ∈ M we define VK = σ∈FK Vσ , and we also denote by
[
VK
V=
S
K∈M
the set of all nodes of the mesh. It S
is then assumed that forPeach face σ ∈ F, there
P exists a socalled “centre” of the face xσ ∈ σ \ e∈Eσ e such that xσ = s∈Vσ βσ,s xs , with
s∈Vσ βσ,s = 1,
and βσ,s ≥ 0 for all s ∈ Vσ ; moreover the face σ is assumed to be defined by the union of the
triangles Tσ,e defined by the face centre xσ and each edge e ∈ Eσ .
The mesh is also supposed to be conforming w.r.t. the fracture
S network Γ in the sense that
for each i ∈ I there exists a subset FΓi of F such that Γi = σ∈FΓ σ. We will denote by FΓ
i
S
the set of fracture faces i∈I FΓi . The following notations will be used for convenience:
Ms = {K ∈ M | s ∈ VK },
Mσ = {K ∈ M | σ ∈ FK },
FΓ,s = {σ ∈ FΓ | s ∈ Vσ },
and
FΓ,K = FK ∩ FΓ .
This geometrical discretization of Ω and Γ is denoted in the following by D.
The VAG discretization was introduced in [19] for diffusive problems on heterogeneous,
anisotropic media. Its extension to the hybrid dimensional Darcy model is based on the following vector space of unknowns:
XD = {vK , vs , vσ ∈ R, K ∈ M, s ∈ V, σ ∈ FΓ },
7
and its subspace with homogeneous Dirichlet boundary conditions on ∂Ω:
XD0 = {v ∈ XD | vs = 0 for s ∈ Vext }.
where Vext = V ∩ ∂Ω denotes the set of boundary nodes, and Vint = V \ ∂Ω denotes the set of
interior nodes.
A finite element discretization of V is built using a tetrahedral sub-mesh of M and a second
order interpolation at the face centres xσ , σ ∈ F \ FΓ defined by the operator Iσ : XD → R
such that
X
Iσ (v) =
βσ,s vs .
s∈Vσ
The tetrahedral sub-mesh is defined by T = {TK,σ,e , e ∈ Eσ , σ ∈ FK , K ∈ M} where TK,σ,e is
the tetrahedron joining the cell centre xK to the triangle Tσ,e (see Figure 2 for examples of such
tetrahedra).
Figure 2: For a cell K with one fracture face σ in bold: cell unknown vK located at xK , fractureface unknown vσ located at xσ , node unknowns vs , vs0 , face centre xσ0 of face σ 0 , triangle Tσ,e
(convex hull of e and xσ ), triangle Tσ0 ,e0 (convex hull of e0 and xσ0 ) and tetrahedron TK,σ0 ,e0
(convex hull of xK , xσ0 and e0 ).
For a given v ∈ XD , we define the function πT v ∈ V as the continuous piecewise affine
function on each tetrahedron of T such that πT v(xK ) = vK , πT v(s) = vs , πT v(xσ ) = vσ , and
πT v(xσ0 ) = Iσ0 (v) for all K ∈ M, s ∈ V, σ ∈ FΓ , and σ 0 ∈ F \ FΓ . The nodal basis of this
finite element discretization will be denoted by ηK , ηs , ησ , for K ∈ M, s ∈ V, σ ∈ FΓ .
The VAG discretization of the hybrid dimensional Darcy flow model (2) is based on its weak
formulation (3). Given ūs , s ∈ Vext , it amounts to: find uD ∈ XD with us = ūs for all s ∈ Vext
and such that for all vD ∈ XD0 one has
Z
Z
Λf (x)
Λm (x)
∇πT uD (x) · ∇πT vD (x)dx + df (x)
∇τ γπT uD (x) · ∇τ γπT vD (x)dτ (x) = 0. (5)
µ
µ
Ω
Γ
Following [12], this Galerkin Finite Element formulation (5) can be reformulated in terms
of discrete conservation laws using the following definition of the VAG fluxes. For all vD ∈ XD ,
the VAG matrix fluxes connect the cell K ∈ M to its nodes and fracture faces ν ∈ VK ∪ FΓ,K :
Z
X
Λm (x)
0
FK,ν (vD ) = −
∇πT vD (x) · ∇ην (x)dx =
aνK,ν (vK − vν 0 )
(6)
µ
K
0
ν ∈V ∪F
K
8
Γ,K
R
0
with aνK,ν = K Λmµ(x) ∇ην (x) · ∇ην 0 (x)dx. The VAG fracture fluxes connect the face σ ∈ FΓ to
its nodes s ∈ Vσ :
Z
X 0
Λf (x)
∇τ γπT vD (x) · ∇τ γηs (x)dτ (x) =
asσ,s (vσ − vs0 )
(7)
Fσ,s (vD ) = − df (x)
µ
σ
s0 ∈V
σ
0
with asσ,s =
R
σ
df (x)
Λf (x)
∇τ γηs (x)
µ
· ∇τ γηs0 (x)dτ (x).
Figure 3: Matrix fluxes (in blue) and fracture fluxes (in red) inside a cell K with a fracture
face σ (in bold). The matrix fluxes FK,ν connect the cell K to its nodes and fracture faces
ν ∈ VK ∪ FΓ,K . The fracture fluxes Fσ,s connect the face σ to the nodes s ∈ Vσ of σ.
Then, the Galerkin Finite Element formulation (5) is equivalent to: find uD ∈ XD satisfying
the following set of discrete conservation equations and Dirichlet boundary conditions:
X
X
F
(u
)
+
FK,σ (uD ) = 0, K ∈ M
K,s D
s∈VK
σ∈FΓ,K
X
X
Fσ,s (uD ) +
−FK,σ (uD ) = 0, σ ∈ FΓ
s∈V
K∈M
σ
σ
X
X
−F
(u
)
+
−Fσ,s (uD ) = 0, s ∈ Vint ,
K,s D
σ∈FΓ,s
K∈Ms
us = ūs , s ∈ Vext .
3.2
3.2.1
First order upwind discretization of the transport model
Definition of control volumes
The VAG discretization of the hybrid dimensional transport model combines the VAG matrix
and fracture fluxes (6), (7) with the following definition of the control volumes based on partitions of the cells and of the fracture faces. These partitions are respectively denoted, for all
K ∈ M, by
[
K = ωK ∪
ω K,s
s∈VK \(Vext ∪VΓ )
and, for all σ ∈ FΓ , by
σ = ωσ ∪
[
s∈Vσ \Vext
9
ω σ,s .
Then, the control volumes are defined by ωK for all cells K ∈ M, by ωσ for all fracture faces
σ ∈ FΓ , and by
[
ωs =
ω K,s ,
K∈Ms
for all nodes s ∈ Vint \ VΓ , and by
[
ωs =
ω σ,s ,
σ∈FΓ,s
for all nodes s ∈ VΓ \ Vext . Note that this definition avoid the mixing of the fracture and
matrix rocktypes at the control volumes s ∈ VΓ \ Vext and σ ∈ FΓ . This is exhibited in Figure
4 in comparison with an alternative choice mixing the matrix and fracture rocktypes which
artificially enlarges the fractures. We refer to [12] for numerical comparisons on a two phase
flow model of these two types of choices of the control volumes.
Figure 4: Example of choices of the control volumes at cells, fracture face, and nodes, in the
case of two cells K and L splitted by one fracture face σ (the width of the fracture has been
enlarged in this figure). The left figure exhibits the good choice with no mixing of fracture
and matrix rocktypes while the right figure exhibits the bad choice enlarging artificially the
fracture.
The same idea is applied for all nodes located at different rocktype interfaces. In practice,
for such a node s ∈ Vint \ VΓ (resp. s ∈ VΓ \ Vext ), the set ωK,s (resp. ωσ,s ) should be non
empty only for the cell(s) K (resp. fracture face(s) σ) with the largest permeability among
those around the node s (see [20] for details).
In practice, the above partitions of the cells and fracture faces does not need to be built. It
is sufficient to define the matrix volume fractions
R
dx
ω
αK,s = R K,s , s ∈ VK \ (Vext ∪ VΓ ), K ∈ M,
dx
K
P
constrained to satisfy αK,s ≥ 0, and s∈VK \(Vext ∪VΓ ) αK,s ≤ 1, as well as the fracture volume
fractions
R
df (x)dτ (x)
ω
ασ,s = Rσ,s
, s ∈ Vσ \ Vext , σ ∈ FΓ ,
d
(x)dτ
(x)
f
σ
P
such that ασ,s ≥ 0, and s∈Vσ \Vext ασ,s ≤ 1. Then, the porous volumes of the control volumes
are set to
Z
X
φK = (1 −
αK,s )
φm (x)dx, K ∈ M
K
s∈VK \(VΓ ∪Vext )
φσ = (1 −
X
Z
φf (x)dτ (x), σ ∈ FΓ ,
ασ,s )df,σ
σ
s∈Vσ \Vext
10
φs =
X
Z
σ
σ∈FΓ,s
φs =
Z
X
with df,σ =
3.2.2
σ
φm (x)dx, s ∈ V \ (Vext ∪ VΓ ),
αK,s
K∈Ms
R
φf (x)dτ (x), s ∈ VΓ \ Vext ,
ασ,s df,σ
K
df (x)dτ (x)
R
.
dτ (x)
σ
Time integration
For N ∈ N∗ , let us consider the time discretization t0 = 0 < t1 < · · · < tn−1 < tn · · · < tN = T
of the time interval [0, T ]. We denote the time steps by ∆tn = tn+1 − tn for all n = 0, · · · , N − 1.
Given c̄s , s ∈ Vext with arbitrary values on the set of ouput boundary nodes
+
Vext
= {s ∈ Vext | FK,s (uD ) ≥ 0 ∀K ∈ Ms and Fσ,s (uD ) ≥ 0 ∀σ ∈ FΓ,s },
and c0D ∈ XD such that c0s = c̄s for all s ∈ Vext , the transport discrete model amounts to find
cn+1
∈ XD for all n = 0, · · · , N − 1 satisfying the following discrete conservation laws and
D
Dirichlet input conditions
X
X
cn+1
− cnK
n
K
+
H
(c
)
+
HK,σ (cnD ) = 0, K ∈ M,
φ
K,s
K
D
n
∆t
s∈VK
σ∈FΓ,K
n
n+1
X
X
−
c
c
σ
φ σ
+
Hσ,s (cnD ) −
HK,σ (cnD ) = 0, σ ∈ FΓ ,
σ
∆tn
s∈Vσ
K∈Mσ
n
n+1
X
X
−
c
c
s
s
n
Hσ,s (cnD ) = 0, s ∈ Vint ,
H
(c
)
−
−
φ
K,s
s
D
n
∆t
σ∈FΓ,s
K∈Ms
n+1
cs = c̄s , s ∈ Vext ,
with the following explicit upwind two point fluxes
HK,ν (cnD ) = cnK FK,ν (uD )+ + cnν FK,ν (uD )−
(8)
Hσ,s (cnD ) = cnσ Fσ,s (uD )+ + cns Fσ,s (uD )− .
The solution of this explicit upwind scheme classically satisfies the following maximum principle
+
≤ M for all µ ∈ V ∪ FΓ ∪ M \ Vext
,
m ≤ cn+1
µ
with
M=
max
+
µ∈V∪FΓ ∪M\Vext
c0µ and m =
min
+
µ∈V∪FΓ ∪M\Vext
c0µ ,
provided that the following Courant-Friedrichs-Lewy (CFL) condition
∆tn ≤ min(∆tM , ∆tFΓ , ∆tV ),
is satisfied with
φK
P
∆tM = min P
,
+
+
K∈M
s∈VK FK,s (uD ) +
σ∈FΓ,K FK,σ (uD )
φσ
P
∆tFΓ = min P
,
++
+
σ∈FΓ
F
(u
)
(−F
(u
))
σ,s
D
K,σ
D
s∈V
K∈M
σ
σ
φs
P
P
.
∆tV = min
+
+
s∈Vint
K∈Ms (−FK,s (uD )) +
σ∈FΓ,s (−Fσ,s (uD ))
11
(9)
4
Parallel implementation in ComPASS
The hybrid dimensional Darcy flow and transport discrete model is implemented in the framework of the code ComPASS (Computing Parallel Architecture to Speed up Simulations) [15],
which focuses on parallel high performance simulation (distributed memory, Message Parsing
Interface - MPI) adapted to general unstructured polyhedral meshes (see [21]).
4.1
Mesh non overlapping and overlapping decompositions
Let us denote by Np the number of MPI processes. The set of cells M is partitioned into Np
subsets Mp , p = 1, ..., Np using the library METIS [32]. The partitioning of the set of nodes
V and of the set of fracture faces FΓ is defined as follows: assuming we have defined a global
index of the cells K ∈ M let us denote by K(s), s ∈ V (resp. K(σ), σ ∈ FΓ ) the cell with the
smallest global index among those of Ms (resp. Mσ ). Then we set
V p = {s ∈ V | K(s) ∈ Mp },
and
FΓp = {σ ∈ FΓ | K(σ) ∈ Mp }.
The overlapping decomposition of M into the sets
p
M , p = 1, ..., Np ,
is chosen in such a way that any compact finite volume scheme such as the VAG scheme can
p
be assembled locally on each process. Hence, as exhibited in Figure 5, M is defined as the set
of cells sharing a node with a cell of Mp . The overlapping decompositions of the set of nodes
and of the set of fracture faces follow from this definition:
[
p
V =
VK , p = 1, · · · , Np ,
K∈M
and
p
FΓ =
[
p
FK ∩ FΓ , p = 1, · · · , Np .
p
K∈M
Figure 5: Example of mesh decomposition.
12
The partitioning of the mesh is performed by the master process (process 1), and then, each
local mesh is distributed to its process. Therefore, each MPI process contains the local mesh
p
p
p
(M , V , F Γ ), p = 1, 2, ..., Np which is splitted into two parts:
own mesh: (Mp , V p , FΓp ),
p
p
p
ghost mesh: (M \Mp , V \V p , F Γ \FΓp ).
We now turn to the parallel implementation of the discrete hybrid dimensional Darcy flow
model (2) and transport model (4).
4.2
Parallelization of the discrete hybrid dimensional Darcy flow
On each process p = 1, ..., Np , the local problem of the discrete hybrid dimensional Darcy flow
p
p
p
(2) is defined by the set of unknowns uµ , µ ∈ V ∪ F Γ ∪ M and the set of equations
X
X
p
F
(u
)
+
FK,σ (uD ) = 0, K ∈ M ,
K,s
D
s∈VK
σ∈FΓ,K
X
X
Fσ,s (uD ) +
−FK,σ (uD ) = 0, σ ∈ FΓp ,
(10)
s∈V
K∈MσX
σ
X
p
−FK,s (uD ) +
−Fσ,s (uD ) = 0, s ∈ Vint ∩ V ,
K∈M
σ∈F
s
Γ,s
us = ūs , s ∈ Vext ∩ V p .
Note that this includes the equations of the nodes s ∈ V p , of the fracture faces σ ∈ FΓp and of
p
p
the cells K ∈ M , both those own cells in K ∈ Mp and the ghost cells K ∈ M \ Mp . The
set of equations can be rewritten as the following rectangular linear system
p
p p
Avv Apvf Apvc
bv
Uv
Apf v Apf f Apf c U pf = bpf
(11)
p
p
Apcv Apcf Apcc
bc
Uc
p
p
p
p
p
p
where U v ∈ RV , U f ∈ RF Γ and U c ∈ RM denote the vector of process p own and ghost
unknowns at nodes, fracture faces and cells respectively. The above matrices have the following
sizes
Apvv ∈ RV
p ×V p
p
p
, Apvf ∈ RV
p ×F p
Γ
p
FΓp ×F Γ
Apf v ∈ RFΓ ×V , Apf f ∈ R
p
Apcv ∈ RM
p
p
p
×V
p
p
, Apcf ∈ RM
, Apvc ∈ RV
p ×Mp
,
p
p
, Apf c ∈ RFΓ ×M ,
p
×F Γ
p
, Apcc ∈ RM
×M
p
.
p
and bpv ∈ RV , bpf ∈ RFΓ , bc ∈ RM denote the corresponding right hand side vectors. The
matrix Apcc is a non singular diagonal matrix and the cell unknowns can be easily eliminated
without fill-in leading to the following Schur complement system
p
p Uv
A
= bp
(12)
p
Uf
with
p
A :=
Apvv Apvf
Apf v Apf f
p p
Apvc
b
Avc
p
p
p −1
p
Apcv Acf , b := pv −
−
(Acc )
(Apcc )−1 bc ,
Apf c
bf
Apf c
13
and
p
p
p
U c = (Apcc )−1 (bpc − Apcv U v − Apcf U f ).
(13)
The linear system (12) is built locally on each process p and transfered to the parallel
linear solver library PETSc [7] or Trilinos [27]. The parallel matrix and the parallel vector in
PETSc or Trilinos are stored in a distributed manner, i.e. each process stores its own rows. We
construct the following parallel linear system
AU = b,
(14)
with
A :=
A1 R1
A2 R2
..
.
ANp RNp
process 1
process 2
..
.
process Np
,
Uv1
process 1
U 1
f
2
U := Uv
process 2 ,
U 2
f
..
..
.
.
b :=
b1
b2
..
.
bN
p
process 1
process 2
..
.
process Np
where Rp , p = 1, 2, ..., Np is a restriction matrix satisfying
p
Uv
p
.
R U=
p
Uf
p
U
v
The matrix Ap Rp , the vector
and the vector bp are stored in process p.
Ufp
The parallel linear system (14) is solved using the GMRES or BiCGStab algorithm preconditioned by different type of preconditioners as discussed in the numerical
p section. The
Uv
of own node
solution of the linear system provides on each process p the solution vector
Ufp
p
and fracture-face unknowns. Then, the ghost node unknowns uµ , µ ∈ (V \V p ) and the ghost
p
fracture-face unknowns uµ , µ ∈ (F Γ \FΓp ) are recovered by a synchronization step with MPI
communications. This synchronization is efficiently implemented using a PETSc or Trilinos
matrix vector product
U = SU
(15)
where
1
Uv
1
U f
2
U :=
U 2v
U
f
..
.
is the vector of own and ghost node and fracture-face unknowns on all processes. The matrix S,
containing only 0 and 1 entries, is assembled once and for all at the beginning of the simulation.
p
Finally, thanks to (13), the vector of own and ghost cell unknowns U c is computed locally
on each process p.
In conclusion, the parallel implementation of the discrete hybrid dimensional Darcy flow
can be summarized as:
14
Algorithm 1: Parallel implementation of the discrete hybrid dimensional Darcy flow
1: Assemble locally on each process the rectangular linear system (11),
2: Compute locally on each process the Schur complement (12) of (11),
3: Construct the parallel linear system (14) in PETSc or Trilinos,
4: Solve the parallel linear system (14) to obtain the solution at own nodes
and fracture faces,
5: Communicate the solution at ghost nodes and fracture faces from (15),
6: Compute locally on each process the solution at own and ghost cells from
(13).
4.3
Parallelization of the discrete hybrid dimensional transport model
The parallel implementation of the transport model (4) with an explicit upwind discretization
of the fluxes consists of the following four steps.
1. Compute the Darcy matrix and fracture fluxes defined by (6) and (7).
2. Compute the maximum time step ∆t satisfying the CFL condition (9) and set ∆tn = ∆t
T
e.
for all n = 0, · · · , N − 2, and ∆tN −1 = T − (N − 1)∆t with N = d ∆t
3. For each time step n = 0, 1, ..., N − 2,
p
p
p
and cn+1
, cn+1
3a. Compute cn+1
σ
s
K , s ∈ Vint ∩V , σ ∈ FΓ , K ∈ M solution of the following
explicit equations
X
X
cn+1
− cnK
p
n
n
K
H
(c
)
=
0,
K
∈
H
(c
)
+
M
,
φ
+
K,σ
K,s
K
D
D
∆t
σ∈FΓ,K
s∈VK
n
n+1
X
X
−
c
c
σ
φ σ
HK,σ (cnD ) = 0, σ ∈ FΓp ,
Hσ,s (cnD ) −
+
σ
∆t
(16)
s∈Vσ
K∈Mσ
n+1
n
X
X
cs − cs
n
φ
−
H
(c
)
−
Hσ,s (cnD ) = 0, s ∈ Vint ∩ V p ,
s
K,s
D
∆t
K∈Ms
σ∈FΓ,s
p
cs = c̄s , s ∈ Vext ∩ V .
p
p
3b. Get the node and fracture-face ghost unknowns cn+1
, cn+1
s
σ , s ∈ Vint ∩ (V \V ), σ ∈
p
p
F Γ \FΓ using the PETSc or Trilinos matrix vector product with the matrix S defined
in (15), as was done for the Darcy flow solution U :
1
1
Cv
Cv
1
1
C f
C f
2
C v = S Cv2
2
C 2
C
f
f
..
..
.
.
Thanks to our mesh decomposition, step 1 and step 3a are performed locally on each process.
For step 2, the maximum time step ∆tp is computed locally on each process p, then the time
step ∆t is obtained using the MPI reduce operation.
15
5
Numerical experiments
All the numerical tests have been implemented on the Cicada cluster of the University Nice
Sophia Antipolis consisting of 72 nodes (16 cores/node, Intel Sandy Bridge E5-2670, 64GB/node).
We always fix 1 core per process and 16 processes per node. The communications are handled
by OpenMPI 1.8.2 (GCC 4.9) and PETSc 3.5.3.
The first two test cases are designed in order to validate the Darcy fluxes and the convergence
of the transport model discretization on two analytical solutions including one fracture for the
first test case and four intersecting fractures for the second test case. In the remaining test
cases, the parallel scalability of our Darcy flow and transport solvers is assessed with different
types of fracture networks and meshes and different matrix fracture permeability ratios. In
particular, the last test case applies our algorithm to a complex fracture network with hundreds
of fractures.
5.1
Numerical convergence for an analytical solution with one fracture
Let us set Ω = (0, 1)2 , and denote by (x, y) the Cartesian coordinates of x. We then define
x1 = (0, 41 ), θ ∈ (0, arctan( 34 )), x2 = (1, 14 + tan(θ)). Let Ω1 = {(x, y) ∈ Ω | y > 41 + x tan(θ)},
and Ω2 = Ω \ Ω1 . We consider a single fracture defined by Γ = (x1 , x2 ) = ∂Ω1 ∩ ∂Ω2 with
tangential permeability Λf > 0, and width df > 0. The matrix permeability is isotropic and set
to Λm = 1, the matrix and fracture porosities are set to φm = φf = 1, and the fluid viscosity is
set to µ = 1. The pressure solution is fixed to u(x, y) = 1 − x. In this case, the transport model
(4) reduces to the following system of equations which specifies our choice of the boundary and
initial conditions:
∂t cm,α (x, y, t) + ∂x cm,α (x, y, t) = 0
on Ωα × (0, T ), α = 1, 2,
cm,α (x, y, 0) = 0
on Ωα , α = 1, 2,
cm,1 (0, y, t) = 1
on ( 14 , 1) × (0, T ),
cm,2 (0, y, t) = 1
on (0, 14 ) × (0, T ),
(17)
cm,2 (x, 41 + x tan(θ), t) = cf (x, t)
on (0, 1) × (0, T ),
Lcf (x, t) = βcm,1 (x, 41 + x tan(θ), t) on (0, 1) × (0, T ),
cf (0, t) = 1
on (0, T ),
cf (x, 0) = 0
on (0, 1),
and k = Λf cos2 (θ). It is assumed that k > 1. This
where L = ∂t + k∂x + β with β = sin(θ)
df
system can be integrated along the characteristics of the matrix and fracture velocity fields
leading to the following analytical solution:
0 if t < x,
cm,1 (x, y, t) =
1 if t > x,
cf (x, t) =
0
e
cm,2 (x, y, t) =
if y ∈ (0, 14 )
if y ∈
( 14 , 14
β
− k−1
(x−t)
1
if t < xk ,
if xk < t < x,
if t > x,
0 if t < x,
1 if t(> x,
+ tan(θ))
0
4y−1
cf ( 4 tan(θ) , t +
16
4y−1
4 tan(θ)
if t < x −
− x) if t > x −
4y−1
,
4 tan(θ)
4y−1
.
4 tan(θ)
In the following numerical experiments the parameters are set to tan(θ) = 12 , Λf = 20 and
df = 0.01. The mesh is a topologically Cartesian nx × nx grid. Figure 6 shows an example
of the mesh with nx = 20 as well as the analytical solution in the matrix obtained at time
tf = 0.5 chosen as the final time of the simulation. The time step is defined by the maximum
time step allowed by the CFL condition (9). Figure 8 exhibits the convergence of the relative
L1 errors between the analytical solution and the numerical solution at time tf both in the
matrix domain and in the fracture as a function of the grid size nx = 100, 200, 400, 800, 1600.
Figure 7 shows the analytical solution and the numerical solutions obtained at time tf along
the fracture. In both cases, we can observe the expected convergence of the numerical solution
to the analytical solution with a higher order of convergence in the fracture due to the fact that
at time tf the analytical solution in the fracture is continuous as exhibited in Figure 7.
Figure 6: Left: example of mesh with nx = 20 where the red line is the fracture. Right:
analytical solution of (17) at time tf = 0.5.
1.05
nx =100
nx =200
nx =400
nx =800
nx =1600
exact
1.0
0.8
1.00
0.95
cf
cf
0.6
nx = 100
nx = 200
nx = 400
nx = 800
nx = 1600
exact
0.90
0.4
0.85
0.2
0.00.0
0.2
0.4
x
0.6
0.8
0.800.40
1.0
0.45
0.50
x
0.55
0.60
Figure 7: Left: analytical solution and numerical solutions along the fracture at time tf with
nx = 100, 200, 400, 800, 1600. Right: zoom view of left figure around x = 0.5.
17
matrix domain
fracture
slope 1
slope 1/2
L 1 Error
10 -1
10 -2
10 -3
10 -4
100
200
400
nx
800
1600
Figure 8: Relative L1 errors and references in the matrix domain and in the fracture at time
tf between the analytical solution and the numerical solutions as a function of the grid size
nx = 100, 200, 400, 800, 1600.
5.2
Numerical convergence for an analytical solution with four intersecting fractures
Let Ω = (0, 1)2 , x1 = (0, 14 ), θ1 ∈ (0, arctan( 43 )), x2 = (1, 14 + tan(θ1 )), x3 = ( 34 , 0), x4 =
( 34 − tan(θ2 ), 1), and the intersection of x1 x2 and x3 x4 equal to
x0 = (x0 , y0 ) =
1
(3 − tan(θ2 ), 1 + 3 tan(θ1 )).
4(1 + tan(θ1 ) tan(θ2 ))
It is assumed that θ1 , θ2 ∈ (0, arctan( 43 )).
We consider the four fractures Γ1 = (x1 , x0 ), Γ2 = (x0 , x2 ), Γ3 = (x3 , x0 ), Γ4 = (x0 , x4 ), with
tangential permeabilities Λf,1 = Λf,2 > 0, and Λf,3 = Λf,4 > 0, and with widths df,1 = df,2 > 0,
and df,3 = df,4 > 0. It is assumed that Λm = 1.
The fractures partition the domain Ω in the following four subdomains
Ω1 = {x = (x, y) ∈ Ω | y >
3
1
+ x tan(θ1 ), x < − y tan(θ2 )},
4
4
1
+ x tan(θ1 ), x >
4
1
Ω3 = {x = (x, y) ∈ Ω | y < + x tan(θ1 ), x <
4
1
Ω4 = {x = (x, y) ∈ Ω | y < + x tan(θ1 ), x >
4
Ω2 = {x = (x, y) ∈ Ω | y >
3
− y tan(θ2 )},
4
3
− y tan(θ2 )},
4
3
− y tan(θ2 )}.
4
Λ
d
sin(θ )
2
f,3 f,3
1)
2)
Let us set β1 = sin(θ
, k1 = Λf,1 cos2 (θ1 ), β2 = cos(θ
, k2 = Λf,3 cos(θ2 ) sin(θ2 ), r = Λf,1
.
df,1
df,3
df,1 cos(θ1 )
It is assumed that k1 > 1 and k2 tan(θ2 ) > 1. The matrix and fracture porosities are set to
φm = φf = 1 and the fluid viscosity is set to µ = 1.
The pressure solution is set to u(x, y) = 1 − x. In that case, the transport model (4) reduces
to the following system of equations which specifies our choice of the boundary and initial
conditions: find cm,α (x, y, t), α = 1, · · · , 4, cf,1 (x, t), cf,2 (x, t), cf,3 (y, t), cf,4 (y, t), and c0 (t) such
18
that
∂t cm,α (x, y, t) + ∂x cm,α (x, y, t)
cm,α (x, y, 0)
cm,1 (0, y, t)
cm,3 (0, y, t)
3
−
y
tan(θ
c
(
2 ), y, t)
m,2 4
3
cm,4 ( 4 − y tan(θ2 ), y, t)
cm,3 (x, 41 + x tan(θ1 ), t)
cm,4 (x, 41 + x tan(θ1 ), t)
L1 cf,1 (x, t) − β1 cm,1 (x, 41 + x tan(θ1 ), t)
L1 cf,2 (x, t) − β1 cm,2 (x, 14 + x tan(θ1 ), t)
L2 cf,3 (y, t) − β2 cm,3 ( 43 − y tan(θ2 ), y, t)
L2 cf,4 (y, t) − β2 cm,1 ( 34 − y tan(θ2 ), y, t)
cf,2 (x0 , t) = cf,3 (y0 , t)
(r + 1)c0 (t) − cf,1 (x0 , t) − rcf,4 (y0 , t)
cf,1 (0, t) = cf,4 (1, t)
cf,1 (x, 0)
cf,2 (x, 0)
cf,3 (y, 0)
cf,4 (y, 0)
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
0
0
0
0
cf,4 (y, t)
cf,3 (y, t)
cf,1 (x, t)
cf,2 (x, t)
0
0
0
0
c0 (t)
0
1
0
0
0
0
on
on
on
on
on
on
on
on
on
on
on
on
on
on
on
on
on
on
on
Ωα × (0, T ), α = 1, · · · , 4,
Ωα , α = 1, · · · , 4,
( 14 , 1) × (0, T ),
(0, 14 ) × (0, T ),
(y0 , 1) × (0, T ),
(0, y0 ) × (0, T ),
(0, x0 ) × (0, T ),
(x0 , 1) × (0, T ),
(0, x0 ) × (0, T ),
(x0 , 1) × (0, T ),
(0, y0 ) × (0, T ),
(y0 , 1) × (0, T ),
(0, T ),
(0, T ),
(0, T ),
(0, x0 ),
(x0 , 1),
(0, y0 ),
(y0 , 1),
(18)
where L1 = ∂t + k1 ∂x + β1 and L2 = ∂t − k2 ∂y + β2 .
This system can also be integrated analytically along the characteristics of the matrix and
fracture velocity fields, but it leads to complex computations. It is much easier to obtain the
stationary solution of this system which is defined in the fractures by
β
cf,1 (x) = e
− k1 x
1
,
β
− k2 (1−y)
cf,4 (y) = e
β
− k 1 x0
2
−
β2
,
(1−y )
0
+ re k2
c0 =
,
r+1
β1
β
3β
2)
β1 (r1 x− 3β
− 1x
x
(r x − 2 )
4k2
cf,2 (x) = e k1 c0 e k1 0 +
e
− e 1 0 4k2
,
k1 r1
β2
r
β
β1
β1
)
(−r2 y0 − 4k tan(θ
)
e k2 y c0 e− k22 y0 + β2 e(− 42 − 4k1 tan(θ
1) − e
1
1)
, if y < 14 ,
k2 r2
cf,3 (y) =
β
β
β1
β1
)
(−r2 y0 − 4k tan(θ
)
e k22 y c0 e− k22 y0 + β2 e(−r2 y− 4k1 tan(θ
1) − e
1
1)
if y > 41 ,
k2 r2
e
with r1 =
β1
k1
+
β2
k2
tan(θ1 ) and r2 =
β2
k2
+
1
β1
,
k1 tan(θ1 )
and in the matrix by
cm,1 (x, y) = 0,
cm,2 (x, y) = cf,4 (y),
(
0 1 if
cm,3 (x, y) =
y−
cf,1 tan(θ41 ) if
(
cf,3 (y) if
cm,4 (x, y) =
y− 1
cf,2 tan(θ41 ) if
19
y < 41 ,
y > 14 ,
y < y0 ,
y > y0 .
In the following numerical experiments the parameters are set to tan(θ1 ) = 58 , tan(θ2 ) = 14 ,
Λf,1 = 200, Λf,3 = 400, and df,1 = df,3 = 0.01. The mesh is, as for the previous test case, a
topologically Cartesian nx × nx grid exhibited in Figure 9 for nx = 20. Figure 9 also shows the
stationary analytical solution in the matrix. The time step is again defined by the maximum
time step allowed by the CFL condition (9) and the simulation time is chosen large enough to
obtain the numerical stationary solution.
Figure 10 exhibits the convergence of the relative L1 errors between the stationary analytical
and the numerical solutions both in the matrix domain and in the fracture as a function of
the grid size nx = 100, 200, 400, 800. We can again observe the expected convergence of the
numerical solution to the analytical solution with a higher order of convergence in the fracture
network due to the fact that the solution is continuous on each individual fracture as exhibited
in Figure 10. This property is always true when looking at the solution at the matrix time scale
and could be exploited in a future work by using an implicit time integration in the fracture
coupled to an explicit time integration in the matrix domain with a higher order discretization
in space in the spirit of what has been done in [29].
Figure 9: Left: example of mesh with nx = 20 where the red lines account for the four fractures.
Right: stationary analytical solution of (18).
10 0
matrix domain
fracture
slope 1
slope 1/2
L 1 Error
10 -1
10 -2
10 -3 50
100
200
nx
400
800
Figure 10: Left: stationary analytical solution in the four fractures as a function of the x
coordinate. Right: relative L1 errors and references in the matrix domain and in the fracture
network between the stationary analytical and numerical solutions as a function of the grid size
nx = 100, 200, 400, 800.
20
5.3
Fracture network with hexahedral meshes
The objective of this subsection and of the next subsection is to investigate the parallel scalability of the algorithms described in Section 4. In this subsection we consider a topologically
Cartesian mesh of size nx × nx × nx of the cubic domain (0, 1)3 exhibited in Figure 11 for
nx = 32. The mesh is exponentially refined at the interface between the matrix and the fracture network exhibited in Figure 12. The permeabilities are isotropic and set to Λf = 20 in the
fracture network and to Λm = 1 in the matrix. The porosities are set to φm = φf = 1 and the
fluid viscosity is set to µ = 1. The initial concentration is set to 0 both in the matrix domain
and in the fracture network and a value of 1 of the concentration is injected on the bottom
boundaries of the matrix and of the fracture network. The pressure is fixed to u = 1 and γu = 1
on the bottom boundary and to u = 0 and γu = 0 on the top boundary. The remaining lateral
boundaries are considered impervious. Figure 12 exhibits the tracer concentrations obtained
with the mesh nx = 128 at times t = 0, t = 0.2, t = 0.4 and at the final simulation time
tf = 0.5.
Figure 11: Hexahedral mesh in the matrix domain (left) and in the fracture network (right)
with nx = 32.
Table 1 presents the numbers of linear solver iterations for the stationary pressure solution
for a number of MPI processes ranging from Np = 2 to Np = 512 and with the mesh size
nx = 128 corresponding to roughly 2.1 × 106 cells, 2.1 × 106 nodes and 5.2 × 104 fractures faces.
Both the GMRES and BiCGStab linear solvers from the PETSc library are tested combined
with either the Boomer AMG preconditioner from the Hypre library [30], the Aggregation
AMG preconditioner from the Trilinos library [27] or the block Jacobi ILU(0) preconditioner
from the Euclid library. No restart is used for the GMRES linear solver. Table 2 shows the
corresponding computation times both for the setup phase of the preconditioner and for the
solve phase of the linear solver.
According to these tables, the GMRES and the BiCGStab linear solvers combined with the
Boomer AMG preconditioner are good choices for a number of processes Np 6 128, while the
BiCGStab linear solver combined with the block Jacobi ILU(0) preconditioner is more efficient
for this test case for Np = 256 and Np = 512. This was expected since the Boomer AMG
preconditioner requires a sufficiently large number of unknowns per core to maintain a good
parallel scalability due to the high level of communications in particular in the setup phase
of the preconditioner. For this linear system, the number of unknowns per MPI process is
roughly 4100 for Np = 512 which is too small for this type of preconditioner while the block
Jacobi preconditioner still maintains a good parallel scalability for such a number of unknowns
21
(a) t = 0.0
(b) t = 0.2
(c) t = 0.4
(d) t = 0.5
Figure 12: Concentration in the matrix domain and in the fracture network obtained at different
times with the mesh nx = 128. A threshold concentration of 0.2 is used in the matrix domain.
22
per MPI process. On the other hand, as shown in Table 3, Boomer AMG exhibits an optimal
scalability while ILU(0) is not scalable in terms of iteration count with respect to the ratio
Λf
between the fracture and matrix permeabilities. The same remark also holds in terms of
Λm
scalability with respect to the mesh size which means that the ILU(0) preconditioner is only
advantageous for small size and mildly heterogeneous problems.
Tables 1 and 2 also clearly show that the BiCGStab linear solver outperforms the GMRES
linear solver for cases requiring a large number of iterations due to the fact that the cost of
the orthogonalization procedure increases with the dimension of the Krylov subspace. The Aggregation AMG preconditioner yields a larger number of iterations compared with the Boomer
AMG preconditioner but has a much lower setup time resulting for this test case in a total
lower CPU time. However, this implementation of the Aggregation AMG preconditioner seems
to lack robustness with respect to the matrix fracture permeability ratio as exhibited in Table
3.
Table 1: Number of linear solver iterations vs. the number of MPI processes obtained with
different linear solvers and preconditioners for the mesh size nx = 128.
Np
GMRES + Boomer AMG
GMRES + Aggregation AMG
GMRES + ILU(0)
BiCGStab + Boomer AMG
BiCGStab + ILU(0)
2
15
59
751
9
508
4
15
78
707
9
476
8
15
65
655
9
484
16
15
39
644
9
503
32
15
65
648
9
473
64
16
54
634
10
513
128
15
73
633
9
491
256
15
53
624
9
487
512
15
62
613
10
484
Table 2: Linear solver setup phase and solve phase computation times vs. the number of MPI
processes obtained with different linear solvers and preconditioners for the mesh size nx = 128.
Np
GMRES
Boomer AMG
GMRES
Aggregation AMG
GMRES
ILU(0)
BiCGStab
Boomer AMG
BiCGStab
ILU(0)
Setup
Solve
Setup
Solve
Setup
Solve
Setup
Solve
Setup
Solve
2
34.1
26.3
4.7
45.1
16.9
672.3
38.0
37.1
18.9
179.4
4
20.1
15.6
1.9
20.9
21.3
590.9
23.3
21.3
19.9
111.7
8
16.3
14.8
1.6
17.0
16.3
281.6
15.0
11.5
16.5
86.0
16
11.7
7.2
1.5
9.7
23.2
163.9
10.3
7.4
22.1
59.9
32
11.3
5.2
2.3
5.2
14.6
71.4
9.1
4.1
14.3
27.9
64
11.9
3.8
2.9
2.5
11.0
30.7
9.4
2.9
12.4
15.4
128
12.1
2.5
4.4
2.3
9.7
16.7
12.8
2.5
9.4
8.0
256
19.2
5.2
6.6
1.5
6.0
8.3
14.8
4.4
5.8
4.2
512
29.3
9.6
11.3
3.2
4.8
4.0
23.8
10.0
3.9
2.2
Next, Figure 13 plots the total (Darcy flow and transport models) computation time and
the computation time for the transport model only as a function of the number of processes. In
these runs the GMRES linear solver is used combined with the Boomer AMG preconditioner
for Np 6 128 and with the ILU(0) preconditioner for Np = 256, 512. For the range 2 − 512 of
the number of processes, it appears that the computation time of the Darcy flow linear system
solution remains small compared with the transport model computation time. This can be
checked by comparison of the computation times in Table 2 and in Figure 13. This explains
the good scalability obtained for both the total and transport computation times thanks to the
explicit nature of the time integration scheme.
23
Table 3: Number of linear solver iterations vs. the matrix fracture permeability ratio
nx = 128 and Np = 2, 128.
Λf /Λm
GMRES + Boomer AMG
GMRES + Aggregation AMG
GMRES + ILU(0)
20
15
59
751
Np = 2
100 1000
15
16
-
20
15
73
633
Λf
Λm
for
Np = 128
100 1000
15
15
-
10
4
10
3
10
2
10
1
Total time
Linear speedup
Time (s)
Time (s)
-: The solver doesn’t converge in 1200 iterations.
2
10
4
10
3
10
2
Transport time
Linear speedup
2
4
8 16 32 64 128 256 512
Number of MPI processes
4
8 16 32 64 128 256 512
Number of MPI processes
Figure 13: Total computation time (left) and computation time for the transport model (right)
vs. the number of MPI processes for the mesh size nx = 128.
5.4
Fracture network with tetrahedral meshes
This test case considers tetrahedral meshes of the cubic domain (0, 1)3 conforming to the
fracture network. An example of tetrahedral mesh showing both the matrix domain and the
fracture network is exhibited in Figure 14. All the physical parameters, initial and boundary
conditions are the same as for the previous test case. The mesh used in this subsection contains
about 6.2 × 106 cells, 9.7 × 105 nodes and 7.1 × 104 fracture faces. Figure 15 exhibits the tracer
concentrations obtained with this tetrahedral mesh at times t = 0, t = 0.2, t = 0.4 and at the
final simulation time tf = 0.5.
As for the previous test case, Tables 4, 5 and 6 investigate the performance of the Darcy flow
system linear solution for both the GMRES and BiCGStab linear solvers and for the same three
preconditioners as in the previous test case. The conclusions are basically the same as for the
hexahedral mesh test case. The Boomer AMG preconditioner exhibits an optimal robustness
Λ
with respect to the matrix fracture permeability ratio Λmf . On the other hand it requires a
rather high number of unknowns per MPI process to maintain a good parallel scalability due to
the high level of communications in particular in the setup phase. The ILU(0) preconditioner
can be an interesting alternative but only for small size and midly heterogeneous problems.
The aggregation AMG preconditioner from the Trilinos library used in our test cases seems to
lack robustness and we did not manage to make it work better through tuning its parameters.
Figure 16 plots the total (Darcy flow and transport models) computation time and the
computation time of the transport model only as a function of the number of processes. In
these runs the GMRES linear solver is used combined with the Boomer AMG preconditioner
for Np 6 128 and with the ILU(0) preconditioner for Np = 256, 512. Compared with the
previous subsection, an even better parallel scalability of the transport model computation time
24
Figure 14: Example of tetrahedral mesh of the matrix domain (left) conforming to the fracture
network (right).
(a) t = 0.0
(b) t = 0.2
(c) t = 0.4
(d) t = 0.5
Figure 15: Concentration in the matrix domain and in the fracture network obtained at different
times for the tetrahedral mesh. A threshold of 0.2 is used in the matrix domain.
25
Table 4: Number of linear solver iterations vs. the number of MPI processes obtained with
different linear solvers and preconditioners for the tetrahedral mesh.
Np
GMRES + Boomer AMG
GMRES + Aggregation AMG
GMRES + ILU(0)
BiCGStab + Boomer AMG
BiCGStab + ILU(0)
2
11
38
1003
8
565
4
12
78
725
7
513
8
12
40
717
8
527
16
12
39
682
8
544
32
12
52
667
8
535
64
12
656
8
483
128
12
35
644
8
489
256
12
629
8
483
512
12
52
612
8
473
-: The relative residual norm stagnates after a few iterations.
-: Some future investigations are necessary.
Table 5: Linear solver setup phase and solve phase computation times vs. the number of MPI
processes obtained with different linear solvers and preconditioners for the tetrahedral mesh.
Np
GMRES
Boomer AMG
GMRES
Aggregation AMG
GMRES
ILU(0)
BiCGStab
Boomer AMG
BiCGStab
ILU(0)
Setup time
Solve time
Setup time
Solve time
Setup time
Solve time
Setup time
Solve time
Setup time
Solve time
2
12.4
8.0
3.7
19.7
5.7
560.6
21.5
24.0
5.8
110.4
4
7.8
5.5
1.9
20.9
7.4
254.4
14.5
10.2
6.4
63.0
8
4.9
2.9
1.2
5.1
7.2
150.0
9.9
6.1
6.4
39.0
16
5.0
1.7
1.8
2.7
5.6
66.5
6.5
3.5
5.4
19.2
32
4.3
1.1
2.1
2.0
4.7
30.1
5.3
1.8
4.7
11.6
64
6.2
0.9
1.6
5.2
15.2
5.9
1.5
5.0
5.4
128
7.2
1.4
2.9
1.5
3.4
7.7
8.2
2.1
3.4
2.8
256
13.5
3.1
3.3
2.8
4.1
12.4
4.3
2.6
1.4
512
22.4
6.9
4.7
3.0
1.8
2.8
19.7
9.5
1.8
1.2
-: The residual norm stagnates after a few iterations.
Table 6: Number of linear solver iterations vs. the matrix fracture permeability ratio
the tetrahedral mesh and Np = 2, 128.
Λf /Λm
GMRES + Boomer AMG
GMRES + Aggregation AMG
GMRES + ILU(0)
Np = 2
20
100 1000
11
13
12
38
1002
-
-: The solver doesn’t converge in 1200 iterations.
26
Np = 128
20 100 1000
12
13
12
35
644
-
Λf
Λm
for
10
3
10
2
10
1
10
0
10
3
Total time
Linear speedup
Transport time
Linear speedup
10
2
10
1
10
0
Time (s)
Time (s)
is observed in the right Figure 16. This can be explained by the ratio of roughly 6 between
the number of cells and the number of nodes typical of a tetrahedral mesh. For a topologically
Cartesian mesh, this ratio is roughly 1. Since the cell concentrations are computed locally in
each process, this explains the better scalability observed for this tetrahedral mesh compared
with the previous hexahedral mesh. On the left Figure 16, it is observed that the linear system
solution computation time is no longer small compared with the transport computation time
for Np = 256 and 512. Hence, it significantly reduces the parallel efficiency of the simulation
for a large number of processes, say Np = 256, 512 in this test case.
2
4
8 18 32 64 128 256 512
Number of MPI processes
2
4
8 16 32 64 128 256 512
Number of MPI processes
Figure 16: Total computation time (left) and computation time for transport model (right) vs.
number of MPI processes with tetrahedral mesh.
5.5
Application to a complex fracture network
In this subsection, our algorithm is applied to a complex fracture network kindly provided by
M. Karimi-Fard and A. Lapène from Stanford University and TOTAL. Figure 17 exhibits the
mesh of the domain Ω = (0, 5888.75) × (0, 3157.5) × (0, 250) (m) which contains about 1.2 × 107
prismatic cells, 6.5 × 106 nodes and 5.13 × 105 fracture faces. This 3D mesh is defined by the
tensor product of a triangular 2D mesh with a uniform vertical 1D mesh with 24 intervals.
The fracture network exhibited in Figure 18 contains 581 connected components. It is a set of
21376 × 24 faces of the 3D mesh defined by the tensor product of a subset of 21376 edges of
the triangular 2D mesh with the 1D vertical mesh. The 2D triangular mesh contains 517540
cells and is refined in the neighbourhood of the fracture network down to an average size of
3.5 m. Figure 18 also shows the location of the injection well and of the two production wells.
Each well is vertical of radius rw = 0.1 m and its centre in the horizontal plane is located at
the middle of a fracture edge in the 2D triangular mesh. In the vertical direction, only the 12
fracture faces at the center of the 1D mesh are perforated. The permeabilities are isotropic and
set to Λf = 10−11 (m2 ) in the fracture network and to Λm = 10−15 (m2 ) in the matrix domain.
The porosities are set to φm = φf = 0.1, the fracture width to df = 1 m and the fluid viscosity
to µ = 10−3 Pa.s−1 .
The initial concentration is set to 0 both in the matrix domain and in the fracture network. A
total volume of 5.0×106 m3 is injected in one year at the injector well with a tracer concentration
of 1. The pressures of each perforated fracture face σ of the producer wells are fixed to pw = 0
and the flow rates are given by the Peaceman model
qσ = W Iσ (pσ − pw ),
27
where pσ is the pressure in the fracture face and W Iσ the well index of the fracture face.
This well index is computed following Peaceman methodology [36], [37], [14] by expanding the
fracture face as a box of size dx × df × dz. The analytical pressure solution obtained for a
vertical well with the well pressure pw , the well radius rw and the well flow rate qw per unit
length is imposed at the 8 corners of the box. Then, the flow rate qw dz is imposed at the
box center and the pressure pc at the box center can be computed analytically using the VAG
scheme. We deduce the well index
qw dz
WI =
pc − pw
leading in this simple case to the analytical formula
WI =
2πdzΛf
log( rrw0 )
with
r0 = D exp(−
and
2πdz
),
C
4 dxdz dxdf
dzdf
C= (
+
+
),
3 df
dz
dx
q
D = 0.5 dx2 + d2f .
The production lasts 8 years.
Figure 17: Prismatic mesh of the domain Ω defined by the tensor product of a vertical 1D
uniform mesh with a 2D triangular mesh.
Figure 19 plots the mean tracer concentration in each well as a function of time as well as the
total volume of tracer as a function of time in the matrix, in the fracture network and their sum.
Figure 20 exhibit the pressure solution in the matrix domain and Figures 21 and 22 shows the
tracer concentration after one year of injection and at final time both in the matrix domain and
in the fracture network. Figure 23 shows the total computation times with different number
of MPI processes Np = 16, 32, 64, 128, 256, 512. It is observed that the total computation time
exhibits a rather good scalability. In addition, the linear solver (GMRES+Boomer AMG) for
the pressure converges in no more than 25 iterations whatever the number of MPI processes.
Also the comparison of the total and transport computation times in Figure 23 shows that the
time for the pressure solution remains small compared with the transport computation time up
to Np = 512.
28
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.000
5.0e+06
Producer (2938,1500)
Producer (3002,1256)
4.0e+06
Tracer amount
cf
Figure 18: Fracture network showing the location of the single injection well and of the two
production wells.
matrix
fracture
matrix+fracture
3.0e+06
2.0e+06
1.0e+06
1
2
3
4 5
Years
6
7
0.0e+000
8
1
2
3
4 5
Years
6
7
8
Figure 19: Mean tracer concentration in both production wells as a function of time (left)
defined as the ratio between the well tracer flow rate and the well fluid flow rate (equal in our
case to the well fracture-face tracer concentration). Volume of tracer as a function of time in
the matrix domain, in the fracture network and their sum (right).
Figure 20: Pressure on the matrix domain.
29
Figure 21: Tracer concentration after one year of injection in the matrix domain (left) and in
the fracture network (right).
Figure 22: Tracer concentration at final time in the matrix domain (left) and in the fracture
network (right).
10 2
10 2
Total time
Linear speedup
Transport time
Linear speedup
10 1
Time (h)
Time (h)
10 1
10 0
10 -1
16
32
64
128
256
Number of MPI processes
10 0
10 -1
512
16
32
64
128
256
Number of MPI processes
512
Figure 23: Total computation time in hours (left) and computation time for transport model
(right) vs. number of MPI processes with prismatic mesh.
30
6
Conclusion
This paper introduced a parallel VAG scheme for the simulation of a hybrid dimensional Darcy
flow and transport model in a discrete fracture network taking into account the mass exchanges
with the matrix. The convergence of the scheme was validated on two original analytical
solutions for a flow and transport model that includes fractures. The parallel efficiency of
the algorithm was studied for different complexities of fracture networks, and a large range
of matrix fracture permeability ratios and different type of meshes. The numerical results
exhibit a very good parallel strong scalability as expected from the explicit nature of the time
integration of the transport model with a better result on tetrahedral meshes thanks to the
communication free computation of the cell unknowns. The Darcy flow solution is remarkably
robust using the Boomer AMG preconditioner on all types of fracture networks, meshes and
for all permeability ratios that have been tested. On the other hand, it requires as usual a
rather high number of unknowns per process to maintain a good parallel scalability. Future
work includes the extension of the parallel algorithm to hybrid dimensional multiphase flow
models and the use of a more accurate second order MUSCL scheme for the transport model.
Acknowledgments
This work is supported by a joint project between INRIA and BRGM Carnot institutes (ANR,
INRIA, BRGM). This work was also granted access to the HPC and visualization resources
of “Centre de Calcul Interactif” hosted by University Nice Sophia Antipolis. This work was
partially funded by ADEME BRGM research project Orbou (convention 1305C0131). We
would like also to thank Mohammad Karimi-Fard and Alexandre Lapène for kindly providing
us the mesh of Section 5.5.
References
[1] Ahmed, R., Edwards, M.G., Lamine, S., Huisman, B.A.H., Pal, M.: Control-volume distributed multi-point flux approximation coupled with a lower-dimensional fracture model.
Journal of Computational Physics 284, 462-489 (2015)
[2] Ahmed, R., Edwards, M.G., Lamine, S., Huisman, B.A.H., Pal, M.: Threedimensional control-volume distributed multi-point flux approximation coupled with a
lower-dimensional surface fracture model. Journal of Computational Physics 303, 470-497
(2015)
[3] Alboin, C., Jaffré, J., Roberts, J.E., Serres, C.: Modeling fractures as interfaces for flow
and transport in porous media. Fluid flow and transport in porous media 295, 13-24 (2002)
[4] Amir, L., Kern, M., Martin, V., Roberts, J.E.: Décomposition de domaine et
préconditionnement pour un modèle 3D en milieu poreux fracturé Proceeding of JANO 8.
8th conference on Numerical Analysis and Optimization, (2005)
[5] Angot, P., Boyer, F., Hubert, F.: Asymptotic and numerical modelling of flows in fractured
porous media. ESAIM Mathematical Modelling and Numerical Analysis 23, 239-275 (2009)
[6] Antonietti, P.F., Formaggia, L., Scotti, A., Verani, M., Verzott, N.: Mimetic finite difference approximation of flows in fractured porous media. ESAIM Mathematical Modelling
and Numerical Analysis 50, 809-832 (2016)
31
[7] Balay, S., Adams, M., Brown, J., Brune, P., Buschelman, K., Eijkhout, V., Zhang, H.:
PETSc Users Manual. Revision 3.5 (2015)
[8] Berrone, S., Pieraccini, S., Scialò, S.: An optimization approach for large scale simulations
of discrete fracture network flows. Journal of Computational Physics 256, 838-853 (2014)
[9] Bogdanov, I.I., Mourzenko, V.V., Thovert, J.-F., Adler, P.M.: Two-phase flow through
fractured porous media. Physical Review E 68, 026703 (2003)
[10] Brenner, K., Masson, R.: Convergence of a vertex centered discretization of two-phase
Darcy flows on general meshes. International Journal of Finite Volume Methods, june
(2013)
[11] Brenner, K., Groza, M., Guichard, C., Masson, R.: Vertex approximate gradient scheme
for hybrid dimensional two-phase Darcy flows in fractured porous media. ESAIM Mathematical Modelling and Numerical Analysis 49, 303-330 (2015)
[12] Brenner, K., Groza, M., Guichard, C., Lebeau, G., Masson, R.: Gradient discretization of
hybrid dimensional Darcy flows in fractured porous media. Numerische Mathematik, 1-41
(2015)
[13] Brenner, K., Hennicker, J., Masson, R., Samier, P.: Gradient discretization of hybrid
dimensional Darcy flows in fractured porous media with discontinuous pressure at matrix
fracture interfaces. IMA Journal of Numerical Analysis, published online september 28
(2016)
[14] Chen, Z., Zhang, Y.: Well flow models for various numerical methods. International Journal of Numerical Analysis and Modeling 6(3), 375-388 (2009)
[15] Dalissier, E., Guichard, C., Havé, P., Masson, R., Yang, C.: ComPASS: a tool for distributed parallel finite volume discretizations on general unstructured polyhedral meshes.
ESAIM: Proceedings 43, 147-163 (2013)
[16] D’Angelo, C., Scotti, A.: A mixed finite element method for Darcy flow in fractured porous
media with non-matching grids. ESAIM Mathematical Modelling and Numerical Analysis
46(2) 465-489 (2012)
[17] Droniou, J., Eymard, R., Gallouët, T., Herbin, R.: Gradient schemes: A generic framework
for the discretisation of linear, nonlinear and nonlocal elliptic and parabolic equations.
Mathematical Models and Methods in Applied Sciences 23(13), 2395-2432 (2013)
[18] Droniou, J., Eymard, R., Gallouët, T., Guichard, C., Herbin, R.: The Gradient discretization method: A framework for the discretization of linear and nonlinear elliptic
and parabolic problems. Preprint https://hal.archives-ouvertes.fr/hal-01382358,
(2016)
[19] Eymard, R., Guichard, C., Herbin, R.: Small-stencil 3D schemes for diffusive flows in
porous media. Mathematical Modelling and Numerical Analysis 46, 265-290 (2010)
[20] Eymard, R., Herbin, R., Guichard, C., Masson, R.: Vertex centred discretization of compositional multiphase Darcy flows on general meshes. Computational Geosciences 16(4)
987-1005 (2012)
32
[21] Eymard, R., Guichard, C., Masson, R.: High performance computing linear algorithms for
two-phase flow in porous media. FVCA 7 Proceedings, (2014)
[22] Flauraud, E., Nataf, F., Faille, I., Masson, R.: Domain Decomposition for an asymptotic
geological fault modeling, Comptes Rendus à l’académie des Sciences de Mécanique 331,
849-855 (2003)
[23] Faille, I., Fumagalli, A., Jaffré, J., Roberts, J.E.: Model reduction and discretization using
hybrid finite volumes of flow in porous media containing faults. Computational Geosciences
20, 317-339 (2016)
[24] Formaggia, L., Fumagalli, A., Scotti, A., Ruffo, P.: A reduced model for Darcy’s problem
in networks of fractures. ESAIM Mathematical Modelling and Numerical Analysis 48(4),
1089-1116 (2014)
[25] Fumagalli, A., Scotti, A.: A reduced model for flow and transport in fractured porous
media with non-matching grids. Numerical Mathematics and Advanced Applications 2011,
499-507 (2013)
[26] Haegland, H.: Streamline methods with application to flow and transport in fractured
media. PhD thesis, University of Bergen (2009)
[27] Heroux M. A., Willenbring J. M.: Trilinos Users Guide. (2003)
[28] Geiger, S., Huangfu, Q., Reid, F., Matthai, S., Coumou D., Belayneh, M., Fricke, C.,
Schmid, K.: Massively parallel sector scale discrete fracture and matrix simulation. Society
of Petroleum Engineers (2009)
[29] Hoteit, J., Firoozabadi, A.: An efficient numerical model for incompressible two-phase flow
in fracture media. Advances in Water Resources 31, 891-905, (2008)
[30] Hypre - Parallel high performance preconditioners, http://acts.nersc.gov/hypre.
[31] Karimi-Fard, M., Durlofsky, L.J., Aziz, K.: An efficient discrete-fracture model applicable
for general-purpose reservoir simulators. Society of Petroleum Engineers, (2004)
[32] Karypis, G., Kumar, V.: A fast and high quality multilevel scheme for partitioning irregular
graphs. SIAM Journal on Scientific Computing 20(1), 359-392 (1998)
[33] Martin, V., Jaffré, J., Roberts, J.E.: Modeling fractures and barriers as interfaces for flow
in porous media. SIAM Journal on Scientific Computing 26(5), 1667-1691 (2005)
[34] Matthai, S.K., Mezentsev, M.A, Belayneh, M.A: Finite element-node-centred finite-volume
two-phase-flow experiments with fractured rock represented by hybrid-element. SPE Reservoir Evaluation & Engineering 12, 740-756 (2007)
[35] Monteagudu, J., Firoozabadi, A.: Control-volume model for simulation of water injection
in fractured media: incorporating matrix heterogeneity and reservoir wettability effects.
SPE Journal 12(03), 355-366 (2007)
[36] Peaceman, D.W.: Interpretation of well-block pressures in numerical reservoir simulation.
SPE Journal 18(03), 183-94, (1978)
33
[37] Peaceman, D.W.: Interpretation of well-block pressures in numerical reservoir simulation
with nonsquare grid blocks and anisotropic permeability. SPE Journal 23(03), 531-43,
(1983)
[38] Reichenberger, V., Jakobs, H., Bastian, P., Helmig, R.: A mixed-dimensional finite volume
method for multiphase flow in fractured porous media. Advances in Water Resources 29(7),
1020-1036 (2006)
[39] Sandve, T.H., Berre, I., Nordbotten, J.M.: An efficient multi-point flux approximation
method for discrete fracture-matrix simulations. Journal of Computational Physics 231,
3784-3800 (2012)
[40] Schwenck, N., Flemisch, B., Helmig, R., Wohlmuth, B.: Dimensionally reduced flow models
in fractured porous media: crossings and boundaries. Computational Geosciences 19, 12191230 (2015)
[41] Si, H.: http://tetgen.org
[42] Tunc, X., Faille, I., Gallouët, T., Cacas, M.C., Havé, P.: A model for conductive faults
with non matching grids. Computational Geosciences 16, 277-296 (2012)
34
| 5cs.CE
|
Regularization for Unsupervised Deep Neural Nets
Baiyang Wang, Diego Klabjan
arXiv:1608.04426v4 [cs.LG] 17 Feb 2017
Department of Industrial Engineering and Management Sciences,
Northwestern University, 2145 Sheridan Road, C210
Evanston, Illinois 60208
Abstract
Unsupervised neural networks, such as restricted Boltzmann
machines (RBMs) and deep belief networks (DBNs), are
powerful tools for feature selection and pattern recognition
tasks. We demonstrate that overfitting occurs in such models just as in deep feedforward neural networks, and discuss
possible regularization methods to reduce overfitting. We also
propose a “partial” approach to improve the efficiency of
Dropout/DropConnect in this scenario, and discuss the theoretical justification of these methods from model convergence
and likelihood bounds. Finally, we compare the performance
of these methods based on their likelihood and classification
error rates for various pattern recognition data sets.
1
Introduction
Unsupervised neural networks assume unlabeled data to be
generated from a neural network structure, and have been
applied extensively to pattern analysis and recognition. The
most basic one is the restricted Boltzmann machine (RBM)
(Salakhutdinov, Mnih, and Hinton 2007), an energy-based
model with a layer of hidden nodes and a layer of visible
nodes. With such a basic structure, we can stack multiple
layers of RBMs to create an unsupervised deep neural network structure, such as the deep belief network (DBN) and
the deep Boltzmann machine (DBM) (Hinton, Osindero, and
Teh 2006; Salakhutdinov and Hinton 2009a). These models
can be calibrated with a combination of the stochastic gradient descent and the contrastive divergence (CD) algorithm or
the PCD algorithm (Salakhutdinov, Mnih, and Hinton 2007;
Tieleman 2008). Once we learn the parameters of a model,
we can retrieve the values of the hidden nodes from the visible nodes, thus applying unsupervised neural networks for
feature selection. Alternatively, we may consider applying
the parameters obtained from an unsupervised deep neural network to initialize a deep feedforward neural network
(FFNN), thus improving supervised learning.
One essential question for such models is to adjust for the
high-dimensionality of their parameters and avoid overfitting. In FFNNs, the simplest regularization is arguably the
early stopping method, which stops the gradient descent algorithm before the validation error rate goes up. The weight
Copyright c 2017, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
decay method, or Ls regularization, is also commonly used
(Witten, Frank, and Hall 2011). Recently Dropout is proposed, which optimizes the parameters over an average of
exponentially many models with a subset of all nodes (Srivastava et al. 2014). It has been shown to outperform weight
decay regularization in many situations.
For regularizing unsupervised neural networks, sparseRBM-type models encourage a smaller proportion of 1valued hidden nodes (Cho, Ilin, and Taiko 2012; Lee,
Ekanadham, and Ng 2007). DBNs are regularized in Goh et
al. (2013) with outcome labels. While these works tend to be
goal-specific, we consider regularization for unsupervised
neural networks in a more general setting. Our work and
contributions are as follows: (1) we extend common regularization methods to unsupervised deep neural networks, and
explain their underlying mechanisms; (2) we propose partial
Dropout/DropConnect which can improve the performance
of Dropout/DropConnect; (3) we compare the performance
of different regularization methods on real data sets, thus
providing suggestions on regularizing unsupervised neural
networks. We note that this is the very first study illustrating
the mechanisms of various regularization methods for unsupervised neural nets with model convergence and likelihood bounds, including the effective newly proposed partial
Dropout/DropConnect.
Section 2 reviews recent works for regularizing neural
networks, and Section 3 exhibits RBM regularization as a
basis for regularizing deeper networks. Section 4 discusses
the model convergence of each regularization method. Section 5 extends regularization to unsupervised deep neural
nets. Section 6 presents a numerical comparison of different regularization methods on RBM, DBN, DBM, RSM
(Salakhutdinov and Hinton 2009b) and Gaussian RBM
(Salakhutdinov, Mnih, and Hinton 2007). Section 7 discusses potential future research and concludes the paper.
2
Related Works
To begin with, we consider a simple FFNN with a single
layer of input ı = (ı1 , . . . , ıI )T and a single layer of output
o = (o1 , . . . , oJ )T ∈ {0, 1}J . The weight matrix W is of
size J × I. We assume the relation
E(o) = a(W · ı),
(1)
where a(·) is the activation function, such as the sigmoid
function σ(x) = 1/(1 + e−x ) applied element-wise. Equation (1) has the modified form in Srivastava et al. (2014),
(
E(o|m) = a(m ? (W · ı)),
(2)
iid
m = (m1 , . . . , mJ )T ∼ Ber(p),
where ? denotes element-wise multiplication, and Ber(·)
denotes the Bernoulli distribution, thereby achieving the
Dropout (DO) regularization for neural networks. In
Dropout, we minimize the objective function
−lDO (W ) = −
N
X
Em [log p(o(n) |ı(n) , W, m)],
(3)
n=1
which can be achieved by a stochastic gradient descent algorithm, sampling a different mask m per data example
(o(n) , ı(n) ) and per iteration. We observe that this can be
readily extended to deep FFNNs. Dropout regularizes neural networks because it incorporates prediction based on any
subset of all the nodes, therefore penalizing the likelihood.
A theoretical explanation is provided in Wager, Wang, and
Liang (2013) for Dropout, noting that it can be viewed as
feature noising for GLMs, and we have the relation
N
X
.
log p(o(n) |ı(n) , W ) + Rq (W ).
−lDO (W ) = −
(4)
n=1
p PN
Here J = 1 for simplicity, and Rq (W ) = 12 1−p
n=1
PI
(n) 2
00
(n)
2
A
(W
ı
)(ı
)
W
,
where
A(·)
is
the
log-partition
i
i
i=1
function of a GLM. Therefore, Dropout can be viewed approximately as the adaptive L2 regularization (Baldi and
Sadowski 2013; Wager, Wang, and Liang 2013). A recursive approximation of Dropout is provided in Baldi and Sadowski (2013) using normalized weighted geometric means
to study its averaging properties.
An intuitive extension of Dropout is DropConnect (DC)
(Wan et al. 2013), which has the form below
(
E(o|m) = a((m ? W ) · ı),
(5)
iid
m = (mij )J×I ∼ Ber(p),
and thus masks the weights rather than the nodes. The objective lDC (W ) has the same form as in (3). There are a
number of related model averaging regularization methods,
each of which averages over subsets of the original model.
For instance, Standout varies Dropout probabilities for different nodes which constitute a binary belief network (Ba
and Frey 2013). Shakeout adds additional noise to Dropout
so that it approximates elastic-net regularization (Kang, Li,
and Tao 2016). Fast Dropout accelerates Dropout with Gaussian approximation (Wang and Manning 2013). Variational
Dropout applies variational Bayes to infer the Dropout function (Kingma, Salimans, and Welling 2015).
We note that while Dropout has been discussed for RBMs
(Srivastava et al. 2014), to the best of our knowledge, there
is no literature extending common regularization methods to
RBMs and unsupervised deep neural networks; for instance,
adaptive Ls regularization and DropConnect as mentioned.
Therefore, below we discuss their implementations and examine their empirical performance. In addition to studying
model convergence and likelihood bounds, we propose partial Dropout/DropConnect which iteratively drops a subset
of nodes or edges based on a given calibrated model, therefore improving robustness in many situations.
3
RBM Regularization
For a Restricted Boltzmann machine, we assume that v =
(v1 , · · · , vJ )T ∈ {0, 1}J denotes the visible vector, and h =
(h1 , · · · , hI )T ∈ {0, 1}I denotes the hidden vector. Each
vj , j = 1, . . . , J is a visible node and each hi , i = 1, . . . , I
is a hidden node. The joint probability is
(
P
P (v, h) = e−E(v,h) / ν,η e−E(ν,η) ,
(6)
E(v, h) = −bT v − cT h − hT W v.
We let the parameters ϑ = (b, c, W ) ∈ Θ, which is a vector
containing all components of b, c, and W . To calibrate the
PN
model is to find θ̂ = arg max n=1 log P (v (n) |ϑ).
ϑ∈Θ
An RBM is a neural network because we have the following conditional probabilities
P (hi = 1|v) = σ(ci + Wi· v),
(7)
P (vj = 1|h) = σ(bj + W·jT h),
where Wi· and W·j represent, respectively, the i-th row and
j-th column of W . The gradient descent algorithm is applied
to calibration. The gradient of the log-likelihood can be expressed in the following form
−
∂ log P (v (n) )
∂F(v (n) )
=
−
∂ϑ
∂ϑ
X
v∈{0,1}J
P (v)
∂F(v)
, (8)
∂ϑ
PI
where F(v) = −bT v − i=1 log(1 + eci +Wi· v ) is the free
energy. The right-hand side of (8) is approximated by contrastive divergence with k steps of Gibbs sampling (CD-k)
(Salakhutdinov, Mnih, and Hinton 2007).
3.1
Weight Decay Regularization
Weight decay, or Ls regularization, adds the term λkW kss
to the negative log-likelihood of an RBM. The most commonly used is L2 (ridge regression), or L1 (LASSO). In all
situations, we do not regularize biases for simplicity.
Here we consider a more general form. Suppose we have
a trained set of weights W from CD with no regularization. Instead of adding the term λkW kss , we add the term
µ P
s
s
i,j |Wij | /|Ŵij | to the negative log-likelihood. ApIJ
parently this adjusts for the different scales of the components of W . We refer to this approach as adaptive Ls . We
note that adaptive L1 is the adaptive LASSO (Zou 2006),
and adaptive L2 plus L1 is the elastic-net (Zou and Hastie
2005). We consider the performance of L2 regularization
plus adaptive L1 regularization (L2 + AL1 ) below.
3.2
ind
Model Averaging Regularization
As discussed in Srivastava et al. (2014), to characterize a
Dropout (DO) RBM, we simply need to apply the following
conditional distributions
PDO (hi = 1|v, m) = mi · σ(ci + Wi· v),
(9)
PDO (vj = 1|h, m) = σ(bj + W·jT h).
Therefore, given a fixed mask m ∈ {0, 1}I , we actually
obtain an RBM with all visible nodes v and hidden nodes
{hi : mi = 1}. Hidden nodes {hi : mi = 0} are fixed
to zero so they have no influence on the conditional RBM.
Apart from replacing (7) with (9), the only other change
needed is to replace F(v) with FDO (v|m) = −bT v −
PI
ci +Wi· v
). In terms of training, we sugi=1 mi log(1 + e
gest sampling a different mask per data example v (n) and
per iteration as in Srivastava et al. (2014).
A DropConnect (DC) RBM is closely related; given a
mask m = {0, 1}IJ on weights W , W in a plain RBM is
replaced by m ∗ W everywhere. We suggest sampling a different mask m per mini-batch since it is usually much larger
than a mask in a Dropout RBM.
3.3
Network Pruning Regularization
There are typically many nodes or weights which are of little importance in a neural network. In network pruning, such
unimportant nodes or weights are discarded, and the neural
network is retrained. This process can be conducted iteratively (Reed 1993). Now we consider two variants of network pruning for RBMs. For an trained set of weights Ŵ
with no regularization, we consider implementing a fixed
mask m = (mij )I×J where
mij = 1|Ŵij |≥Q , Q = Q100(1−p)% (|Ŵ |),
(10)
i.e. Q is the 100(1 − p)%-th left percentile of all |Ŵij |,
and p ∈ (0, 1) is some fixed proportion of retained weights.
We then recalibrate the weights and biases fixing mask m,
leading to a simple network pruning (SNP) procedure which
deletes 100(1 − p)% of all weights. We may also consider
deleting 100(1−p)/r% of all weights at a time, and conduct
the above process r times, leading to an iterative network
pruning (INP) procedure.
3.4
Hybrid Regularization
We may consider combining some of the above approaches.
For instance, Srivastava et al. (2014) considered a combination of Ls and Dropout. We introduce two new hybrid
approaches, namely partial DropConnect (PDC) presented
in Algorithm 1 and partial Dropout (PDO), which generalizes DropConnect and Dropout, and borrows from network
pruning. The rationale comes from some of the model convergence results exhibited later.
As before, suppose we have a trained set of weights Ŵ
with no regularization. Instead of implementing a fixed mask
m, we perform DropConnect regularization with different
retaining probabilities pij for each weight Wij . We let the
quantile Q = Q100(1−q)% (|Ŵ |), and
pij = 1|Ŵij |≥Q + p0 · 1|Ŵij |<Q .
(11)
Therefore, we sample a different m = (mij )I×J ∼
Ber(pij ) per mini-batch, which means that we always keep
100q% of all the weights, and randomly drop the remaining
weights with probability 100(1 − p0 )%. The mask m can be
resampled iteratively. Intuitively, we are trying to maximize
the following
max
ϑ∈Θ,pij ∈{p0 ,1}
ind
Em [log P (data|ϑ, m)].
such that mij ∼ Ber(pij ), and
P
(12)
1pij =1 = qIJ.
Algorithm 1. (Partial DropConnect)
1. Initialize θ̂p = θ̂, the unregularized trained parameters for an RBM.
2. Find retaining rates p = (pij )I×J from (11).
3. Retrain weights ϑ with DropConnect for a given
number of iterations, and then update θ̂p .
4. If maximum number of iterations reached, stop and
obtain θ̂p ; otherwise, go back to Step 2.
This technique is proposed because we hypothesize that
some weights could be more important than others a posteriori, so dropping them could cause much variation among
the models being averaged. From (11), in partial Dropout,
we tend to drop weights which have smaller magnitude,
since setting larger weights to zero may substantially alter
the structure of a neural network. Experiments on real data
show that this technique can effectively improve the performance of plain DropConnect.
We denote lp (ϑ) = Em∼Ber(p) [log P (data|ϑ, m)], and
l(ϑ) = log P (data|ϑ). From first-order Taylor’s expansion,
|lp (θ̂p ) − l(θ̂)| ≤ |lp (θ̂p ) − l(θ̂p )| + |l(θ̂p ) − l(θ̂)|
∂
l(θ̃)(1 − pij )Ŵij,p + |l(θ̂p ) − l(θ̂)|
∂Wij
i,j
X
≤K
(1 − pij )|Ŵij,p | + |l(θ̂p ) − l(θ̂)|.
(13)
=
X
Here θ̃ lies between θ̂p and m ∗ θ̂p from Taylor’s expansion,
∂
l(ϑ)k∞ is a Lipschitz constant.
and K = supϑ∈Θ k ∂ϑ
Note that given p and θ̂p , Step 2 in Algorithm 1 lowP
ers the term (1 − pij )|Ŵij,p | by assigning (1 − p0 ) to
weights of smaller magnitude, reducing an upper bound of
|lp (θ̂p ) − l(θ̂)|. Step 3 further increases lp (θ̂p ) and reduces
the gap |lp (θ̂p ) − l(θ̂)|. Therefore, each iteration of Algorithm 1 tends to increase lp (θ̂p ), and hence Algorithm 1 provides an intuitive solution to problem (12).
We also consider a partial Dropout approach which is
analogous to partial DropConnect and keeps some important nodes rather than weights. We set a mask for nodes
ind
m = (m1 , . . . , mI ), mi ∼ Ber(pi ), where
(
pi = 1kŴi· k≥Q + p0 · 1kŴi· k<Q ,
(14)
Q = Q100(1−q)% (kŴi· k).
This algorithm protects more important hidden nodes from
being dropped in order to reduce variation. We also evaluate
its empirical performance later.
4
also based on Proposition 3, i.e. we do not drop the more important components of θ, therefore possibly reducing variation caused by dropping influential weights. Partial Dropout
follows from the same reasoning.
More Theoretical Considerations
Here we discuss the model convergence properties of different regularization methods when the number of data examples N → ∞. We mark all regularization coefficients
and parameter estimates with (N ) when there are N data
examples. We assume ϑ = (b, c, W ) ∈ Θ, which is compact, dim(Θ) = D, P (v|ϑ) is unique for each ϑ ∈ Θ,
and v (1) , . . . , v (N ) are i.i.d. generated from an RBM with a
“true” set of parameters θ. We denote each regularized calibrated set of parameters as θ̃(N ) .
Let A = {d : θd 6= 0} and θA = {θd : d ∈ A}. (Zou
2006) showed that AL1 guarantees asymptotic normality
and identification of set A for linear regression. We demonstrate that similar results hold for L2 + AL1 for RBMs. We
(N )
(N )
(N )
(N )
let λ(N ) = (λ1 , . . . , λD ) and µ(N ) = (µ1 , . . . , µD )
2
1
be the L and L regularization coefficients for each component. The proofs of all propositions and corollaries below
are in the supplementary material (Wang and Klabjan 2016).
Proposition 1. (a) If λ(N ) /N → 0, µ(N ) /N → 0 as N →
P
(N ) √
∞, then the estimate θ̃(N ) → θ; (b) if also, µd / N →
√
√ (N )
d
0·1θd 6=0 +∞·1θd =0 , λ(N ) / N → 0, then n(θ̃A −θA ) →
N (0, I −1 (θA )), where I is the Fisher information matrix;
(N )
P (Â(N ) = A) → 1, where Â(N ) = {d : θ̃d 6= 0}.
For Dropout and DropConnect RBMs, we also assume
that the data is generated from a plain RBM structure. We
assume p(N ) is of size I × J as in (11) for DropConnect
and of length I as in (14) for Dropout, therefore covering
the cases of both original and partial Dropout/DropConnect
with a fixed set of dropping rates. With a decreasing dropping rate 1 − p(N ) → 0 with N → ∞, we obtain the following convergence result.
P
Proposition 2. If p(N ) → 1 as N → ∞, then θ̃(N ) → θ.
For network pruning, we show that as the number of data
examples increase, if the retained proportion of parameters
p(N ) = p can cover all nonzero components of θ, we will
not miss any important component.
Proposition 3. Assume p > p0 := |A|/D. Then for simP
ple network pruning, as N → ∞, (a) θ̃(N ) → θ; (b) for
sufficiently large N , there exists ρ > 0 such that P (A ∈
Â(N ) ) ≥ 1 − e−ρN .
Corollary 1. The above results also hold for iterative network pruning.
We note that for all regularization methods, under the
above conditions, the calibrated weights converge to the
“true” set of parameters θ, which indicates consistency.
Also, adding L1 regularization guarantees that we can identify components of zero value with infinitely many examples. The major benefits of Dropout come from the facts
that it makes L2 regularization adaptive, and also encourages more confident prediction of the outcomes (Wager,
Wang, and Liang 2013). We propose partial DropConnect
5
5.1
Extension to Other Networks
Deep Belief Networks
We consider the multilayer network below,
P (v, h1 , . . . , hL ) P (v|h1 )
L−1
Y
P (hl |hl+1 )P (hL ),
(15)
l=1
where each probability on the right-hand side is from
an RBM. To train the weights of RBM(v, h1 ), . . .,
RBM(hL−1 , hL ), we only need to carry out a greedy layerwise training approach, i.e. we first train the weights of
RBM(v, h1 ), and then use E(h1 |v) to train RBM(h1 , h2 ),
etc. The weights of the RBMs are used to initialize a deep
FFNN which is finetuned with gradient descent. RBM regularization is applicable to each layer of a DBN.
Here we show that adding layers to a Dropout/DropConnect DBN improves the likelihood given symmetry of
the weights of two adjacent layers. Similar results for plain
DBN are in Hinton, Osindero, and Teh (2006) and Bengio
(2007). We demonstrate this by using likelihood bounds.
We let DBNL denote an L-layer DBN and DBNL+1 denote an (L + 1)-layer DBN with the first L layers being the
same as in DBNL . For a data example of a visible vector v,
the log-likelihood is bounded as follows,
Em [log PDBNL+1 (v|m, m∗ )]
X
≥ Em [HPDBNL (hL |v,m) ] +
Em,m∗ {PDBNL (hL |v, m)
hL
L
∗
· [log PRBML+1 (h |m ) + log PDBNL (v|hL , m)]}.
(16)
Here, H is the entropy function, and the derivation is
analogous to Section 11 in Bengio (2007). Mask m is for
DBNL , and mask m∗ is for the new (L + 1)-th layer. Note
that after we have trained the first L layers, and initialized
the (L + 1)-th layer symmetric to the L-th layer, assuming a
constant dropping probability, we have
Em∗ [log PRBML+1 (hL |m∗ )] = Em [log PDBNL (hL |m)],
(17)
so DBNL+1 has the same log-likelihood bound as DBNL .
Training RBML+1 , Em∗ [log PRBML+1 (hL |m∗ )] is guaranteed to increase, and therefore the likelihood of DBNL+1 is
expected to improve. As a result, for regularized unsupervised deep neural nets, adding layers also tend to elevate
the explanatory power of the network. Adding nodes has the
same effect, providing a rationale for deep and large-scale
networks. We present the following proposition.
Proposition 4. Adding nodes or layers (preserving weight
symmetry) to a Dropout/DropConnect DBN continually improves the likelihood; also, adding layers of size J ≤ H 1 ≤
H 2 ≤ · · · continually improves the likelihood.
5.2
Other RBM Variants
More descriptions of DBMs, RSMs, and Gaussian RBMs
are in the supplementary material (Wang and Klabjan 2016).
RBM regularization can be extended to all these situations.
6
Data Studies
In this section, we compare the empirical performance of
the aforementioned regularization methods on the following data sets: MNIST, NORB (image recognition); 20 Newsgroups, Reuters21578 (text classification); ISOLET (speech
recognition). All results are obtained using GeForce GTX
TITAN X in Theano.
6.1
Experiment Settings
We consider the following unsupervised neural network
structures: DBN/DBM for MNIST; DBN for NORB;
RSM plus logistic regression for 20 Newsgroups and
Reuters21578; GRBM for ISOLET. CD-1 is performed for
the rest of the paper. The following regularization methods are considered: None (no regularization); DO; DC; L2 ;
L2 + AL1 ; SNP; INP(r = 3); PDO; PDC. The number of
pretraining epochs is 100 per layer and the number of finetuning epochs is 300, with a finetuning learning rate of 0.1.
For L2 + AL1 , SNP, and INP which need re-calibration, we
cut the 100 epochs into two halves (4 quarters for INP). For
regularization parameters, we apply the following ranges:
p = 0.8 ∼ 0.9 for DO/DC/SNP/INP; λ = 10−5 ∼ 10−4 for
L2 , similar to Hinton (2010); µ = 0.01 ∼ 0.1 for L2 + AL1 ;
p0 = 0.5, q = 0.7 ∼ 0.9 or the reverse for PDO/PDC.
We only make one update to the “partial” dropping rates to
maintain simplicity. From the results, we note that unsupervised neural networks tend to need less regularization than
FFNNs. We choose the best iteration and regularization parameters over a fixed set of parameter values according to
the validation error rates.
6.2
The MNIST Data Set
The MNIST data set consists of 282 pixels of handwritten
0-9 digits. There are 50,000 training examples, 10,000 validation and 10,000 testing examples. We first consider the
likelihood of the testing data of an RBM with 500 nodes
for MNIST. There are two model fitting evaluation criteria:
pseudo-likelihood and AIS-likelihood (Salakhutdinov and
Murray 2008). The former is a sum of conditional likelihoods, while the latter directly estimates P (v) with AIS.
In Figure 1 below using log-scale, p = 0.9 for DO,
and λ = 10−4 for L2 . These figures tend to be representative of the model fitting process. The pseudo-likelihood
is a more optimistic estimate of the model fitting. We observe that Dropout outperforms the other two after about 50
epochs, and L2 regularization does not improve the pseudolikelihood. In terms of the AIS-likelihood, which is a much
more conservative estimate of the model fitting, the fitting
process seems to have three stages: (1) initial fitting; (2)
“overfitting”; (3) re-fitting. We observe that L2 improves the
likelihood significantly, while Dropout catches up at about
300 epochs. Therefore, Dropout tends to improve model fitting according to both likelihood criteria.
Figure 1: Left: Pseudo-likelihood of the RBM over 500 pretraining epochs. Right: AIS-likelihood of the RBM over 500
pretraining epochs.
In Figure 2, we can observe that more nodes increase the
pseudo-likelihood, which is consistent with Proposition 4,
but exhibit “overfitting” for the AIS-likelihood. However,
such “overfitting” does not exist for pretraining purposes as
well. Thus we suggest the pseudo-likelihood, and the AISlikelihood should be viewed as too conservative.
Figure 2: Left: Pseudo-likelihood of the Dropout RBM
over 500 pretraining epochs. Right: AIS-likelihood of the
Dropout RBM over 500 pretraining epochs.
Classification error rates tend to be a more practical measure. We first consider a 3-hidden-layer DBN with 1,000
nodes per layer, pretraining learning rate 0.01, and batch
size 10; see Table 1. We tried DBNs of 1, 2, and 4 hidden
layers and found the aforementioned structure to perform
best with None as baseline. The same was done for all other
structures. We calculate the means of the classification errors for each regularization method averaged over 5 random
replicates and their standard deviations. In each table, we
stress in bold the top 3 performers with ties broken by deviation. We note that most of the regularization methods tend
to improve the classification error rates, with DC and PDO
yielding slightly higher error rates than no regularization.
m.
sd.
m.
sd.
None
1.35%
0.02%
SNP
1.30%
0.03%
DO
1.30%
0.04%
INP
1.32%
0.03%
DC
1.37%
0.02%
PDO
1.36%
0.04%
L2
1.31%
0.03%
PDC
1.30%
0.03%
L2 + AL1
1.35%
0.01%
Table 1: Classification errors for a 3-layer DBN for the
MNIST data set.
In Table 2, we consider a 3-hidden-layer DBM with 1,000
nodes per layer. For simplicity, we only classify based on the
original features. We let the pretraining learning rate be 0.03
and the batch size be 10.
m.
sd.
m.
sd.
None
1.22%
0.02%
SNP
1.18%
0.03%
DO
1.21%
0.02%
INP
1.26%
0.02%
DC
1.20%
0.02%
PDO
1.21%
0.03%
L2
1.14%
0.02%
PDC
1.12%
0.02%
L2 + AL1
1.15%
0.04%
Table 2: Classification errors for a 3-layer DBM for the
MNIST data set.
It can be observed that regularization tends to yield more
improvement for DBM than DBN, possibly because a DBM
doubles both the visible layer and the third hidden layer,
resulting in a “larger” neural network structure in general.
Only INP proves to be unsuitable for the DBM; all other regularization methods work better, with PDC being the best.
6.3
The NORB Data Set
The NORB data set has 5 categories of images of 3D objects.
There are 24,300 training examples, with 2,300 validation
examples held out, and 24,300 testing examples. We follow
preprocessing of Nair and Hinton (2009), and apply a sparse
two-hidden-layer DBN with 4,000 nodes per layer as in Lee,
Ekanadham, and Ng (2007) with a sparsity regularization
coefficient of 10.0 and the first hidden layer being a Gaussian RBM. The pretraining learning rates are 0.001 and 0.01
for the first and second hidden layer, and the batch sizes for
pretraining and finetuning are 100 and 20. Because the validation error often goes to zero, we choose the 300-th epoch
and fix the regularization parameters as follows based on the
best values of other data sets: p = 0.9 for DO/DC/SNP/INP,
λ = 10−4 for L2 , µ = 0.1 for L2 +AL1 , (p0 , q) = (0.5, 0.8)
for PDO and (p0 , q) = (0.8, 0.5) for PDC. In Table 3, only
weight decay and PDO/PDC perform better than None, with
PDC again being the best.
batch size is 50. We apply logistic regression to classify the
trained features, i.e. hidden values of the RSM, as in Srivastava, Salakhutdinov, and Hinton (2013). This setting is
quite challenging for unsupervised neural networks. In Table
4, Dropout performs best with other regularization methods
yielding improvements except DropConnect.
m.
sd.
m.
sd.
m.
sd.
m.
sd.
Table 3: Classification errors for the NORB data set.
6.4
DO
28.8%
0.23%
INP
29.7%
0.48%
DC
35.2%
0.91%
PDO
30.1%
0.71%
L2
30.1%
0.30%
PDC
29.7%
0.34%
L2 + AL1
30.1%
0.65%
Table 4: Classification errors for the trained features of RSM
for the 20 Newsgroups data set.
6.5
The Reuters21578 Data Set
The Reuters21578 data set is a collection of newswire articles. We adopt the stemmed R-52 version which has 52
categories, 6,532 training examples, from which 1,032 validation examples are randomly held out, and 2,568 testing
examples. We retain the most common 2,000 words, and
train an RSM with 500 hidden nodes in a single layer. The
pretraining learning rate is 0.1 and the batch size is 50. We
make the learning rate large because the cost function is
quite bumpy. From Table 5, we note that PDC works best,
and PDO improves the performance of Dropout.
None
m. 10.50%
sd. 0.64%
SNP
m. 9.99%
sd. 0.27%
DO
11.91%
0.70%
INP
10.10%
0.30%
DC
10.10%
0.32%
PDO
9.98%
0.24%
L2
L2 + AL1
10.06%
9.99%
0.28%
0.41%
PDC
9.84%
0.23%
Table 5: Classification errors for the trained features from
RSM and the Reuters21578 data set.
6.6
None
DO
DC
L2
L2 + AL1
11.00% 11.15% 11.19% 10.93% 10.91%
0.15% 0.12% 0.10% 0.18%
0.17%
SNP
INP
PDO
PDC
11.04% 11.14% 10.95% 10.81%
0.18% 0.20% 0.15% 0.13%
None
30.8%
0.70%
SNP
29.7%
0.26%
The ISOLET Data Set
The ISOLET data set consists of voice recordings of the
Latin alphabet (a-z). There are 6,138 training examples,
from which 638 validation examples are randomly held out,
and 1,559 testing examples. We train a 1,000-hidden-node
Gaussian RBM with pretraining learning rate 0.005, batch
size 20, and initialize a FFNN, which can be viewed as a
single-hidden-layer DBN. From Table 6, it is evident that
all regularization methods work better then None, with PDC
again being the best.
The 20 Newsgroups Data Set
The 20 Newsgroups data set is a collection of news documents with 20 categories. There are 11,293 training examples, from which 6,293 validation examples are randomly
held out, and 7,528 testing examples. We adopt the stemmed
version, retain the most common 5,000 words, and train an
RSM with 1,000 hidden nodes in a single layer. We consider this as a simple case of deep learning since it is a twostep procedure. The pretraining learning rate is 0.02 and the
m.
sd.
m.
sd.
None
3.98%
0.09%
SNP
3.86%
0.07%
DO
3.87%
0.06%
INP
3.86%
0.10%
DC
3.88%
0.11%
PDO
3.96%
0.08%
L2
3.83%
0.10%
PDC
3.78%
0.05%
L2 + AL1
3.86%
0.07%
Table 6: Classification errors for the ISOLET data set.
6.7
Summary
From the above results, we observe that regularization does
improve the structure of unsupervised deep neural networks
and yields lower classification error rates for each data set
studied herein. The most robust methods which yield improvements for all six instances are L2 , L2 +AL1 , and PDC.
SNP is also acceptable, and preferable over INP. PDO can
yield improvements for Dropout when Dropout is unsuitable
for the network structure. PDC turns out to be the most stable method of all, and thus the recommended choice.
7
Conclusion
Regularization for deep learning has aroused much interest,
and in this paper, we extend regularization to unsupervised
deep learning, i.e. for DBNs and DBMs. We proposed several approaches, demonstrated their performance, and empirically compared the different techniques. For the future,
we suggest that it would be of interest to consider more variants of model averaging regularization for supervised deep
learning as well as novel methods of unsupervised learning;
for instance, Kingma and Welling (2015) provided an interesting variational Bayesian auto-encoder approach.
References
Ba, L., and Frey, B. 2013. Adaptive dropout for training
deep neural networks. In Advances in Neural Information
Processing Systems 26. MIT Press.
Baldi, P., and Sadowski, P. 2013. Understanding dropout.
In Advances in Neural Information Processing Systems 26.
MIT Press.
Bengio, Y. 2007. Learning deep architectures for ai.
https://www.iro.umontreal.ca/˜lisa/pointeurs/TR1312.pdf.
Cho, K.; Ilin, A.; and Taiko, T. 2012. Tikhonov-type regularization for restricted boltzmann machines. In 22nd International Conference on Artificial Neural Networks.
Goh, H.; Thome, N.; Cord, M.; and Lim, J. 2013. Topdown regularization of deep belief networks. In Advances in
Neural Information Processing Systems 26. MIT Press.
Hinton, G.; Osindero, S.; and Teh, Y. 2006. A fast learning algorithm for deep belief nets. Neural Computation
18:1527–1554.
Hinton, G. 2010. A practical guide to training restricted
boltzmann machines. https://www.cs.toronto.edu/˜hinton/
absps/guideTR.pdf.
Kang, G.; Li, J.; and Tao, D. 2016. Shakeout: A new regularized deep neural network training scheme. In 30th AAAI
Conference on Artificial Intelligence.
Kingma, D., and Welling, M. 2014. Auto-encoding variational bayes. In International Conference on Learning Representations.
Kingma, D.; Salimans, T.; and Welling, M. 2015. Variational
dropout and the local reparameterization trick. In Advances
in Neural Information Processing Systems 28.
Lee, H.; Ekanadham, C.; and Ng, A. 2007. Sparse deep
belief net model for visual area v2. In Advances in Neural
Information Processing Systems 20. MIT Press.
Nair, V., and Hinton, G. 2009. 3d object recognition with
deep belief nets. In Advances in Neural Information Processing Systems 22. MIT Press.
Reed, R. 1993. Pruning algorithms: a survey. IEEE Transactions on Neural Networks 4:740–747.
Salakhutdinov, R., and Hinton, G. 2009a. Deep boltzmann
machines. In 12th International Conference on Artificial Intelligence and Statistics.
Salakhutdinov, R., and Hinton, G. 2009b. Replicated softmax: an undirected topic model. In Advances in Neural Information Processing Systems 22. MIT Press.
Salakhutdinov, R., and Murray, I. 2008. On the quantitative analysis of deep belief networks. In 25th International
Conference on Machine Learning.
Salakhutdinov, R.; Mnih, A.; and Hinton, G. 2007. Restricted boltzmann machines for collaborative filtering. In
24th International Conference on Machine Learning.
Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; and
Salakhutdinov, R. 2014. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine
Learning Research 14:1929–1958.
Srivastava, N.; Salakhutdinov, R.; and Hinton, G. 2013.
Modeling documents with a deep boltzmann machine. In
29th Conference on Uncertainty in Artificial Intelligence.
Tieleman, T. 2008. Training restricted boltzmann machines
using approximations to the likelihood gradient. In 25th International Conference on Machine Learning.
Wager, S.; Wang, S.; and Liang, P. 2013. Dropout training as
adaptive regularization. In Advances in Neural Information
Processing Systems 26. MIT Press.
Wan, L.; Zeiler, M.; Zhang, S.; LeCun, Y.; and Fergus, R.
2013. Regularization of neural networks using dropconnect.
In 30th International Conference on Machine Learning.
Wang, B., and Klabjan, D. 2016. Supplementary material for “regularization for unsupervised deep neural nets”.
http://www.dynresmanagement.com/publications.html.
Wang, S., and Manning, C. 2013. Fast dropout training. In
30th International Conference on Machine Learning.
Witten, I.; Frank, E.; and Hall, M. 2011. Data Mining: Practical Machine Learning Tools and Techniques. Burlington,
Massachusetts, USA: Morgan Kaufmann Publishers.
Zou, H., and Hastie, T. 2005. Regularization and variable
selection via the elastic net. Journal of the Royal Statistical
Society: Series B (Statistical Methodology) 67:301320.
Zou, H. 2006. The adaptive lasso and its oracle properties.
Journal of the American Statistical Association 101:1418–
1429.
| 9cs.NE
|
Convergence of the randomized Kaczmarz method for phase
retrieval
arXiv:1706.10291v2 [math.NA] 24 Jul 2017
Halyun Jeong∗ and C. Sinan Güntürk†
Courant Institute, NYU
June 30, 2017; revised July 16, 2017
Abstract
The classical Kaczmarz iteration and its randomized variants are popular tools for fast
inversion of linear overdetermined systems. This method extends naturally to the setting of the
phase retrieval problem via substituting at each iteration the phase of any measurement of the
available approximate solution for the unknown phase of the measurement of the true solution.
Despite the simplicity of the method, rigorous convergence guarantees that are available for the
classical linear setting have not been established so far for the phase retrieval setting. In this
short note, we provide a convergence result for the randomized Kaczmarz method for phase
retrieval in Rd . We show that with high probability a random measurement system of size
m d will be admissible for this method in the sense that convergence in the mean square sense
is guaranteed with any prescribed probability. The convergence is exponential and comparable
to the linear setting.
1
Introduction
The classical Kaczmarz iteration is a popular and convenient method for the recovery of any real
or complex d-dimensional vector x from a collection of sufficient linear measurements yi := x · φi ,
i = 1, . . . , m, where u · v denotes the Euclidean inner product of u and v. Starting with any initial
point x0 , the algorithm produces a succession of iterates (xk )∞
0 defined by
xk+1 = xk + (yt − xk · φt )
φt
,
kφt k2
(1)
where t := t(k) ∈ {1, ..., m} is the index of the selected vector (and the corresponding measurement)
at time k. This equation has a simple interpretation: xk+1 is the orthogonal projection of xk on
the solution hyperplane {u : u · φt = yt }. In other words, the update xk+1 − xk is the orthogonal
projection of the error x − xk on the chosen direction φt . Kaczmarz’s original scheme cycles
through the indices periodically, but it has been shown that random selection generally yields
faster convergence. For this and other results, see [11, 8, 7, 2].
∗
†
email: [email protected]
email: [email protected]
1
This method can be adapted in a straightforward manner to the phase retrieval problem where
we only have access to the intensities {|yi |}m
i=1 : By simply using the sign (phase) of the approximate
measurement xk · φt in place of that of yt , we get the phase-adapting Kaczmarz iteration
xk+1 = xk + (σ(xk · φt )|yt | − xk · φt )
φt
,
kφt k2
(2)
where σ(w) is the sign (or phase) of the scalar w, defined by the relation w = σ(w)|w|. We will
assume the convention that σ(0) = 1. This method has been proposed by various authors (e.g.
[14, 6]) and has been observed to perform well in practice. For general theory and some other main
approaches to the phase retrieval problem, such as PhaseLift and PhaseCut, see [4, 1, 13].
Intuitively, this scheme has the biggest chance of success if the iterates can be guaranteed to
stay reasonably close to one of the solutions of the phaseless equations so that the approximate
signs σ(xk · φt ) have a chance to frequently match (or approximate, in the complex case) the true
signs and make progress. Each time there is a phase mismatch, the iterate gets an update in the
wrong direction, so it is important that this event does not happen too frequently. Hence, unlike
the linear classical Kaczmarz scheme (1) which is not susceptible to the initial condition, a good
initialization is needed for the nonlinear phase-adapting version (2). There are now good methods
for this, such as the truncated spectral initialization [3].
1.1
Contribution
This paper will be about the real case, i.e. both x and the φi are in Rd . Without loss of generality,
we assume that the φi are of unit norm, since we can always run the iteration (2) with normalized
vectors φ̂i := φi /kφi k and intensity measurements |ŷi | := |yi |/kφi k. Hence we will work with the
iteration
xk+1 = xk + (σ(xk · φt )|x · φt | − xk · φt ) φt .
(3)
There will be two sources of randomness in this paper. The first and the primary source of randomness is the following: Given any measurement system Φ := (φ1 , . . . , φm ), we will assume that
the indices t are chosen uniformly and independently from {1, . . . , m}. We will call the resulting
method phase-adapting randomized Kaczmarz iteration, irrespective of how Φ may have been chosen. In Section 3, we present a certain deterministic condition on Φ called “δ-admissibility”(which
consists of four individual properties), and show that with a δ-admissible Φ (and for a sufficiently
small δ), if the starting relative error is less than δ, then after one iteration the error shrinks in
conditional expectation (with respect to the random choice of t). We then carry out a probabilistic
analysis of convergence in Section 4 via “drift analysis” and “hitting-time” bounds.
The secondary source of randomness will come into play when we want to show that most
measurement systems are δ-admissible in the m d regime. To achieve this, we will assume that
the φi are chosen independently from the uniform distribution on the unit sphere Sd−1 in Rd . The
standard Gaussian distribution on Rd can also be used.
For convenience, we state here a summarized theorem combining these two types of randomness.
Individual (and stronger) results are stated separately in Sections 3 and 4. We use the notation
dist(u, v) := min(ku − vk, ku + vk)
to denote the distance between u and v up to a global phase.
2
Theorem 1.1. Let φ1 , . . . , φm be chosen independently and uniformly on Sd−1 . There exist absolute
positive constants C, c, and δ0 such that if m ≥ Cd, then with probability 1 − exp(−cm) the system
Φ := (φ1 , . . . , φm ) satisfies the following property:
For any 0 < ε < 1, if the phase-adapting randomized Kaczmarz method with respect to Φ is
applied to any initial point x0 satisfying the relative error bound
dist(x, x0 )
≤ δ0 ε,
kxk
then the stability event
Σ :=
dist(x, xk )
≤ δ0 for all k ≥ 1
kxk
holds with probability at least 1 − ε2 , and conditioned on this event the expected squared error decays
exponentially. More precisely, we have
E dist2 (x, xk )1Σ ≤ e−k/4d dist2 (x, x0 )
for all k ≥ 1.
We prove this theorem at the end of Section 4. Some remarks are in order:
• As is the case for the randomized Kaczmarz method for linear inverse problems, the exponential convergence of xk to x is achieved in the mean-squared sense. However, an important
distinction is that this is conditional on a stability event. (In the linear case, this event is
automatic due to the fact that error decreases deterministically.) We handle this problem
using methods that are known as “drift analysis” (see [5]).
• The above stated probability lower bound for the stability event is not tight. Furthermore, our
preliminary calculations suggest that the methods of this paper can be extended to achieve
an improved probabilistic guarantee of the form 1 − O(ε2p ) for any fixed p ≥ 1. For the sake
of exposition we do not pursue this extension in this manuscript.
• We have left out performance guarantees regarding the initialization procedure from the above
theorem because we have no new results to offer here. One may simply use the truncated
spectral method [3] which is capable of providing the kind of guarantee that is compatible
with the above theorem in that for any accuracy guarantee it can operate in the regime m d
and succeed with probability 1 − exp(−Ω(m)).
Note for the revision: We would like to note here that simultaneously with the initial posting
of this paper, Y. Shuo Tan and R. Vershynin posted a manuscript (see [10]) on the randomized
Kaczmarz method for phase retrieval, with results that are somewhat similar to ours, but established
using different methods. Subsequently, we were also informed that Zhang et al. [15] had previously
established a conditional error contractivity result for the Gaussian measurement model and using
the so-called “reshaped Wirtinger flow” method.
3
2
Basic relations
Let zk := x − xk . Then (3) can be rewritten as
zk+1 = zk − (σ(xk · φt )|x · φt | − xk · φt ) φt
= zk − (zk · φt )φt + [σ(x · φt ) − σ(xk · φt )] |x · φt |φt .
(4)
Since zk − (zk · φt )φt and φt are orthogonal, we obtain
kzk+1 k2 = kzk − (zk · φt )φt k2 + |σ(x · φt ) − σ(xk · φt )|2 |x · φt |2
= kzk k2 − |zk · φt |2 + |σ(x · φt ) − σ(xk · φt )|2 |x · φt |2 .
(5)
When x · φt and xk · φt have opposite signs we have |x · φt | ≤ |(x − xk ) · φt | so that
|σ(x · φt ) − σ(xk · φt )| |x · φt | ≤ |σ(x · φt ) − σ(xk · φt )| |zk · φt |
is always valid. Hence (5) implies
h
i
kzk+1 k2 ≤ kzk k2 + |σ(x · φt ) − σ(xk · φt )|2 − 1 |zk · φt |2 .
(6)
Note that (3) is invariant under the transformation x 7→ −x. Hence we actually have
h
i
k ± x − xk+1 k2 ≤ k ± x − xk k2 + |σ(±x · φt ) − σ(xk · φt )|2 − 1 |(±x − xk ) · φt |2 ,
(7)
i.e. the analysis is identical for x and −x. For convenience of notation and without loss of generality
we will work to analyze kx − xk k and make our initial condition assumption on kx − x0 k.
2.1
Heuristic for convergence
Let φ be uniformly distributed on Sd−1 . It is a standard fact that
E |z · φ|2 =
1
kzk2 ,
d
and an easy calculation (see Appendix) yields
E |z · φ|4 =
3
kzk4 .
d(d + 2)
It can also be checked easily that for any two nonzero x and y we have
P{σ(x · φ) 6= σ(y · φ)} =
1
θx,y =: d(x̂, ŷ),
π
where θx,y ∈ [0, π] is the angle between x and y, and therefore d(x̂, ŷ) is the normalized geodesic
distance on Sd−1 between x̂ and ŷ. Hence, by Cauchy-Schwarz inequality, we obtain
E |σ(x · φ) − σ(y · φ)|2 |(x − y) · φ|2 ≤ 4 (P{σ(x · φ) 6= σ(y · φ)})1/2 E|(x − y) · φ|4
4 3θx,y 1/2
≤
kx − yk2 .
d
π
4
1/2
(8)
Hence, if θx,y is sufficiently small (e.g., less than 1/64), then
1
E |σ(x · φ) − σ(y · φ)|2 − 1 |(x − y) · φ|2 ≤ − kx − yk2 .
2d
Guided by these calculations, we turn to the error bound (6). We see that if θx,xk is sufficiently
small (which, for a fixed x, would be guaranteed by a sufficiently small zk ) and if we were to choose
each φt uniformly and independently on the unit sphere, then we would have
1
1
2
E kzk+1 k Fk , θx,xk <
≤ 1−
kzk k2 ,
64
2d
where Fk is the sigma-algebra generated by φt(0) , . . . , φt(k−1) , and for any event E, {Fk , E} is the
sigma-algebra in E formed by intersecting elements of Fk with E.
Hence the stochastic process (kzk k2 )∞
k=0 is contractive in conditional expectation which is also
conditional on the size of kzk k. Without the size condition on kzk k, the analysis would have been
fairly straightforward, similar to the situation of the randomized Kaczmarz iteration for linear
systems. As we will see, this condition makes the task non-trivial.
However, we must also establish a similar contractivity result (conditional and in expectation)
for the actual random model used in this paper, i.e., when φt is chosen uniformly from a fixed
collection Φ := (φ1 , . . . , φm ). This collection itself may also have been chosen randomly, though
with the above observation we can now define certain deterministic properties of Φ that are needed
for the algorithm to work.
3
Admissible measurement systems
Let δ ∈ (0, 1) and Φ := (φ1 , . . . , φm ) be a given collection of nonzero vectors in Rd . Following [9],
⊥
we say that Φ, or more appropriately, the linear hyperspaces (φ⊥
1 , . . . , φm ) produce a δ-uniform
d−1
d−1
tessellation of S
if for all x and y in S , we have
1
card 1 ≤ i ≤ m : σ(x · φi ) 6= σ(y · φi ) − d(x, y) < δ.
m
(9)
Then by Theorem 1.2 of [9], there exists two positive absolute constants C and c such that if
m ≥ Cδ −6 d and the φi are chosen independently from the uniform distribution on Sd−1 , then with
probability at least 1 − 2 exp(−cδ 2 m), we get a δ-uniform tessellation of Sd−1 .
If φ is chosen from the collection Φ := (φ1 , . . . , φm ) uniformly at random and f is any function
on Φ, then we define the empirical mean
m
Eφ∼Φ
1 X
f (φ) :=
f (φi ).
m
i=1
With a Φ that yields a δ-uniform tessellation, we have that the empirical mean Eφ∼Φ 1{σ(x·φ)6=σ(y·φ)}
is within δ of the ensemble mean d(x̂, ŷ). The upper part of this bound obviously yields
Eφ∼Φ
1{σ(x·φ)6=σ(y·φ)} ≤ δ + d(x̂, ŷ)
for all x, y.
(10)
The above result provides a pathway for mimicking the argument in Section 2.1 with E replaced
by Eφ∼Φ . Under the same random model for Φ, a useful concentration result (i.e. for the regime
5
m d) holds for the empirical mean Eφ∼Φ |z ·φ|2 . Indeed as it follows from [12, Theorem 5.39], there
exist absolute positive constants C and c such that for m ≥ Cd and with probability 1 − exp(−cm)
we have
kzk2
3kzk2
≤ Eφ∼Φ |z · φ|2 ≤
for all z.
(11)
2d
2d
(If desired, the constants 1/2 and 3/2 can be chosen closer to 1 without changing the form of this
statement.)
In order to continue on the same path, one would wish to have Eφ∼Φ |z · φ|4 . kzk4 /d2 with
high probability. As it turns out,1 this is impossible in the regime m d. We will circumvent
this obstacle by tightening the Cauchy-Schwarz argument of Section 2.1: In order to do this, we
will invoke (10) coupled with the Cauchy-Schwarz inequality only in the event |z · φ|2 does not
exceed a fixed multiple of its mean value kzk2 /d, and show that the above desirable upper bound
is then achievable with high probability. At the same time, we will show that the second moment
contribution from the large values is in fact small, so in this event we will only invoke the trivial
bound on |σ(x · φ) − σ(y · φ)|2 .
To this end, given δ ∈ (0, 1), consider the alternative weaker conditions
Eφ∼Φ |z · φ|4 1{|z·φ|2 ≤kzk2 /δd} ≤
4kzk4
d2
for all z,
(12)
and
4δkzk2
for all z.
(13)
d
We will say that Φ is δ-admissible if all of the four conditions (10), (11), (12), and (13) hold. Note
that all of these are deterministic conditions on Φ. We will show in Lemma 3.2 that a random
measurement system Φ is δ-admissible with high probability when m ≥ Cd, but first let us show
how these two alternative conditions are used instead of a bound on Eφ∼Φ |z · φ|4 . Suppose Φ is
δ-admissible. Noting that |σ(x · φ) − σ(y · φ)|2 /4 = 1{σ(x·φ)6=σ(y·φ)} , we have
Eφ∼Φ |z · φ|2 1{|z·φ|2 >kzk2 /δd} ≤
Eφ∼Φ |σ(x · φ) − σ(y · φ)|2 |(x − y) · φ|2
1{σ(x·φ)6=σ(y·φ)} |(x − y) · φ|2 1{|(x−y)·φ|2 ≤kx−yk2 /δd}
+ 4 Eφ∼Φ 1{σ(x·φ)6=σ(y·φ)} |(x − y) · φ|2 1{|(x−y)·φ|2 >kx−yk2 /δd}
= 4 Eφ∼Φ
1/2
≤ 4 (δ + d(x̂, ŷ))1/2 Eφ∼Φ |(x − y) · φ|4 1{|(x−y)·φ|2 ≤kx−yk2 /δd}
+ 4 Eφ∼Φ |(x − y) · φ|2 1{|(x−y)·φ|2 >kx−yk2 /δd} .
(14)
Invoking (12) and (13), we get that
Eφ∼Φ |σ(x · φ) − σ(y · φ)|2 |(x − y) · φ|2 ≤
8(δ + d(x̂, ŷ))1/2 + 16δ
kx − yk2
d
1
≤
kx − yk2 ,
(15)
4d
√
provided d(x̂, ŷ) ≤ δ and δ is sufficiently small (e.g. 8 2δ + 16δ ≤ 1/4). Hence, together with the
lower bound of (11), we have
1
Eφ∼Φ |σ(x · φ) − σ(y · φ)|2 − 1 |(x − y) · φ|2 ≤ − kx − yk2
4d
1
We thank Y. Shuo Tan and R. Vershynin for bringing this fact to our attention.
6
and therefore
h
i
1
E kzk+1 k2 Fk , d(x̂, x̂k ) ≤ δ ≤ 1 −
kzk k2 ,
4d
where again Fk is the sigma-algebra generated by φt(0) , . . . , φt(k−1) .
At this point, it will be helpful to replace the condition d(x̂, x̂k ) ≤ δ by a size condition on zk .
Note that for any two nonzero vectors x and y, we have
(x − y)kyk + y(kyk − kxk)
1
1 x
y
kx − yk
d(x̂, ŷ) ≤ kx̂ − ŷk =
−
=
≤
2
2 kxk kyk
2kxkkyk
kxk
so that the condition kzk k ≤ δkxk implies d(x̂, x̂k ) ≤ δ. Therefore we have
i
h
1
kzk k2 .
E kzk+1 k2 Fk , kzk k ≤ δkxk ≤ 1 −
4d
(16)
With the above discussion we have established the following result:
Lemma 3.1. There exists δ0 > 0 such that, if δ ≤ δ0 and Φ is δ-admissible, then
h
i
E kzk+1 k2 Fk , kzk k ≤ b ≤ ρkzk k2 for all k ≥ 0,
where b := δkxk and ρ := 1 −
(17)
1
4d .
We now show that a random Φ is δ-admissible with high probability in the regime m d.
Lemma 3.2. For every δ ∈ (0, 1), there exists positive constants C and c depending only on δ such
that if m ≥ Cd, then a random measurement system Φ := (φ1 , . . . , φm ) that is chosen independently
from the uniform distribution on Sd−1 is δ-admissible with probability at least 1 − exp(−cm).
Proof. The property (10) is proven in [9] and (11) is covered by [12, Theorem 5.39]. Hence we only
need to establish (12) and (13). Note that by homogeneity we may assume kzk = 1.
We start with (12). As is standard in this type of question, we would like to establish the stated
inequality for fixed z first (with high probability) and then use approximation over an -net of Sd−1
to achieve uniformity over z. However 1{|z·φ|2 ≤kzk2 /δd} is a discontinuous function of the random
variable |z · φ|, presenting a difficulty for the approximation argument. The solution will follow by
incorporating a suitable Lipschitz extension, as also done in [3].
For this purpose, let γ1 : [0, ∞) → [0, ∞) be defined by
2
s ≤ δ −1 ,
s ,
(2δ −1 − s)δ −1 , δ −1 < s ≤ 2δ −1 ,
γ1 (s) :=
(18)
0,
2δ −1 < s.
Then γ1 is a Lipschitz function with Lipschitz constant 2δ −1 . Furthermore,
s2 χ[0,δ−1 ] (s) ≤ γ1 (s) ≤ s2
so that for any z we have
Eφ∼Φ |z · φ|4 1{|z·φ|2 ≤1/δd} =
≤
√
√
1
Eφ∼Φ |z · ( dφ)|4 χ[0,δ−1 ] (|z · ( dφ)|2 )
2
d
√
1
E
γ
(|z
·
(
dφ)|2 ).
1
φ∼Φ
d2
7
(19)
√
Now, let φ denote the random vector uniformly distributed on Sd−1 so that dφ is a spherical
random vector in Rd (see [12, Section 5.2.5]). Let k · kψ1 stand for the sub-exponential norm and
k · kψ2 the sub-Gaussian norm (see [12, Section 5.2.3 and 5.2.4]). Noting that γ1 (s) ≤ δ −1 s, we have
√
γ1 (|z · ( dφ)|2 )
ψ1
√
≤ δ −1 |z · ( dφ)|2
ψ1
√
≤ 2δ −1 |z · ( dφ)|
2
ψ2
. δ −1
where in the second step we have used [12, Lemma 5.14]) and in the last step the fact that the subGaussian norm of a spherical random vector is bounded by an absolute constant (see [12, Section
5.2.5]; a direct computation is also possible).
Hence, by the Bernstein-type inequality [12, Proposition 5.16], there is an absolute constant
c1 > 0 such that for any t > 0 we have, with probability at least 1 − exp(−c1 min(δt, δ 2 t2 )m),
√
√
Eφ∼Φ γ1 (|z · ( dφ)|2 ) ≤ E γ1 (|z · ( dφ)|2 ) + t
√
≤ E |z · ( dφ)|4 + t
≤ 3 + t,
(20)
where in the second step we have used γ1 (s) ≤ s2 instead.
Now pick an -net N of the unit sphere Sd−1 of cardinality at most (3/)d where < 1. For
each z 0 ∈ Sd−1 and z ∈ N such that kz 0 − zk < , we have
√
√
√
√
Eφ∼Φ γ1 (|z 0 · ( dφ)|2 ) − γ1 (|z · ( dφ)|2 ) ≤ 2δ −1 Eφ∼Φ |z 0 · ( dφ)|2 − |z · ( dφ)|2
√
√
= 2δ −1 Eφ∼Φ [(z 0 − z) · ( dφ)][(z 0 + z) · ( dφ)]
≤ 6δ −1 ,
(21)
where in the first step we have utilized the Lipschitz continuity of γ1 , and in the last step CauchySchwarz inequality coupled with the upper bound of (11). Combining (19), (20), and (21), we find
that with probability at least 1 − (3/)d exp(−c1 min(δt, δ 2 t2 )m), we have
Eφ∼Φ |z 0 · φ|4 1{|z 0 ·φ|2 ≤1/δd} ≤
1
(3 + t + 6δ −1 )
d2
for every z 0 ∈ Sd−1 . We may choose t = 1/2 and = δ/12 so that 3 + t + 6δ −1 = 4 and therefore
(12) holds with probability at least 1 − exp(−c1 δ 2 m/8) provided c1 δ 2 m/8 ≥ d log(36/δ).
We continue with (13). We will use the same method, but with a different Lipschitz function.
Let γ2 : [0, ∞) → [0, ∞) be defined by
2
δs , s ≤ δ −1
γ2 (s) :=
(22)
s,
s > δ −1 .
Then γ2 is a Lipschitz function that fixes 0 with Lipschitz constant 2. We have
sχ(δ−1 ,∞) (s) ≤ γ2 (s) = min(δs2 , s)
so that for any fixed z we have
Eφ∼Φ |z · φ|2 1{|z·φ|2 >1/δd} =
≤
√
√
1
Eφ∼Φ |z · ( dφ)|2 χ(δ−1 ,∞) (|z · ( dφ)|2 )
d
√
1
Eφ∼Φ γ2 (|z · ( dφ)|2 ).
d
8
(23)
Noting that γ2 (s) ≤ s, we now have
√
γ2 (|z · ( dφ)|2 )
ψ1
.1
so that by the Bernstein-type inequality (and reducing the value of c1 if necessary), for any t ∈ (0, 1)
we have, with probability at least 1 − exp(−c1 t2 m),
√
√
Eφ∼Φ γ2 (|z · ( dφ)|2 ) ≤ E γ2 (|z · ( dφ)|2 ) + t
√
≤ δ E |z · ( dφ)|4 + t
≤ 3δ + t,
(24)
where in the second step we have used γ2 (s) ≤ δs2 instead. We again pick an -net N of the unit
sphere Sd−1 of cardinality at most (3/)d . For each z 0 ∈ Sd−1 and z ∈ N such that kz 0 − zk < ,
this time we have
√
√
√
√
Eφ∼Φ γ2 (|z 0 · ( dφ)|2 ) − γ2 (|z · ( dφ)|2 ) ≤ 2 Eφ∼Φ |z 0 · ( dφ)|2 − |z · ( dφ)|2
≤ 6.
(25)
Hence by the union bound, with probability at least 1 − (3/)d exp(−c1 t2 m) we have
Eφ∼Φ |z 0 · φ|2 1{|z 0 ·φ|2 >1/δd} ≤
1
(3δ + t + 6)
d
for every z 0 ∈ Sd−1 . We may choose t = δ/2 and = δ/12 so that 3δ + t + 6 = 4δ and therefore
(13) holds with probability at least 1 − exp(−c1 δ 2 m/8) provided c1 δ 2 m/8 ≥ d log(36/δ).
4
Probabilistic analysis of the error sequence
Our goal in this section will be to bound the probability that kzk k exceeds b at some point and to
obtain probabilistic guarantees on the exponential decay of kzk k. Lemma 3.1 uses the randomness
present in the selection of φt(k) only. To be able to iterate this result recursively we need to condition
on the event Ωk := {kz0 k ≤ b, kz1 k ≤ b, . . . , kzk k ≤ b}. We define the “hitting time”
τb := min{j ≥ 0 : kzj k > b}.
Hence Ωk is the same as {τb > k} and the event {τb = ∞} means kzk k ≤ b for all k.
Lemma 4.1. Suppose (17) holds. Then
h
i
h
i
E kzk+1 k2 τb > k ≤ ρ E kzk k2 τb > k−1
(26)
and therefore
h
i
E kzk k2 τb > k−1 ≤ ρk kz0 k2
for all k ≥ 0.
9
(27)
Proof. Note that F0 is the trivial sigma-algebra. We have
i
i
h
h
E kzk+1 k2 τb > k = E kzk+1 k2 τb > k, F0
h h
i
i
= E E kzk+1 k2 τb > k, Fk τb > k, F0
i
h h
i
≤ ρ E E kzk k2 τb > k, Fk τb > k, F0
h
i
2
= ρ E kzk k τb > k .
Note also that
i
i
h
h
E kzk k2 τb > k ≤ b2 ≤ E kzk k2 τb = k .
Hence we have
h
i
h
i
E kzk k2 τb > k = P kzk k ≤ b τb > k−1 E kzk k2 τb > k
h
i
+ P kzk k > b τb > k−1 E kzk k2 τb > k
h
i
≤ P kzk k ≤ b τb > k−1 E kzk k2 τb > k
h
i
+ P kzk k > b τb > k−1 E kzk k2 τb = k
h
i
= P kzk k ≤ b τb > k−1 E kzk k2 kzk k ≤ b, τb > k−1
h
i
+ P kzk k > b τb > k−1 E kzk k2 kzk k > b, τb > k−1
h
i
h
i
= E kzk k2 1{kzk k≤b} τb > k−1 + E kzk k2 1{kzk k>b} τb > k−1
h
i
= E kzk k2 τb > k−1 .
(28)
(29)
(30)
The result now follows by combining (28) and (30).
We can now use Lemma 4.1 to control (i) the probability of the event that the error kzk k exceeds
b at some point (i.e. {τb < ∞}), and (ii) the expected decay of squared error kzk k2 conditional on
the event that the error remains bounded by b (i.e. {τb = ∞}).
Lemma 4.2. Suppose (17) holds. Then for any k ≥ 1 we have
h
i
E kzk k2 τb = ∞ ≤
1
ρk kz0 k2 .
1 − P (τb < ∞)
(31)
and for any a > 0
k/2
P kzk k ≥ ρ
a ≤
kz0 k
a
2
+ P (τb < ∞) .
Proof. For the first claim it suffices to observe that {τb = ∞} ⊂ {τb > k−1} so that
h
i
h
i
P (τb = ∞) E kzk k2 τb = ∞ ≤ P (τb > k−1) E kzk k2 τb > k−1 .
The result follows by bounding the right hand side of this inequality using (27).
10
(32)
For the second claim, note that
P (kzk k ≥ ) = P kzk k ≥ τb > k−1 P (τb > k−1) + P kzk k ≥ τb ≤ k−1 P (τb ≤ k−1)
≤ P kzk k ≥ τb > k−1 + P (τb < ∞)
i
h
(33)
≤ −2 E kzk k2 τb > k−1 + P (τb < ∞) .
The result follows by setting = ρk/2 a and using (27) again.
Next we give a bound on P (τb < ∞).
Lemma 4.3. Suppose (17) holds. Then
b2 P (τb < ∞) + (ρ−1 − 1)
∞
X
i
h
P (τb > k) E kzk+1 k2 τb > k ≤ ρkz0 k2 ,
(34)
k=1
in particular we have
P (τb < ∞) ≤ ρ
Proof. We start by noting that
h
i
P (τb > k−1) E kzk k2 τb > k−1
=
≥
≥
≥
kz0 k
b
2
.
(35)
h
i
P (kzk k > b, τb > k−1) E kzk k2 kzk k > b, τb > k−1
h
i
+ P (kzk k ≤ b, τb > k−1) E kzk k2 kzk k ≤ b, τb > k−1
h
i
P (τb = k) b2 + P (τb > k) E kzk k2 τb > k
h
i
P (τb = k) b2 + P (τb > k) ρ−1 E kzk+1 k2 τb > k
h
i
P (τb = k) b2 + P (τb > k) E kzk+1 k2 τb > k
h
i
+ (ρ−1 − 1)P (τb > k) E kzk+1 k2 τb > k
(36)
where in the first
h inequality weihave used (the second inequality of) (29) and in the second equality
(28). Since E kzk k2 τb > k−1 has exponential decay we can sum both sides from k = 1 to ∞.
After cancelling the common term (the sum from k = 2 to ∞) we obtain
∞
∞
h
i
h
i
X
X
P (τb > 0) E kz1 k2 τb > 0 ≥ b2
P (τb = k) + (ρ−1 − 1)
P (τb > k) E kzk+1 k2 τb > k .
k=1
k=1
The result follows by noting that the left hand side is bounded above by ρkz0 k2 .
We are now ready to prove Theorem 1.1.
11
Proof of Theorem 1.1. We choose δ0 as implied by Lemma 3.1 and apply Lemma 3.1 and 3.2
together where we set the values of C and c according to the choice δ = δ0 . Hence (17) is valid
with b = δ0 kxk.
Let ε ∈ (0, 1) and x0 such that dist(x, x0 ) ≤ δ0 ε be be given. Let us assume kx−x0 k ≤ k−x−x0 k
so that dist(x, x0 ) = kx − x0 k. Otherwise we replace x by −x below. With (17), we have the
conclusions of Lemmas 4.2 and 4.3 at our disposal. The stability event is simply equal to {τb = ∞}.
We bound P(τb < ∞) by ρ(δ0 εkxk/δ0 kxk)2 ≤ ε2 . Lemma 4.2 then yields the decay bound
E kzk k2 1{τb =∞} ≤ e−k/4d kz0 k2 ,
where we have used ρ = 1 −
1
4d
≤ e−1/4d .
Appendix: A simple moment calculation
Let the random variable φ be uniformly distributed on Sd−1 and e1 , . . . , ed be the standard basis
of Rd . By the unitary invariance of the uniform distribution on Sd−1 , E |z · φ|2 independent of
z ∈ Sd−1 . Hence we have
1 = E kφk2 = E
d
X
|ej · φ|2 = d E |e1 · φ|2
j=1
2
d
so that E |e1 · φ|2 = d1 , and therefore E |z · φ|2 = kzk
d for general z ∈ R .
Let us now consider E |z · φ|4 which is also independent of z ∈ Sd−1 . Call its common value α.
It is also clear by symmetry that for i 6= j, the quantity E |ei · φ|2 |ej · φ|2 is independent of the pair
(i, j). Call its common value β.
We have
d
X
1
= E |e1 · φ|2 = E |e1 · φ|2 kφk2 = E |e1 · φ|2
|ej · φ|2 = α + (d − 1)β.
d
j=1
We also have
2α = E
4
e1 + e2
e1 − e2
√
√
·φ +E
·φ
2
2
4
=
1
E |e1 · φ|4 + E |e2 · φ|4 + 6E |e1 · φ|2 |e2 · φ|2 = α + 3β,
2
so that α = 3β. Solving these two equations we get α =
general z ∈
3
d(d+2) ,
and therefore E |z · φ|4 =
3kzk4
d(d+2)
for
Rd .
Acknowledgement The authors gratefully acknowledge Y. Shuo Tan and R. Vershynin for pointing out a mistake in the first version of this paper. H.J. thanks Yuri Bakhtin and Afonso Bandeira
for various valuable discussions.
References
[1] Emmanuel J Candès, Thomas Strohmer, and Vladislav Voroninski. Phaselift: Exact and stable
signal recovery from magnitude measurements via convex programming. Communications on
Pure and Applied Mathematics, 66(8):1241–1274, 2013.
12
[2] Xuemei Chen and Alexander M Powell. Almost sure convergence of the Kaczmarz algorithm
with random measurements. Journal of Fourier Analysis and Applications, 6:1195–1214, 2012.
[3] Yuxin Chen and Emmanuel Candès. Solving random quadratic systems of equations is nearly as
easy as solving linear systems. Communications on Pure and Applied Mathematics, 70(5):822–
883, 2017.
[4] Yonina C Eldar and Shahar Mendelson. Phase retrieval: Stability and recovery guarantees.
Applied and Computational Harmonic Analysis, 36(3):473–494, 2014.
[5] Bruce Hajek. Hitting-time and occupation-time bounds implied by drift analysis with applications. Advances in applied probability, 14(3):502–525, 1982.
[6] G. Li, Y. Gu, and Y. M. Lu. Phase retrieval using iterative projections: Dynamics in the
large systems limit. In 2015 53rd Annual Allerton Conference on Communication, Control,
and Computing (Allerton), pages 1114–1118, Sept 2015.
[7] Deanna Needell, Nathan Srebro, and Rachel Ward. Stochastic gradient descent, weighted
sampling, and the randomized Kaczmarz algorithm. Mathematical Programming, 155(1-2):549–
573, 2016.
[8] Deanna Needell and Joel A Tropp. Paved with good intentions: analysis of a randomized block
Kaczmarz method. Linear Algebra and its Applications, 441:199–221, 2014.
[9] Yaniv Plan and Roman Vershynin. Dimension reduction by random hyperplane tessellations.
Discrete and computational geometry, 51(2):438–461, 2014.
[10] Y. Shuo Tan and R. Vershynin. Phase Retrieval via Randomized Kaczmarz: Theoretical
Guarantees. arXiv/1706.09993.
[11] Thomas Strohmer and Roman Vershynin. A randomized Kaczmarz algorithm with exponential
convergence. Journal of Fourier Analysis and Applications, 15(2):262–278, 2009.
[12] R. Vershynin. Introduction to the non-asymptotic analysis of random matrices. In Compressed
sensing, pages 210–268. Cambridge Univ. Press, Cambridge, 2012.
[13] Irène Waldspurger, Alexandre D’aspremont, and Stéphane Mallat. Phase recovery, maxcut
and complex semidefinite programming. Mathematical Programming, 149(1-2):47, 2015.
[14] Ke Wei. Solving systems of phaseless equations via Kaczmarz methods: a proof of concept
study. Inverse Problems, 31(12):125008, 2015.
[15] H. Zhang, Y. Zhou, Y. Liang, and Y. Chi. Reshaped Wirtinger Flow and Incremental Algorithm for Solving Quadratic System of Equations. arXiv/1605.07719.
13
| 7cs.IT
|
arXiv:1610.02080v3 [math.ST] 21 Mar 2017
On Shapley value for measuring importance of
dependent inputs
Art B. Owen
Stanford University
Clémentine Prieur
Université Grenoble Alpes, CNRS, LJK, F-38000 Grenoble, France
Inria project/team AIRSEA
Orig: October 2016
This: March 2017
Abstract
This paper makes the case for using Shapley value to quantify the
importance of random input variables to a function. Alternatives based
on the ANOVA decomposition can run into conceptual and computational
problems when the input variables are dependent. Our main goal here is
to show that Shapley value removes the conceptual problems. We do
this with some simple examples where Shapley value leads to intuitively
reasonable nearly closed form answers.
1
Introduction
The importance of inputs to a function is commonly measured via Sobol’ indices. Those are defined in terms of the functional analysis of variance (ANOVA)
decomposition, which is conventionally defined with respect to statistically independent inputs. In applications to computer experiments, it is common that
the input space is constrained to a non-rectangular region, or that the input
variables have some other known form of dependence, such as a general Gaussian distribution. When the inputs are described by an empirical distribution
on observational data it is extremely rare that the variables are statistically independent. Even designed experiments avoid having independent inputs (i.e., a
Cartesian product of input levels) when the dimension is moderately large (Wu
and Hamada, 2011).
A common way to address dependence is to build on work by Stone (1994)
and Hooker (2012) who define an ANOVA for dependent inputs and then define
variable importance through that generalization of ANOVA. This is the method
taken by Chastaing et al. (2012) for computer experiments.
1
The dependent-variable ANOVA leads to importance measures with two
conceptual problems:
1) the needed ANOVA is only defined when the random x has a distribution
with a density (or mass function) uniformly bounded below by a positive constant times another density/mass function that has independent
margins, and
2) the resulting importance of a variable can be negative (Chastaing et al.,
2015).
The first condition is very problematic. It fails even for Gaussian x with
nonzero correlation. It fails for inputs constrained to a simplex. It fails when
the empirical distribution of say (xi1 , xi2 ) is such that some input combinations
are never observed or, by definition, cannot possibly be observed.
The second condition is also conceptually problematic. A variable on which
the function does not depend at all will get importance zero and thus be more
important than one that the function truly does depend on in a way that gave
it negative importance.
The Shapley value, from economics, provides an alternative way to define
variable importance. As we describe below, Shapley value provides a way to
attribute the value created by a team to its individual members. In our context
the members are individual input variables. Owen (2014) derived Shapley value
importance for independent inputs where the value is variance explained. The
Shapley value of a variable turns out to be bracketed between two different
Sobol’ indices. Song et al. (2016) recently advocated the use of Shapley value for
the case of dependent inputs. They report that it is more suitable than Sobol’
indices for such problems. They use the term “Shapley effects” to describe
variance based Shapley values.
The Shapley value provides an importance measure that avoids the two problems mentioned above: It is available for any function in L2 of the appropriate
domain and it never gives negative importance.
Although Shapley value solves the conceptual problems, computational problems remain a serious challenge (Castro et al., 2009). The Shapley value is
defined in terms of 2d − 1 models where d is the dimension of x. Song et al.
(2016) presented a Monte Carlo algorithm to estimate Shapley importance and
they apply it to detailed real-world problems. We address only the conceptual
appropriateness of Shapley value to variable importance, not computational issues.
The outline of this paper is as follows. Section 2 introduces our notation, defines the functional ANOVA and the Sobol’ indices and presents the dependentvariable ANOVA. Section 3 presents the Shapley value and its use for variable
importance. From the definition there it is clear that Shapley value for variance
explained will never be negative. Section 4 gives several examples of simple cases
and exceptional corner cases where we can derive the Shapley value of variable
importance and verify that it is reasonable. Section 5 has brief conclusions.
Section 6 contains the longer proofs.
2
2
Notation
We consider real valued functions f defined on a space X . The point x ∈ X has
d components, and we write x = (x1 , . . . , xd ) where xj ∈ Xj . The individual
Xj are ordinarily interval subsets of R but each of them may be much more
general (regions in Euclidean space, functions on [0, 1], or even images, sounds,
and video). What we must assume is that x follows a distribution P chosen by
the user, and that f (x) is then a random variable with E(f (x)2 ) < ∞.
When the components of x are independent, then Sobol’ indices (Sobol’,
1990, 1993) provide ways to measure the importance of individual components of
x as well as sets of them. They are based on a functional ANOVA decomposition.
For details and references on the functional ANOVA, see Owen (2013).
2.1
ANOVA for independent variables
Here is a brief summary of the ANOVA to introduce our notation. For simplicity
we will take f ∈ L2 [0, 1]d with the argument x = (x1 , . . . , xd ) of f uniformly disQd
tributed on [0, 1]d , but the approach extends straightforwardly to L2 ( j=1 Xj )
with independent not necessarily uniform xj ∈ Xj .
The set {1, 2, . . . , d} is written 1:d. For u ⊆ 1:d, |u| denotes cardinality
and −u is the complement {1 6 j 6 d | j 6∈ u}. If u = (j1 , j2 , . . . , j|u| ) then
Q
xu = (xj1 , xj2 , . . . , xj|u| ) ∈ [0, 1]|u| and dxu = j∈u dxj . We use u + v as a
shortcut for u ∪ v when u ∩ v = ∅, especially in subscripts.
The P
ANOVA is defined via functions fu ∈ L2 [0, 1]d . These functions
satisfy
R
f (x) = u⊆1:d fu (x). They are defined as follows. First, f∅ = f (x) dx and
then
Z
X
fu (x) =
f (x) −
fv (x) dx−u
(1)
v(u
for |u| > 0. The integral in (1) is over [0, 1]d−|u| and it yields Ra function fu that
depends on x only through xu . The effects fu are orthogonal: fu (x)fv (x) dx =
0 when u 6= v.
R
The variance component for the set u is σu2 = Pfu (x)2 dx for |u| > 0 and
2
σ∅
= 0. The variance of f for x ∼ U[0, 1]d is σ 2 = u⊆1:d σu2 .
We can define the importance of a set of variables by how much of the
variance of f is explained by those variables. The best prediction of f (x) given
xu is
X
f[u] (x) ≡ E(f (x) | xu ) =
fv (x).
v⊆u
This prediction explains
τ 2u ≡
X
v⊆u
3
σv2 ,
(2)
of the variance in f . This is one of Sobol’s global sensitivity indices. His other
index is
X
τ 2u ≡
σv2 = σ 2 − τ 2−u .
v∩u6=∅
It is more conventional to use normalized versions τ 2u /σ 2 and τ 2u /σ 2 but unnormalized ones are simpler for our purposes. The importance of an individual
variable xj is sometimes defined through τ 2{j} or τ 2{j} . If τ 2{j} is large then xj is
important and if τ 2{j} is small then xj is unimportant.
2.2
ANOVA for dependent variables
Now suppose that f is defined on Rd but the argument x does not have independent components. Instead x has distribution P . We could generalize (1) to
the Stone-Hooker ANOVA
Z
X
fu (x) =
f (x) −
fv (x) dP (x−u )
(3)
v(u
but the result would not generally have orthogonal
effects. To take a basic
example, suppose that P is the N ( 00 ) , ρ1 ρ1 distribution for 0 < ρ < 1 and
let f (x) = β1 x1 + β2 x2 . Then (3) yields
f∅ (x) = 0,
f{1} (x) = (β1 + β2 ρ)x1 ,
f{2} (x) = (β2 + β1 ρ)x2
and f{1,2} (x) = −β2 ρx1 − β1 ρx2 . These effects are not orthogonal under P and
their mean squares do not sum to the variance of f (x) forP
x ∼ P.
It is however possible to get a decomposition f (x) = u⊆1:d fu (x) with a
hierarchical orthogonality property
Z
fu (x)fv (x) dP (x) = 0, ∀v ( u.
(4)
Chastaing et al. (2012) give conditions under which a decomposition of f satisfying (4) exists and they use it to define variable importance.
They assume that the joint distribution P is absolutely continuous
with reQ
spect to a product probability measure ν. That is P (dx) = p(x) j∈1:d νj (dxj )
for a density function p. They require also that this density satisfies
∃ 0 < M 6 1,
∀u ⊆ 1:d,
p(dx) > M p(dxu )p(dx−u ),
ν − a.e.
(5)
The joint density is bounded below by a product of two marginal densities.
Among other things, this criterion forbids ‘holes’ in the support of P . There
cannot be regions Ru ∈ Ru and R−u ∈ R−u with P (Ru × R−u ) = 0 while
min(P (Ru × R−u ), P (Ru × R−u )) > 0.
4
2.3
Challenges with dependent variable ANOVA
The no holes condition (5) is problematic in many applications. For example,
when x is uniformly distributed on the triangle
{(x1 , x2 ) ∈ [0, 1]2 | x1 6 x2 }
then (5) is violated. More generally, Gilquin et al. (2015) and Kucherenko et al.
(2016) consider functions on non-rectangular regions defined by linear inequality
constraints. These and similar regions arise in many engineering problems where
safety or costs impose constraints on design parameters.
The simplest distribution with a hole is one with positive probability on the
points
{(0, 0), (0, 1), (1, 0)}
and no others. Sobol’s ‘pick-freeze’ methods (Sobol’, 1990, 1993) estimate variable importance by freezing the level of some inputs and then picking new values
for the others. For the example here, setting x1 = 1 implies that x2 cannot be
changed at all, which is a severe problem for a pick-freeze approach with dependent inputs.
It is not just probability zero holes that cause a problem for dependent variable ANOVA. When x is normally distributed with some nonzero correlations,
then (5) does not hold, and then as we mentioned in the introduction, the
dependent-variable ANOVA is unavailable. The second problem we mentioned
there is that the dependent variable ANOVA can yield negative estimates of
importance.
3
Shapley value
Shapley value is a way to attribute the economic output of a team to the indivitual members of that team. In our case, the team will be the set of variables
x1 , x2 , . . . , xd . Given any subset u ⊆ 1:d of variables, the value that subset
creates on its own is its explanatory power. A convenient way to measure explanatory power is via
val(u) = τ 2u ≡ var(E(f (x) | xu )).
(6)
Here, the empty set creates no value and the entire team contributes σ 2 which
we must now partition among the xj .
There are four very compelling properties that an attribution method should
have. The following list is based on the account in Winter (2002). Let val(u) ∈ R
be the value attained by the subset u ⊆ {1, . . . , d} ≡ 1:d. It is always assumed
that val(∅) = 0, which holds in our variance explained setting. The values
φj = φj (val) should satisfy these properties:
Pd
1) (Efficiency) j=1 φj = val(1:d).
2) (Symmetry) If val(u ∪ {i}) = val(u ∪ {j}) for all u ⊆ 1:d − {i, j}, then
φi = φj .
5
3) (Dummy) If val(u ∪ {i}) = val(u) for all u ⊆ 1:d, then φi = 0.
4) (Additivity) If val and val0 have Shapley values φ and φ0 respectively then
the game with value val + val0 has Shapley value φj + φ0j for j ∈ 1:d.
Shapley (1953) showed that the unique valuation φ that satisfies these axioms
attributes value
−1
1 X
d−1
val(u ∪ {j}) − val(u)
φj =
d
|u|
u⊆−{j}
to variable j. Defining the value via (6) we get
1
φj =
d
X d − 1−1
(τ 2u+{j} − τ 2u ).
|u|
(7)
u⊆−{j}
From (7) we see that the Shapley value is defined for any function for which
var(E(f (x) | xu )) is always defined. The components xj do not have to be real
valued, though f (x) must be. Holes in the domain X do not make it impossible
to define a Shapley value. Next, because xu+{j} always has at least as much
explanatory power as xu has, we see that φj > 0. That is, no variable has
a negative Shapley value. As a result, the Shapley value addresses the two
conceptual problems mentioned in the introduction.
Song et al. (2016) show that the same Shapley value arises if we use val(u) =
E(var(f (x) | x−u )). That provides an alternative way to compute Shapley value.
The Shapley value simplifies for independent inputs.
Theorem 1. Let the ANOVA decomposition of a function with d independent
inputs have variance components σu2 for u ⊆ 1:d. If the value of a subset u of
variables is val(u) = τ 2u , then the Shapley value of variable j is
X
φj =
σu2 /|u|.
u⊆1:d, j∈u
Proof. Owen (2014).
It follows from Theorem 1 that τ 2{j} 6 φj 6 τ 2{j} . This is how the Sobol’
indices bracket the Shapley value.
4
Special cases
Here we consider some special case distributions and toy functions where we
can work out the Shapley value in a closed or nearly closed form. The point of
these examples is to show that Shapley gives sensible answers in both regular
cases and corner cases. Because σ 2 = var(E(f (x) | xu )) + E(var(f (x) | xu )) we
may use
τ 2u = σ 2 − E(var(f (x) | xu )).
6
(8)
4.1
Linear functions
Pd
Let f (x) = β0 + j=1 βj xj where xj are independent with variances σj2 . It
is then easy to find that φj = βj2 σj2 . If we reparameterize xj to cxj for c 6= 0
then βj becomes βj /c and the importance of this variable remains unchanged
as it should. Dependence among the xj complicates the expression for Shapley
effects in linear settings.
Shapley value for linear functions has historically been used to partition the
R2 quantity (proportion of sample variance explained) from a regression on d
variables among those d variables. Taking the value of a subset u of variables
to be Ru2 , the R2 value when regressing a response on predictors xj for j ∈ u,
yields Shapley value
φj =
1
d
X d − 1−1
2
(Ru+{j}
− Ru2 ).
|u|
(9)
u⊆−{j}
This is the LMG measure of variable importance, named after the authors of Lindeman et al. (1980). If we rearrange the d variables into all d! orders, find the
improvement in R2 that comes at the moment the j’th variable is added to
the regression, then (9) is the average of all those improvements. The LMG
reference is difficult to obtain. Genizi (1993) is another reference, having (9) as
equation (1). Grömping (2007) cites several more references on partitioning R2
in regression and discusses alternative measures and criteria for choosing. It is
clear that (9) is expensive for large d.
Here we consider a population/distribution version of partitioning variance
explained among a set of variables acting linearly. We suppose that x ∼ N (µ, Σ)
where Σ ∈ Rd×d is a positive semi-definite symmetric matrix. The function of
interest is f (x) = β0 + xT β where β = (β1 , . . . , βd ) ∈ Rd . If there is an error
term as in a linear regression on noisy data, then we can let xd be that error
variable with a corresponding βd = 1.
If Σ is not diagonal then the Stone-Hooker ANOVA is not available because (5) does not hold. Shapley value gives an interpretable expression for
general d.
Theorem 2. If f (x) = β0 + β T x for x ∼ N (µ, Σ) where Σ ∈ Rd×d has full
rank, then the Shapley effect for variable j is
2
−1
cov xj , xT
1 X d−1
−u β−u | xu
φj =
.
d
|u|
var(xj | xu )
u⊆−j
Proof. See Section 6.1.
A variable with βj = 0 can still have φj > 0. For instance if Σ = ρ0 ρ0 and
f (x) = x1 , then we can find directly from (7) that φ2 = ρ2 /2 and φ1 = 1 − ρ2 /2.
For ρ = ±1 we already know this by bijection.
7
The Shapley value works with conditional variances and the Gaussian distribution makes these very convenient. For non-Gaussian distributions the conditional covariance of xv and xw given xu may depend on the specific value of
xu , while in the Gaussian case it is simply Σvw − Σvu Σ−1
Σuw for all xu .
uuP
In P
a related problem, if we define val(u) to be var( j∈u xj ), instead of
var(E( j xjP
| xu )), then the Shapley value of variable j is φj = cov(xj , S),
where S =
j∈1:d xj . See Colini-Baldeschi et al. (2016). This quantity can
be negative. For instance, if d = 2, then φ1 = var(x1 ) + cov(x1 , x2 ) which is
negative when x1 and x2 are negatively correlated and x2 has much greater
variance than x1 .
4.2
Transformations, bijections and invariance
We can generalize the linear example
to independent random variables that
Pd
contribute additively: f (x) = j=1 gj (xj ). Then φj = var(gj (xj )). Replacing
xj by a bijection τj (xj ) and adjusting gj to gj ◦ τj−1 leaves φj unchanged.
More generally, suppose that y = f (x) and we transform the variables xj
into zj by bijections: zj = τj (xj ), xj = τj−1 (zj ), for j = 1, . . . , d. Now define
f 0 (z) = f (τ1−1 (z1 ), . . . , τd−1 (zd )) and let φ0j be the Shapley importance of zj as
a predictor of y 0 = f 0 (z). Because var(E(f 0 (z) | z u )) = var(E(f (x) | xu )), we
find that φ0j = φj for j = 1, . . . , d, where φj is the Shapley importance of xj as
a predictor of y. As a result we can apply invertible transformations to any or
all of the xj without changing the Shapley values.
Now lets revisit the linear setting with an extreme example: f (x1 , x2 ) =
106 x1 + x2 with x1 = 106 x2 where x2 (and hence x1 ) has a finite positive
variance. Because ∂f /∂x1 ∂f /∂x2 > 0 and var(x1 ) var(x2 ) one might
expect x1 to be the more important variable. However, the Shapley formula
easily yields φ1 = φ2 ; these variables are equally important. This is quite
reasonable because f is a function of x1 alone and equally a function of x2
alone.
More generally, for d > 2, if there is a bijection between any two of the xj
then those two variables have the same Shapley value. To see this, let x1 =
g1 (x2 ) and x2 = g2 (x1 ), both with probability one then for any u ⊂ 1:d with
u ∩ {1, 2} = ∅ we have
E(f (x) | xu+{1} ) = E(f (x) | xu+{2} ).
It follows that τ 2u+{1} −τ 2u = τ 2u+{2} −τ 2u and therefore φ1 = φ2 by the symmetry
property of Shapley value.
To summarize:
1) Shapley value is preserved under invertible transformations, and
2) a bijection between variables implies that they have the same Shapley
value.
8
4.3
Bivariate settings
When d = 2 we can get some simpler formulas for the importance of the two
variables.
Proposition 1. Let f (x) have finite variance σ 2 > 0 for random x = (x1 , x2 ).
Then from (7),
1
var(E(Y | x1 )) − var(E(Y | x2 )
φ1
=
1
+
σ2
2
σ2
1
E(var(Y | x2 )) − E(var(Y | x1 ))
=
1+
,
2
σ2
φ1
var(E(Y | x1 )) + E(var(Y | x2 ))
=
.
φ2
var(E(Y | x2 )) + E(var(Y | x1 ))
(10)
and
(11)
(12)
Proof. Using τ 2{1,2} = σ 2 and τ 2∅ = 0, we find that
φ1 =
1
1 2
τ
+ σ 2 − τ 2{2} = σ 2 + var(E(Y | x1 )) − var(E(Y | x2 )),
2 {1}
2
which gives us (10). The others are algebraic rearrangements.
We can use Proposition 1 to get analogous expressions for φ2 /σ 2 and φ2 /φ1
by exchanging indices.
4.3.1
Farlie-Gumbel-Morgenstern copula for d = 2
Here we focus on the case where the dependence between both components x1
and x2 is explicitly described by some copula. There exist simple conditional
expectation formulas when considering some classical classes of copulas (see e.g.,
Crane and Hoek (2008) and references therein). Starting from such formulas, it
is possible to derive explicit computations for Shapley values in a linear model.
In this section, we state explicit results for the Farlie-Gumbel-Morgenstern family of copulas.
The Farlie-Gumbel-Morgenstern copula describes a random vector x ∈ [0, 1]2
with each component xj ∼ U[0, 1] and joint probability density function
cθ (x1 , x2 ) = 1 + θ(1 − 2x1 )(1 − 2x2 ),
−1 6 θ 6 1.
(13)
One can show that cor(x1 , x2 ) = θ/3. Lai (1978) proved that, for 0 ≤ θ ≤ 1, x1
and x2 are positively quadrant dependent and positively regression dependent.
Moreover,
E(x2 | x1 ) =
1 θ
θ
x1 +
−
.
3
2 6
(14)
The linearity above is very useful for our purpose, as it will allow an explicit
computation for Shapley values in that model.
9
Proposition 2. Let f (x) = xT β for x, β ∈ R2 and x ∼ cθ (x1 , x2 ), with −1 ≤
θ ≤ 1. Then
1
φ1
θ2 β12 − β22
=
,
1
+
1
−
σ2
2
9
12σ 2
with σ 2 = (β12 + β22 )/12 + β1 β2 θ/18.
Proof. From the linearity of the regression function (14),
1 θ
θ
E(f (x) | x1 ) = x1 β1 + β2 + β2
,
−
3
2 6
thus
1
θ 2
β1 + β2 .
12
3
Symmetry gets us the corresponding expression for var(E(f (x) | x2 )). Then
Proposition 1 establishes the expression for φ1 /σ 2 . Finally, because var(xj ) =
1/12 and cor(x1 , x2 ) = θ/3, we get σ 2 = (β12 + β22 )/12 + β1 β2 θ/18.
var(E(f (x) | x1 )) =
Now we consider the Farlie-Gumbel-Morgenstern copula, but we assume xj
has as cumulative distribution function Fj , and probability density function Fj0 ,
not necessarily from the uniform distribution.
Lemma 1. Let x ∈ R2 have probability density F10 (x1 )F20 (x2 )cθ (F1 (x1 ), F2 (x2 )),
with −1 ≤ θ ≤ 1. Then
Z
E(x2 | x1 ) = E(x2 ) + θ(1 − 2F1 (x1 )) y(1 − 2F2 (y))F20 (y) dy.
R
For exponential xj with Fj (xj ) = 1 − exp(−λj xj ) for λj > 0, we get
E(x2 | x1 ) =
1
θ
+
(1 − 2e−λ1 x1 ).
λ2
2λ2
(15)
Proof. Crane and Hoek (2008).
Next we assume that x has exponential margins and we transform these
margins to be unit exponential by making a corresponding scale adjustment to
β. From Section 4.2, we know that such transformations do not change the
Shapley value.
Proposition 3. Let f (x) = xT β for x, β ∈ R2 where x has probability density
function e−x1 −x2 cθ (1 − e−x1 , 1 − e−x2 ), where −1 ≤ θ ≤ 1. Then
θ2 β12 − β22
φ1
1
1+ 1−
=
2
σ
2
12
σ2
with σ 2 = β12 + β22 + θβ1 β2 /2.
10
(16)
Proof. From Lemma 1, E(x2 | x1 ) = 1 + θ/2 − θe−x1 so
E(f (x) | x1 ) = β1 x1 + β2 (1 + θ/2 − θe−x1 ).
Therefore
var(E(f (x) | x1 )) = β12 + β22 θ2 var(e−x1 ) − 2β1 β2 θcov(x1 , e−x1 ).
Now var(e−x1 ) = E(e−2x1 ) − E(e−x1 )2 = 1/12 and
Z ∞
1
1
xe−2x dx − = − ,
cov(x1 , e−x1 ) =
2
4
0
so var(E(f (x) | x1 )) = β12 + β22 θ2 /12 + β1 β2 θ/2. This establishes (16) by Proposition 1.
Suppose that β1 > β2 > 0. Then of course φ1 /σ 2 > 1/2. Equation (16)
shows that φ1 /σ 2 decreases as θ increases from 0 to 1. It does not approach 1/2
because even at θ = 1, x2 is not a deterministic function of x1 .
4.3.2
Gaussian variables, exponential f , d = 2
Pd
Let x ∼ N (µ, Σ) and take Y = eβ0 + j=1 xj βj . The effect of β0 and µj is
simply to scale Y and so we can take β0 = 0 and µ = 0 without affecting
φj /σ 2 . Next we suppose that the diagonal elements of Σ are nonzero. By the
transformation result in Section 4.2 we can replace each xj by xj /Σjj if need be
without changing φj and so we suppose that each xj ∼ N (0, 1). Here we find
variable importances for d = 2.
T
Proposition
4.
Let
f
(x)
=
exp
x
β
for x, β ∈ R2 and x ∼ N (0, Σ), for
1 ρ
Σ=
. Then
ρ 1
2
2
φ1
1
e(β1 +β2 ρ) − e(β2 +β1 ρ)
=
1
+
,
2
2
σ2
2
eβ1 +β2 +2ρβ1 β2 − 1
(17)
where the variance of f (x) is
2
2
2
2
σ 2 = eβ1 +β2 +2ρβ1 β2 (eβ1 +β2 +2ρβ1 β2 − 1).
(18)
2
Proof. Recall the lognormal moments: if Z ∼ N (µ, σ 2 ) then E(eZ ) = eµ+σ /2
2
2
and var(eZ ) = (eσ − 1)e2µ+σ . Taking Z = xT β we find that Y = eZ has
variance σ 2 given by (18).
The distribution of x2 β2 given x1 is N (ρx1 β2 , (1 − ρ2 )β22 ). Therefore
2
2
E(Y | x1 ) = e(β1 +ρβ2 )x1 +β2 (1−ρ
var(E(Y | x1 )) = e
β22 (1−ρ2 )
e
11
2
(β1 +ρβ2 )
)/2
(e
,
and so
(β1 +ρβ2 )2
− 1)
1.0
0.9
0.8
0.7
0.5
0.6
Importance
0.0
0.2
0.4
0.6
0.8
1.0
Correlation
Figure 1: Relative importance φ1 /σ 2 versus correlation |ρ| from Proposition 2.
From top to bottom, β T is (8, 1), (4, 1), and (2, 1).
= eβ
T
Σβ
2
(e(β1 +ρβ2 ) − 1).
T
2
Similarly, var(E(Y | x2 )) = eβ Σβ (e(β2 +ρβ1 ) − 1). Then applying Proposition 1
T
and noticing that the lead factor eβ Σβ appears also in σ 2 , yields the result.
If ρ = ±1 then φ1 /σ 2 = 1/2 as it must because there is then a bijection
between the variables. The value of φ1 /σ 2 in (17) is unchanged if we replace
ρ by −ρ. The formula is not obviously symmetric, but the fraction within
parentheses there can be divided by the corresponding one for −ρ and the
ratio reduces to 1. More directly, we know from Section 4.2 that making the
transformation x2 → −x2 and β2 → −β2 would leave the variable importances
unchanged while switching ρ → −ρ.
It is clear that for β1 > β2 we must have φ1 /σ 2 > 1/2. Even with the closed
form (17), it is not obvious how φ1 /σ 2 should depend on ρ or on β. Figure 1
shows that increasing |ρ| from zero generally raises the importance of x1 until at
some high correlation level the relative importance quickly drops down to 1/2.
Also, for ρ = 0 the effect of β1 over the range 2 6 β1 6 8 is quite small when
β2 = 1.
The lognormal case is different from the bivariate normal case. There, the
value of φ1 converges monotonically towards 1/2 as |ρ| increases from 0 to 1.
12
p
x1
x2
y
p0
p1
p2
0
1
0
0
0
1
y0
y1
y2
Table 1: The random variable y = f (x) is the given function of x = (x1 , x2 ).
That vector takes three values with the probabilities in this table. For example,
Pr(x = (1, 0)) = p1 and then y = y1 .
4.4
Holes
Here we consider the simplest setting where there is an unreachable part of the x
space. We consider two binary variables x1 and x2 but x1 = x2 = 1 never occurs.
For instance f could be the weight of a sea turtle, x1 could be 1 iff the turtle is
bearing eggs and x2 could be 1 iff the turtle is male. It may seem unreasonable
to even attempt to compare the importance of these variables (male/female
versus eggs/none) but Shapley value does provide such a comparison based on
compelling axioms in the event that we do seek a comparison.
This simplest setting is depicted in Table 1 where p0 + p1 + p2 = 1. We
assume that p1 > 0 and p2 > 0 for otherwise the function does not have two
input variables.
Theorem 3. Let y be a function of the random vector x as given in Table 1.
Assume that σ 2 = var(y) > 0, and min(p1 , p2 ) > 0. Then the Shapley relative
importance of variable x1 is
p0
p1 (1 − p1 )ȳ12 − p2 (1 − p2 )ȳ22
1
1+ 2 ×
2
σ
(1 − p1 )(1 − p2 )
(19)
where ȳj = yj − y0 for j = 1, 2.
Proof. See section 6.2.
We see that when p0 = 0, then the Shapley relative importance of x1 is 1/2.
That is what it must be because there is then a bijection between x1 and x2 via
x1 + x2 = 1.
Now suppose that ȳ1 = ȳ2 . For instance y1 = y2 = 1 while y0 = 0. Then the
more important variable is the one with the larger variance. That is x1 is more
important if p1 (1 − p1 ) > p2 (1 − p2 ). This can only happen if p1 > p2 . So the
more probable input is the more important one in this case.
4.5
Maximum of exponential random variables
Keinan et al. (2004) considered a network of neurons e1 , . . . , ed where the ej have
independent lifetimes xj that are exponentially distributed with mean 1/λj . In
their setting the value of a set of neurons is φ(u) = E(maxj∈u xj ), that is the
13
expected amount of time that at least part of that subset survives. For d = 3,
they give a Shapley value of
φj =
1
1
1
1
1
1
1
−
−
+
,
λ1
2 λ1 + λ2
2 λ1 + λ3
3 λ1 + λ2 + λ3
but they do not give a proof. While value in this example is not based on
prediction error, we include it because it is another example of a closed form
for Shapley value based on random variables. We prove their formula here and
generalize it to any d > 1.
Theorem 4. Let the value of a set u ⊆ 1:d be val(u) = E(maxj∈u xj ) where
x1 , . . . , xd are independent exponential random variables with E(xj ) = 1/λj .
Then
d
X
X
1
(−1)r−1
P
.
φj =
r
`∈w λ`
r=1
w⊆1:d,j∈w,|w|=r
Proof. See section 6.3.
5
Conclusions
The Shapley value from economics remedies the conceptual difficulties in measuring importance of dependent variables via ANOVA. Like ANOVA it uses
variances, but unlike the dependent data ANOVA, Shapley value never goes
negative and it can be defined without onerous assumptions on the input distribution.
We find that Shapley value has useful properties. When two variables are
functionally equivalent, then they get equal Shapley value. When an invertible
transformation is made to a variable, it retains its Shapley value. We thus
conclude that Song et al. (2016) had the right idea proposing Shapley value for
dependent inputs. Computation of Shapley values remains a challenge outside
of special cases like the ones we discuss here.
A potential application that we find interesting is measuring the importance
of parameters in a Bayesian context. When the parameter vector β has an approximate Gaussian posterior distribution, as the central limit theorem often
provides, then Theorem 2 yields a measure φj (x0 ) for the importance of parameter βj for the posterior uncertainty of the prediction xT
0 β. We hasten to
add that parameter independence is quite different from variable importance,
which is a more common goal. By this measure an important parameter is one
whose uncertainty dominates uncertainty in xT
0 β. The corresponding variable
may or may not be important. Another potential application is in modeling
the importance of order statistics. They naturally belong to a non-rectangular
set Lebrun and Dutfoy (2014).
14
Acknowledgments
This work was supported by grant DMS-1521145 from the U.S. National Science
Foundation. We thank Marco Scarsini, Jiangming Xiang, Bertrand Iooss, two
anonymous referees and an associate editor for valuable comments.
References
Castro, J., Gómez, D., and Tejada, J. (2009). Polynomial calculation of
the Shapley value based on sampling. Computers & Operations Research,
36(5):1726–1730.
Chastaing, G., Gamboa, F., and Prieur, C. (2012). Generalized Hoeffding-Sobol’
decomposition for dependent variables – applications to sensitivity analysis.
Electronic Journal of Statistics, 6:2420–2448.
Chastaing, G., Gamboa, F., and Prieur, C. (2015). Generalized Sobol’ sensitivity
indices for dependent variables: Numerical methods. Journal of Statistical
Computation and Simulation, 85(7):1306–1333.
Colini-Baldeschi, R., Scarsini, M., and Vaccari, S. (2016). Variance allocation
and Shapley value. Methodology and Computing in Applied Probability, pages
1–15.
Crane, G. J. and Hoek, J. v. d. (2008). Conditional expectation formulae for
copulas. Australian & New Zealand Journal of Statistics, 50(1):53–67.
Genizi, A. (1993). Decomposition of r2 in multiple regression with correlated
regressors. Statistica Sinica, pages 407–420.
Gilquin, L., Prieur, C., and Arnaud, E. (2015). Replication procedure for
grouped Sobol’ indices estimation in dependent uncertainty spaces. Information and Inference, 4(4):354–379.
Grömping, U. (2007). Estimators of relative importance in linear regression
based on variance decomposition. The American Statistician, 61(2).
Hooker, G. (2012). Generalized functional ANOVA diagnostics for highdimensional functions of dependent variables. Journal of Computational and
Graphical Statistics.
Keinan, A., Hilgetag, C. C., Meilijson, I., and Ruppin, E. (2004). Causal localization of neural function: the Shapley value method. Neurocomputing,
58:215–222.
Kucherenko, S., Klymenko, O. V., and Shah, N. (2016). Sobol’ indices for problems defined in non-rectangular domains. Technical report, arXiv:1605.05069.
15
Lai, C. D. (1978). Morgenstern’s bivariate distribution and its application
to point processes. Journal of Mathematical Analysis and Applications,
65(2):247–256.
Lebrun, R. and Dutfoy, A. (2014). Copulas for order statistics with prescribed
margins. Journal of Multivariate Analysis, 128:120–133.
Lindeman, R. H., Merenda, P. F., and Gold, R. Z. (1980). Introduction to
bivariate and multivariate analysis. Scott Foresman and Company, Glenview,
IL.
Owen, A. B. (2013). Variance components and generalized Sobol’ indices. Journal of Uncertainty Quantification, 1(1):19–41.
Owen, A. B. (2014). Sobol’ indices and Shapley value. Journal on Uncertainty
Quantification, 2:245–251.
Shapley, L. S. (1953). A value for n-person games. In Kuhn, H. W. and Tucker,
A. W., editors, Contribution to the Theory of Games II (Annals of Mathematics Studies 28), pages 307–317. Princeton University Press, Princeton,
NJ.
Sobol’, I. M. (1990). On sensitivity estimation for nonlinear mathematical models. Matematicheskoe Modelirovanie, 2(1):112–118. (In Russian).
Sobol’, I. M. (1993). Sensitivity estimates for nonlinear mathematical models.
Mathematical Modeling and Computational Experiment, 1:407–414.
Song, E., Nelson, B. L., and Staum, J. (2016). Shapley effects for global sensitivity analysis: Theory and computation. SIAM/ASA Journal on Uncertainty
Quantification, 4(1):1060–1083.
Stone, C. J. (1994). The use of polynomial splines and their tensor products in
multivariate function estimation. The Annals of Statistics, 22(1):118–184.
Winter, E. (2002). The Shapley value. Handbook of game theory with economic
applications, 3:2025–2054.
Wu, C. F. J. and Hamada, M. S. (2011). Experiments: planning, analysis, and
optimization. John Wiley & Sons.
6
6.1
Proofs
Proof of Theorem 2
Recall that f (x) = xT β where x ∼ N (µ, Σ). We also assumed that Σ is of full
rank. Now var(x−u | xu ) = Σ−u,−u − Σ−u,u Σ−1
u,u Σu,−u , and so
T
var(f (x) | xu ) = var(xT
u βu + x−u β−u | xu )
16
= var(xT
−u β−u | xu )
T
= β−u
Σ−u,−u − Σ−u,u Σ−1
u,u Σu,−u β−u .
We will use v = v(j, u) ≡ −u − {j}. It
covariance matrix
Σuu Σuj
Σ = Σju Σjj
Σvu Σvj
helps to visualize the partitioned
Σuv
Σjv
Σvv
if the indices have been ordered for those in u to precede j which precedes those
in v. For this section only, we make a further notational compression shortening
u + {j} to u + j. Next
τ 2u+j − τ 2u = var(f (x) | xu ) − var(f (x) | xu+j )
T
= β−u
Σ−u,−u − Σ−u,u Σ−1
u,u Σu,−u β−u
− βvT Σvv − Σv,u+j Σ−1
u+j,u+j Σu+j,v βv .
Using the formula for the inverse of a partitioned matrix, we find that
−1
−1
−Σ−1
Σuu + Σ−1
−1
uu Σuj Dj (u)
uu Σuj Dj (u)Σju Σuu
,
Σu+j,u+j =
−Dj (u)Σju Σ−1
Dj (u)
uu
−1
where Dj (u) = (Σjj − Σju Σ−1
= var(xj | xu )−1 , which exists because Σ
uu Σuj )
has full rank. Continuing,
Σv,u+j Σ−1
u+j,u+j Σu+j,v
−1
−1
Σ−1
−Σ−1
Σuv
uu + Σuu Σuj Dj (u)Σju Σuu
uu Σuj Dj (u)
= Σvu Σvj
−Dj (u)Σju Σ−1
Dj (u)
Σjv
uu
−1
−1
−1
−1
Σuu Σuv + Σuu Σuj Dj (u)Σju Σuu Σuv − Σuu Σuj Dj (u)Σjv
= Σvu Σvj
−Dj (u)Σju Σ−1
uu Σuv + Dj (u)Σjv
−1
−1
−1
= Σvu Σ−1
uu Σuv + Σvu Σuu Σuj Dj (u)Σju Σuu Σuv − Σvu Σuu Σuj Dj (u)Σjv
− Σvj Dj (u)Σju Σ−1
uu Σuv + Σvj Dj (u)Σjv
−1
−1
= Σvu Σuu Σuv + Dj (u) Σvu Σ−1
uu Σuj − Σvj Σju Σuu Σuv − Σjv
= Σvu Σ−1
uu Σuv + Dj (u)cov(xv , xj | xu )cov(xj , xv | xu )
recalling that Dj (u) is a scalar.
Now τ 2u+j − τ 2u is
T
β−u
cov(x−u | xu )β−u − βvT Σvv βv
+ βvT Σvu Σ−1
uu Σuv + Dj (u)cov(xv , xj | xu )cov(xj , xv | xu ) βv
T
= β−u
cov(x−u | xu )β−u − βvT cov(xv | xu )βv
+ Dj (u)βvT cov(xv , xj | xu )cov(xj , xv | xu )βv
= Σjj βj2 + βj Σjv βv + βvT Σvj βj
17
−1
T
−1
− βj2 Σju Σ−1
uu Σuj − βj Σju Σuu Σuv βv − βv Σvu Σuu Σuj βj
+ Dj (u)βvT cov(xv , xj | xu )cov(xj , xv | xu )βv
= βj2 var(xj | xu ) + 2βj cov(xj , xv | xu )βv
+ Dj (u)βvT cov(xv , xj | xu )cov(xj , xv | xu )βv .
Putting this together, the Shapley value of variable j is
φj =
−1
1 X d−1
βj2 var(xj | xu ) + 2βj cov(xj , x−u−j | xu )β−u−j
d
|u|
u⊆−j
T
+ var(xj | xu )−1 β−u−j
cov(x−u−j , xj | xu )cov(xj , x−u−j | xu )β−u−j .
(20)
Writing
T
cov(xj , xT
−u β−u | xu ) = cov(xj , x−u−j β−u−j | xu ) + βj var(xj | xu )
2
we then find that cov(xj , xT
−u β−u | xu ) /var(xj | xu ) equals the factor to the
d−1
right of |u| in (20).
6.2
Proof of Theorem 3
Without loss of generality take y0 = 0. Then µ = p1 y1 + p2 y2 and σ 2 =
p1 y12 + p2 y22 − µ2 .
Now with y0 = 0,
2
y p
2 2
− µ + p1 (y1 − µ)2
var(E(y | x1 )) = (p0 + p2 )
p0 + p2
y p
2
2 2
= (1 − p1 )
− µ + p1 (y1 − µ)2
1 − p1
2 2
p y
= 2 2 − 2µy2 p2 + µ2 (1 − p1 ) + p1 (y1 − µ)2
1 − p1
p2 y 2
= 2 2 − 2(p1 y1 + p2 y2 )y2 p2 + (p1 y1 + p2 y2 )2 (1 − p1 ) + p1 (y1 (1 − p1 ) − p2 y2 )2
1 − p1
p2
2
− 2p22 + p22 (1 − p1 ) + p1 p22
= y22
1 − p1
2
+ y1 p21 (1 − p1 ) + p1 (1 − p1 )2
+ y1 y2 −2p1 p2 + 2p1 p2 (1 − p1 ) − 2p1 p2 (1 − p1 )
p2
2
= y22
− p22 + y12 p1 (1 − p1 ) − 2y1 y2 p1 p2
1 − p1
2
p
1 p2
= y22
+ y12 p1 (1 − p1 ) − 2y1 y2 p1 p2 .
1 − p1
18
Then var(E(y | x1 )) − var(E(y | x2 )) equals
p1 p22
p2 p21
+ y12 p1 (1 − p1 ) − y12
− y22 p2 (1 − p2 )
1 − p1
1 − p2
p p2
p2 p21
1 2
= y22
− p2 (1 − p2 ) + y12 p1 (1 − p1 ) −
1−p
1 − p2
p p
p p 1
0 2
0 1
2
2
− y2
.
= y1
1 − p2
1 − p1
y22
Finally, the relative importance of variable x1 is
p0 p1
p0 p2
y12 1−p
− y22 1−p
1
1
p0 y12 p1 (1 − p1 ) − y22 p2 (1 − p2 )
2
1
1+
=
1
+
2
σ2
2
σ2
(1 − p1 )(1 − p2 )
1
p0 p1 y12
p2 y22
=
1+ 2
−
.
2
σ 1 − p2
1 − p1
6.3
Proof of Theorem 4
Recall that the random vector x ∈ [0, ∞)d has independent components xj .
They are exponentially distributed and E(xj ) = 1/λj for 0 < λj < ∞. Let
Mu = maxj∈u xj and define value val(u) = E(Mu ). Our first step is to evaluate
the expected value of a maximum of independent not identically distributed
exponential random variables.
Proposition 5.
X
E(Mu ) =
Q
Pr(xj < x) =
j∈u
∞
Z
0
0
=
(1 − e−λj x ) dx
X
X
X
(−e−λj x ) dx
v⊆u
(−1)|v|−1
Z
∞
e−x
P
j∈v
0
∅6=v⊆u
=
− e−λj x ). Then,
Y
1 −
=
j∈u (1
.
j∈u
∞
Z
Q
1 −
E(Mu ) =
λj
j∈v
∅6=v⊆u
Proof. First Pr(Mu < x) =
1
(−1)|v|−1 P
(−1)|v|−1 P
1
j∈v
∅6=v⊆u
Using Proposition 5 we get Shapley value
−1
1 X
d−1
(val(u + j) − val(u))
φj =
d
|u|
u⊆−{j}
19
λj
.
λj
dx
!
X d − 1−1
X
X
|v|−1 P 1
|v|−1 P 1
−
(−1)
(−1)
|u|
`∈v λ`
`∈v λ`
∅6=v⊆u
∅6=v⊆u+j
u⊆−{j}
−1 X
1 X
d−1
1
=
(−1)|w| P
.
|u|
d
`∈w+j λ`
1
=
d
w⊆u
u⊆−{j}
Introducing the ‘slack variable’ v with u = v + w,
1
φj =
d
=
1
d
X
(−1)
|w| P
(−1)|w| P
λ`
v⊆−{j}−w
d−1−|w|
1
X
`∈w+j
w⊆−{j}
X
`∈w+j
w⊆−{j}
X
1
λ`
r=0
d−1
|v + w|
−1
d − 1 − |w|
r
.
d−1
r + |w|
d−|w|
X d − |w| . d − 1
1 X
|w|−1 P 1
(−1)
.
=
d w:j∈w
r
r − 1 + |w|
`∈w λ` r=0
The following diagonal sum identity for binomial coefficients will be useful:
A
X
L+r
r=0
r
=
A+L+1
.
A
Using that identity at the third step below,
d−|w|
X
r=0
d − |w|
r
d−|w|
X (d − |w|)! .
(d − 1)!
d−1
=
r!
(r
−
1 + |w|)!
r − 1 + |w|
r=0
.
d−|w|
X r − 1 + |w|
(d − |w|)!
(|w| − 1)!
r
(d − 1)!
r=0
(d − |w|)!
d
=
(|w| − 1)!
(d − 1)!
d − |w|
d
=
.
|w|
=
As a result,
φj =
X
w:j∈w
1
1
(−1)|w|−1 P
|w|
`∈w λ`
which after slight rearrangement gives the conclusion of Theorem 4.
20
| 10math.ST
|
arXiv:1211.5520v1 [cs.CE] 23 Nov 2012
Accurate Demarcation of Protein Domain Linkers
based on Structural Analysis of Linker Probable
Region
Vivekanand V. Samant∗
Arvind Hulgeri†
Alfonso Valencia‡
Ashish V. Tendulkar§¶
Abstract
In multi-domain proteins, the domains are connected by a flexible
unstructured region called as protein domain linker. The accurate demarcation of these linkers holds a key to understanding of their biochemical
and evolutionary attributes. This knowledge helps in designing a suitable linker for engineering stable multi-domain chimeric proteins. Here
we propose a novel method for the demarcation of the linker based on
a three-dimensional protein structure and a domain definition. The proposed method is based on biological knowledge about structural flexibility
of the linkers. We performed structural analysis on a linker probable region (LPR) around domain boundary points of known SCOP domains.
The LPR was described using a set of overlapping peptide fragments of
fixed size. Each peptide fragment was then described by geometric invariants (GIs) and subjected to clustering process where the fragments
corresponding to actual linker come up as outliers. We then discover the
actual linkers by finding the longest continuous stretch of outlier fragments from LPRs. This method was evaluated on a benchmark dataset of
51 continuous multi-domain proteins, where it achieves F1 score of 0.745
(0.83 precision and 0.66 recall). When the method was applied on 725 continuous multi-domain proteins, it was able to identify novel linkers that
were not reported previously. This method can be used in combination
with supervised / sequence based linker prediction methods for accurate
linker demarcation.
1
Introduction
Complex proteins are made up of several domains that work independently or in
tandem with the neighboring domains to perform the intended functions in the
∗ Persistent
Systems Ltd. Pune
Systems Ltd. Pune
‡ Spanish National Cancer Research Center
§ Tata Institute of Fundamental Research
¶ Corresponding Author: [email protected]
† Persistent
1
cellular processes [18]. The domains are linked by means of flexible structures
known as domain linkers. The linkers perform a key role in cooperative interdomain interactions, function regulation, protein stability, folding rates, and
domain-domain orientation [18, 16]. The linkers are known to possess special
biochemical properties such as high solvent accessibility and a typical amino
acid composition, due to their role and location in the protein structure. To
further our understanding on these fronts, a systematic analysis of the known
linkers needs to be performed. The progress is hampered by lack of availability of
known and reliable linkers. For instance, there is no database of experimentally
characterized linkers something that would be of immense importance in such
studies. The improved understanding of linkers and their biochemical properties
is crucial in designing linkers for engineering stable multi-domain proteins.
The most reliable and accurate linker demarcation can be obtained using
protein structure analysis. Crystallographers usually perform such analysis to
identify domains and linkers while determining the structure of multi-domain
proteins. However, in many cases, the domain linkers are not reported explicitly and we need to employ computational methods to demarcate the linkers
based on sequence or structure of the protein. State of the art sequence based
methods [37, 4, 11] can be used to identify a list of putative linkers and these
need to be processed further using available structural features to determine the
actual linkers. These methods take amino acid sequence as an input and predict domain boundaries and linkers using a domain linker index computed from
amino acid propensities in the known linker region. Miyazaki and co- workers have proposed neural network [26] and support vector machine [12] based
techniques using amino acid propensities to distinguish intra-domain loops from
the inter-domain ones. Tanaka and co-workers used predicted secondary structure in addition to amino acid propensities to identify loops, which are further
distinguished between linker and non-linker loops. Domain prediction methods are also used to predict linkers by carving out a stretch of residues in the
inter domain region around domain boundary points [14, 25]. These methods
tend to provide multiple linker predictions with liberal allowance for the linker
boundaries and hence are not very useful for accurate protein linker demarcation. Besides, most of these methods are unable to predict helical linkers due
to their assumption about linkers being loops.
George and Heringa conducted a systematic study of biochemical properties of the linkers extracted from three dimensional structures of multi-domain
proteins [16, 15]. They first identify structural domains using Taylors method
[41] and then extract linkers by branching out from domain boundaries until
the branches become buried within the core of the domain or till the branch
becomes 40 residues long. This method takes into account biochemical properties of linkers for their demarcation without using any of the structural features.
It is well known that the linkers assume unique structures due to their placement in the protein structure [2, 33] and this forms the basis of our method.
The proposed method performs accurate demarcation of linkers given a three
dimensional structure and its domain definition. It first extracts a linker probable region (LPR) around domain boundary point and then performs structural
2
analysis of the LPR to demarcate actual linker. We perform a rigorous assessment of the method using a benchmark dataset of known linkers extracted from
the literature.
The rest of the paper is organized as follows: (i) The Method section explains
the proposed technique in detail, (ii) The Results section presents findings and
representative linkers demarcated by the proposed method. It also reports accuracy of the method and its performance vis-a-vis other state of the art methods,
(iii) The Discussion section documents key contributions of the method.
2
Proposed method
The method takes a set of protein structures and the corresponding domain
definitions as input and identifies the corresponding domain linkers.
It can be broadly divided into the following four steps, as depicted in Figure
1, (i) Construction of linker probable regions (LPRs), (ii) Parameterization of
LPRs, (iii) Generation of structure profiles by clustering LPRs, and (iv) Demarcation of actual linkers by applying dynamic programming on the structure
profiles. Note that the current version of the method works only with continuous
multi-domain proteins. The algorithm is given below:
We explain each of these steps in greater detail in the rest of the section.
2.1
Construction and parameterization of LPRs
Line 1–6 in our algorithm constructs a set of linker probable regions (LPRs), R,
from the input set of protein structures along with their domain definitions S.
We then represent each LPR i.e. r ∈ R using a set of overlapping tertrahedrons
as described below.
Let S be the set of protein structures along with their domain definitions.
Each element of S is an ordered pair of structure and its domain definition.
Note that since we consider only continuous multidomain proteins in our analysis, we are in a position to define domains using the position of the last amino
acid residue in the domain. We will refer to the position of last amino acid as
the endpoint of that domain. The set D in ordered pair (S, D) ∈ S specifies
endpoints of each domain in S. Thus, for a given structure S with e domains,
D = {d1 , d2 , . . . , de }. Here dj is the endpoint of domain j. Note that the first
domain starts at the first position and the last domain ends at the last position
in the protein. Any other domain j with j > 1 starts at position dj−1 + 1 and
ends at position dj in the structure. With this background, we are in a position
to define LPR.
DEFINITION 1: Linker probable region (LPR) between domain i and j of
protein s is a substructure starting at position di − k + 1 and ending at di + k in
s. It is denoted as LPR(s, i, j). Note that LPR(s, i, j) contains the end poisition
di of domain i and its length is 2k. The parameter k is chosen based on the
average linker length as reported in literature [26, 15, 39]. LPR is the basic
3
Algorithm 1 Linker Demarcation
Require: S = {(S1 , D1 ), (S2 , D2 ), . . . , (Sn , Dn )}: Set of n protein structures
with domain definition; and k: Number of positions from one domain to be
included in linker probable region
Ensure: L = {(S1 , L1 ), (S2 , L2 ), . . . , (Sn , Ln )}.
1: for each (Si , Di ) ∈ S do
2:
R = ExtractLPR(Si , Di );
3:
T = DiscretizeLPR(R);
4:
T =T +T
5:
R=R+R
6: end for
7: I = InvariantList()
8: for each T ∈ T do
9:
X = X + InvariantGeneration(T )
10: end for
11: Xz = Standardize(X )
12: Xpc = PCA(Xz )
13: C = Cluster(Xpc )
14: for each C ∈ C do
15:
E = AssignEval(C)
16: end for
17: for each T ∈ T do
18:
U = ComputeSUS(T , E)
19: end for
20: for each R ∈ R do
21:
LPRProfile = ConstructProfile(T , U)
22:
L = GetMaximalScoringSubsequence(LPRProfile);
23:
L=L+L
24: end for
25: return L
4
unit of our analysis.
EXAMPLE 1: Let sj be the protein structure with two domains. Let Dsj =
{d1 , d2 }. sj has exactly one LPR that starts at position d1 − k + 1 and ends at
position d1 + k.
We construct a set of LPRs, R by extracting LPRs from S, the input set of
proteins and their domain definitions (line – 2 in algorithm). Note that all
the LPRs in R are of equal length 2k. The backbone structure of each LPR
is approximated with its Cα coordinates [42]. Thus, R is a set of 2k amino
acid residues along with their positions in the structure as given by the x, y, z
coordinates. Now we will describe a procedure DiscretizeLPR (line – 3 in our
algorithm). Each r ∈ R is discretized into a sequence of 2k − 3 overlapping
tetrapeptides T = t1,2,3,4 , t2,3,4,5 , . . . , t(2k−3),(2k−2),(2k−1),(2k) . Note that the
consecutive tetrapeptides ti and ti+1 in sequence T share an overlap of three
amino acid residues. Each tetrapeptide in T is added to T , which is a global
set of tetrapeptides obtained by discretizing LPRs (line – 4 in algorithm).
Each tetrapeptide t ∈ T represents a tetrahedral geometry and is described
by a fixed suite of g descriptors, which are invariant under transformations such
as rotation and translation [28, 45, 44, 42]. These descriptors are referred to
as geometric invariants (GIs) in the subsequent text. The suite of invariants is
carefully chosen after extensive trial and error on training data to address the
following two issues: (a) for superimposable tetrapeptides, the invariants must
be similar within a tolerance limit δ; and (b) for a pair of non-superimposable
tetrapeptides t1 and t2 , there must be at least one geometric invariant such that
f (t1 ) is not similar to f (t2 ). Here f is a function that calculates a specific GI.
We represent each tetrapeptide t ∈ T with a suite of fifteen GIs (line – 7). The
detailed method for calculating these GIs is given in our previous work [44, 42]
(line – 9).
1. Nine GIs are calculated based on the tetrahedral geometry of t and they
represent signed volume and perimeter of t, length of each edge in t. Since
there are in all six edges so we have six GIs corresponding to the length.
One more invariant is computed based on the sum of distance of each
vertex of t from the centroid of all the vertices. Let Vt be the set of all
vertices in t. The i-th vertex vi ∈ Vt gives x, y, z coordinate position of
i-th amino acid residue in t. The centroid is calculated as follows:
1 X
v
µ=
n
∀v∈Vt
and the sum of distance from centroid is calculated as
X
(v − µ)
∀v∈Vt
2. The remaining six invariants for t are calculated by forming three triangles
using vertices in Vt . The three traingles are as follows: v1 , v2 , v3 , v1 , v3 , v4
5
and v1 , v2 , v4 . We calculate area and perimeter for each of these traingles,
thus accounting for the six remaining invariants.
.
Further we standardize X to zero mean and unit standard deviation values.
Let Xz be the set of standardized GIs for T (Line – 11). We perform Principal
Component Analysis (PCA) to get rid of correlations between GIs [43]. PCA
gives a new set of orthogonal dimensions, which are linear combinations of the
original dimensions (GIs in this case). We selected first m significant principal components (PCs) to represent the tetrapeptides. Let Xpc be the set of
tetrapeptides represented using m PCs (Line – 12).
2.2
Structural profiling of LPRs
Since the linkers are unstructured regions, we believe that the corresponding
tetrapeptides share structural similarity with fewer other tetrapeptides. On
the other hand, the tetrapeptides from non-linker region are expected to share
structural similarity with a large number of other tetrapeptides. Our objective is
to determine the groups of structurally similar tetrapeptides through clustering
process and utilize this knowledge towards the demarcation of actual linkers.
We perform clustering of a set of tetrapeptides Xpc , which are represented using
m PCs as explained earlier (Line – 13). The clustering is carried out via Matlab
implementation of hierarchical agglomerative clustering (HAC) [38]. We use
euclidean distance as a measure of similarity and ward linkage [38] for merging
the nodes in the clustering tree. The optimal cut in the resulting dendrogram
is determined by using inconsistency parameter, which leads to the discovery
of a set of clusters C. The inconsistency parameter compares each link in the
cluster hierarchy with the adjacent links to determine natural cluster division
in the dataset [38]. The clustering process assigns each tetrapeptide to exactly
one cluster.
Once the clustering process is over, we obtain the distribution of cluster sizes,
which is used to assign e-value to each cluster based on its size (Lines 14–16).
The e-value for a cluster C ∈ C with size |C| is calculated as α/|C|, where α be
the number of clusters in C with size greater than |C| and |C| is the total number
of clusters. Note that the large clusters are expected to contain tetrapeptides
corresponding to the non-linker regions, while the smaller clusters are more
likely to contain tetrapeptides corresponding to the linker region. The e-values
are normalized to zero mean and unit standard deviation to obtain structural
uniqueness score (SUS) of the cluster. The large clusters have lower SUS, while
the smaller clusters have higher SUS. The smallest SUS is assigned to the largest
clusters, while the largest SUS is assigned to the singleton clusters. Thus, the
SUS indicates the structural uniqueness of the cluster and its propensity to be
a part of the actual linker. Each tetrapeptide in the cluster is assigned the SUS
of that cluster (Lines 17–19). The structural profile of an LPR is represented
using the SUS of its constituent tetrapeptides (Line 21).
6
2.3
Protein domain linker demarcation
Given the structural profile of an LPR, we are interested in finding the longest
continuous stretch of tetrapeptides with the highest cumulative SUS. Note that
such tetrapeptides; being highly unstructured; appear as outliers in the clustering process. Hence, the stretch is demarcated as a linker between the domains.
We are required to enumerate all possible stretches in order to find the one with
the greatest cumulative SUS. The problem is tackled by using linear time dynamic programming algorithm proposed by Ruzzo and Tompa [?] (Line – 22).
The algorithm takes a sequence of real numbers as input and generates nonoverlapping, contiguous subsequences having greatest total score. Here, the
algorithm takes the structural profile of an LPR as input, which is a sequence
of nine SUS scores of the constituent tetrapeptides (u1 , u2 , . . . , u2k−3 ), where
ui ∈ R. Let Q be the set of all possible subsequences of tetrapeptides. The
cumulative SUS for each subsequence is obtained by simply summing the SUS
of the constituent tetrapeptides. Let CumSUS(q) be the function that gives cumulative score for the subsequence q ∈ Q. The GetMaximalScoringSubsequence
procedure finds the subsequence with the greatest cumulative SUS (Line – 22).
We declare such a subsequence as a domain linker. In case of a tie, a subsequence with the closest proximity to the domain boundary is declared as a
domain linker. Thus,
L = argmaxq∈Q CumSUS(q)
2.4
Evaluation of proposed method
It is of interest to evaluate the accuracy of the proposed method. In the absence
of linker database and due to a lot of subjectivity in linker detection by visual
examination, we decided to extract experimentally reported linkers from the
literature. We first selected research papers based on PDB reference record
of each protein in our input dataset. We then manually read the literature
for extracting information about experimentally detected linkers. We succeded
in extracting linker information about 51 proteins out of 725 proteins in the
input set (Supplementary Table 1). These linkers form an evaluation set for the
benchmark studies.
After demarcating the linkers using the proposed method, we compare them
with the literature reported linkers. We compute the accuracy of demarcation
residue wise as follows: If the residue marked as a part of the linker also happens
to be the part of the literature reported linker, we count it as a true positive
match, else it is counted as a false positive match. If the residue that is part
of the literature reported linker, but is not present in the linker marked by
the proposed method, it is counted as a false negative match. Let T P denotes
the number of correctly demarcated linker residues, F P denotes the number of
incorrectly demarcated linker residues, which are actually non-linker residues
and F N denotes the number of actual linker residues, which were not included
in demarcated linker region. Based on T P, T N and F N , we compute precision
7
and recall of the proposed method as follows:
TP
TP + FN
TP
Precision =
TP + FP
Recall =
We also compute F 1 measure, which is harmonic mean of precision and recall,
for the proposed method.
3
3.1
Results
Dataset preparation and structural profiling of LPRs
We have selected 610 continuous multi-domain proteins from the ASTRAL 40
[8] dataset version 1.69. Out of 610 selected proteins, we have 505 two domain,
95 three domain and 10 four domain proteins. Based on the SCOP [29] domain
definition, we selected a stretch of 6 amino acids on either side of the domain
boundary point to extract LPR of length 12 for each domain connection. Thus,
we obtain 725 LPRs from the input protein domains. Each LPR is represented
by nine overlapping tetrapeptides with an overlap of three residues between the
consecutive tetrapeptides. Each tetrapeptide is represented with fifteen geometric invariants (GIs) as described earlier. Thus, we obtain 6525 tetrapeptides
represented in 15 dimensional space spanned by GIs. This dataset is subjected
to PCA, which reveals that the first 8 PCs cover 99% variance in the data. The
tetrapeptides were then transformed into a reduced dimensional space spanned
by first 8 PCs. The transformed dataset of tetrapeptides is subjected to hierarchical clustering algorithm (Matlab implementation). The resulting dendrogram
was cut based on inconsistency parameter to obtain 2188 clusters. The distribution of clusters in terms of their size is shown in Table 1. Note that we obtain
a large number of smaller clusters, approximately 50%, with size less than three
members. The largest cluster contains 14 tetrapeptides.
The larger clusters are assigned smaller e-values, while the smaller clusters
are assigned larger e-values. We then constructed the structural profile of LPRs
using the SUS of the corresponding tetrapeptides. The structural profiles, each
of length nine, were subjected to a maximally scoring subsequence finding algorithm to demarcate the actual linkers. We were able to demarcate 692 domain
linkers from 725 input LPRs. In the remaining 33 cases, we observed that these
LPRs contain tetrapeptides with lower SUS. The distribution of linker lengths
is shown in Table 2. We found that the average length of the linker detected by
the proposed method is 5.3 residues.
3.2
Comparison with other methods
We were interested in comparing the proposed method against the state of the
art methods to assess its performance. We used 51 literature reported linkers for the comparative analysis. The same set was used for evaluating the
8
Cluster Size
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Number of Clusters
207
899
520
269
131
58
45
23
13
9
3
4
4
3
Table 1: Distribution of clusters of tetrapeptides according to their size.
Linker Length
Linker Count
4
286
5
178
6
93
7
67
8
31
9
19
10
11
11
7
Table 2: Distribution of linkers according to their lengths. The linkers are obtained via the
proposed method
proposed method. We have selected the following methods in the comparison
study: Ebina et al. [12], GM[14] and CHOP [25]. Note that the direct comparison is inappropriate since most of these methods predict putative domain
linkers from the sequence characteristics, while our method demarcates domain
linkers by analyzing structural characteristics of LPRs. Moreover most of these
methods predict multiple putative linkers with certain flexibility on start and
end positions. From these predictions, we selected the most appropriate linker
based on the known domain definition and used it for the comparison. A representative examples of the linkers identified by different methods on the input
set are reported in Table 3. The complete list can be obtained from Supplementary Table 1. These predictions are matched with the actual linkers and
the accuracy is calculated in terms of F1 score, which is the harmonic mean
of precision and recall. The comparative performance of the proposed method
is given in Table 4. The proposed method achieves overall recall of 0.66 and
precision of 0.83 on the benchmark dataset. It significantly outperforms state
of the art methods in terms of the number of linkers identified as well as the
accuracy of the predictions.
We further compared our method against DomCut, which predicts the domain cut point based on domain linker index. Note that DomCut does not
predict the start and the end position of linker. The DomCut prediction is
9
taken as a correct prediction if the predicted domain cut point falls within the
actual linker. Out of 51 linkers, we found that DomCut predicts correctly in 13
cases and does not predict the domain cut point in 14 cases. In the remaining
cases, the predicted cut point does not fall inside the actual linker.
We computed the agreement between our method against the linker database
of George and Heringa [16, 15], which gives linker predictions for 79 proteins in
our input dataset. Note that linker prediction is available for few proteins from
evaluation set used earlier and hence it is not used for the comparison between
the two methods. We found that the predictions partially agree in 55 cases and
completely disagree in 24 cases. Reasonable agreement (> 75%) was obtained
in 5 cases, while medium agreement (< 75% and > 40%) was obtained in 29
cases and weak agreement was observed in the remaining cases.
3.3
Representative linkers
The examples of demarcated linkers by the proposed method are shown in Figures 2A2J. Since the method use LPR for demarcation, the entire 3-D structure
of corresponding domains is not shown. Instead, a stretch of 16 amino acids
on either side of the domain boundary point is shown to maintain clarity of
representation. Here we describe a few representative linkers.
Example of literature reported linkers, also demarcated by our method
1. Streptococcus pneumonia SP14.3 (PDB Code: 1IB8)
Streptococcis pneumonia is a deadly human pathogen causing high mortality and morbidity rates [49]. SP14.3 is a key protein responsible for
growth of the pathogene. The three-dimensional structure of SP14.3 contains a very short linker of size 3 between residues 88-90 with a moderate
flexibility. Our method predicts the linker exactly at the same place as
reported by Yu et. al. [49]. The linker plays role in relative orientation
of domains and maintaining rotational cooorelation of domains.
2. Yeast Sec18p (PDB Code: 1CR5)
Yeast Sec18p is a hexameric ATPase with a central role in vesicle trafficking [3]. The reported linker is located between residues 104-113 and
is flexible in its structure. The SNAP binding site is located opposite to
the linker. It connects two beta rich sub domains and is likely to facilitate
different sub-domain orientation. Our method detected the linker between
residues 104-111, which is enclosed within the literature reported linker
(Figure 2A).
3. TnsA (PDB Code: 1F1Z)
TnsA carries out DNA breakage at 5 end of transposon. It contains a
six sized loop linker between residues 165-170 that connects two domains
of homodimeric endonuclease enzymes [20]. Our method predicted eight
sized linker between residues 164-171 (Figure 2B). The linker is likely to
play a role in cooperative domain binding.
10
Examples of novel linkers that are not reported in literature
Figure 2G shows the novel linker in Ornithine transcarbamoylase (1DUV) which
is not reported in the literature. Fascin (1DFC) has three domains, and the two
linkers delimiting these domains were demarcated accurately (Figures 2H and
2I). Spectrin beta chain (1S35) is an all alpha protein with an alpha-helical
linker between two domains (Figure 2J). A helical conformation of the linker
region is compatible with a variety of different twist angles between Spectrin
repeats, leading to many different conformations.
4
Discussion
We proposed a novel objective method for accurate demarcation of linkers. The
method takes three dimensional protein structure and domain definition as an
input and provides accurate linker demarcation. This is the first instance where
structural aspects are rigorously analyzed in the linker demarcation task. The
earlier methods have used biochemical and sequence properties for the same
task. Since the proposed method provides structural perspective in the demarcation, it can be used in tandem with the other methods reported in literature.
As stated earlier, accurate domain linker demarcation is a key to understand their biochemical properties. Given a three dimensional structure and
its domain definition, the linkers can be demarcated either through direct visualization or through objective automated methods reported in the literature
[16, 26, 37, 4, 11, 12]. Visualization methods are often subjective, while automated methods demarcate linkers only approximately. The proposed method
demarcates the linker more precisely than the other methods as demonstrated
on the benchmark dataset. Our method outperforms other methods with F1
score of 0.745 (precision 0.83 and recall 0.66) on the benchmark dataset.
The method is also the first of its kind in exploiting biological knowledge
about structural uniqueness of linkers. Since the linkers possess flexible structure, their constituent fragments are unique and appear as outliers during clustering process [47]. Since the outliers are assigned the maximum SUS, the
stretch of fragments with maximum cumulative SUS corresponds to the actual
linker. The discovery of such stretch was performed using an efficient linear
time dynamic programming algorithm.
The proposed method has following configurable parameters, which can be
adjusted to achieve desired results: (i) k which affects the length of LPRs and
(ii) the length of the peptide fragment. The current study uses LPRs of length
12, tetrahedron as a choice for local structure and clusters of tetrahedrons in
LPRs. The length of LPR was decided based on the prior reports of average
linker lengths [16, 26, 39]. Flexibility can be added to LPR selection with
the help of other methods reported in the literature. For instance, we can
use Taylors method [41] as applied by [16] to come up with more appropriate
LPRs. The LPRs thus obtained can be processed further with the help of
the proposed method to demarcate accurate linkers. In the present study, we
extract tetra-peptides from LPRs and perform the clustering. The clusters are
11
then assigned SUS based on cluster size distribution. Our method is able to
find linkers with irregular or unique structure. It is also able to detect linkers
containing alpha-helices and beta-strands with structural purturbations. Due
to these purturbations, these tetrapeptides are part of smaller clusters, which
are often assigned higher SUS and hence our method is able to detect linkers
containing such structures. However, our method is unable to detect linkers
made up of regular alpha-helices and beta-strands, since the regular structures
tend to form larger clusters and usually have smaller SUS compared to the
irregular structures.
Finally, we plan to construct a database of linkers demarcated via the proposed method. The database will help to further our understanding of biochemical properties of linkers and help to design better linkers while engineering
multi-domain proteins. We are planning to use insights about linkers from this
work to develop a sequence based linker prediction method. This can be useful
in predicting protein domains by virtue of linker prediction.
5
Acknowledgement
We are grateful to Prof. Pramod Wangikar - IIT Bombay and Mr. Nivruti
Hinge for their useful discussions and timely help. We thank Michael Tress for
his useful comments on the manuscript. This work is supported by Innovative
Young Biotechnologist Award (IYBA) grant from Department of Biotechnology
of Government of India to AVT.
References
[1] A. Aliverti, R. Faber, C. M. Finnerty, C. Ferioli, V. Pandini, A. Negri,
P. A. Karplus, and G. Zanetti. Biochemical and crystallographic characterization of ferredoxin-NADP(+) reductase from nonphotosynthetic tissues.
Biochemistry, 40:14501–14508, 2001.
[2] P. Argos. An investigation of oligopeptides linking domains in protein
tertiary structures and possible candidates for general gene fusion. J Mol
Biol, 211:943–958, 1990.
[3] S. M. Babor and D. Fass. Crystal structure of the Sec18p N-terminal
domain. Proc Natl Acad Sci U S A, 96:14759–14764, 1999.
[4] K. Bae, B. K. Mallick, and C. G. Elsik. Prediction of protein interdomain
linker regions by a hidden Markov model. Bioinformatics, 21:2264–2270,
2005.
[5] J. J. Barycki, L. K. O’Brien, A. W. Strauss, and L. J. Banaszak. Sequestration of the active site by interdomain shifting. Crystallographic and
spectroscopic evidence for distinct conformations of L-3-hydroxyacyl-CoA
dehydrogenase. J Biol Chem, 275:27186–27196, 2000.
12
[6] P. G. Board, M. Coggan, G. Chelvanayagam, S. Easteal, L. S. Jermiin,
G. K. Schulte, D. E. Danley, L. R. Hoth, M. C. Griffor, A. V. Kamath, M. H.
Rosner, B. A. Chrunyk, D. E. Perregaux, C. A. Gabel, K. F. Geoghegan,
and J. Pandit. Identification, characterization, and crystal structure of the
Omega class glutathione transferases. J Biol Chem, 275:24798–24806, 2000.
[7] L. Bousset, H. Belrhali, R. Melki, and S. Morera. Crystal structures of
the yeast prion Ure2p functional region in complex with glutathione and
related compounds. Biochemistry, 40:13564–13573, 2001.
[8] S. E. Brenner, P. Koehl, and M. Levitt. The ASTRAL compendium for
protein structure and sequence analysis. Nucleic Acids Res, 28:254–256,
2000.
[9] Z. W. Chen, M. Koh, G. Van Driessche, J. J. Van Beeumen, R. G. Bartsch,
T. E. Meyer, M. A. Cusanovich, and F. S. Mathews. The structure of flavocytochrome c sulfide dehydrogenase from a purple phototrophic bacterium.
Science, 266:430–432, 1994.
[10] S. C. Cordell, R. E. Anderson, and J. Löwe. Crystal structure of the
bacterial cell division inhibitor MinC. EMBO J, 20:2454–2461, 2001.
[11] M. Dumontier, R. Yao, H. J. Feldman, and C. W. V. Hogue. Armadillo:
domain boundary prediction by amino acid composition. J Mol Biol,
350:1061–1073, 2005.
[12] T. Ebina, H. Toh, and Y. Kuroda. Loop-length-dependent SVM prediction
of domain linkers for high-throughput structural proteomics. Biopolymers,
92:1–8, 2009.
[13] M. D. Feese, B. P. Ingason, J. Goranson-Siekierke, R. K. Holmes, and W. G.
Hol. Crystal structure of the iron-dependent regulator from Mycobacterium
tuberculosis at 2.0-A resolution reveals the Src homology domain 3-like fold
and metal binding function of the third domain. J Biol Chem, 276:5959–
5966, 2001.
[14] O. V. Galzitskaya and B. S. Melnik. Prediction of protein domain boundaries from sequence alone. Protein Sci, 12:696–701, 2003.
[15] R. A. George and J. Heringa. An analysis of protein domain linkers: Their
classification and role in protein folding. Protein Eng, 15:871–879, 2002.
[16] R. A. George and J. Heringa. SnapDRAGON: a method to delineate protein
structural domains from sequence data. J Mol Biol, 316:839–851, 2002.
[17] C. Gibbons, M. G. Montgomery, A. G. Leslie, and J. E. Walker. The
structure of the central stalk in bovine F(1)-ATPase at 2.4 A resolution.
Nat Struct Biol, 7:1055–1061, 2000.
13
[18] R. S. Gokhale and C. Khosla. Role of linkers in communication between
protein modules. Curr Opin Chem Biol, 4:22–27, 2000.
[19] S. J. Harrop, M. Z. DeMaere, W. D. Fairlie, T. Reztsova, S. M. Valenzuela,
M. Mazzanti, R. Tonini, M. R. Qiu, L. Jankova, K. Warton, A. R. Bauskin,
W. M. Wu, S. Pankhurst, T. J. Campbell, S. N. Breit, and P. M. Curmi.
Crystal structure of a soluble form of the intracellular chloride ion channel
CLIC1 (NCC27) at 1.4-A resolution. J Biol Chem, 276:44993–45000, 2001.
[20] A. B. Hickman, Y. Li, S. V. Mathew, E. W. May, N. L. Craig, and F. Dyda.
Unexpected structural diversity in dna recombination: the restriction endonuclease connection. Mol Cell, 6:1025–1034, 2000.
[21] R. H. Jacobson, A. G. Ladurner, D. S. King, and R. Tjian. Structure
and function of a human TAFII250 double bromodomain module. Science,
288:1422–1425, 2000.
[22] P. D. Jeffrey, L. Tong, and N. P. Pavletich. Structural basis of inhibition
of CDK-cyclin complexes by INK4 inhibitors. Genes Dev, 14:3115–3125,
2000.
[23] L. Jin, B. Stec, and E. R. Kantrowitz. A cis-proline to alanine mutant of
E. coli aspartate transcarbamoylase: Kinetic studies and three-dimensional
crystal structures. Biochemistry, 39:8058–8066, 2000.
[24] P. Li, D. L. Morris, B. E. Willcox, A. Steinle, T. Spies, and R. K. Strong.
Complex structure of the activating immunoreceptor NKG2D and its MHC
class I-like ligand MICA. Nat Immunol, 2:443–451, 2001.
[25] J. Liu and B. Rost. CHOP proteins into structural domain-like fragments.
Proteins, 55:678–688, 2004.
[26] S. Miyazaki, Y. Kuroda, and S. Yokoyama. Characterization and prediction
of linker sequences of multi-domain proteins by a neural network. J Struct
Funct Genomics, 2:37–51, 2002.
[27] J. Moser, W. D. Schubert, V. Beier, I. Bringemeier, D. Jahn, and D. W.
Heinz. V-shaped structure of glutamyl-tRNA reductase, the first enzyme of
tRNA-dependent tetrapyrrole biosynthesis. EMBO J, 20:6583–6590, 2001.
[28] D. Mumford, J. Fogarty, and F. Kirwan. Geometric invariant theory.
Ergebnisse der Mathematik und ihrer Grenzgebiete. Springer-Verlag, 1994.
[29] A. G. Murzin, S. E. Brenner, T. Hubbard, and C. Chothia. SCOP: a structural classification of proteins database for the investigation of sequences
and structures. J Mol Biol, 247:536–540, 1995.
[30] N. Nandhagopal, A. A. Simpson, J. R. Gurnon, X. Yan, T. S. Baker, M. V.
Graves, J. L. Van Etten, and M. G. Rossmann. The structure and evolution
of the major capsid protein of a large, lipid-containing DNA virus. Proc
Natl Acad Sci U S A, 99:14758–14763, 2002.
14
[31] F. Park, K. Gajiwala, G. Eroshkina, E. Furlong, D. He, Y. Batiyenko,
R. Romero, J. Christopher, J. Badger, J. Hendle, J. Lin, T. Peat, and
S. Buchanan. Crystal structure of YIGZ, a conserved hypothetical protein
from Escherichia coli k12 with a novel fold. Proteins, 55:775–777, 2004.
[32] P. Reinemer, L. Prade, P. Hof, T. Neuefeind, R. Huber, R. Zettl, K. Palme,
J. Schell, I. Koelln, H. D. Bartunik, and B. Bieseler. Three-dimensional
structure of glutathione S-transferase from Arabidopsis thaliana at 2.2 A
resolution: Structural characterization of herbicide-conjugating plant glutathione S-transferases and a novel active site architecture. J Mol Biol,
255:289–309, 1996.
[33] C. R. Robinson and R. T. Sauer. Optimizing the stability of single-chain
proteins by linker length and composition mutagenesis. Proc Natl Acad Sci
U S A, 95:5929–5934, 1998.
[34] M. A. Schumacher, M. C. Miller, S. Grkovic, M. H. Brown, R. A. Skurray, and R. G. Brennan. Structural mechanisms of QacR induction and
multidrug recognition. Science, 294:2158–2163, 2001.
[35] R. M. Story, H. Li, and J. N. Abelson. Crystal structure of a DEAD box
protein from the hyperthermophile Methanococcus jannaschii. Proc Natl
Acad Sci U S A, 98:1465–1470, 2001.
[36] X. Sun, J. A. Cross, A. L. Bognar, E. N. Baker, and C. A. Smith. Folatebinding triggers the activation of folylpolyglutamate synthetase. J Mol
Biol, 310:1067–1078, 2001.
[37] M. Suyama and O. Ohara. DomCut: prediction of inter-domain linker
regions in amino acid sequences. Bioinformatics, 19:673–674, 2003.
[38] P. Tan, M. Steinbach, and V. Kumar. Introduction to data mining. Pearson
Addison Wesley, 2006.
[39] T. Tanaka, Y. Kuroda, and S. Yokoyama. Characteristics and prediction
of domain linker sequences in multi-domain proteins. J Struct Funct Genomics, 4:79–85, 2003.
[40] Y. Tanaka, O. Nureki, H. Kurumizaka, S. Fukai, S. Kawaguchi, M. Ikuta,
J. Iwahara, T. Okazaki, and S. Yokoyama. Crystal structure of the CENPB protein-DNA complex: the DNA-binding domains of CENP-B induce
kinks in the CENP-B box DNA. EMBO J, 20:6612–6618, 2001.
[41] W. R. Taylor. Protein structural domain identification. Protein Eng.,
3:203–216, 1999.
[42] A. V. Tendulkar, A. A. Joshi, M. A. Sohoni, and P. P. Wangikar. Clustering
of protein structural fragments reveals modular building block approach of
nature. J Mol Biol, 338:611–629, 2004.
15
[43] A. V. Tendulkar, M. A. Sohoni, B. Ogunnaike, and P. P. Wangikar. A geometric invariant-based framework for the analysis of protein conformational
space. Bioinformatics, 21:3622–3628, 2005.
[44] A. V. Tendulkar, P. P. Wangikar, M. A. Sohoni, V. V. Samant, and C. Y.
Mone. Parameterization and classification of the protein universe via geometric techniques. J Mol Biol, 334:157–172, 2003.
[45] H. Weyl. The classical groups: their invariants and representations. Princeton landmarks in mathematics and physics. Princeton University Press,
1997.
[46] P. Williams, Y. Chaudhry, I. G. Goodfellow, J. Billington, R. Powell, O. B.
Spiller, D. J. Evans, and S. Lea. Mapping CD55 function. The structure of
two pathogen-binding domains at 1.7 A. J Biol Chem, 278:10691–10696,
2003.
[47] W. Wriggers, S. Chakravarty, and P. A. Jennings. Control of protein functional dynamics by peptide linkers. Biopolymers, 80:736–746, 2005.
[48] B. A. Wurzburg, S. C. Garman, and T. S. Jardetzky. Structure of the
human IgE-Fc C epsilon 3-C epsilon 4 reveals conformational flexibility in
the antibody effector domains. Immunity, 13:375–385, 2000.
[49] L. Yu, A. H. Gunasekera, J. Mack, E. T. Olejniczak, L. E. Chovan, X. Ruan,
D. L. Towne, C. G. Lerner, and S. W. Fesik. Solution structure and function of a conserved protein SP14.3 encoded by an essential Streptococcus
pneumoniae gene. J Mol Biol, 311:593–604, 2001.
[50] R. Zhang, T. Pappas, J. L. Brace, P. C. Miller, T. Oulmassov,
J. M. Molyneaux, J. C. Anderson, J. K. Bashkin, S. C. Winans, and
A. Joachimiak. Structure of a bacterial quorum-sensing transcription factor
complexed with pheromone and DNA. Nature, 417:971–974, 2002.
[51] X. Zhu, X. Zhao, W. F. Burkholder, A. Gragerov, C. M. Ogata, M. E.
Gottesman, and W. A. Hendrickson. Structural analysis of substrate binding by the molecular chaperone DnaK. Science, 272:1606–1614, 1996.
16
!
Figure 1: Schematic for protein linker demarcation: (i) Based on SCOP domain definition;
we identify domain boundary point in the protein structure. A twelve residue long linker
"#$%&'!()!!"#$%&'(")*+,)-,+'$(.)/(.0$,)1$%&,"&'(+.2!3(4)5&6$1)+.)!789)1+%&(.)1$*(.('(+.:);$)
probable region (LPR) is carved out by taking six amino acids each from either domain. (ii)
(1$.'(*<)1+%&(.)=+>.1&,<)-+(.')(.)'#$)-,+'$(.)6',>"'>,$?))@)';$/A$),$6(1>$)/+.B)/(.0$,)-,+=&=/$)
LPR is represented with nine overlapping tetrapeptides. (iii) Each tetrapeptide is described
with a set of 15 geometric invariants (GIs). The geometric invariants are standardized to
,$B(+.)3C9D4)(6)"&,A$1)+>')=<)'&0(.B)6(E)&%(.+)&"(16)$&"#)*,+%)$('#$,)1+%&(.?)3((4)C9D)(6)
zero mean and unit standard deviation. The number of dimensions is reduced via principal
,$-,$6$.'$1);('#).(.$)+A$,/&--(.B)'$',&#$1,+.6?)3(((4)F&"#)'$',&#$1,+.)(6)1$6",(=$1);('#)&)6$')+*)
component analysis (PCA). We select first 8 PCs to represent tetrapeptides. The tetrapepGH)B$+%$',(")(.A&,(&.'6)3IJ64?)K#$)B$+%$',(")(.A&,(&.'6)&,$)6'&.1&,1(L$1)'+)L$,+)%$&.)&.1)>.(')
tides are then transformed into PC space. (iv) The hierarchical agglomerative clustering is
performed to identify clusters of similar tetrapeptides. The clusters are assigned E-values
6'&.1&,1)1$A(&'(+.?))K#$).>%=$,)+*)1(%$.6(+.6)(6),$1>"$1)A(&)-,(."(-&/)"+%-+.$.')&.&/<6(6)
based on the cluster size distribution. The E-values are standardized to zero mean and unit
397@4?))M$)6$/$"')*(,6')N)976)'+),$-,$6$.')'$',&#$1,+.6?)K#$)'$',&#$1,+.6)&,$)'#$.)',&.6*+,%$1)
standard deviation, yielding structural uniqueness scores (SUS). (v) Based on the membership
of a tetrapeptide to a particular cluster, we construct LPR structural profiles using SUS of
(.'+)97)6-&"$?)3(A4)K#$)#($,&,"#("&/)&BB/+%$,&'(A$)"/>6'$,(.B)(6)-$,*+,%$1)'+)(1$.'(*<)"/>6'$,6)+*)
the respective clusters. (vi) We perform maximally scoring subsequence discovery on top of
6(%(/&,)'$',&#$1,+.6?))K#$)"/>6'$,6)&,$)&66(B.$1)FOA&/>$6)=&6$1)+.)'#$)"/>6'$,)6(L$)1(6',(=>'(+.?))
LPR structural profile to identify the continuous stretch with maximum cumulative SUS. This
stretch of tetrapeptides corresponds to the actual linker.
K#$)FOA&/>$6)&,$)6'&.1&,1(L$1)'+)L$,+)%$&.)&.1)>.(')6'&.1&,1)1$A(&'(+.P)<($/1(.B)6',>"'>,&/)
>.(Q>$.$66)6"+,$6)3!R!4?)3A4)5&6$1)+.)'#$)%$%=$,6#(-)+*)&)'$',&#$1,+.)'+)&)-&,'(">/&,)"/>6'$,P)
;$)"+.6',>"')C9D)6',>"'>,&/)-,+*(/$6)>6(.B)!R!)+*)'#$),$6-$"'(A$)"/>6'$,6?)3A(4)M$)-$,*+,%)
%&E(%&//<)6"+,(.B)6>=6$Q>$."$)1(6"+A$,<)+.)'+-)+*)C9D)6',>"'>,&/)-,+*(/$)'+)(1$.'(*<)'#$)
"+.'(.>+>6)6',$'"#);('#)%&E(%>%)">%>/&'(A$)!R!?)K#(6)6',$'"#)+*)'$',&#$1,+.6)"+,,$6-+.16)
'+)'#$)&"'>&/)/(.0$,?)
17
)
)
)
)
PDB
ID
Lit.
Ref.
Actual
Linker
Proposed
Method
1h03
1eqf
[46]
[21]
64-70
1492-1503
1fcd
1vi7
1fp5
1fx7
1f1b
1f14
1jt6
1m3y
1eem
1g3n
1l3l
1gnw
1dkz
1e79
1gpj
1hf2
1hlv
1hv8
1hyr
1jb9
1jbw
1k0d
1k0m
[9]
[31]
[48]
[13]
[23]
[5]
[34]
[30]
[6]
[22]
[50]
[32]
[51]
[17]
[27]
[10]
[40]
[35]
[24]
[1]
[36]
[7]
[19]
65-68
14951502
76-84
134-139
436-440
141-150
95-101
201-206
72-74
213-224
98-107
142-150
163-174
78-92
503-508
95-105
142-148
96-102
65-74
208-214
175-184
156-164
295-300
197-205
89-100
77-80
133-138
438-442
143-146
95-102
201-204
70-73
220-226
100-104
149-152
168-172
84-87
501-512
96-101
141-144
95-99
64-67
207-210
178-181
158-161
291-300
195-198
92-97
Ebina
et.
al. [12]
47-73
15211540
75-83
132-138
420-461
139-147
93-102
193-206
73-76
213-226
97-112
133-136
166-170
106-113
535-540
66-80
136-142
82-87
56-57
249-253
175-194
159-164
278-288
175-186
90-95
GM [14] CHOP [25]
50-70
15041524
97-117
156-176
436-456
137-157
88-108
249-269
123-143
244-264
82-102
193-213
146-166
107-127
459-479
61-81
153-173
132-152
70-90
300-320
185-205
155-175
345-365
195-215
96-116
66-67
14961515
78-79
137-138
446-447
139-151
100-101
218-222
51-52
223-224
125-126
150-151
171-172
88-92
541-542
117-118
142-146
100-102
66-72
208-211
183-184
158-159
298-299
210-211
97-98
Table 3: The table contains a representative examples of linkers extracted by the proposed
method. We have also shown actual linker as extracted from the literature as well as the
linkers predicted by state of the art methods. The column Lit. Ref. provides the literature
reference for the actual linker.
Method
Proposed Method
Ebina et. al. [12]
CHOP [25]
GM [14]
Precision
0.83
0.48
0.35
0.19
Recall
0.66
0.57
0.39
0.50
F1 Score
0.74
0.52
0.37
0.27
No. of Predictions
51
29
36
22
Table 4: Performance of various methods on the benchmark dataset of 51 linkers
18
Figure 2: Different categories of substructures predicted as linkers. The segment of the Nterminal domain is shown in cyan whereas the segment of the C-terminal domain is shown
in blue. The segment of the structure demarcated as a linker is shown in red. (A) 1CR5
[N–terminal domain of Sec18p], helical linker [LEU 104 - GLN 111]; (B) 1F1Z [TnsA endonuclease], linker [THR 164 - VAL 171]; (C) 1VI7 [Hypothetical protein YigZ], beta linker
[THR 133 - PRO 138]; (D) 1P2F [Response regulator DrrB from Thermotoga maritime], loop
linker [GLU 118 - GLY 121]; (E) 1R89 [tRNA nucleotidyltransferase], loop linker [GLY 139
- GLY 143]; (F) 1DT9 [Eukaryotic peptide chain release factor subunit 1], novel linker [LEU
140 - SER 144]; (G) 1DUV [Ornithine transcarbamoylase], novel linker [LEU 148 - ALA 152];
(H) 1DFC [Fascin], loop linker between domain 1 and domain 2 [HIS 1135 - GLN 1141]; (I)
1DFC [Fascin], loop linker between domain 2 and domain 3 [SER 1259 - GLN 1262]; (J) 1S35
[Spectrin beta chain], helical linker [THR 1163 PHE 1170]. The figures are prepared using
PyMOL (http://www.pymol.org).
Figure 3: Since the method use LPR for demarcation, the entire
three dimensional structures of corresponding domains is not
shown. Instead, a stretch of 16 amino acids on either side of the
domain boundary point is shown to maintain clarity of
representation. The segment of the N-terminal domain is shown
in cyan whereas the segment of the C-terminal domain is shown
in blue. The segment of the structure demarcated as a linker is
shown in red. Different categories of substructures predicted as
linkers (A)-(E). (A) 1CR5 [N-terminal domain of Sec18p], helical
linker [LEU 104 – GLN 111]; (B) 1F1Z [TnsA endonuclease],
linker [THR 164 – VAL 171]; (C) 1VI7 [Hypothetical protein
YigZ], beta linker [THR 133 – PRO 138]; (D) 1P2F [Response
regulator DrrB from Thermotoga maritime], loop linker [GLU
19 nucleotidyltransferase], loop
118 – GLY 121]; (E) 1R89 [tRNA
linker [GLY 139 – GLY 143]; novel linkers (F)-(J); (F) 1DT9
[Eukaryotic peptide chain release factor subunit 1], novel linker
[LEU 140 – SER 144]; (G) 1DUV [Ornithine
transcarbamoylase], novel linker [LEU 148 – ALA 152]; (H)
1DFC [Fascin], loop linker between domain 1 and domain 2 [HIS
visualization or
reported in the l
Miyazaki et al.,
2005; Dumontie
Visualization m
automated meth
The proposed m
than the other m
dataset. Our me
precision of 84%
The method
biological know
linkers. Since t
constituent frag
during clusterin
al., 2000). Sinc
SUS, the stretch
SUS correspond
such stretch wa
dynamic progra
The propose
parameters, whi
results: (i) LPR
peptide fragmen
12, tetrahedron
of tetrahedrons
based on the pri
(Miyazaki et al.
et al., 2003). Fl
with the help of
For instance, we
as applied by G
more appropria
processed furthe
| 5cs.CE
|
arXiv:1009.1342v1 [cs.PL] 7 Sep 2010
Simplifying Complex Software Assembly: The
Component Retrieval Language and Implementation
Eric L. Seidel
Gabrielle Allen
Steven Brandt
City College of New York
New York, NY 10031
(Presenting author)
Department of Computer
Science
Center for Computation &
Technology
Louisiana State University
Baton Rouge, LA 70803
Center for Computation &
Technology
Louisiana State University
Baton Rouge, LA 70803
Frank Löffler
Erik Schnetter
Center for Computation &
Technology
Louisiana State University
Baton Rouge, LA 70803
Department of Physics &
Astronomy
Center for Computation &
Technology
Louisiana State University
Baton Rouge, LA 70803
ABSTRACT
1.
Assembling simulation software along with the associated tools and utilities is a challenging endeavor, particularly when the components are distributed across
multiple source code versioning systems. It is problematic for researchers compiling and running the software across many different supercomputers, as well as
for novices in a field who are often presented with a bewildering list of software to collect and install.
In this paper, we describe a language (CRL) for specifying software components with the details needed to
obtain them from source code repositories. The language supports public and private access. We describe
a tool called GetComponents which implements CRL and
can be used to assemble software.
We demonstrate the tool for application scenarios
with the Cactus Framework on the NSF TeraGrid resources. The tool itself is distributed with an open
source license and freely available from our web page.
Compute resources, along with their associated data
storage and network connectivity, are growing ever more
powerful. The current computational environment provided by the National Science Foundation to support
its academic research agenda includes several petascale
machines as part of the distributed TeraGrid facility
and the multi-petaflop “Blue Waters” machine which
should be operational in 2011. This increase in compute capacity is needed to satisfy the requirements of
software applications that are being developed to model
Grand Challenge scientific problems with unprecedented
fidelity in fields such as climate change, nuclear fusion,
astrophysics, material science as well as non-traditional
applications in social sciences and humanities. As these
applications grow in size they are also growing more
complex; coupling together different physical models
across varying spatial and temporal scales, and involving distributed teams of interdisciplinary researchers,
heralding a new era of collaborative multi-scale and
multi-model simulation codes.
One approach to developing application codes in an efficient, sustainable and extensible manner is through the
use of application-level component frameworks or programming environments. Component frameworks can
support reuse and community development of software
by encapsulating common tools or methods within a domain or set of domains. Cactus, a component framework for high performance computing [12, 10], provided
the motivation for the work described in this paper. As
we describe in Section 2, Cactus users typically assemble their simulation codes from many different software
modules distributed from different locations, providing
a number of challenges for users in both describing the
needed modules and actually retrieving them (Figure 1).
Version control systems, such as CVS, Subversion, or
Git, are used to manage and maintain the modules that
Categories and Subject Descriptors
D.2.7 [Software Engineering]: Distribution, Maintenance, and Enhancement—Version Control, Extensibility; D.3.2 [Programming Languages]: Language
Classifications—Specialized application languages
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
TeraGrid ‘10, August 2-5, 2010, Pittsburgh, PA, USA.
Copyright 2010 ACM 978-1-60558-818-6/10/08 ...$10.00.
INTRODUCTION
Figure 1: Applications such as the Einstein Toolkit (Section 7) built from component frameworks such
as Cactus can involve assembling hundreds of modules from distributed, heterogeneous source code
repositories.
make up these frameworks. Such systems track changes
to the source code and allow developers to recover a
stable version of their software, should an error be introduced. There are a large number of version control
systems in use, and while some are relatively compatible
(tools exist to convert a CVS repository to Subversion
or Git), many are not.1 This can create issues when
users want to assemble, and then maintain, a component framework that includes modules from a variety of
systems. A complex framework like Cactus would be
very difficult to maintain without some way of automating the checkout/update process.
To address this issue in a general manner for complex
code assembly for any application, we have designed a
new language, the Component Retrieval Language (or
CRL) that can be used to describe modules along with
information needed for their retrieval from remote, centralized repositories. We have implemented a tool based
on this language that is now being used by Cactus users
for large scale code assembly.
This paper starts by describing the Cactus framework 2, which provides the motivation for the Component Retrieval Language. Then it describes related
work in Section 3 before detailing the design issues for
the component retrieval language in Section 4. Section 5
describes the grammar of the new component retrieval
language, and Section 6 discusses the GetComponents
tool that has been written to implement this language.
Section 7 provides an example showcasing the use of
GetComponents on the resources of the NSF TeraGrid
1
An in-depth comparison of version control system can
be found at http://en.wikipedia.org/wiki/Comparison
of revision control software
for a community of Cactus users. Section 8 describes
planned future work in improving code assembly for
complex software efforts in scientific computing before
concluding in Section 9.
2.
CACTUS EXAMPLE FOR DISTRIBUTED CODE ASSEMBLY
Cactus [12, 10] is an open-source framework designed
for the collaborative development of large scale simulation codes in science and engineering. Computational
toolkits distributed with Cactus already provide a broad
range of capabilities for solving initial value problems
in a parallel environment. The Cactus Computational
Toolkit includes modules for I/O, setting up coordinate
systems, outer and symmetry boundary conditions, domain decomposition and message passing, standard reduction and interpolation operators, numerical methods
such as method of lines, as well as tools for debugging,
remote steering, and profiling. Cactus is supported and
used on all the major NSF TeraGrid machines, as well
as others outside the TeraGrid, and is included in the
advanced tools development for the NSF Blue Waters
facility.
Cactus is used by applications in areas including
relativistic astrophysics, computational fluid dynamics,
reservoir simulations, quantum gravity, coastal science
and computer science. Cactus users assemble their
codes from a variety of independent components (called
thorns) which are typically developed and distributed
from different source code repositories which are geographically, institutionally and politically varied. Source
code repositories can be public (with anonymous read
access), private (with authentication by user or group);
they are of different types (e.g. CVS, SVN, darcs, git,
Mercurial); and the location of thorns within a repository varies. Cactus simulations, for example in the field
of numerical relativity, can involve some 200 thorns from
some ten different repository servers around the world.
In addition, Cactus users typically use other tools or
utilities that are not part of the actual simulation code,
such as the Simulation Factory [1] for building and deploying, or visualization clients and shared parameter
files.
3.
RELATED WORK
Cactus already included a tool for assembling codes
from thorns; GetCactus that was released in 1999 with
the first general release of Cactus and addressed several
of the issues alluded to in the introduction. GetCactus was written specifically to check out Cactus thorns,
with a rudimentary syntax [2] that built on the existing
concept of a Cactus thorn list. When GetCactus was
designed and implemented, in addition to being specific
for Cactus, it only supported the use of CVS repositories. One issue that has become more serious as thorn
lists have become longer is the difficulty in distributing
thorn lists to others, since editing of the thorn list is
required to change authentication details that are user
specific.
A rudimentary syntax for code assembly is also provided by the NMI Build and Test Lab [3] that provides
infrastructure for automated downloading, building, and
testing of complex applications and software infrastructures on a set of commonly used architectures. NMI
provides access to actual hardware on which the test is
run, focuses on reproducible results by providing welldefined test systems, and offers a web-based user interface to browse and examine test results. To download
codes for testing, NMI supports CVS and SVN repositories directly, but only simple scenarios are supported,
and each component’s location has to be described in a
separate file. In addition, NMI supports a generic fetch
stage where a user-defined script can execute arbitrary
code to download components. To build and test Cactus
at NMI, we first download GetComponents and a thorn
list via SVN, and then run GetComponents to retrieve
Cactus and the desired components.
ETICS (eInfrastructure for Testing, Integration and
Configuration of Software) and its successor ETICS 2 are
similar to NMI. They focus on dependencies between
packages, testing, and reproducibility and certification
of results. That is, the emphasis is on testing a snapshot
of project in very well defined environment (e.g. “test on
MacOS X 10.4 with a 32-bit PowerPC processor, kernel
version 8.8.0, and using gcc 4.0.1”). This addresses the
needs of integrators and managers, who can assume that
a project is releasing software in a shrink-wrapped manner. GetComponents, on the other hand, addresses the
needs of software developers that need to handle and assemble components long before the shrink-wrap stage of
a project has been reached. (In fact, in a research environment, software is often never publicly released since
the potential user base is too small; instead, it is only
informally shared among colleagues.)
BuildBot [4] is a Python-based system to automate
building and testing. It is much simpler than NMI or
ETICS, and consists only of software that the user installs, without providing actual testing hardware. Being
Python based, software is checked out via commands in
a Python script. BuildBot provides some abstraction to
access CVS, SVN, etc. repositories, but the download
process is described in a procedural manner as sequence
of commands, not in a descriptive manner. This means
that the information that has been specified to download the software is “hidden” in the Python script and
is not accessible to other tools.
Debian, Red Hat, SUSE etc. are Linux distributions
where a complete installation consists of a set of packages. These packages are available in a specific format
(e.g. deb, rpm) which contains their source code (or binary code) as well as metadata describing e.g. package
dependencies and installation procedures. Usually, these
packages are available from a single, centralized source
(e.g. the distributor itself), and they thus do not need
to address the issues that GetComponents addresses.
Ubiqis uses a naming system where components are
completely identified (including their location and version) by means of a uniquely constructed package
name [9]. References to this package can be automatically detected and downloaded in response to file system
access using FUSE [5]. In principle, this means that versioning information can automatically be sorted out for
any component distribution. In cases where dependencies are not automatic, it makes it possible to search the
community component space in a convenient way. However, it requires that its referents be immutable, and it
downloads packages from the web instead of communicating with source code control systems. While it addresses somewhat different issues than the current work,
there is synergy and the possibility exists to make use
of some Ubiqis or some variant of it in the future.
4.
DESIGN ISSUES
Based on our experiences with the Cactus Framework
and its different user communities, we identified the following needs for the component retrieval language:
• Easy distribution of component lists. The component lists (or CRL files) should be able to be
constructed such that they can be distributed and
used without editing. For Cactus users, this had
been a growing issue with the GetCactus format
where each entry in files would typically need to
be edited to change the username for each repository. The ability to easily distribute and pass on
Cactus thorn lists is a crucial step in simplifying
Cactus for new users.
• Support for both anonymous and authenticated retrieval of components. Authenticated checkout of
components is important for developers that will
be committing changes back to a software repository, or for software that is restricted in access.
Such a capability is important for the Cactus community where many users are also developing components. Authentication is handled differently by
different versioning systems (for example, CVS
requires an ”anonymous username/password” for
users to perform an anonymous checkout, whereas
Subversion and Git do not), further users can have
Table 1: CRL directives (outside component
block)
CRL Directive Description
CRL_VERSION
Currently 1.0. Also indicates that
the file is in CRL format, so it must
be the first non-comment line in
each component list.
DEFINE
User-defined terms that will be reflected throughout the rest of the
component list.
different authentications for different systems (e.g.
different usernames and passwords).
• Support for different repository and distribution
types: Cactus thorns across the community are
currently distributed from CVS, Subversion (SVN)
and git source code repositories, with Mercurial
being a likely choice in the future. Other common
distribution mechanisms for software components
include Darcs and simple HTTP/FTP downloads.
In addition to supporting these features of the CRL,
the implemented tool should:
• Support updating components: Source code repositories using CVS, SVN, etc, support updating
of software, making it possible for developers to
merge changes from others with their own code
changes. The retrieval tool should handle updates
in a manner suitable for developers.
• Handle multiple component lists: This allows a
community to share a common component list,
which can be extended via additional component
lists for a research group and/or individual.
• Handle distributed version control systems: The
nature of distributed systems such as git or mercurial require that one ’clone’ an entire repository
instead of retrieving individual components. This
is inconvenient when trying assemble complex software frameworks that only require a few components from a distributed repository. The retrieval
tool should be able to process an entire repository
while presenting only the components that have
been requested by the user.
5.
THE COMPONENT
LANGUAGE
RETRIEVAL
This section provides a formal description of the Component Retrieval Language (CRL). In designing the
CRL we did not seek to replicate all possible features of
existing version control systems, but to encapsulate the
functionality required by our considered use cases and
allow for future extensibility. Further, a careful distinction was kept between the underlying language and implementation specific details in the GetComponents tool.
The resulting Component Retrieval Language has
eleven different directives which are described in Tables 1 and 2. Files written using the CRL are structured with a header section that defines the version of
Table 2: CRL directives (inside component
block)
CRL Directive Description
TARGET
Placing of component relative to
the current directory.
TYPE
Tool used to checkout the component.
URL
Repository location.
AUTH_URL
Repository location for authenticated access. Only needs to be set if
the URL for authenticated access is
different from the URL for anonymous access.
ANON_USER
Username associated with an
anonymous checkout.
ANON_PASS
Password associated with an anonymous checkout. !ANON_PASS must
be set if !ANON_USER is.
REPO_PATH
Prefix for retrieving components
from a git or mercurial repository,
when a directory structure different
than that provided by the repository is needed.
CHECKOUT
Components to be retrieved from
a repository. Multiple components
are separated by one or more newlines.
NAME
Alternate name for checkout directory if required.
the CRL in the file, which serves the dual purpose of
identifying the file as a CRL list, and providing a way
to determine compatibility with future updates of the
language. It also sets up user defined variables that can
simplify maintaining a long component list. Following
the header section, the rest of the file consists of component blocks, with each block of components having a
common repository description.
The Component Retrieval Language also has an associated grammar written in Bison [6] (a variant of the
Backus-Naur Form (BNF)), which is shown in Figure 2.
While the grammar is fairly simple, it is nonetheless useful to provide a formal specification. This provides assurance that the grammar is unambiguous, and provides
a complete and succinct (albeit somewhat mathematical) form of documentation for the syntax.
6.
GETCOMPONENTS: A CRL IMPLEMENTATION
This section describes a Perl script, called GetComponents, which was developed to process CRL files and
retrieve the indicated components. Perl was chosen because it is quick, lightweight, and it has a very powerful
regular expression engine to parse the component lists.
GetComponents can currently retrieve components from
CVS, Subversion, Git, Darcs and Mercurial repositories,
as well as via http and ftp downloads. It provides multiple command line options as seen in Table 3, including
anonymous mode, automatic updates, two levels of verbosity, and overriding the root directory for the components. Anonymous mode will force all checkouts to use
# NAME is an alphanumeric or ’.’ character
DOCUMENT : DIRECTIVES ;
DIRECTIVE :
|
|
|
|
|
DEFINE NAME ’=’ PATH EOL
CHECKOUT ’=’ COMPONENTLIST EOL
CHECKOUT ’=’ EOL COMPONENTLIST EOL
REPO_LOC ’=’ LOC EOL
AUTH_LOC ’=’ LOC EOL
PATH_DIRECTIVE ’=’ PATH EOL
# !REPO_PATH, !CHECKOUT, !TARGET,
# !ANON_PASS, !NAME
| NAME_DIRECTIVE ’=’ NAME EOL
# !CRL_VERSION, !AUTH_USER,
# !ANON_USER, !TYPE
;
DIRECTIVES : DIRECTIVE
| DIRECTIVES DIRECTIVE
;
LOC : PSERVER PATH
| NAME ’:’ ’/’ ’/’ PATH
| NAME ’@’ NAME ’:’ PATH
;
# CVS repository
# Git/SVN repository
# Git repository
PATH : NAME
| ’/’ NAME
| PATH ’/’ NAME
;
COMPONENTLIST : PATH
| COMPONENTLIST EOL PATH ;
Figure 2: Grammar for the CRL in Bison format
anonymous methods. The auto-update option will bypass the user prompt and update any components that
have been previously checked out, this allows GetComponents to be safely called by another program as a
background process.
Authentication and updates are handled by the underlying version control tools, with GetComponents providing a uniform layer between the user and the underlying tools. Figure 3 shows the general authentication
process used by GetComponents, which is called once
for each component block, unless anonymous mode has
been selected. It first checks for !AUTH_URL, which specifies authenticated access to the repository. It then attempts to match the AUTH_URL to the GetComponents
users file (located by default in $HOME/.crl/users). If
a match is found, GetComponents will use the associated username and then proceed to processing the next
component block. If no match is found, GetComponents
will prompt the user for their username, and attempt to
login to the repository using the appropriate command
(eg. cvs login), after which it will save the username
and URL in the users file. This has the security benefit
of keeping passwords visible only to the actual retrieval
tools. The user may also specify a ’-’ at this prompt
to indicate they wish to perform an anonymous checkout for all components in the block. GetComponents will
store this as well in the users file, so the user is not
forced to specify anonymous access repeatedly. If the
user mistakenly entered the wrong username, or wishes
to change access methods, they may specify the -resetauthentication option, which will delete the users file
and allow the user to reenter their usernames.
If errors occur during the checkout process, GetComponents stores the name of the component that caused the
error, and prints out a list of all components that had errors before exiting. In addition any error will be logged,
including the exact command that was called, and the
error that was returned by the checkout tool. GetComponents will also time the entire checkout/update process
and print the total time elapsed before exiting.
Multiple component lists may be specified together, in
which case GetComponents will concatenate the lists and
process them as one. The component list may also be
specified as an URL, which GetComponents will download and then process normally. This further simplifies
the code assembly process, as the user must only download GetComponents to initiate the assembly. In addition, the anonymous checkout process is shortened by
performing a shallow checkout of git repositories. As a
distributed versioning system, cloning a git repository
requires one to clone the entire repository, along with
the full history of the repository. Over time, this history
accumulates, and can consume a large amount of disk
space. A shallow checkout of a git repository only clones
the most recent changeset, thereby reducing (sometimes
greatly) the size of the resulting local copy, for example
the Carpet repository can be reduced from 115MB to
76MB by performing a shallow checkout.
GetComponents was written to be very modular, and it
can easily be extended to include other versioning tools.
All of the tools are handled by their own subroutine, and
are pointed to by a single hash, which GetComponents
compares with the !TYPE directive in each component.
To add new functionality, one would only have to write
a subroutine for the new tool, and add an entry to the
checkout_types hash.
7.
EXAMPLE: EINSTEIN TOOLKIT
The Einstein Toolkit [7] is a collection of software
components and tools for simulating and analyzing general relativistic astrophysical systems. Such systems include gravitational wave space-times, collisions of compact objects such as black holes or neutron stars, accretion onto compact objects, supernovae core collapse
and gamma-ray bursts. Different research teams typically use the Einstein Toolkit as the basis of their group
codes where they supplement the toolkit with additional
modules for initial data, evolution, analysis etc.
The Einstein Toolkit uses a distributed development
model where its software modules are either developed,
distributed and supported by the core maintainers team,
or by individual groups. Where modules are provided
by external groups, the Einstein Toolkit maintainers
provide quality control for modules for inclusion in the
toolkit and coordinate support and releases. While the
core of the toolkit is a set of Cactus thorns (distributed
from different repositories), the toolkit also contains example parameter files, documentation, and tools for visualization, debugging, and simulation deployment.
The component list (einsteintoolkit.th2 ) for the
Einstein Toolkit uses the CRL for distribution of its currently 130 different software components. All the components of the Einstein Toolkit are available by anony2
https://svn.einsteintoolkit.org/manifest/einsteintoolkit.th
Command-line Option
-help
-man
-verbose
-debug
-anonymous
-update
-root
-reset-authentication
Description
Print a brief help message and exit.
Print the full man page and exit.
Print all system commands as they are executed by script. A second level of verbosity, declared by -v -v, will also display the output from the system commands.
Print a list of components that will be checked out or updated, along with the
total number of components in the list.
Override any stored login credentials and use anonymous checkouts for all components.
Override the update prompt and process all updates.
Override the root directory in the component list. This allows checking out into
an arbitrary directory.
Delete any CRL authentication files before processing the component list.
Table 3: The command-line options for GetComponents.
mous authentication as well as private authentication
for the toolkit developers.
The toolkit currently is distributed using SVN (Cactus Computational Toolkit, core Einstein Toolkit, parameter files, Simulation Factory), git (Carpet AMR
driver), and CVS (components at CCT). A sample from
the CRL file for the Einstein Toolkit is shown in Figure 4.
The GetComponents tool was tested using the Einstein
Toolkit component on the resources of the NSF TeraGrid
and the built-in timing mechanism was used to illustrate
the time needed for both checking out and updating the
full list of components (Figure 5). While the testing was
mostly successful, there were some issues. Notably, the
Frost supercomputer at NCAR was using an outdated
default version of Subversion, which was unable to process components using http or https protocols3 . It was
also difficult to establish a reliable connection to one of
the CVS servers at CCT, so the tests did not include
the two components from this repository, and they will
likely be moved to Subversion in the near future.
8.
FUTURE WORK
As illustrated in the results in Figure 5 assembling the
Einstein Toolkit requires over 9 minutes on average. The
time required for this could be reduced by introducing
the concurrent checkout of different components.
The CRL and GetComponents support the checkout
and update of components. Source code versioning systems support many other options, including commits,
tagging, and updates by date, version or tag. All of
these features could be supported by extending the language to support distributed software development. For
example, in the Einstein Toolkit consortium it could be
helpful to remotely tag all the involved source code in
releases of the Einstein Toolkit. One option that is currently being added is the ability to checkout or update
the source code to a given date, to allow developers to
more easily isolate the time and location at which a bug
was introduced into the code base. With such a feature
in GetComponents a wrapper script could call GetComponents and run a regression suite repeatedly, to determine when software errors were introduced.
Including provenance information is becoming a press3
The version of Subversion on Frost was updated to a
working version just before submission of this paper.
!CRL_VERSION = 1.0
!DEFINE ROOT = Cactus
!DEFINE ARR = $ROOT/arrangements
# Cactus thorns
!TARGET
= $ARR
!TYPE
= svn
!AUTH_URL =
https://svn.cactuscode.org/arrangements/$1/$2/trunk
!URL
=
http://svn.cactuscode.org/arrangements/$1/$2/trunk
!CHECKOUT =
CactusArchive/ADM
CactusBase/Boundary
CactusBase/CartGrid3D
CactusBase/CoordBase
CactusBase/Fortran
CactusBase/IOASCII
CactusBase/IOBasic
CactusBase/IOUtil
CactusBase/InitBase
CactusBase/LocalInterp
CactusBase/LocalReduce
CactusBase/SymBase
CactusBase/Time
# McLachlan, the spacetime code
!TARGET
= $ARR
!TYPE
= git
!URL
=
git://carpetcode.dyndns.org/McLachlan
!AUTH_URL =
[email protected]:McLachlan
!REPO_PATH= $2
!CHECKOUT =
McLachlan/ML_BSSN
McLachlan/ML_BSSN_Helper
McLachlan/ML_BSSN_O2
McLachlan/ML_BSSN_O2_Helper
McLachlan/ML_ADMConstraints
McLachlan/ML_ADMQuantities
Figure 4: Part of the CRL component list for the
Einstein Toolkit.
Anonymous
mode
selected?
yes
no
Are components
available
anonymously?
Is username
for URL
known?
no
yes
Use known
username
yes
Prompt for
username
Use
anonymous
checkout
no
Print error
and ignore
component
Verify
access
Checkout
components
Figure 3: Process for authentication implemented in the GetComponents tool. Authentication is
defined on component blocks. Shaded areas indicate user interaction.
ing challenge for scientific simulations. It is important
that the code that produced published results can be
easily reconstructed and rerun to reproduce or further
analyze data. Currently, the Cactus Computational
Toolkit contains a module, Formaline, which saves a
copy of the complete simulation source code with the
output data of a simulation. GetComponents could be
extended to complement this by outputting a CRL file
from a checkout that includes the information needed to
recreate that checkout (depending on the particular system, this could for example be a version number or date
associated with each component block or component).
One issue that the CRL language and GetComponents
do not address is how to construct a CRL file to solve
a particular scientific problem. A future step will be to
develop and implement a description language that describes the capability of components in such a way that
users can query for components with a particular functionality. Further, the language should also be capable
of describing dependencies between components.
One possibility for identifying dependencies between
components would be the use of an !INCLUDE directive,
which would function similarly to the equivalent C directive. This extra directive would allow users to create an individual component list for each project, and
use !INCLUDE to create a more logical structure for the
framework, as opposed to listing every component in one
large file or forcing users to always specify multiple files.
9.
CONCLUSIONS
This paper presented a language (CRL) that fully
describes distribution mechanisms for software components for scientific codes. The GetComponents tool that
implements the CRL supports multiple source code versioning systems and other access methods, has the ability to checkout and update components and allows users
to distribute and share component lists.
The open source GetComponents tool should be of interest to collaborative teams of researchers with complex
code bases using the NSF TeraGrid and other resources.
GetComponents is now in production use with the Cactus
Framework and the Einstein Toolkit.
1,300
Complete Checkout
Update
Time (seconds)
975
650
325
0
Abe
Frost
Kraken
Lincoln
LoneStar
Longhorn Queen Bee
Ranger
Spur
Steele
TeraGrid Machine
Figure 5: Time taken for a complete checkout, and update, of the Einstein Toolkit with the GetComponents tool on different resources of the NSF TeraGrid.
10.
DISTRIBUTION
GetComponents is released under an open
source license and is freely downloadable from
http://www.eseidel.org/projects/getcomponents/.
Acknowledgments
This work was supported by NSF #0904015 (CIGR)
[13], NSF #0725070 (Blue Waters), and NSF #0721915
(Alpaca) [11, 8]. We used TeraGrid resources at LONI,
NCAR, NCSA, Purdue, and TACC under allocation
TG-MCA02N014. The CRL and GetComponents implementation build on GetCactus which was originally
developed by Gabrielle Allen and Tom Goodale. We
gratefully acknowledge suggestions provided by our colleagues in the Cactus and Einstein Toolkit projects, in
particular Roland Haas for contributing darcs support
to GetComponents. We thank Edward Seidel for providing comments on the paper.
11.
REFERENCES
[1] The Simulation Factory, URL
http://www.cct.lsu.edu/˜eschnett/SimFactory/.
[2] GetCactus language and implementaton for thorn
assembly, URL http://cactuscode.org/
documentation/tutorials/managementHowTo.txt.
[3] Software assembly with the NMI Build and Test
Lab, URL https://nmi.cs.wisc.edu/node/1797.
[4] BuildBot web site, URL http://buildbot.net/trac.
[5] FUSE: Filesystem in Userspace, URL
http://fuse.sourceforce.net.
[6] Bison - GNU parser generator, URL
http://www.gnu.org/software/bison.
[7] The Einstein Toolkit, URL
http://www.einsteintoolkit.org.
[8] Alpaca: Cactus tools for Application-Level
Profiling and Correctness Analysis, URL
http://www.cct.lsu.edu/˜eschnett/Alpaca/.
[9] S. B. R. Brandt, UBIQIS (UBIQuitous InStall),
Proceedings of the 15th ACM Mardi Gras
conference: From lightweight mash-ups to lambda
[10]
[11]
[12]
[13]
grids: Understanding the spectrum of distributed
computing requirements, applications, tools,
infrastructures, interoperability, and the
incremental adoption of key capabilities, ACM,
2008, p. 20.
T. Goodale, G. Allen, G. Lanfermann, J. Massó,
T. Radke, E. Seidel, and J. Shalf, The Cactus
framework and toolkit: Design and applications,
Vector and Parallel Processing – VECPAR’2002,
5th International Conference, Lecture Notes in
Computer Science (Berlin), Springer, 2003.
Erik Schnetter, Gabrielle Allen, Tom Goodale,
and Mayank Tyagi, Alpaca: Cactus tools for
application level performance and correctness
analysis, Tech. Report CCT-TR-2008-2, Louisiana
State University, 2008, URL http:
//www.cct.lsu.edu/CCT-TR/CCT-TR-2008-2.
Cactus Computational Toolkit, URL
http://www.cactuscode.org.
XiRel: Cyberinfrastructure for Numerical
Relativity, URL http://www.cct.lsu.edu/xirel/.
| 6cs.PL
|
Computing with Tangles∗
Martin Grohe and Pascal Schweitzer
RWTH Aachen University
arXiv:1503.00190v3 [cs.DM] 2 Mar 2016
{grohe,schweitzer}@informatik.rwth-aachen.de
Abstract
Tangles of graphs have been introduced by Robertson and Seymour in the context of their
graph minor theory. Tangles may be viewed as describing “k-connected components” of a
graph (though in a twisted way). An interesting aspect of tangles is that they cannot only
be defined for graphs, but more generally for arbitrary connectivity functions (that is, integervalued submodular and symmetric set functions).
However, tangles are difficult to deal with algorithmically. To start with, it is unclear how
to represent them, because they are families of separations and as such may be exponentially
large. Our first contribution is a data structure for representing and accessing all tangles of a
graph up to some fixed order.
Using this data structure, we can prove an algorithmic version of a very general structure
theorem due to Carmesin, Diestel, Hamann and Hundertmark (for graphs) and Hundertmark
(for arbitrary connectivity functions) that yields a canonical tree decomposition whose parts
correspond to the maximal tangles. This may be viewed as a generalisation of the decomposition
of a graph into its 3-connected components.
1
Introduction
Tangles are strange objects—yet they are very useful. Tangles of graphs have been introduced
by Robertson and Seymour [21], and they play an important role in their graph minor theory
(see, e.g. [22]). Intuitively, tangles of order k may be viewed as descriptions of the “k-connected
components” of a graph. Recall that every graph has a nice and simple decomposition into its 2connected components, which are induced subgraphs of the graph. A graph also has a well-defined
decomposition into its 3-connected components. However, the 3-connected components are not
necessarily subgraphs of the graph, but only topological subgraphs; they may contain so-called
“virtual edges” not present in the graph. It is not surprising that for 4-connected components
the situation becomes even more complicated. In fact, for k ≥ 4 there is no clear-cut notion
of a k-connected component of a graph. Tangles may be seen as one attempt towards such a
notion. (Another one, which is related to tangles, has recently been proposed by Carmesin, Diestel,
Hundertmark and Stein [3].)
The idea of tangles is to describe a “region” in a graph (maybe the presumed k-connected component) by pointing to it. Formally, this means that a tangle assigns to each low-order separation
of the graph (separation of order less than k in the case of k-connected components) a “big side”,
which is the side where the “region” described by the tangle is to be found. To turn this into a
meaningful definition, the assignment of “big sides” to the separations has to satisfy certain consistency conditions. Note, however, that a “region” described in this way is elusive and does not
∗
A preliminary version of this paper appeared in the Proceedings of the 47th Annual Symposium on the Theory
of Computing (STOC 2015).
1
necessarily correspond to a subgraph or even just a subset of the vertex set or edge set of the graph,
because the intersection of the “big sides” of all separations may be empty. This way of describing
a region may be viewed as dual to a more direct description as (something resembling) a subgraph.
Indeed, tangles are dual to a form of graph decompositions known as branch decompositions, which
are closely related to the better-known tree decompositions, in a precise technical sense (see the
Duality Lemma 2.11 due to [21]).
Carmesin, Diestel, Hamann, and Hundertmark [1], extending earlier work of Robertson and
Seymour [21] (also see Reed [20]), proved that every graph has a canonical decomposition into
parts corresponding to its tangles of order at most k, that is, a canonical decomposition into its
k-connected components. Here “canonical” means that an isomorphism between two graphs can
be extended to an isomorphism between their decompositions.
What makes tangles even more interesting, and was actually our motivation to start this work,
is that they can be defined in a very abstract setting that applies to various “connectivity measures”
not only for graphs but also for other structures such as hypergraphs and matroids. Two examples
are the “cut-rank” measure for graphs, which leads to the notion of “rank width” [15, 17], and
the connectivity function of a matroid (see, for example, [19]). Such “connectivity measures” can
be specified by a symmetric and submodular function defined on the subsets of some ground set.
Tangles give us an abstract notion of “k-connected components” with respect to these connectivity
measures. Hundertmark [12] generalised the decomposition theorem from graphs to this abstract
setting, giving us canonical decompositions of structures into parts corresponding to the tangles
with respect to arbitrary connectivity measures. (Earlier, Geelen, Gerards, and Whittle [7] had
already shown the existence of such decompositions, but not canonical ones.)
However, these decomposition theorems are pure existence theorems; they are not algorithmic.1
In fact, it is not clear how to efficiently compute with tangles at all. Tangles are defined as families
of separations of a ground set (such as the edge set of a graph) and as such may be exponentially
large; this means that a priori there is only a doubly-exponential upper bound on the number of
tangles. Remarkably, the decomposition theorems mentioned above imply that for each k there is
only a linear number of tangles of order k. But this still does not tell us how to identify them. Our
first main contribution is a data structure that represents all tangles of some order k and provides
us with a “membership oracle” for these tangles, that is, for a given separation we can ask which
side of the separation is the “big side” with respect to a given tangle. The data structure as well
as the membership oracle can be implemented in polynomial time for every fixed order k.
Using this data structure, we can then prove that a canonical decomposition of a structure
into parts corresponding to the tangles of order at most k can be computed in polynomial time
(again for fixed k). This is our second main result. Proving these results, we devise a number
of algorithmic subroutines that may be useful in other contexts. For example, we show how to
find canonical “leftmost minimum separations” between a set and a tangle. All our results apply
in the most general setting where we are only given oracle access to an arbitrary integer-valued
symmetric and submodular connectivity function. Our algorithms rely on the minimisation of
submodular functions [13, 23]. We build on algorithmic ideas for computing branch decompositions
of connectivity functions due to Oum and Seymour [16, 17, 18]. Furthermore, the duality between
tangles and branch decompositions plays an important role in our proofs.
Our main motivation for this work is isomorphism testing and canonisation. It is almost selfevident that a canonical decomposition of a structure into highly connected parts may be useful
in this context, in a similar way as the decomposition of a graph into its 3-connected components
1
The “algorithms” in the title of [1] refer to general strategies for choosing the set of separations in a decomposition,
they are not implementable algorithms.
2
is essential for planar graph isomorphism testing. Tangles and the decompositions of graphs they
induce have already played an important role in the recent polynomial-time isomorphism test for
graph classes with excluded topological subgraphs by Marx and the first author of this paper [9].
However, there it was sufficient to only work with a specific type of tangles that can be represented
by so called “well-linked” or “unbreakable” sets of bounded size; this way the computational problems addressed in the present paper could be circumvented. In other settings, this is impossible.
In [10], we use a “directed version” of our canonical decomposition (Theorem 5.1) to design a
polynomial isomorphism test for graph classes of bounded rank width.
A different, more speculative, application of the techniques developed here may be the logical
definability of decompositions, which is related to well-known open problems such as Seese’s conjecture [24, 5] and the definability of tree decompositions in monadic second-order logic (see [4]).
Definable decompositions are necessarily canonical, because logics define isomorphism-invariant
objects. Even though our algorithms may not translate into logical definitions directly, our constructive arguments may help to come up with such definitions.
While these applications rely on the canonicity of the decompositions we obtain, our data
structure for tangles (which is not canonical) may have other algorithmic applications. Tangles
have already played a role in algorithmic structural graph theory (for example in [6, 14, 9, 8]), but
due to the non-constructive nature of tangles, this role is usually indirect or implicit. Having direct
access to all tangles may facilitate new applications.
2
Connectivity Functions, Tangles, and Branch Decompositions
In this section we introduce connectivity functions, branch decompositions, and tangles and prove
some basic results about them. All lemmas in this section, except the Exactness Lemma and the
Duality Lemma from [21], are simple and can be proved by standard techniques.
A connectivity function on a finite set U is a symmetric and submodular function κ : 2U → N
with κ(∅) = 0. Symmetric means that κ(X) = κ(X) for all X ⊆ U ; here and whenever the ground
set U is clear from the context we write X to denote U \ X, the complement of X. Submodular
means that κ(X) + κ(Y ) ≥ κ(X ∩ Y ) + κ(X ∪ Y ) for all X, Y ⊆ U . Observe that a symmetric and
submodular set function is also posimodular, that is, it satisfies κ(X) + κ(Y ) ≥ κ(X \ Y ) + κ(Y \ X)
(apply submodularity to X and Y ).
Note that it is no real restriction to assume that the range of κ is the set N of nonnegative
integers instead of the set Z of all integers and that κ(∅) = 0. For an arbitrary symmetric and
submodular function κ′ : 2U → Z, we can work with the function κ defined by κ(X) := κ′ (X)−κ′ (∅)
instead. It clearly satisfies κ(∅) = 0. To see that it is nonnegative, observe that for all X ⊆ U ,
1
1 ′
κ (X) + κ′ (X) ≥ κ′ (∅) + κ′ (U ) = κ′ (∅).
2
2
Here the first an third inequalities hold by symmetry, and the second inequality holds by submodularity.
κ′ (X) =
Example 2.1. Maybe the most obvious example of a connectivity function is the “cut function” in
a graph G. For all subsets X, Y ⊆ V (G), we let EX,Y be the set of all edges of G with one endvertex
in X and one endvertex in Y . We define a connectivity function γG on V (G) by γG (X) = |EX,X |.
y
Example 2.2 (Robertson and Seymour [21]). Let G be a graph. The boundary ∂X of an edge
set X ⊆ E(G) is the set of all vertices of G incident with an edge in X and an edge in X = E(G)\X.
We define a connectivity function κG on E(G) by κG (X) = |∂X|. This is the connectivity function
Robertson and Seymour [21] originally developed their theory of tangles for.
y
3
Example 2.3 (Oum and Seymour [17]). Let G be a graph. For all subsets X, Y ⊆ V (G), we
let M = MG (X, Y ) be the X × Y -matrix over the 2-element field F2 with entries Mxy = 1 ⇐⇒
xy ∈ E(G). Now we define a connectivity function ρG on V (G) by letting ρG (X), known as the cut
rank of X, be the row rank of the matrix MG (X, X). This connectivity function was introduced
by Oum and Seymour to define the rank width of graphs, which approximates the clique width, but
has better algorithmic properties.
y
Example 2.4. Let M be a matroid with ground set E and rank function r. (The rank of a set
X ⊆ E is defined to be the maximum size of an independent set contained in X.) The connectivity
function of M is the set function κM : 2E → N defined by κM (X) = r(X) + r(X) − r(E) (see, for
example, [19]).
y
For the rest of this section, let κ be a connectivity function on a finite set U . We often think
of a subset Z ⊆ U as a separation of U into Z and Z and of κ(Z) as the order of this separation;
consequently, we also refer to κ(Z) as the order of Z. For disjoint sets X, Y ⊆ U , an (X, Y )separation is a set Z ⊆ U such that X ⊆ Z ⊆ Y . Such a separation Z is minimum if its order is
minimum, that is, if κ(Z) ≤ κ(Z ′ ) for all (X, Y )-separations Z ′ . It is an easy consequence of the
submodularity of κ that there is a unique minimum (X, Y )-separation Z such that Z ⊆ Z ′ for all
other minimum (X, Y )-separations Z ′ . We call Z the leftmost minimum (X, Y )-separation. There
is also a unique rightmost minimum (X, Y )-separation, which is easily seen to be the complement
of the leftmost minimum (Y, X)-separation.
2.1 Tangles
A κ-tangle of order k ≥ 0 is a set T ⊆ 2U satisfying the following conditions.2
(T.0) κ(X) < k for all X ∈ T ,
(T.1) For all X ⊆ U with κ(X) < k, either X ∈ T or X ∈ T .
(T.2) X1 ∩ X2 ∩ X3 6= ∅ for all X1 , X2 , X3 ∈ T .
(T.3) T does not contain any singletons, that is, {a} 6∈ T for all a ∈ U .
We denote the order of a κ-tangle T by ord(T ).
Example 2.5. Let G be a graph and H a 3-connected subgraph of G. For a set X ⊆ E(G) we let
V (X) be the set of all the endvertices of edges in X. So ∂(X) = V (X) ∩ V (X). Let T be the set
of all subsets X ⊆ E(G) such that κG (X) = |∂(X)| < 3 (see Example 2.2) and V (H) ⊆ V (X). It
is not difficult to prove that T is a κG -tangle of order 3 in G.
Actually, essentially the same argument works if H is a subdivision of a 3-connected graph and
T the set of all X ⊆ E(G) such that κG (X) = |∂(X)| < 3 and X contains all vertices of H of
degree at least 3.
y
Example 2.6 (Robertson and Seymour [21, Section 7]). Let G be a graph and H ⊆ G a
(k × k)-grid. Let T be the set of all X ⊆ E(G) such that κG (X) < k and X contains all edges of
some row of the grid. Then T is a κG -tangle of order k.
y
Let T , T ′ be κ-tangles. If T ′ ⊆ T , we say that T is an extension of T ′ . The tangles T and T ′
are incomparable (we write T ⊥T ′ ) if neither is an extension of the other. The truncation of T to
2
Our definition of tangle is “dual” to the one mostly found in the literature, e.g. [7, 12]. In our definition, the “big
side” of a separation belongs to the tangle, which seems natural if one thinks of a tangle as “pointing to a region”
(as described in the introduction). But of course the definitions yield equivalent theories.
4
order k ≤ ord(T ) is the set {X ∈ T | κ(X) < k}, which is obviously a tangle of order k. Observe
that if T is an extension of T ′ , then ord(T ′ ) ≤ ord(T ), and T ′ is the truncation of T to order
ord(T ′ ).
Remark 2.7. There is a small technical issue that one needs to be aware of, but that never causes
any real problems: if we view tangles as families of sets, then their order is not always well-defined.
Indeed, if there is no set X of order κ(X) = k − 1, then a tangle of order k contains exactly the
same sets as its truncation to order k − 1. In such a situation, we have to explicitly annotate a
tangle with its order, formally viewing a tangle as a pair (T , k) where T ⊆ 2U and k ≥ 0. We
always view a tangle of order k and its truncation to order k − 1 as distinct tangles, even if they
contain exactly the same sets.
A (T , T ′ )-separation is a set Z ⊆ U such that Z ∈ T and Z ∈ T ′ . Obviously, if Z is a (T , T ′ )separation then Z is a (T ′ , T )-separation. Observe that there is a (T , T ′ )-separation if and only if
T and T ′ are incomparable. The order of a (T , T ′ )-separation Z is κ(Z). A (T , T ′ )-separation Z
is minimum if its order is minimum.
Lemma 2.8. Let κ be a connectivity function on a set U , and let T , T ′ be incomparable tangles.
Then there is a (unique) minimum (T , T ′ )-separation Z(T , T ′ ) such that Z(T , T ′ ) ⊆ Z ′ for all
minimum (T , T ′ )-separations Z ′ .
We call Z(T , T ′ ) the leftmost minimum (T , T ′ )-separation.
Proof. Let Z = Z(T , T ′ ) be a minimum (T , T ′ )-separation of minimum size |Z|, and let Z ′ be
another minimum (T , T ′ )-separation. We shall prove that Z ⊆ Z ′ .
Let k := κ(Z) = κ(Z ′ ) < min{ord(T ), ord(T ′ )}. We claim that
κ(Z ∪ Z ′ ) ≥ k.
(2.A)
Suppose for contradiction that κ(Z ∪ Z ′ ) < k. Then Z ∪ Z ′ ∈ T , because κ(Z ∪ Z ′ ) < ord(T )
and Z ∪ Z ′ ∩ Z = ∅. Furthermore, Z ∪ Z ′ ∈ T ′ , because κ(Z ∪ Z ′ ) = κ(Z ∪ Z ′ ) < ord(T ′ ) and
′
Z ∩ Z ∩ (Z ∪ Z ′ ) = ∅. Thus Z ∪ Z ′ is a (T , T ′ )-separation of smaller order than Z. This contradicts
the minimality of Z.
By submodularity,
κ(Z ∩ Z ′ ) ≤ k.
(2.B)
A similar argument as above shows that Z ∩ Z ′ is a (T , T ′ )-separation. Thus, by the minimality
of Z, we have κ(Z ∩ Z ′ ) = k, and by the minimality of |Z| we have |Z| ≤ |Z ∩ Z ′ |. This implies
Z = Z ∩ Z ′ and thus Z ⊆ Z ′ .
Now that we have defined (X, Y )-separations for sets X, Y and (T , T ′ )-separations for tangles
T , T ′ , we also need to define combinations of both. For a κ-tangle T and a set X ⊆ U a (T , X)separation is a set Z ∈ T such that Z ⊆ X. Note that even if X 6∈ T , such a separation does not
always exist. A (T , X)-separation Z is minimum if its order is minimum. Z is leftmost minimum
(rightmost minimum) if Z ⊆ Z ′ for all minimum (T , X)-separations Z ′ . By essentially the same
submodularity arguments as those in the proof of Lemma 2.8, it is easy to prove that if there
is a (T , X)-separation then there is a unique leftmost minimum (T , X)-separation and a unique
rightmost minimum (T , X)-separation.
5
2.2 Branch Decompositions and Branch Width
A cubic tree is a tree where every node that is not a leaf has degree 3. An oriented edge of a tree
−
→
T is a pair (s, t), where st ∈ E(T ). We denote the set of all oriented edges of T by E (T ) and the
e where T is a cubic tree
set of leaves of T by L(T ). A partial decomposition of κ is a pair (T, ξ),
−
→
U
e s) for all st ∈ E(T ) and ξ(s,
e t1 ) ∪ ξ(s,
e t2 ) ∪ ξ(s,
e t3 ) = U
e t) = ξ(t,
and ξe: E (T ) → 2 such that ξ(s,
for all non-leaf nodes s ∈ V (T ) with neighbours t1 , t2 , t3 . The partial decomposition is exact if for
e t ), ξ(s,
e t ), ξ(s,
e t ) are mutually
all non-leaf nodes s ∈ V (T ) with neighbours t1 , t2 , t3 the sets ξ(s,
1
2
3
e u), where t is the
disjoint. To simplify the notation, for all leaves u ∈ L(T ) we let ξ(u) := ξ(t,
neighbour of u in T . Note that for an exact partial decomposition, the values ξ(u) at the leaves u
e t), and the sets ξ(u) for the leaves u form a partition of U . We say
determine all other values ξ(s,
e is over a set X ⊆ 2U if ξ(u) ∈ X for all leaves u ∈ L(T ). Finally,
that a partial decomposition (T, ξ)
a branch decomposition of κ is an exact partial decomposition over the set SU = {{u} | u ∈ U } of
all singletons.3
e is the maximum of the values κ(ξ(s,
e t)) for (s, t) ∈
The width of a partial decomposition (T, ξ)
−
→
E (T ). The branch width of κ is the minimum of the widths of all branch decompositions of κ.
Example 2.9. The branch width of κG (Example 2.2) is known as the branch width of the graph
G. The branch width of ρG (Example 2.3) is known as the rank width of G. The branch width of
κM (Example 2.4) is known as the branch width of the matroid M .
y
The following Exactness Lemma and Duality Lemma due to Robertson and Seymour [21] are
fundamental results relating decompositions and tangles. For the reader’s convenience, we include
proofs of these lemmas.
e be a partial decomposition of κ. Then there
Lemma 2.10 (Exactness Lemma [21]). Let (T, ξ)
−
→
is a function ξe′ : E (T ) → 2U such that (T, ξe′ ) is an exact partial decomposition of κ satisfying
→
e t)) for all (s, t) ∈ −
κ(ξe′ (s, t)) ≤ κ(ξ(s,
E (T ) and ξ ′ (t) ⊆ ξ(t) for all leaves t ∈ L(T ).
Proof. It is convenient in this proof to work with binary rooted trees instead of cubic trees. To turn
T into a binary tree T b , we choose an arbitrary edge, say, sb tb ∈ E(T ), and subdivide it, inserting
a new node r b . We make r b the root of our new binary tree T b . Now we can “push” the mapping
ξe from the edges of the tree to the nodes. That is, we define a mapping ξ b : V (T ) → 2U by
• ξ b (r b ) := U ;
e b , sb ) and ξ b (tb ) := ξ(s
e b , tb );
• ξ b (sb ) := ξ(t
e t) for all nodes t ∈ V (T b ) \ {r b , sb , tb } with parent s.
• ξ b (t) := ξ(s,
Observe that for all nodes s ∈ V (T b ) with children t1 , t2 it holds that ξ b (s) ⊆ ξ b (t1 ) ∪ ξ b (t2 ).
Furthermore, the trees T and T b have the same leaves and for all leaves t we have ξ b (t) = ξ(t).
In general, if (T ′ , r ′ ) is rooted binary tree and ξ ′ : V (T ′ ) → 2U such that ξ ′ (r ′ ) = U and ξ ′
satisfies
ξ ′ (s) ⊆ ξ ′ (t1 ) ∪ ξ ′ (t2 ),
(2.C)
we call (T ′ , r ′ , ξ ′ ) a binary partial decomposition. The decomposition is exact at a node s with
children t1 , t2 if (2.C) can be strengthened to
ξ ′ (s) = ξ ′ (t1 ) ⊎ ξ ′ (t2 ),
(2.D)
3
Since a branch decomposition is exact, it suffices to specify the values ξ(t) at the leaves, and as these values are
singletons, we can also define ξ as a mapping from L(T ) to U , which must be bijective.
6
where ⊎ denotes disjoint union. The decomposition (T ′ , r ′ , ξ ′ ) is exact if it is exact at all inner
nodes. We have seen how we can turn a partial decomposition into a binary partial decomposition.
It is also easy to reverse the construction and turn an exact binary partial decomposition (T ′ , r ′ , ξ ′ )
into an exact partial decomposition (T ′′ , ξe′′ ).
We will iteratively construct a sequence ξ1 , . . . , ξm of mappings from V (T b ) to 2U such that
b
(T , r b , ξ1 ), . . . , (T b , r b , ξm ) are binary partial decompositions satisfying the following invariants for
all i ∈ [m − 1] and nodes t ∈ V (T b ):
(i) κ(ξi+1 (t)) ≤ κ(ξi (t));
(ii) either ξi+1 (t) ⊆ ξi (t) or κ(ξi+1 (t)) < κ(ξi (t));
(iii) if t is a leaf of T b then ξi+1 (t) ⊆ ξi (t).
Furthermore, the decomposition (T b , r b , ξm ) will be exact.
We let ξ1 := ξ b . In the inductive step, we assume that we have defined ξi . If (T b , r b , ξi ) is
exact, we let m := i and stop the construction. Otherwise, we pick an arbitrary node s ∈ V (T b )
with children t1 , t2 such that (T b , r b , ξi ) is not exact at s, that is, either ξi (s) ⊂ ξi (t1 ) ∪ ξi (t2 ) or
ξi (t1 ) ∩ ξi (t2 ) 6= ∅. We let X := ξi (s) and Yp := ξi (tp ) for p = 1, 2.
In each of the following cases, we only modify ξi at the nodes s, t1 , t2 and let ξi+1 (u) := ξi (u)
for all u ∈ V (T b ) \ {s, t1 , t2 }.
Case 1: X ⊂ Y1 ∪ Y2 .
Case 1a: κ(X ∩ Yp ) ≤ κ(Yp ) for p = 1, 2.
We let ξi+1 (s) := ξi (s) and ξi+1 (tp ) := X ∩ Yp for p = 1, 2.
Note that in this case we have κ(ξi+1 (u)) ≤ κ(ξi (u)) and ξi+1 (u) ⊆ ξi (u) for all nodes u and
either ξi+1 (t1 ) ⊂ ξi (t1 ) or ξi+1 (t2 ) ⊂ ξi (t2 ).
Case 1b: κ(X ∩ Yp ) > κ(Yp ) for some p ∈ {1, 2}.
For p = 1, 2, we let ξi+1 (tp ) := Yp .
By submodularity we have κ(X ∪ Yp ) < κ(X) for some p ∈ {1, 2}. If κ(X ∪ Y1 ) < κ(X) we let
ξi+1 (s) := X ∪ Y1 , and otherwise we let ξi+1 (s) := X ∪ Y2 .
Note that in this case we have κ(ξi+1 (u)) ≤ κ(ξi (u)) for all nodes u and κ(ξi+1 (s)) < κ(ξi (s))
and ξi+1 (u) = ξi (u) for all nodes u 6= s.
Also note that invariant (iii) is preserved, because s is not a leaf of the tree.
Case 2: X = Y1 ∪ Y2 and Y1 ∩ Y2 6= ∅.
We let ξi+1 (s) := ξi (s).
By posimodularity, either κ(Y1 \ Y2 ) ≤ κ(Y1 ) or κ(Y2 \ Y1 ) ≤ κ(Y2 ). If κ(Y1 \ Y2 ) ≤ κ(Y1 ), we
let ξi+1 (t1 ) := Y1 \ Y2 and ξi+1 (t2 ) := Y2 . Otherwise, we let ξi+1 (t1 ) := Y1 and ξi+1 (t2 ) := Y2 \ Y1 .
Note that in this case we have κ(ξi+1 (u)) ≤ κ(ξi (u)) and ξi+1 (u) ⊆ ξi (u) for all nodes u and
either ξi+1 (t1 ) ⊂ ξi (t1 ) or ξi+1 (t2 ) ⊂ ξi (t2 ).
This completes the description of the construction. To see that it terminates, we say that the
P
P
total weight of ξi is t∈V (T b ) κ(ξi (t)) and the total size of ξi is t∈V (T b ) |ξi (t)|. Now observe that
in each step of the construction either the total weight decreases or the total weight stays the same
and the total size decreases. This proves termination.
To see that for all i the triple (T b , r b , ξi ) is indeed as partial decomposition, observe first that
ξi (r b ) = U , because the root can only occur as the parent node s in the construction above, and
the set at the parent node either stays the same (in Cases 1a and 2) or increases (in Case 1b).
Moreover, it is easy to check that for all nodes s′ with children t′1 , t′2 we have ξi (s′ ) ⊆ ξi (t′1 ) ∪ ξi (t′2 ).
This follows immediately from the construction if s = s′ . If s′ is the parent of s = t′i , it follows
7
because the set at s can only increase. If s′ = ti , it follows because the set at ti can only decrease.
Otherwise, all the sets at s′ , t′1 , t′2 remain unchanged. Note that the invariant (iii) is preserved,
because leaves can only occur as the child nodes ti in the construction above, and the set at the
cild nodes either decrease (in Cases 1a and 2) or stay the same (in Case 1b).
Now we turn the exact binary partial decomposition (T b , r b , ξm ) into an exact partial decomposition (T ′ , ξe′ ). Invariants (i) and (iii) guarantee that (T ′ , ξe′ ) has the desired properties.
e be an exact partial decomposition of κ over some set X ⊆ 2U . Observe that we
Let (T, ξ)
can easily eliminate leaves u ∈ L(T ) with ξ(u) = ∅ (we call them empty leaves) by deleting u and
contracting the edge from the sibling of u to its parent. Doing this repeatedly, we can turn the
decomposition into a decomposition of at most the same width over X \ {∅}.
We say that a tangle T avoids a set X ⊆ 2U if T ∩ X = ∅. Note that, by (T.3), every tangle
avoids the set SU of all singletons.
Lemma 2.11 (Duality Lemma, [21]). Let X ⊆ 2U be closed under taking subsets. Then there
is a partial decomposition of width less than k over X ∪ SU if and only if there is no κ-tangle of
order k that avoids X .
e a partial decomposition of κ of width less than k over
Proof. For the forward direction, let (T, ξ)
X ∪ SU . Suppose for contradiction that T is a κ-tangle of order k that avoids X . For every edge
e t) = ξ(t,
e s) ∈ T . As T is a
e t) ∈ T and towards s if ξ(s,
st ∈ E(T ), we orient st towards t if ξ(s,
−
→
e t)) < k for all (s, t) ∈ E (T ), every edge gets an orientation. As T is a
tangle of order k and κ(ξ(s,
tree, there is a node s ∈ V (T ) such that all edges incident with s are oriented towards s. If s is a
e
leaf, then ξ(s) ∈ T and thus ξ(s) 6∈ X ∪SU by (T.3) and because T avoids X . This contradicts (T, ξ)
being a decomposition over X ∪ SU . Thus s is an inner node, say, with neighbours t1 , t2 , t3 . Then
e i , s) ∈ T and thus ξ(t
e 1 , s) ∩ ξ(t
e 2 , s) ∩ ξ(t
e 3 , s) 6= ∅. This implies ξ(s,
e t1 ) ∪ ξ(s,
e t2 ) ∪ ξ(s,
e t3 ) 6= U ,
ξ(t
e
which contradicts (T, ξ) being a partial decomposition.
The proof of the backward direction is by induction on the number of nonempty sets X ⊂ U
with κ(X) < k such that neither X nor X are in X .
For the base case, let us assume that for all X ⊂ U with κ(X) < k either X ∈ X or X ∈ X .
Let
Y = {X | X ∈ X with κ(X) < k}.
Then Y trivially satisfies the tangle axiom (T.0). It satisfies (T.1) by our assumption that either
X ∈ X or X ∈ X for all X ⊂ U with κ(X) < k.
If Y violates (T.2), then there are sets Y1 , Y2 , Y3 ∈ Y with Y1 ∩ Y2 ∩ Y3 = ∅. We let T be the
e i , s) := Yi
tree with vertex set V (T ) = {s, t1 , t2 , t3 } and edge set {st1 , st2 , st3 }, and we define ξ(t
e
e
and ξ(s, ti ) := Y i ∈ X . Then (T, ξ) is a partial branch decomposition over X of width less than k.
So let us assume that Y satisfies (T.2). As there is no tangle of order k, the set Y must violate
(T.3). Thus for some x ∈ U we have {x} ∈ Y and thus {x} ∈ X and κ({x}) = κ({x}) < k. We let
e t) = {x}, ξ(s,
e t) = {x}.
T ∼
= K2 be a tree consisting of just one edge, say, st, and define ξe by ξ(s,
This yields a partial branch decomposition over X ∪ SU of width less than k.
For the inductive step, suppose that there is no partial decomposition of κ of width less than
k over X . Let X ⊂ U such that κ(X) < k and neither X ∈ X nor X ∈ X and such that |X| is
minimum subject to these conditions. Let X 1 := X ∪ 2X and X 2 := X ∪ 2X . Then by the inductive
hypothesis, for i = 1, 2 there is a partial decomposition (T i , ξei ) of width less than k over X i . As
there is no partial decomposition of width less than k over X , there is a leaf ti of T i with ξ(ti ) 6∈ X .
8
Consider (T 1 , ξe1 ). By the Exactness Lemma and since X 1 is closed under taking subgraphs,
we may assume that (T 1 , ξe1 ) is exact. This implies that the sets ξ 1 (t) for the leaves t ∈ L(T 1 ) are
′
mutually disjoint. Let t1 be a leaf of T 1 with X ′ := ξ 1 (t1 ) 6∈ X . Then X ′ ⊆ X and X ⊆ X , and
′
as X 6∈ X and X is closed under taking subsets, it follows that X 6∈ X . By the minimality of X,
this implies X ′ = X. Furthermore, as the decomposition (T 1 , ξe1 ) is exact, there is exactly one leaf
t1 of T 1 with ξ 1 (t1 ) = X, and for all other leaves t we have ξ 1 (t1 ) ∈ X . Let s1 be the neighbour of
t1 ∈ T 1 .
Now consider (T 2 , ξe2 ). Let t21 , . . . , t2m be an enumeration of all leaves t2 ∈ L(T 2 ) with ξ 2 (t2 ) 6∈ X .
Then ξ(t2i ) ⊆ X for all i ∈ [m]. Without loss of generality we may assume that ξ(t2i ) = X, because
increasing a set ξ 2 (t) preserves the property of being a partial decomposition. (We do not assume
(T 2 , ξe2 ) to be exact.) For every i ∈ [m], let s2i be the neighbour of t2i in T 2 .
e of κ over X , we take m disjoint copies
To construct a partial decomposition (T, ξ)
1 e1
(T11 , ξe11 ), . . . , (Tm
, ξm )
of (T 1 , ξe1 ). For each node t ∈ V (T 1 ), we denote its copy in Ti1 by ti . Then for every edge st ∈ E(T 1 )
we have ξei1 (si , ti ) = ξe1 (s, t). In particular, ξei1 (s1i , t1i ) = ξe1 (s1 , t1 ) = X. We let T be the tree obtained
1 , T 2 by deleting the nodes t1 , t2 and adding edges s1 s2 for all
from the disjoint union of T11 , . . . , Tm
i i
i i
U
e
i ∈ [m]. We define ξ : V (T ) → 2 by
e t) :=
ξ(s,
X
X
ξei1 (s, t)
e2
ξ (s, t)
if
if
if
if
(s, t) = (s1i , s2i ) for some i ∈ [m],
(s, t) = (s2i , s1i ) for some i ∈ [m],
st ∈ E(Ti1 ),
st ∈ E(T 2 ).
e is a partial decomposition of width less than k over X . This is a
It is easy to see that (T, ξ)
contradiction.
Since we may assume decompositions to be exact, the Duality Lemma implies that κ has a
branch decomposition of width less than k if and only if there is no κ-tangle of order k, in other
words: the branch width of κ is exactly the maximum order of a κ-tangle.
The following somewhat surprising fact follows from the existence claim of Theorem 4.12 (see
Remark 4.13). However, the fact is needed to construct our tangle data structure (Theorem 3.4)
and hence to prove the algorithmic claim of Theorem 4.12.4
Fact 2.12 ([21]). Let k ≥ 0, and let κ be a a connectivity function on a set U of size |U | = n.
Then there are at most n κ-tangles of order k.
Actually, Corollary (10.4) of [21] only states that there are at most n maximal tangles. But as
every maximal tangles contains at most one tangle of order k, the fact follows.
2.3 Bases
For disjoint sets X, Y ⊆ U we define κmin (X, Y ) := min{κ(Z) | X ⊆ Z ⊆ Y }. Note that for all
X, Y the two functions X ′ 7→ κmin (X ′ , Y ) and Y ′ 7→ κmin (X, Y ′ ) are monotone and submodular.
For sets Y ⊆ X, we say that a set Y is free in X if κmin (Y, X) = κ(X) and |Y | ≤ κ(X). Let
us remark that if κ(∅) = 0 and κ({x}) = 1 for all x ∈ X, then the function Y 7→ κmin (Y, X) is the
4
Thus a reader interested in a self-contained treatment may skip Section 3 at first reading and immediately jump
to Section 4, but ignore all algorithmic parts there, then go back to Section 3, and finally to the algorithmic parts of
Section 4. See Remark 4.13 for further details.
9
rank function of a matroid on the set X, and a set Y is free in X if and only if it is a base for this
matroid.
The following simple and well-known lemma (see, for example, [16]) thus generalises the fact
that every matroid, defined by its rank function, has a base.
Lemma 2.13. For every X ⊆ U there is a set Y that is free in X.
Proof. Let Y ⊆ X be an inclusion-wise minimal set with κmin (Y, X) = κ(X). Such a set exists because κmin (X, X) = κ(X). We claim that |Y | ≤ κ(X). Suppose otherwise and let Y = {y1 , . . . , yℓ }
with ℓ > κ(X). Define Yi = {y1 , . . . , yi } for all i ∈ [ℓ]. Then κmin (Yi , X) ≤ κmin (Yi+1 , X)
since κmin is monotone in the first argument. This implies that there exists a j ∈ [ℓ − 1] such
that κmin (Yj , X) = κmin (Yj+1 , X). By the submodularity of κmin in the first argument,
κmin (Y \ {yj+1 }, X) + κmin (Yj+1 , X) ≥ κmin (Yj , X) + κmin (Y, X).
This implies κmin (Y \ {yj+1 }, X) = κmin (Y, X) = κ(X) contradicting the minimality of Y .
For disjoint sets B1 , B2 , we let L(B1 , B2 ) be the set of all X ⊆ U such that B1 is free in X and
B2 is free in X. If X ∈ L(B1 , B2 ), then we say that (B1 , B2 ) is a base for X.
Corollary 2.14. Every set X ⊆ U has a base.
For every X ⊆ U , we let B(X) be the set of all bases for X. Then (B1 , B2 ) ∈ B(X) ⇐⇒ X ∈
S
L(B1 , B2 ). We let B = X⊆U B(X) be the set of all bases. Obviously, B is the set of all pairs
(B1 , B2 ) such that B1 , B2 ⊆ U are disjoint and L(B1 , B2 ) 6= ∅.
Lemma 2.15. Let B1 , B2 ⊆ U be disjoint and X ∈ L(B1 , B2 ). Then κ(X) = κmin (B1 , B2 ).
Proof. As B1 ⊆ X ⊆ B 2 , we have κ(X) ≥ κmin (B1 , B2 ). Suppose for contradiction that κ(X) >
κmin (B1 , B2 ), and let Z ⊆ U such that B1 ⊆ Z ⊆ B 2 and κ(Z) = κmin (B1 , B2 ). Then κ(Z) < κ(X),
and by submodularity this implies κ(X ∩ Z) < κ(X) or κ(X ∪ Z) < κ(X). However, we have
B1 ⊆ X ∩ Z ⊆ X and thus κ(X ∩ Z) ≥ κmin (B1 , X) = κ(X). Similarly, B2 ⊆ X ∩ Z ⊆ X and thus
κ(X ∪ Z) = κ(X ∩ Z) ≥ κmin (B2 , X) = κ(X). This is a contradiction.
We call κmin (B1 , B2 ) the order of a base (B1 , B2 ) ∈ B. For every k ≥ 0, let B≤k denote the set
of all bases of order at most k.
Lemma 2.16. Let B1 , B2 ⊆ U be disjoint. Then (B1 , B2 ) ∈ B if and only if κmin (B1 , B2 ) ≥
max{|B1 |, |B2 |}.
Proof. For the forward direction, let X ∈ L(B1 , B2 ). Then κmin (B1 , B2 ) = κ(X) ≥ max{|B1 |, |B2 |},
where the equality follows from Lemma 2.15 and the inequality because B1 is free in X and B2 is
free in X.
For the backward direction, suppose that κmin (B1 , B2 ) ≥ max{|B1 |, |B2 |}, and let X ⊆ U such
that B1 ⊆ X ⊆ B 2 and κ(X) = κmin (B1 , B2 ). Then κ(X) ≥ |B1 | and κ(X) ≥ κmin (B1 , X) ≥
κmin (B1 , B2 ) = κ(X) and thus κ(X) = κmin (B1 , X). Similarly, κ(X) ≥ |B2 | and κ(X) =
κmin (B2 , X). Thus (B1 , B2 ) is a base for X.
Lemma 2.17. Let B1 , B2 ⊆ U be disjoint and X1 , X2 ∈ L(B1 , B2 ). Then X1 ∩ X2 , X1 ∪ X2 ∈
L(B1 , B2 ).
10
Proof. As B1 ⊆ X1 ∩ X2 , X1 ∪ X2 ⊆ B 2 , by Lemma 2.15 we have
κ(X1 ∩ X2 ), κ(X1 ∪ X2 ) ≥ κmin (B1 , B2 ) = κ(X1 ) = κ(X2 )
By submodularity, this implies
κmin (B1 , B2 ) = κ(X1 ) = κ(X2 ) = κ(X1 ∩ X2 ) = κ(X1 ∪ X2 ).
By the monotonicity of κmin in the second argument, we have
κ(X1 ∩ X2 ) ≥ κmin (B1 , X1 ∩ X2 ) ≥ κmin (B1 , B2 ) = κ(X1 ∩ X2 )
and thus κ(X1 ∩ X2 ) = κmin (B1 , X1 ∩ X2 ). Moreover, κ(X1 ∩ X2 ) = κ(X1 ) ≥ |B1 |, and thus B1 is
free in X1 ∩ X2 . Similarly,
κ(X1 ∩ X2 ) ≥ κmin (B2 , X1 ∩ X2 ) ≥ κmin (B2 , B1 )
= κmin (B1 , B2 ) = κ(X1 ∩ X2 ) = κ(X1 ∩ X2 )
and thus κ(X1 ∩ X2 ) = κmin (B2 , X1 ∩ X2 ). Moreover, κ(X1 ∩ X2 ) = κ(X1 ∩ X2 ) = κ(X2 ) ≥ |B2 |,
and thus B2 is free in X1 ∩ X2 .
Hence X1 ∩ X2 ∈ L(B1 , B2 ). By symmetry, X1 ∪ X2 ∈ L(B1 , B2 ).
The lemma shows that L(B1 , B2 ) has a lattice structure—hence the letter L. In particular, the
lemma has the following consequence.
Corollary 2.18. For every base (B1 , B2 ) ∈ B there is a unique minimal X⊥ (B1 , B2 ) ∈ L(B1 , B2 )
and a unique maximal X⊤ (B1 , B2 ) ∈ L(B1 , B2 ).
Note that in contrast to the leftmost separations defined earlier, the element X⊥ (B1 , B2 ) is
inclusion-wise minimal among all sets in L(B1 , B2 ) and not just among those of minimum rank.
2.4 Computing with tangles
Algorithms expecting a set function κ : 2U → N as input are given the ground set U as actual input
(say, as a list of objects), and they are given an oracle that returns for X ⊆ U the value of κ(X).
The running time of such algorithms is measured in terms of the size |U | of the ground set. We
assume this computation model whenever we say that an algorithm is given oracle access to a set
function κ.
An important fact underlying most of our algorithms is that, under this model of computation,
submodular functions can be efficiently minimised.
Fact 2.19 (Iwata, Fleischer, Fujishige [13], Schrijver [23]). There is a polynomial time
algorithm that, given oracle access to a submodular function λ : 2U → N, returns a set X ⊆ U that
minimises λ.
Observe that this implies that, given oracle access to a connectivity function κ, the function κmin
can also be evaluated in polynomial time. In fact, for given arguments X, Y we can compute a Z
such that X ⊆ Z ⊆ Y and κ(Z) = κmin (X, Y ) in polynomial time.
A membership oracle for a family S ⊆ 2U of subsets of a ground set U is an oracle that when
queried with a set X ⊆ U determines whether X ∈ S. Most often, we will use such membership
oracles for tangles; in the following lemma we use them for the union of a family of tangles.
11
X′
X
X∗
x
Y1
Y2
Z
Figure 1. Proof of Lemma 2.20
Lemma 2.20. Let k ≥ 1. There is a polynomial time algorithm that, given oracle access to a
connectivity function κ : 2U → N, a membership oracle for set S ⊆ 2U such that S = T1 ∪ · · · ∪ Tn
is a union of κ-tangles of order k, and sets Y1 , Y2 ⊆ U ,
• either computes a set X ′ satisfying
(1) X ′ ∈ T1 ∪ · · · ∪ Tn and
(2) Y1 ⊆ X ′ ⊆ Y2
such that X ′ is inclusion-wise minimal among all sets with these two properties
• or determines that no such set exists.
Proof. Starting by initially setting X = Y2 , it suffices to find repeatedly a proper subset X ′ ⊂ X
satisfying Properties (1) and (2) or determine that no such set exists.
Assume there is some proper subset X ∗ ⊂ X satisfying the properties. Let x be an element
of X \ X ∗ and let Z be a free subset of X ∗ . In particular κmin (Z, X ∗ ) = κ(X ∗ ) and |Z| ≤ κ(X ∗ ).
Such a set exists by Lemma 2.13.
Since |Z| ≤ κ(X ∗ ) ≤ k − 1, the number of choices for x and Z are polynomially bounded,
and by iterating over all choices we can guess x and Z. We compute an inclusion-wise maximal
set X ′ with κ(X ′ ) ≤ k − 1 such that Z ∪ Y1 ⊆ X ′ ⊆ X \ {x} (see Figure 1). Such a set X ′
can be computed by starting with X ′ := Z ∪ Y1 and then iteratively adding elements to X ′ as
long as κmin (X ′ , X \ {x}) ≤ k − 1. Note that Z ⊆ X ′ ∩ X ∗ ⊆ X ∗ and thus κ(X ′ ∩ X ∗ ) ≥
κmin (Z, X ∗ ) = κ(X ∗ ). By submodularity κ(X ∗ ) + κ(X ′ ) ≥ κ(X ′ ∩ X ∗ ) + κ(X ′ ∪ X ∗ ) which implies
that κ(X ′ ∪ X ∗ ) ≤ κ(X ′ ) and thus X ∗ ⊆ X ′ , since X ′ is chosen to be maximal. We conclude
that X ′ ∈ T1 ∪ · · · ∪ Tn . For a correct choice of Z and x, we can use the oracle for T1 ∪ · · · ∪ Tn to
verify that X ′ ∈ T1 ∪ · · · ∪ Tn .
The lemma in particular applies when we are given only one tangle by a membership oracle.
Lemma 2.21. Let T be a tangle of order k and B = (B1 , B2 ) a base such that T ∩ L(B) 6= ∅.
Then there is a unique inclusion-wise minimal element X⊥ (T , B) in T ∩ L(B).
Furthermore, for every fixed k there is a polynomial time algorithm that, given oracle access to
κ, a membership oracle for T , and a base B, decides if T ∩ L(B) 6= ∅ and, in the affirmative case,
computes X⊥ (T , B).
12
Proof. For the uniqueness, it is sufficient to prove that T ∩L(B) is closed under taking intersections.
So let X1 , X2 ∈ T ∩ L(B). Then by Lemma 2.17, X1 ∩ X2 ∈ L(B) and thus by Lemma 2.15,
κ(X1 ∩ X2 ) = κ(X1 ) = κ(X2 ). As X1 ∩ X2 ∩ X1 ∩ X2 = ∅, we have X1 ∩ X2 ∈ T .
The algorithmic claim follows by applying Lemma 2.20 with n = 1, T1 = T , Y1 = B1 and Y2 =
B2 .
3
A Data Structure for Tangles
The following crucial lemma will enable us to iteratively extend our tangle data structure.
Lemma 3.1. There is a polynomial time algorithm that, given oracle access to κ, a membership
oracle for a κ-tangle T0 of order at most k, and sets X1 , . . . , Xn ⊆ U with κ(Xi ) ≤ k, decides if
there is a κ-tangle T ⊇ T0 of order k + 1 that avoids X1 , . . . , Xn .
Proof. Observe that there is a κ-tangle T ⊇ T0 of order k + 1 that avoids X1 , . . . , Xn if and only if
there is a κ-tangle of order k + 1 that avoids
X := {X | X ⊆ Y for some Y ∈ T0 } ∪ {X | X ⊆ Xi for some i ∈ [n]} ∪ SU .
|
{z
=:X0
}
|
{z
=:X1
}
By the Duality Lemma, this is the case if and only if there is no partial decomposition of κ of width
less than k + 1 over X . Our algorithm will test whether such a partial decomposition exists.
e of κ of width
For a set Y ⊆ U , we say that a Y -decomposition is a partial decomposition (T, ξ)
less than k + 1 that has a leaf tY , which we call the Y -leaf, such that ξ(tY ) = Y and ξ(t) ∈ X for
all t ∈ L(T ) \ {tY }. We call a set Y decomposable if there is a Y -decomposition.
Claim 1. Let X, Y, Z ⊆ U such that X and Y are decomposable and Z ⊆ X ∪ Y and κ(Z) ≤ k.
Then Z is decomposable.
Proof. Let (TX , ξeX ) be an X-decomposition and (TY , ξeY ) a Y -decomposition. Let tX be the X-leaf
of TX , and let tY be the Y -leaf of TY . We form a new tree T by taking the disjoint union of TX
and TY , identifying the two leaves tX and tY , and adding a fresh node tZ and an edge between tZ
−
→
and the node sZ corresponding to tX and tY . We define ξe : E (T ) → 2U by
→
ξeX (s, t) if (s, t) ∈ −
E (TX ),
→
ξe (s, t) if (s, t) ∈ −
E (TY ),
Y
e t) :=
ξ(s,
Z
if (s, t) = (tZ , sZ ),
Z
if (s, t) = (sZ , tZ ).
e is a partial decomposition of κ, because X ∪ Y ∪ Z = U . Its width is at most k because
Then (T, ξ)
κ(Z) ≤ k and the width of both (TX , ξeX ) and (TY , ξeY ) is at most k. We have ξ(tZ ) = Z, and for
all leaves t ∈ L(T ) \ {tZ }, either t ∈ L(TX ) \ {tX } and ξ(t) = ξX (t) ∈ X or t ∈ L(TY ) \ {tY } and
e is a Z-decomposition.
ξ(t) = ξY (t) ∈ X . Thus (T, ξ)
y
Claim 2.
Let X, Y ⊆ U such that X and Y are decomposable and X ∪ Y = U . Then there is a
partial decomposition of κ of width at most k over X .
Proof. This follows from Claim 1 by observing that an U -decomposition is a partial decomposition
of κ over X of width at most k. Here we use the facts that U = ∅ ∈ X and that κ(∅) = 0 ≤ k. y
Our algorithm iteratively computes mappings µ1 , . . . , µm : B≤k → 2U such that for all B =
(B1 , B2 ) ∈ B≤k , either µi (B) = ∅ or µi (B) ∈ L(B) and µi (B) is decomposable.
To define µ1 , for every base B = (B1 , B2 ) ∈ B≤k , we first check if T0 ∩ L(B2 , B1 ) 6= ∅ (note the
reversed order in the base) and, if it is, compute X⊥ (T0 , (B2 , B1 )) (see Lemma 2.21).
13
• If T0 ∩ L(B2 , B1 ) 6= ∅, we let Y0 := X⊥ (T0 , (B2 , B1 )). Observe that Y0 is the unique inclusionwise maximal element in L(B) ∩ X0 . Otherwise, we let Y0 := ∅.
• If L(B) ∩ X1 6= ∅, then we let Y1 be the union of all elements of L(B) ∩ X1 . Otherwise, we let
Y1 := ∅.
• If L(B) ∩ SU 6= ∅, then we let Y2 be the union of all elements of L(B) ∩ SU . Otherwise, we
let Y2 := ∅.
We let µ1 (B) := Y0 ∪ Y1 ∪ Y2 . Observing that the elements of X are trivially decomposable, it
follows from Claim 1 and the fact that L(B) is a lattice that if µ1 (B) 6= ∅ then it is decomposable.
Now suppose that µi is defined. Consider the following condition.
(⋆) For all B, C, D ∈ B≤k and Y ∈ L(B), if Y ⊆ µi (C) ∪ µi (D), then Y ⊆ µi (B).
If (⋆) is satisfied, then we let m := i, and the constructions stops.
Otherwise, we find B, C, D ∈ B≤k and Y ∈ L(B) such that Y ⊆ µi (C) ∪ µi (D) and Y 6⊆ µi (B).
(Oum and Seymour [16] explain how to do this in polynomial time: for all B, C, D ∈ B≤k and for
all y ∈ µi (B) we test if there is a Y ∈ L(B) with {y} ⊆ Y ⊆ µi (C) ∪ µi (D).)
We let µi+1 (B) := µi (B) ∪ Y . It follows from Claim 1 that µi+1 (B) is decomposable. For all
′
B ∈ B≤k \ {B}, we let µi+1 (B ′ ) := µi (B ′ ). This completes the description of the construction.
Since in each step of the construction we strictly increase the set µi (B) for some base B and
since the number of bases is polynomially bounded in |U |, the number m of steps of the construction
is polynomially bounded as well. Let µ := µm . Note that µ satisfies (⋆).
Claim 3.
The following are equivalent.
(1) There is a partial decomposition of κ of width at most k over X .
(2) There are B, C ∈ B≤k such that µ(B) ∪ µ(C) = U .
Proof. The implication “(2) =⇒ (1)” follows immediately from Claim 2.
e be an exact partial decomposition of κ of width at most k
To prove “(1) =⇒ (2)”, let (T, ξ)
over X . Let
Y := {Y | κ(Y ) ≤ k and Y ⊆ µ(B) for some B ∈ B≤k }.
→
e t) ∈ Y for all (s, t) ∈ −
By induction on T , we shall prove that ξ(s,
E (T ).
−
→
In the base step, we consider (s, t) ∈ E (T ) where t is a leaf. Then ξ(t) ∈ X . We let B be a
e t) = ξ(t) ∈ Y.
base for ξ(t). Then ξ(t) ⊆ µ1 (B) ⊆ µ(B) and thus ξ(s,
−
→
For the inductive step, consider an edge (s, t) ∈ E (T ) such that t is an inner node and such that
e u), ξ(t,
e v) ∈ Y for the other two neighbours u, v of t. Let C, D ∈ B
e
ξ(t,
≤k such that ξ(t, u) ⊆ µ(C)
e v) ⊆ µ(D). Let B = (B1 , B2 ) be a base for Y := ξ(s,
e t), and note that Y = ξ(t,
e u) ∪ ξ(t,
e v),
and ξ(t,
e
because (T, ξ) is an exact partial decomposition. Hence Y ⊆ µ(C) ∪ µ(D). By (⋆), we have
Y ⊆ µ(B) and thus Y ∈ Y.
→
e t) ∈ Y for all (s, t) ∈ −
Thus ξ(s,
E (T ). Consider an arbitrary edge st ∈ E(T ). Let B, C ∈ B≤k
e t) = ξ(t,
e s) ⊆ µ(C). Then µ(B) ∪ µ(C) = U .
e t) ⊆ µ(B) and ξ(s,
such that ξ(s,
y
Since Condition (2) of Claim 3 can be tested in polynomial time, this completes the proof.
Observe that the following result due to Oum and Seymour [16] follows by applying the theorem
with T0 being the empty tangle (of order 0) and n = 0. In fact, our proof of the lemma builds on
Oum and Seymour’s proof.
14
Corollary 3.2 (Oum and Seymour [16]). For every k there is a polynomial time algorithm
deciding whether there is a κ-tangle of order k.
A comprehensive tangle data structure of order k for a connectivity function κ over a set U
is a data structure D with procedures OrderD , SizeD , TD , TangOrderD , TruncationD , SepD ,
and FindD that provide the following functionalities.
(1) The function OrderD () returns the fixed integer k.
(2) For ℓ ∈ {1, . . . , k} the function SizeD (ℓ) returns the number of κ-tangles of order at most ℓ.
We denote the number of κ-tangles of order at most k by |D|.
(3) For each i ∈ {1, . . . , |D|} the function TD (i, ·) : 2U → {0, 1} is a tangle Ti of order at most k,
(i.e., the function call TD (i, X) determines whether X ∈ Ti ).
We call i the index of the tangle Ti within the data structure.
(4) For i ∈ {1, . . . , |D|} the function TangOrderD (i) returns ord(Ti ).
(5) For i ∈ {1, . . . , |D|} and ℓ ≤ ord(Ti ) the call TruncationD (i, ℓ) returns an integer j such
that Tj is the truncation of Ti to order ℓ. If ℓ > ord(Ti ) the function returns i.
(6) For distinct i, j ∈ {1, . . . , |D|} the call SepD (i, j) outputs a set X ⊆ U such that X is the
leftmost minimum (Ti , Tj )-separation (see Lemma 2.8) or states that no such set exists (in
which case one of the tangles is a truncation of the other).
(7) Given an integer ℓ ∈ {0, . . . , k} and some tangle T ′ of order ℓ (via a membership oracle) the
function FindD (ℓ, T ′ ), returns the index of T ′ , that is, the unique integer i ∈ {1, . . . , |D|}
such that ord(Ti ) = ℓ and T ′ = Ti .
Remark 3.3. Note that, in accordance with Remark 2.7, a tangle data structure considers tangles
as distinct even if they only differ in their order and not as a function 2U → {0, 1}. Recall that a
tangle of order k agrees with its truncation to order k − 1 as a function 2U → {0, 1} if and only
if there is no X ⊆ U of order κ(X) = k − 1. Due to the existence of free sets (Lemma 2.13) it is
possible to check in polynomial time whether such an X exists and thus determine which tangles
coincide.
We say a comprehensive tangle data structure D is efficient if all functions OrderD , SizeD ,
TD , TangOrderD , TruncationD , SepD , and FindD can be evaluated in polynomial time.
Theorem 3.4. For every constant k there is a polynomial time algorithm that, given oracle access
to a connectivity function κ, computes an efficient comprehensive tangle data structure of order k.
Proof. Let k be an integer and let κ : 2U → N be a connectivity function. Note that for k = 0 it
is trivial to construct an efficient comprehensive tangle data structure. We suppose by induction
that we have already constructed an efficient comprehensive tangle data structure of order k − 1
in polynomial time. We first verify that there exists some tangle of order k (using Corollary 3.2).
If this is not the case then the comprehensive tangle data structure of order k − 1 is already a
comprehensive tangle data structure of order k. Otherwise we proceed as follows.
We compute a binary rooted tree T and a function S assigning to every edge (s, t) ∈ E(T ) that
is pointing away from the root a subset S(s, t) ⊆ U with κ(S(s, t)) < k such that the following
properties hold.
15
(1) If s is a vertex of T with children t1 and t2 then S(s, t1 ) = S(s, t2 ).
(2) For each path p1 , . . . , pm form the root to a leaf there is a tangle T of order k such that
S(p1 , p2 ), S(p2 , p3 ), . . . , S(pm−1 , pm ) ∈ T .
(3) For each path p1 , . . . , pm form the root to a leaf there is at most one tangle T of order k
satisfying property (2).
Claim 1. A tree T satisfying properties (1)–(3) can be computed in polynomial time.
Proof. We construct the tree iteratively from subtrees which satisfy properties (1) and (2) but not
necessarily property (3). We say a tangle corresponds to a leaf u if it satisfies property (2) for the
path from the root to u. Note that properties (1) and (2) imply that each tangle corresponds to
exactly one leaf and that property (3) implies that different tangles correspond to different leaves.
We start with the tree T0 that only contains one vertex r. It satisfies properties (1) and (2).
Suppose we have constructed a tree Ti satisfying properties (1) and (2). To construct Ti+1 it suffices
to find a leaf u and a set X with κ(X) < k such that there are tangles T1 and T2 of order at most k
both corresponding to u with X ∈ T1 and X ∈
/ T2 .
Note that for each candidate set X and each leaf u we can determine by Lemma 3.1 in polynomial
time whether there are two tangles corresponding to u that are separated by X. It thus suffices
for us to compute a set of candidates for X among which there is an adequate separator. For a
leaf u we proceed as follows. Let X1 = S(p1 , p2 ), . . . , Xm−1 = S(pm−1 , pm ) where p1 , . . . , pm is
the path from the root to u. (The collection of sets Xi is empty if the tree has only one node.)
Let T1 , . . . , Tn be the tangles of order k which correspond to u. These are exactly the tangles
of order k avoiding X1 , . . . , X m−1 . Lemma 3.1 provides a membership oracle to test for a set X
whether X ∈ T1 ∪ . . . ∪ Tn : to check whether X ∈ T1 ∪ . . . ∪ Tn we test, using Lemma 3.1 for T0 := ∅
(the unique tangle of order 0), whether there is a tangle of order k avoiding X1 , . . . , X m−1 and X.
We have X ∈ T1 ∪ . . . ∪ Tn if and only if the answer is affirmative for some tangle T .
For each base B ∈ B≤k−1 , we compute, if it exists, an inclusion-wise minimal set X ∗ (B) ∈ L(B)
(implying κ(X ∗ (B)) < k) such that there is a tangle Ti (with i ∈ {1, . . . , n}) containing X ∗ (B).
Such a set can be computed in polynomial time by Lemma 2.20.
To prove the claim, it suffices now to show that if there are two tangles corresponding to u then
there is a base B ∈ B≤k−1 for which X ∗ (B) exists and separates two tangles corresponding to u.
e be a (Tj , Tk )-separation. Let B be a base for X.
e By
Let Tj and Tk be two such tangles and let X
construction, there is a tangle Ti with X ∗ (B) ∈ Ti . If X ∗ (B) ∈
/ Tj or X ∗ (B) ∈
/ Tk then X ∗ (B)
separates two tangles corresponding to u (either Ti and Tj or Ti and Tk ). So suppose X ∗ (B) ∈
e because X
e ∈
e ( X ∗ (B). We
Tj and X ∗ (B) ∈ Tk . Then X ∗ (B) 6⊆ X,
/ Tk . Thus X ∗ (B) ∩ X
e ∈ L(B), because L(B) is a lattice. Since X ∗ (B) is inclusion-wise minimal in
have X ∗ (B) ∩ X
e ∈
e ∈ L(B) and all elements of L(B)
L(B) ∩ (T1 ∪ . . . ∪ Tt ), this implies that X ∗ (B) ∩ X
/ Tj . Since X
e ∈ Tj . However the
e = κ(X)
e ≤ k − 1. Thus X ∗ (B) ∩ X
have the same order, we have κ(X ∗ (B) ∩ X)
e have an empty intersection so they cannot all be contained in Tj ,
e and X ∗ (B) ∩ X
sets X ∗ (B), X,
yielding a contradiction.
Repeating the construction we obtain a sequence of trees T0 , T1 , . . .. Since by Fact 2.12 there
is only a linear number of tangles of order at most k and each new tree distinguishes more tangles
than the one before, after a linear number of steps we obtain a tree satisfying properties (1)–(3).y
Let Dk−1 be an efficient comprehensive tangle data structure of order k − 1. We argue that
using Dk−1 together with a tree T satisfying properties (1)–(3) we obtain a comprehensive tangle
data structure of order k.
16
Recall that the tangles of order k are in one to one correspondence to the leaves of the tree T .
Let u1 , . . . , un be an enumeration of the leaves of T .
(1) The function OrderD () simply returns the integer k.
(2) The function SizeD (ℓ) returns SizeDk−1 (ℓ) if ℓ < k and SizeDk−1 (ℓ−1)+n if ℓ = k (where n =
|L(T )| is the number of leaves of T ).
(3) To evaluate TD (i, X) we return TDk−1 (i, X) in case i ≤ |Dk−1 |. Otherwise we suppose
that X1 = S(p1 , p2 ), . . . , Xm−1 = S(pm−1 , pm ) where p1 , . . . , pm is the path from the root
to leaf ui−|Dk−1| . We use Lemma 3.1 to determine whether there exists a tangle avoiding X1 , . . . , X m−1 and X. If this is the case we return 1, otherwise we return 0.
(4) For i ∈ {1, . . . , |D|} the function TangOrderD (i) returns TangOrderDk−1 (i) if i ≤ |Dk−1 |.
Otherwise, it returns k.
(5) To determine TruncationD (i, ℓ), if ℓ = k we return i. Assuming otherwise, if i ≤ |Dk−1 | we
use TruncationDk−1 (i, ℓ). If neither of these cases happens, we modify the function TD (i, X)
to return 0 for all X with κ(X) ≥ ℓ − 1. This provides us with an oracle for a tangle T that
is the ℓ − 1 truncation of Ti . We can then use FindDk−1 (k − 1, T ) to determine the index of
this truncation.
(6) To compute SepD (i, j), if i ≤ |Dk−1 | or j ≤ |Dk−1 | then we can use the truncation and simply
compute SepDk−1 (Truncation(i, k − 1), Truncation(j, k − 1)). Otherwise we let s be the
smallest common ancestor of ui−|Dk−1| and uj−|Dk−1| in T . Let X be the set S(s, t) where t
is the child of s on the path from s to ui−|Dk−1| . Using Lemma 2.20 we compute and return
the minimal set X ′ ⊆ X such that X ′ ∈ Ti .
(7) To determine FindD (ℓ, T ), if ℓ < k we return FindDk−1 (ℓ, T ). Otherwise, the order of T is k,
and there is a leaf u of the tree T such that T is the tangle at this leaf. We start at the root r
of T and traverse towards a leaf as follows. Suppose we are currently at node s with children t1
and t2 . Exactly one of the two sets S(s, t1 ) and S(s, t2 ) = S(s, t1 ) is contained in T and we
traverse to the corresponding child. Once we hit a leaf, uj say, we return |Dk−1 | + j.
4
Canonical Tree Decompositions
In this section we present the canonical tree decomposition of a connectivity function into parts
corresponding to its tangles of order at most k. Our decomposition is more or less the same as the
one presented by Hundertmark [12], but our construction differs in two aspects that are important
for the algorithmic treatment. We exclusively choose leftmost and rightmost minimum separations
in our decomposition. Hundertmark is less restrictive about the separations he uses, which makes it
easier to argue that suitable separations exist, but infeasible to find them algorithmically. Moreover,
our construction is modular: we introduce new connectivity functions during the construction,
decompose them and then merge the decompositions.
In this section, we often speak of “canonical” constructions. The precise technical meaning depends on the context, but in general a construction (or algorithm) is canonical if every isomorphism
between its input objects commutes with an isomorphism between the output objects. More formally, suppose we have a construction (or algorithm) A that associates an output A(I) with every
input I. Then the construction is canonical if for any two inputs I1 and I2 and every isomorphism
17
f from I1 to I2 there is an isomorphism g from A(I1 ) to A(I2 ) such that g(A(I1 )) = A(I2 ), that is,
the following diagram commutes:
I1
f
I2
A
A(I1 )
A
g
.
A(I2 )
We define an isomorphism from a connectivity function κ1 : 2U1 → N to a connectivity function
κ2 : 2U2 → N to be a bijective mapping f : U1 → U2 such that κ(X) = κ(f (X)) for all X ⊆ U1 . For
sets U1 , U2 and families X1 ⊆ 2U1 , X2 ⊆ 2U2 we define an isomorphism from (U1 , X1 ) to (U2 , X2 ) to
be a bijective mapping f : U1 → U2 such that X ∈ X1 ⇐⇒ f (X) ∈ X2 for all X ⊆ X1 . We assume
that the reader is familiar with tree isomorphisms.
It may be worth noting that our construction of comprehensive tangle data structures described
in the previous section is not canonical.
4.1 Tree Decomposition and Nested Separations
Let U be a finite set. We think of U as being the ground set of a connectivity function κ, but
this connectivity function plays no role in this section. A tree decomposition of U is a pair (T, β)
consisting of a tree T and a function β : V (T ) → 2U such that the sets β(t) for t ∈ V (T ) are
mutually disjoint and their union is U . If κ is a connectivity function on U , we also call (T, β) a
tree decomposition of κ.
e t) be the union
Let (T, β) be a tree decomposition of U . For every edge st ∈ E(T ) we let β(s,
′
′
of the sets β(t ) for all nodes t in the connected component of T − st that contains t. Note that
e s). If (T, β) is a tree decomposition of a connectivity function κ on U , we define the
e t) = β(t,
β(s,
→
e t)) | (s, t) ∈ −
adhesion of (T, β) to be max{κ(β(s,
E (T )}. We do not define a “width” for our tree
decompositions.
e corresponds to the tree decomposition (T, β)
Observe that every branch decomposition (T, ξ)
with β(t) = ξ(t) for all t ∈ L(T ) and β(T ) := ∅ for all t ∈ V (T ) \ L(T ). Therefore, we may just
view branch decompositions as special tree decompositions.
Our notion of a tree decomposition of a connectivity function is not new (see for example [7]). It
may be surprising to a reader only familiar with tree decompositions of graphs, because it partitions
the elements of U , whereas the bags of a tree decomposition of a graph may overlap. But note
that if we apply this notion to the connectivity function κG of a graph G (see Example 2.2), we
decompose the edge set and not the vertex set of G. The following example details the relation
between standard tree decompositions of graphs and tree decompositions of κG .
Example 4.1. Let (T, β) be a tree decomposition of a graph G (in the usual sense). It yields a
tree decomposition (T, β ′ ) of κG (in the sense defined above) as follows: for every edge e ∈ E(G),
we arbitrarily choose a node te ∈ V (T ) that covers e. Then for every t ∈ V (T ) we let β ′ (t) := {e ∈
E(G) | t = te }.
Conversely, if we have a tree decomposition (T, β ′ ) of κG , then we can define a tree decomposition
(T, β) of G as follows. For every node v ∈ V (G) we let Uv be the set of all nodes t ∈ V (T ) such
bv be the union of Uv with all nodes t ∈ V (T )
that v is incident with an edge e ∈ β ′ (t). We let U
bv }. We call
appearing on a path between two nodes in Uv . Now we let β(t) = {v ∈ V (G) | t ∈ U
′
(T, β) the tree decomposition of G corresponding to (T, β ).
Note that the construction of a tree decomposition of κG from a tree decomposition of G involves
arbitrary choices, whereas the construction of a tree decomposition of G from a tree decomposition
18
of κG is canonical. Thus the “tree decomposition of a graph corresponding to a tree decomposition
of its edge set” is well-defined.
y
Let (T, β) be a tree decomposition of a set U . We let
e t) | st ∈ E(T )}.
N (T, β) = {β(s,
and call it the set of separations of (T, β). We will now characterise sets of separations that come
from tree decompositions.
Sets (or rather, separations) X, Y ⊆ U are nested if either X ⊆ Y or X ⊆ Y or X ⊆ Y or
X ⊆ Y ; otherwise X and Y cross. Note that X and Y cross if and only if the four sets X ∩ Y ,
X ∩ Y , X ∩ Y , and X ∩ Y are all nonempty. A family N ⊆ 2U is nested if all X, Y ∈ N are nested.
Observe that for every tree decomposition (T, β) of κ the set N (T, β) is nested and closed under
complementation. The following converse of this observation is well-known and goes back (at least)
to [21]. (We include a proof for the reader’s convenience.)
Lemma 4.2. If N ⊆ 2U is nested and closed under complementation, then there is a tree decomposition (T, β) of U such that N = N (T, β).
Furthermore, the construction of (T, β) from U and N is canonical and can be carried out by a
polynomial-time algorithm.
Proof. By induction on |N | we construct a rooted tree (T, r) and a mapping β : V (T ) → 2U such
that (T, β) is a tree decomposition with N (T, β) = N .
In the base step N = ∅, we let T be a tree with one node r and we define β by β(r) := U .
In the inductive step N =
6 ∅, let X1 , . . . , Xm be a list of all inclusion-wise minimal elements of
N (possibly, m = 1 and X1 = ∅). As N is nested, for all i 6= j we have Xi ⊆ X j . This implies that
the sets Xi are mutually disjoint. Let
N ′ := N \ {Xi , X i | i ∈ [m]}.
By the induction hypothesis, there is a rooted tree (T ′ , r ′ ) and a mapping β ′ : V (T ′ ) → 2U such
that (T ′ , β ′ ) is a tree decomposition with N (T ′ , β ′ ) = N ′ . For every i ∈ [m], let ti be a node of
minimum height (where the height of a node is its distance to the root) with Xi ⊆ βe′ (si , ti ) for
the parent si of ti , or ti := r ′ if no such node exists. Observe that there is only one such node ti .
Indeed, if t 6= ti has the same height as ti , then neither ti = r nor t = r. Let s be the parent of t.
Then the edges (si , ti ) and (s, t) are pointing away from each other and thus βe′ (si , ti ) ∩ βe′ (s, t) = ∅.
This implies Xi 6⊆ βe′ (s, t), unless Xi = ∅, in which case ti is the root, and we do not have to worry
about this.
We define a new tree T from T ′ by attaching a fresh leaf ui to ti for every i ∈ [m]. We let
r := r ′ be the root of T . We define β : V (T ) → 2U by
β(t) :=
(
Xi
S
β ′ (t) \ m
i=1 Xi
if t = ui ,
if t ∈ V (T ′ ).
As the sets Xi are mutually disjoint, (T, β) is a tree decomposition of U . We need to prove that
N (T, β) = N .
−
→
e t) = βe′ (s, t).
Claim 1.
For all oriented edges (s, t) ∈ E (T ′ ) we have β(s,
−
→ ′
Proof. Let (s, t) ∈ E (T ). Without loss of generality we assume that t is a child of s. We have
βe′ (s, t) ∈ N ′ ⊆ N . By the minimality of Xi and the nestedness of N , we have Xi ⊆ βe′ (s, t) or
19
Figure 2. The graph of Example 4.3
Xi ∩ βe′ (s, t) = ∅. Moreover, Xi ⊆ βe′ (s, t) if any only if ui is a descendant of t in T . As βe′ (s, t) is
e t) is the union of the sets β(t′ )
the union of the sets β ′ (t′ ) for all descendants t′ of t in T ′ and β(s,
′
for all descendants t of t in T , the claim follows.
y
e t) for some oriented edge (s, t) ∈
To prove that N (T, β) ⊆ N , let Z ∈ N (T, β). Say, Z = β(s,
e
E(T ). If (s, t) = (ti , ui ) for some i ∈ [m], then Z = Xi ∈ N , and if (s, t) = (ui , ti ) then Z = X i ∈ N ,
−
→
because N is closed under complementation. Otherwise, (s, t) ∈ E (T ′ ). Then by Claim 1 we have
Z = βe′ (s, t) ∈ N ′ ⊆ N .
e , u ), and
To prove the converse inclusion, let Z ∈ N . If Z = Xi for some i ∈ [m], then Z = β(t
i i
′
′
e
e
e t) for
if Z = X i , then Z = β(ui , ti ). Otherwise, Z ∈ N , and thus by Claim 1, Z = β (s, t) = β(s,
−
→ ′
some (s, t) ∈ E (T ).
Thus (T, β) is a tree decomposition of U with N (T, β) = N . The construction is obviously
canonical, and it is easy to see that it can be carried out by a polynomial time algorithm.
It is our goal to construct tree decompositions whose parts correspond to tangles and whose
separations separate these tangles. If T is a family of mutually incomparable κ-tangles, then a tree
decomposition for T is a triple (T, β, τ ), where (T, β) is a tree decomposition of κ and τ : T → V (T )
is an injective mapping with the following properties.
−
→
(TD.1) For all distinct T , T ′ ∈ T there is an oriented edge (t, t′ ) ∈ E (T ) on the oriented path
e ′ , t) is a minimum (T , T ′ )-separation.
from τ (T ) to τ (T ′ ) in T such that β(t
−
→
(TD.2) For every oriented edge (t, t′ ) ∈ E (T ) there are tangles T , T ′ ∈ T such that (t, t′ ) appears
e ′ , t) is a minimum (T , T ′ )-separation.
on the oriented path from τ (T ) to τ (T ′ ) and β(t
e ′ , t) ∈ T .
(TD.3) For every tangle T ∈ T and every neighbour t′ of t := τ (T ) in T it holds that β(t
Nodes t ∈ τ (T) are called tangle nodes and the remaining nodes t ∈ V (T ) \ τ (T) are called hub
nodes.
For an arbitrary family of T-tangles, we say that a tree decomposition for T is a tree decomposition for the family Tmax ⊆ T consisting of all inclusion-wise maximal tangles in T.
The following example shows that in general we cannot do without hub nodes if we want to
construct canonical decompositions.
Example 4.3. Let G be the graph consisting of three triangles joined at a single node (see Figure 2). The connectivity function κG has three tangles of order 2 corresponding to the three
triangles.
Let (T, β, τ ) be a canonical tree decomposition for the family of all κG -tangles of order at most
2. Here canonical means that for every automorphism f of the graph G there is an automorphism
20
g of the tree T such that f (β(t)) = β(g(t)) for all t ∈ V (T ). Typically, T would be a star with
three leaves, which are tangle nodes associated with the three tangles of order 2, and one centre,
which is a hub node.
But now suppose for contradiction that the decomposition has no hub nodes. Then T must be
a path of length 2. One of the tangles must be associated with the centre node of this path, and
the other two with the leaves. But as the automorphism group of G acts transitively on the three
tangles, this contradicts the canonicity.
y
Lemma 4.4. Let (T, β, τ ) be a tree decomposition for a family T of mutually incomparable κtangles.
(1) For every tangle node t = τ (T ) and for every oriented edge (u′ , u) pointing towards t, if
e ′ , u)) < ord(T ) then β(u
e ′ , u) ∈ T .
κ(β(u
(2) E(T ) = ∅ if and only if |T| ≤ 1.
(3) If T 6= ∅, all leaves of T are tangle nodes.
(4) The mapping τ : V (T ) → T is uniquely determined by the tree decomposition (T, β). That is,
if τ ′ is an injective mappings from T to V (T ) such that (T, β, τ ′ ) is a tree decomposition for
T, then τ = τ ′ .
Proof. (1) follows from (TD.3) and the fact that for every edge (u, u′ ) pointing towards t there is
e ′ , u) ⊇ β(t
e ′ , t). The
a neighbour t′ of t (the neighbour of t on the path from u′ to t) such that β(u
forward direction of (2) follows from (TD.1) and the backward direction follows from (TD.2). (3)
follows from (TD.2).
To prove (4), suppose for contradiction that τ 6= τ ′ . Let T ∈ T such that t := τ (T ) 6= τ ′ (T ) =: t′ .
Let u, u′ be the neighbours of t, t′ , respectively, on the path from t to t′ in T . Then by (TD.3),
e ′ , t), β(u
e ′ , u) ∈ T . However, β(t
e ′ , t) ∩ β(u
e ′ , u) = ∅. This contradicts T being a tangle.
β(t
The next lemma shows that a canonical tree decomposition for a family T of tangles can be
constructed from a nested family of separations satisfying two extra conditions relating it to T.
Lemma 4.5. Let T be a family of κ-tangles and N ⊆ 2U a family of separations that is nested and
closed under complementation and satisfies the following two conditions.
(TN.1) For all T , T ′ ∈ T with T ⊥T ′ there is a Z ∈ N such that Z is a minimum (T , T ′ )separation.
(TN.2) For all Z ∈ N there are tangles T , T ′ ∈ T with T ⊥T ′ such that Z is a minimum (T , T ′ )separation.
Then for every tree decomposition (T, β) of κ with N (T, β) = N there is a unique injective mapping
τ : Tmax → V (T ) such that (T, β, τ ) is a tree decomposition for T.
Furthermore, given (T, β) and the index set of T in a comprehensive tangle data structure, the
mapping τ can be computed in polynomial time.
Proof. Without loss of generality we may assume that the tangles in T are mutually incomparable;
otherwise we work with Tmax instead of T. Observe that N satisfies conditions (TN.1) and (TN.2)
with respect to T if any only it satisfies the two conditions with respect to Tmax .
We may further assume that |T| ≥ 2. Then N =
6 ∅ by (TN.1). Furthermore, ord(T ) ≥ 1 for
all T ∈ T, because the unique tangle of order 0 is the empty tangle, which is comparable with all
other tangles.
21
Let (T, β) be a tree decomposition of κ with N (T, β) = N . For every k ≥ 1, we let Ek be the
e t′ )) < k.
set of all edges e = tt′ ∈ E(T ) with κ(β(t,
For every tangle T ∈ T of order k we construct a connected subset XT ⊆ V (T ) as follows: we
e t′ ) ∈ T then
orient all edges e = tt′ ∈ Ek in such a way that they point towards T , that is, if β(t,
the orientation of e is (t, t′ ) and otherwise the orientation is (t′ , t). Then there a unique connected
component of T − Ek (the forest obtained from T by deleting all edges in Ek ) such that all oriented
edges point towards this component. We let XT be the node set of this connected component.
It follows from (TN.1) that the sets XT are mutually vertex disjoint. To see this, consider
−
→
distinct T , T ′ ∈ T. Let Z ∈ N be a minimum (T , T ′ ) separation and (t, t′ ) ∈ E (T ) such that
e ′ , t) = Z. Then XT is contained in the connected component of T − tt′ that contains t and X ′
β(t
T
is contained in the connected component of T − tt′ that contains t′ . Hence XT ∩ XT ′ = ∅.
Let T , T ′ ∈ T be distinct, and let Z ∈ N be a minimum (T , T ′ )-separation.
−
→
e ′ , t) = Z.
(1) There is an oriented edge (t, t′ ) ∈ E (T ) such that β(t
−
→
e ′ , t) = Z appears on the oriented path from
(2) Every oriented edge (t, t′ ) ∈ E (T ) such that β(t
XT to XT ′ .
Claim 1.
Proof. (1) follows immediately from N (T, β) = N .
−
→
e ′ , t). As Z ∈ T the oriented edge (t′ , t) points
To prove (2), let (t, t′ ) ∈ E (T ) such that Z = β(t
e t′ ) ∈ T ′ the oriented edge (t, t′ ) points towards X ′ . It follows that
towards XT , and as Z = β(t,
T
−
→
′
y
the oriented edge (t, t ) appears on the oriented path P from XT to XT ′ in T .
Claim 2.
For all T ∈ T it holds that |XT | = 1.
Proof. Suppose for contradiction that |XT | > 1 for some T ∈ T. Let X := XT . As X is
connected, there is an edge e = t1 t2 ∈ E(T ) with both endvertices in X. Then e 6∈ Eord(T ) and
e , t )) ≥ ord(T ).
thus κ(β(t
1 2
e , t ) ∈ N is a minimum (T , T )-separation. Such tangles
Let T1 , T2 ∈ T such that Z := β(t
2 1
1 2
exist by (TN.2). For i = 1, 2, let Xi := XTi . By Claim 1, the oriented edge (t1 , t2 ) appears on the
−
→
oriented path P from X1 to X2 in T .
We have
e 1 , t2 )) = κ(Z) < min{ord(T1 ), ord(T2 )}.
ord(T ) ≤ κ(β(t
Let Z1 ∈ N be a minimum (T1 , T )-separation. Then κ(Z1 ) < ord(T ) ≤ κ(Z). Moreover, by
−
→
e u1 ) =
Claim 1, there is an oriented edge (u1 , u) on the oriented path Q from X1 to X such that β(u,
Z1 .
−
→
We have Z1 ∈ T1 , because Z1 is a (T1 , T )-separation. Since t1 ∈ X, the path Q is an initial
−
→
−
→
segment of the path P , and therefore (u1 , u) is also an edge of P . The edge (u1 , u) occurs before
−
→
e 1 , u) ⊇ β(t
e 1 , t2 ) = Z, and as Z ∈ T2 , this implies Z 1 ∈ T2 .
(t1 , t2 ) on the path P . Thus Z 1 = β(u
Hence Z1 is a (T1 , T2 )-separation. As κ(Z1 ) < κ(Z), this contradicts the minimality of Z.
y
We define τ : T → V (T ) by letting τ (T ) be the unique node in XT , for all T ∈ T. This mapping
is well-defined by Claim 2, and injective, because the sets XT are mutually disjoint.
It follows from (TN.1) and Claim 1 that (T, β, τ ) satisfies (TD.1). It follows from (TN.2) and
N (T, β) = N and Claim 1 that (T, β, τ ) satisfies (TD.2).
By the construction of XT , for all oriented edges (t′ , t) with t ∈ XT and t′ 6∈ XT it holds that
e ′ , t) ∈ T . As XT = {τ (T )}, this implies that (T, β, τ ) satisfies (TD.3).
β(t
The uniqueness of τ follows from Lemma 4.4(4). As τ (T ) is the unique node t of T such that
′
e
β(t , t) ∈ T for all neighbours t′ of t, it is straightforward to compute τ in polynomial time.
22
We call a family N ⊆ 2U that is nested and closed under complementation and satisfies (TN.1)
and (TN.2) a nested family for T. Observe the converse of Lemma 4.5: if (T, β, τ ) is a tree
decomposition for T, then N (T, β) is a nested family for T.
Remark 4.6. It follows from Lemma 4.5 that (TD.1) and (TD.2) imply (TD.3) and that we can
even replace (TD.1) and (TD.2) by the weaker conditions (TN.1) and (TN.2) for N = N (T, β).
The reason that we nevertheless used (TD.1)–(TD.3) is that they state the crucial properties
that we expect from a tree decomposition for a family of tangles.
4.2 Decomposing Coherent Families
Let us call a family T of κ-tangles of order k+1 coherent if all elements of T have the same truncation
to order k. Observe that this condition implies, and is in fact equivalent to, the condition that for
distinct T , T ′ ∈ T the order of a minimum (T , T ′ )-separation is k. The main result of this section,
Lemma 4.8, shows how to compute a tree decomposition for a coherent family of tangles of order
k + 1. In Section 4.3, we will then combine decompositions for different coherent sets of tangles of
different orders.
The family of separations of the tree decomposition our algorithm computes for a given set T
of tangles will be a subset of the set
Z(T) = {Z(T , T ′ ) | T , T ′ ∈ T such that T ⊥T ′ }.
of all leftmost minimum separations of pairs of tangles in T and of their complements. The following
lemma is similar to Lemma 5.3 of [12]. But our proof is different, because we work with different
assumptions and a different set of separations.
Lemma 4.7. Let T be a coherent family of κ-tangles of order k +1, and let Z0 ∈ Z(T) be inclusionwise minimal. Then for all Z ∈ Z(T), either Z0 ⊆ Z or Z0 ⊆ Z.
Proof. Let T0 , T0′ ∈ T such that Z0 = Z(T0 , T0′ ). Moreover, let T , T ′ ∈ T be distinct, and let Z be
a minimum (T , T ′ )-separation. We shall prove that Z0 ⊆ Z or Z0 ⊆ Z. Of course this will imply
the assertion of the lemma, because every Z ∈ Z(T) is a minimum (T , T ′ )-separation for some
T , T ′ ∈ T.
Without loss of generality, we may assume that
Z ∈ T0 .
(4.A)
Otherwise, we swap T and T ′ and take Z instead of Z.
Suppose first that κ(Z0 ∩ Z) ≤ k. Then Z0 ∩ Z ∈ T0 , because Z0 ∩ Z ∩ Z0 ∩ Z = ∅. Moreover,
we have Z0 ∩ Z ∈ T0′ , because (Z0 ∩ Z) ∩ Z 0 = ∅. Thus Z0 ∩ Z is a (T0 , T0′ )-separation. As Z0 is
leftmost minimum, it follows that Z0 ⊆ Z0 ∩ Z and thus Z0 ⊆ Z.
In the following, we assume κ(Z0 ∩ Z) > k = κ(Z). By submodularity, κ(Z0 ∪ Z) < κ(Z0 ). We
have Z0 ∪ Z ∈ T0 , because Z0 ∪ Z ∩ Z0 = ∅. If Z0 ∪ Z was a (T0 , T0′ )-separation, κ(Z0 ∪ Z) < κ(Z0 )
would contradict the minimality of Z0 . Hence Z0 ∪ Z 6∈ T0′ , which implies Z0 ∪ Z ∈ T0′ . As Z 0 ∈ T0′
and Z 0 ∩ Z ∩ (Z0 ∪ Z) = ∅, it follows that
Z ∈ T0′ .
(4.B)
Case 1: Z0 ∈ T ′ .
Then Z0 is a (T ′ , T0′ )-separation. As T ′ , T0′ have the same truncation to order k, Z0 is in
23
fact a minimum (T ′ , T0′ )-separation, and this implies Z(T ′ , T0′ ) ⊆ Z0 . By the inclusion-wise
minimality of Z0 in Z(T), this implies Z0 = Z(T ′ , T0′ )
By (4.B), Z is another (T ′ , T0′ )-separation. As κ(Z) = κ(Z0 ), it follows that Z0 = Z(T ′ , T0′ ) ⊆
Z.
Case 2: Z 0 ∈ T ′ .
Then Z0 is a (T0 , T ′ )-separation, and by a similar argument as in Case 1 it follows that
Z0 = Z(T0 , T ′ ).
By (4.A), Z is another (T0 , T ′ )-separation, and it follows that Z0 = Z(T0 , T ′ ) ⊆ Z.
Lemma 4.8. Let k ≥ 0. There is a polynomial time algorithm that, given a coherent family T of
κ-tangles of order k + 1 for a connectivity function κ on a set U (via a comprehensive tangle data
structure and the set of indices of the tangles in T), computes a canonical nested family for T.
Proof. The idea of the proof is to construct a tree decomposition for T starting from the leaves of the
decomposition tree and then moving towards the centre of the tree. Observe that the separations
of a tree decomposition associated with the edges towards the leafs are precisely the inclusion-wise
minimal separations.
The algorithm inductively computes for all i ∈ N a set Ni ⊆ Z(T) of separations and a family
Ti of tangles.
• N0 := ∅ and T0 := ∅.
• Suppose that Ni and Ti are already computed. Then the algorithm repeatedly queries the
tangle data structure to obtain Z(T \ Ti ).
Ni+1 is the union of Ni with all inclusion-wise minimal Z ∈ Z(T \ Ti ), and Ti+1 is the set of
all tangles T ∈ T such that Z(T , T ′ ) ∈ Ni+1 for some T ′ ∈ T.
S
Let N be the closure of i≥0 Ni under complementation. It is easy to see that N can be computed
in polynomial time. We claim that N is a nested family for T.
It follows from Lemma 4.7 that N is nested: when we add a Z0 to Ni+1 , it is nested with all
S
Z(T \ Ti ) and thus with all Z ∈ j≥i+1 Nj .
The family N trivially satisfies (TN.2), because each element of each Ni is an element of Z(T).
It remains to prove that N satisfies (TN.1). For all i ≥ 0 we prove that for all T ∈ Ti+1 \ Ti ,
T ′ ∈ T \ Ti there is a Z ∈ Ni+1 such that Z is a minimum (T , T ′ )-separation. Let T ∈ Ti+1 \ Ti ,
T ′ ∈ T \ Ti . Let T ′′ ∈ T such that Z = Z(T , T ′′ ) ∈ Ni+1 . Then Z is inclusion-wise minimal in
′
Z(T \ Ti ). Let Z ′ = Z(T , T ′ ). By Lemma 4.7, either Z ⊆ Z ′ or Z ⊆ Z . If Z ⊆ Z ′ , then Z ∈ T ′ ,
′
′
because Z ⊇ Z ∈ T ′ , and thus Z is a (T , T ′ )-separation. If Z ⊆ Z , then Z ∩ Z ′ = ∅, which
contradicts Z, Z ′ ∈ T .
S
S
Now we observe that T \ i≥0 Ti ≤ 1, because otherwise Z(T \ i≥0 Ti ) would be nonempty
and the constructions would not have stopped. So for distinct tangles T , T ′ ∈ T, at least one of
S
them is i≥0 Ti , and for some i ≥ 0 either T ∈ Ti+1 \ Ti and T ′ ∈ T \ Ti or vice versa. We have just
seen that then Ni+1 contains a (T , T ′ )-separation or a (T ′ , T )-separation, and this implies that N ,
which is closed under complementation, contains a (T , T ′ )-separation.
4.3 Decomposing Arbitrary Families
In this section, we will describe how to build a “global” tree decomposition of all tangles of order
at most k + 1 from “local” decompositions for coherent families of tangles. Suppose that we have
24
C5
t1
t5
B
c1
c5
C1
B
t
C4
C2
c4
t4
c3
t2
C3
c2
t3
(a) Node t and its neighbours ti
and the sets B, Ci
(b) Contraction of the Ci
Figure 3. Sets at a node t of a tree decomposition
already built a tree decomposition (T ≤k , β ≤k , τ ≤k ) for the family T≤k of all κ-tangles of order at
≤k (T ) = t.
most k. Consider a tangle node t of this decomposition and let Tt ∈ T≤k
t
max such that τ
Suppose that ord(Tt ) = k, and let Tt be the set of all κ-tangles of order k + 1 whose truncation
to order k is Tt . We want to expand our decomposition to a decomposition over the set T≤k ∪ Tt .
In fact, we want to do this simultaneously for all nodes t of T in a consistent way to obtain a tree
decomposition for T≤k+1 , but let us focus on just one node t first. Suppose that the neighbours of t
e i , t), and let B = β(t) (see Figure 3(a)). Then
in T ≤k are t1 , . . . , tm . For every i ∈ [m], let Ci = β(t
Sm
Tm
B = i=1 Ci and U = B ∪ i=1 C i . Furthermore, the sets C i for i ∈ [m] are mutually disjoint.
We now “contract” each of the sets C i to a single vertex and define a new connectivity function
on the resulting set. We take fresh points c1 , . . . , cm not in U and let U ↓t := B ∪ {c1 , . . . , cm } (see
Figure 3(b)). We call U ↓t the contraction of U at t. We define the expansion of a set X ⊆ U ↓t to be
S
the set X↑t := (X ∩B)∪ ci ∈X C i . We define a set function κ↓t on U ↓t by letting κ↓t (X) := κ(X↑t ).
It is easy to verify that κ↓t is a connectivity function on U ↓t . For every κ-tangle T we let
T ↓t := {X ⊆ U ↓t | X↑t ∈ T }.
T ↓t is not necessarily a κ↓t -tangle, because it may violate tangle axiom (T.3). However, it is easy
to see that T ↓t is a κ↓t -tangle of the same order as T if C i 6∈ T for all i ∈ [m]. By (TD.3), for all
T ∈ Tt we have Ci ∈ Tt ⊆ T and thus C i 6∈ T for all i. Thus ↓t defines a mapping from Tt to the
set of all κ↓t -tangles of order k + 1. Lemma 4.10 below implies that this “contraction mapping” is
injective.
Lemma 4.9. Let X ⊆ U such that there are κ-tangles T , T ′ for which X is a minimum (T , T ′ )separation. Then for every Y ⊆ U , either κ(X ∩ Y ) ≤ κ(Y ) or κ(X \ Y ) ≤ κ(Y ).
25
Proof. Suppose for contradiction that κ(X ∩ Y ) > κ(Y ) and κ(X \ Y ) > κ(Y ). Then by submodularity, κ(X ∪ Y ) < κ(X) and κ(X ∪ Y ) < κ(X).
Now let T , T ′ be tangles T , T ′ such that X is a minimum (T , T ′ )-separation. As X ⊆ X ∪
Y, X ∪ Y and X ∈ T , we have X ∪ Y, X ∪ Y ∈ T . Furthermore, either X ∪ Y ∈ T ′ or X ∪ Y ∈ T ′ ,
because X ∩ (X ∪ Y ) ∩ (X ∪ Y ) = ∅. Thus either X ∪ Y or X ∪ Y is a (T , T ′ )-separation of order
less than κ(X). This contradicts the minimality of X.
Lemma 4.10. Let T , T ′ ⊇ Tt be incomparable. Then T ↓t and T ′ ↓t are incomparable as well, and
for every minimum (T ↓t , T ′ ↓t )-separation Z the expansion Z↑t is a minimum (T , T ′ )-separation.
Proof. We choose a minimum (T , T ′ )-separation Y in such a way that it maximises the number of
i ∈ [m] with Y ∩ C i = ∅ or C i ⊆ Y .
Claim 1.
For all i ∈ [m], either Y ∩ C i = ∅ or C i ⊆ Y .
Proof. Suppose for contradiction that there is some i ∈ [m] such that ∅ ⊂ C i ∩ Y ⊂ C i . By (TD.2),
there are tangles Ti , Ti′ such that Ci = βe≤k (ti , t) is a minimum (Ti , Ti′ )-separation. By Lemma 4.9
(applied to X := Ci and Y ), either κ(Y ∩ Ci ) ≤ κ(Y ) or κ(Y ∩ Ci ) ≤ κ(Y ).
Suppose first that κ(Y ∩ Ci ) ≤ κ(Y ). Then Y ∩ Ci ∈ T , because Y ∈ T and Ci ∈ Tt ⊆ T and
Y ∩ Ci ∩ Y ∩ Ci = ∅. Furthermore, Y ∩ Ci ∈ T ′ , because Y ∩ (Y ∩ Ci ) = ∅. Thus (Y ∩ Ci ) is a
minimum (T , T ′ )-separation as well. Furthermore, (Y ∩ Ci )∩ C i = ∅, and for all j 6= i, if Y ∩ C j = ∅
e tj ) ⊆ β(t
e i , t) = Ci .
then (Y ∩ Ci ) ∩ C j = ∅, and if C j ⊆ Y , then C j ⊆ (Y ∪ Ci ), because C j = β(t,
This contradicts the choice of Y .
Suppose next that κ(Y ∩ Ci ) ≤ κ(Y ). Arguing as above with Y, Y and T , T ′ swapped, we
see that Y ∩ Ci is a minimum (T ′ , T )-separation. Thus Y ∪ C i is a minimum (T , T ′ )-separation.
We have C i ⊆ Y ∪ C i , and for all j 6= i, if C j ⊆ Y then C j ⊆ C i ∪ Y , and if C j ∩ Y = ∅ then
C j ∩ (C i ∪ Y ) = C j ∩ C i = ∅. Again, this contradicts the choice of Y .
y
It follows from Claim 1 that there is a Y ′ ⊆ U ↓t such that Y = Y ′ ↑t . This set Y ′ is a
(T ↓t , T ′ ↓t )-separation. Thus T ↓t ⊥T ′ ↓t , and the order k′ of a minimum (T ↓t , T ′ ↓t )-separation is
at most κ(Y ′ ) = κ(Y ) =: k, the order of a minimum (T , T ′ )-separation. Now let Z ⊆ U ↓t be a
minimum (T ↓t , T ′ ↓t )-separation. Then the expansion Z↑t is a (T , T ′ )-separation. Thus
k ≤ κ(Z↑t ) = κ↓t (Z) = k′ ≤ k.
Hence k = k′ , and Z↑t is a minimum (T , T ′ )-separation.
Corollary 4.11. The “contraction mapping” ↓t on Tt is injective.
Proof. Let T , T ′ ∈ Tt with T =
6 T ′ . Then T ⊥T ′ , because ord(T ) = ord(T ′ ) = k + 1, and it follows
from the lemma that T ↓t 6= T ′ ↓t .
Finally, we are ready to prove the main theorem this section.
Theorem 4.12. Let ℓ ≥ 0. Then there is a polynomial time algorithm that, given oracle access to
a connectivity function κ, computes a canonical tree decomposition for the set of all κ-tangles of
order at most ℓ.
Proof. Our algorithm first computes a comprehensive tangle data structure of order ℓ. Then, by
k
induction on 0 ≤ k ≤ ℓ, it computes a tree decomposition (T ≤k , β ≤ , τ ≤k ) for the set T≤k of all
κ-tangles of order at most k.
The base step k = 0 is trivial, because the only tangle of order 0 is the empty set, and the
trivial one-node tree decomposition is a tree decomposition for this tangle.
26
For the inductive step k → k + 1 (where k < ℓ), we assume that we have already constructed
k
a canonical tree decomposition (T ≤k , β ≤ , τ ≤k ) for T≤k . Let N ≤k := N (T ≤k , β ≤k ) be the set of
separations of this tree decomposition.
For every tangle node t ∈ τ ≤k (T≤k
max ) we do the following:
• We compute the set Tt of all tangles T ⊇ Tt := (τ ≤k )−1 (t) of order k + 1.
• We compute U ↓t ; we can use our oracle for κ to implement an oracle for κ↓t . Then we
compute a comprehensive tangle data structure of order k + 1 for κ↓t .
• Within this data structure, we compute the family T∨
t of all tangles T ↓t for T ∈ Tt .
Note that T∨
t is a coherent family of κ↓t -tangles of order k + 1.
∨
∨
• We apply Lemma 4.8 to κ↓t and T∨
t and obtain a nested family Nt for Tt .
• We compute the set Nt := {X↑t | X ∈ Nt∨ }.
Claim 1.
The family N ≤k+1 := N ≤k ∪
S
t∈V (T ≤t ) Nt of
N ≤k is nested.
separations is nested.
Proof. We already know that the family
Furthermore, for every t ∈ V (T ≤k ) the
family Nt∨ is nested, and this implies that Nt is nested as well.
Thus we need to show that the sets in Nt are nested with all sets in N ≤k as well as all sets in
Nu for u 6= t. So let X ∈ Nt . As before, let t1 , . . . , tm be the neighbours of t in T ≤k . Then for all
e ti ) ⊆ X or X ∩ β(t,
e ti ) = ∅ and thus β(t,
e ti ) ⊆ X or β(t,
e ti ) ⊆ X.
i ∈ [m], either β(t,
−
→
e s′ ) for some (s, s′ ) ∈ E (T ≤k ). Then there is an i ∈ [m]
Every set Z ∈ N ≤k is of the form β(s,
′
e
e
e ′ , s) ⊆ β(t,
e t ). As β(t,
e t ) ⊆ X or β(t,
e t ) ⊆ X, it
such that either Z = β(s, s ) ⊆ β(t, ti ) or Z = β(s
i
i
i
follows that Z and X are nested.
Now consider a set Y ∈ Nu for some node u 6= t. Let ti be the neighbour of t and uj the
e u ) ⊆ Y or β(u,
e
neighbour of u on the path from t to u in T ≤k . Then either β(u,
uj ) ⊆ Y , which
j
e uj ) = β(u
e j , u) ⊆ β(t,
e ti ) or Y ⊆ β(u
e j , u) ⊆ β(t,
e ti ). As β(t,
e ti ) ⊆ X or β(t,
e ti ) ⊆ X,
implies Y ⊆ β(u,
it follows that Y and X are nested.
y
Claim 2.
N ≤k+1 is a nested family for T≤k+1 .
Proof. By Claim 1, the family N ≤k+1 is nested. By construction, it is closed under complementation (instead of going through the construction to check this, we can also just close it under
complementation without any harm). We have to prove that it satisfies (TN.1) and (TN.2) for
T := T≤k+1 .
It satisfies (TN.2), because N ≤k does and by Lemma 4.10, for all t ∈ V (T ≤k ), all Z ∈ Nt are
minimum separations for tangles in Tt ⊆ T≤k+1 .
To see that N ≤k+1 satisfies (TN.1), let T , T ′ ∈ T≤k+1 be incomparable. Let T0 be the truncation
of T to order k if ord(T ) = k + 1 and T0 := T otherwise, and let T0′ be defined similarly from T ′ .
If T0 ⊥T0′ , there is a Z ∈ N ≤k that is a minimum (T0 , T0′ )-separation, and this Z is also a minimum
(T , T ′ )-separation. Otherwise, ord(T ) = ord(T ′ ) = k + 1 and T0 = T0′ . Let t := τ ≤k (T0 ). Then
T , T ′ ∈ Tt , and Nt contains a minimum (T , T ′ )-separation.
y
We use the algorithm of Lemma 4.2 to a compute a canonical tree decomposition (T ≤k+1 , β ≤k+1 )
of κ with N (T ≤k+1 , β ≤k+1 ) = N ≤k+1 , and we use the algorithm of Lemma 4.5 to turn it into a
tree decomposition (T ≤k+1 , β ≤k+1 , τ ≤k+1 ) for T≤k+1 .
Let us say that a κ-tangle is ℓ-maximal if it is inclusion-wise maximal among all κ-tangles of
order at most ℓ. In our previous notation, T≤ℓ
max denotes the set of all ℓ-maximal κ-tangles.
27
Remark 4.13. The existence of a canonical tree decomposition for the set T≤ℓ of all tangles of order
at most ℓ, which of course follows from Theorem 4.12, implies that for every ℓ the number |T≤ℓ
max |
of ℓ-maximal κ-tangles is at most |U | − 1, provided |U | ≥ 2.
To see this, we assume without loss of generality that |T≤ℓ
max | ≥ 2. Let (T, β, τ ) be a tree
≤ℓ
decomposition for T . Then |T | ≥ 2. By Lemma 4.4, all leaves of T are tangle nodes. Let t be
a leaf, s the neighbour of t, and let Tt ∈ T≤ℓ
max be the tangle with τ (Tt ) = t. By (TD.3) we have
e
β(t) = β(s, t) ∈ Tt . By (T.2) and (T.3), this implies |β(t)| > 1.
Now let t ∈ V (T ) be a tangle node of degree 2, say, with neighbours s and u. Again, let
e
e
Tt ∈ T≤ℓ
max be the tangle τ (Tt ) = t. By (TD.3), we have β(u, t), β(s, t) ∈ Tt , which implies β(t) =
e
e
β(u, t) ∩ β(s, t) 6= ∅.
Let n1 , n2 , n≥3 be the numbers of tangle nodes of degree 1, 2, at least 3, respectively. We have
2n1 + n2 ≤ |U |. Furthermore, n≥3 < n1 , because a tree with n1 leaves has less than n1 vertices of
degree at least 3. Thus
|T≤ℓ
y
max | = n1 + n2 + n≥3 < 2n1 + n2 ≤ |U |.
We close this section with another decomposition algorithm that may be useful in some applications. Theorem 4.12 yields a tree decomposition (T, β) where at most one ℓ-maximal κ-tangle
is associated with every node. However, in applications we may want to work with the “local
structure” at the nodes t of the decomposition, and this local structure is represented by the
“contractions” κ↓t on U ↓t . To understand this local structure, we might be more interested in κ↓t tangles than in κ-tangles associated with t. It is not clear whether at every tangle node t there is at
most one ℓ-maximal κ↓t -tangle, and we know even less about the hub nodes of the decomposition.
However, the following theorem shows that we can also construct a decomposition where at every
node t we have at most one ℓ-maximal κ↓t -tangle.
Theorem 4.14. Let ℓ ≥ 0. Then there is a polynomial time algorithm that, given oracle access
to a connectivity function κ, computes a canonical tree decomposition (T, β) of κ of adhesion less
than ℓ such that for all t ∈ V (T ) there is exactly one ℓ-maximal κ↓t -tangle.
Proof. We start by computing the decomposition of (T 0 , β 0 , τ 0 ) of Theorem 4.12. Let N 0 :=
N (T 0 , β 0 ). It follows from (TD.2) that the adhesion of this decomposition is less than ℓ. As every
connectivity function has at least one tangle (the empty tangle of order 0), for every node t ∈ V (T )
there is at least one ℓ-maximal κ↓t -tangle.
Suppose first that T 0 is a star with centre t0 and |β 0 (t)| = 1 for all t ∈ V (T 0 ) \ {t0 }. Then
κ↓t0 = κ (up to renaming of the elements of U in U ↓t0 ). As there is at most one ℓ-maximal κ-tangle
T0 with βe0 (t, t0 ) ∈ T0 for every neighbour t of t0 , there is at most one ℓ-maximal κ↓t0 -tangle. For
all t 6= t0 , we have |U ↓t | = 2, and this implies that there is at most one ℓ-maximal κ↓t -tangle.
So suppose that T 0 is not such a star. Then |U ↓t | < |U | for all t ∈ V (T 0 ). For every node
t ∈ V (T 0 ) such that there is more than one ℓ-maximal κ↓t -tangle, we recursively apply the algorithm
to κ↓t and obtain a tree decomposition (Tt , βt ) of κ↓t . We let Nt := N (Tt , βt ). For nodes t such
that there is only one ℓ-maximal κ↓t -tangle, we let Nt := ∅, and we let
N := N 0 ∪
[
Nt .
t∈V (T 0 )
Using similar arguments as in the proof of Claim 1 in the proof of Theorem 4.12, it is easy to prove
that N is nested and closed under complementation.
We apply the algorithm of Lemma 4.2 to N and obtain a tree decomposition (T, β) of κ with
N (T, β) = N . It is easy to see that this decomposition has the desired properties.
28
5
Directed Decompositions
In this section, we prove a variant of our canonical decomposition theorem (Theorem 4.12) in which
we get rid of the hub nodes at the price weakening the separation properties of the decomposition
and loosing a bit of the canonicity (the decomposition will only be canonical given one tangle).
This version of the decomposition theorem is used in [10] to design a polynomial isomorphism test
for graph classes of bounded rank width.
We work with a directed version of tree decompositions here. A directed tree is an oriented tree
where all edges are directed away from the root. In a directed tree T , by ET , or just E if T is clear
from the context, we denote the “descendant order”, that is, the reflexive transitive closure of the
edge relation. The set of children of node t ∈ V (T ) is denoted by N+T (t) or just N+ (t).
A directed tree decomposition of a set U or a connectivity function κ on U is a pair (T, γ), where
T is a directed tree and γ : V (T ) → 2U such that
• γ(r) = U for the root r of T ;
• γ(t) ⊇ γ(u) for all (t, u) ∈ E(T );
• γ(u1 ) ∩ γ(u2 ) = ∅ for all siblings u1 , u2 .
(We call u1 , u2 siblings if u1 6= u2 and there is a t ∈ V (T ) such that u1 , u2 ∈ N+ (t).)
We call γ(t) the cone of the decomposition at node t. We define β : V (T ) → 2U by
β(t) := γ(t) \
[
γ(u).
(5.A)
u∈N+ (t)
We call β(t) the bag of the decomposition at node t. Observe that the bags are mutually disjoint
and that their union is U . Thus if T◦ is the undirected tree underlying T , then (T◦ , β) is an
(undirected) tree decomposition in the sense defined before. Moreover, for all (s, t) ∈ E(T ) we have
e t).
γ(t) = β(s,
Conversely, let (T◦ , β) be an undirected tree decomposition. Let T a directed tree with underlying undirected tree T◦ (obtained by arbitrarily choosing a root and directing all edges away from
the root) and define γ : V (T ) → 2U by
γ(t) :=
[
β(u).
(5.B)
uDt
Then (T, γ) is a directed tree decomposition.
We always denote the bag function of a directed tree decomposition (T, γ) by β, and we use
implicit naming conventions by which, for example, we denote the bag function of (T ′ , γ ′ ) by β ′ .
Now let T be a family of mutually incomparable κ-tangles. A directed tree decomposition for T
is a triple (T, γ, τ ), where (T, γ) is a directed tree decomposition of κ and τ : T → V (T ) a bijective
mapping such that the following two conditions are satisfied.
(DTD.1) For all nodes t, u ∈ V (T ) with u 6E t there is a minimum (τ −1 (u), τ −1 (t))-separation Y
such that γ(u) ⊆ Y .
(DTD.2) For all nodes t ∈ V (T ) except the root, there is a node u ∈ V (T ) such that t 6E u and
γ(t) is a leftmost minimum (τ −1 (t), τ −1 (u))-separation.
Observe that (DTD.1) implies that for all nodes t ∈ V (T ) and children u ∈ N+ (t) we have γ(u) 6∈
τ −1 (t). Furthermore, (DTD.2) implies that γ(t) ∈ τ −1 (t).
29
Theorem 5.1. Let ℓ ≥ 0. Then there is a polynomial time algorithm that, given oracle access
to a connectivity function κ and a κ-tangle Troot ∈ T≤ℓ
max (via a membership oracle or its index
in a comprehensive tangle data structure for κ), computes a canonical directed tree decomposition
−1 (r) = T
(T, γ, τ ) for the set T≤ℓ
root for the root r of T .
max such that τ
′
′
an
Here canonical means that if κ′ : 2U → N is another connectivity function and Troot
′
′
′
′
ℓ-maximal κ -tangle, and (T , γ , τ ) is the decomposition computed by our algorithm on input
′
′
′ ), then for every isomorphism f from (κ, T
(κ′ , Troot
root ) to (κ , Troot ), that is, bijective mapping
′
for all X ⊆ U ,
f : U → U ′ with κ(X) = κ′ (f (X)) and X ∈ Troot ⇐⇒ f (X) ∈ Troot
′
′
there is an isomorphism g from T to T such that that f (γ(t)) = γ (g(t)) for all t ∈ V (T ) and
X ∈ τ −1 (t) ⇐⇒ f (X) ∈ (τ ′ )−1 (g(t)) for all X ⊆ U, t ∈ V (T ).
≤ℓ
Proof of Theorem 5.1. Without loss of generality we assume that |T≤ℓ
max | ≥ 2. Let Troot ∈ Tmax .
We start our construction from a canonical undirected tree decomposition (T◦ , β◦ , τ◦ ) for T≤ℓ
max ,
which we compute by the algorithm of Theorem 4.12. We let r := τ◦ (Troot ) and henceforth think of
the tree T◦ as being rooted in r. We denote the descendant order in this rooted tree by E◦ . For all
S
t ∈ V (T◦ ), we let γ◦ (t) := uD◦ t β◦ (u). Observe that γ◦ (r) = U and γ◦ (t) = βe◦ (s, t) for all nodes t
with parent s.
We let
≤ℓ
V := τ◦ (T≤ℓ
max ) = {τ◦ (T ) | T ∈ Tmax },
and we let T (0) be the directed tree with vertex set V (T (0) ) := V and edge set
E(T (0) ) := (t, u) ∈ V 2 t ✁◦ u and there is no x ∈ V such that t ✁◦ x ✁◦ u .
Let E(0) be the descendant order in T (0) . Obviously, E(0) is the restriction of E◦ to V .
−1 (t).
We let τ := τ◦ . Note that τ is a bijection between T≤ℓ
max and V . For all t ∈ V we let Tt := τ
Observe that for all t ∈ V (T ) we have γ◦ (t) ∈ Tt : if t = r is the root, we have γ◦ (t) = U ∈ Tt ,
and if s is the parent of t, we have γ◦ (t) = βe◦ (s, t) ∈ Tt by (TD.3).
We define γ : V → 2U as follows: we let γ(r) := U and for every node t ∈ V \ {r} we let γ(t) be
the leftmost minimum (Tt , γ◦ (t))-separation. As γ◦ (t) is a (Tt , γ◦ (t))-separation, there is a unique
leftmost minimum (Tt , γ◦ (t))-separation. Then γ(t) ∈ Tt and κ(γ(t)) ≤ κ(γ◦ (t)) and γ(t) ⊆ γ◦ (t).
Note that we cannot just let γ(t) = γ◦ (t), because γ◦ (t) is not necessarily a leftmost minimum
(Tt , γ◦ (t))-separation, and (DTD.2) requires leftmost minimum separations.
Claim 1.
For all t ∈ V \ {r} there is a t′ ∈ V such that t 6E(0) t′ and γ◦ (t) is a minimum
(Tt , Tt′ )-separation and γ(t) is a leftmost minimum (Tt , Tt′ )-separation
Proof. Let t ∈ V (T )\{r} and T := Tt . Let s be the parent of t in the rooted undirected tree (T◦ , r).
e t). By (TD.2), there are tangles T ′′ , T ′ such that the oriented edge (t, s) appears
Then γ◦ (t) = β(s,
e t) is a minimum
on the oriented path from t′′ := τ (T ′′ ) to t′ := τ (T ′ ) in T◦ , and γ◦ (t) = β(s,
′′
′
′
′′
′′
′
(T , T )-separation. Note that t , t ∈ V (T ) and t E t and t 6E t . Furthermore, γ◦ (t) ∈ T is a
(T , T ′ )-separation.
By (TD.1), there is an edge (v ′ , v) on the oriented path from t′ to t in T◦ such that βe◦ (v ′ , v) is
a minimum (T , T ′ )-separation. Then
e t) = γ (t)
βe◦ (v ′ , v) ⊇ β(s,
◦
and κ(βe◦ (v, v ′ )) ≤ κ(γ◦ (t)).
Suppose for contradiction that κ(βe◦ (v ′ , v)) < κ(γ◦ (t)). As γ◦ (t) ⊆ βe◦ (v ′ , v), this implies
e
β◦ (v ′ , v) ∈ T ′′ , and thus βe◦ (v ′ , v) is also a (T ′′ , T ′ )-separation. This contradicts γ◦ (t) being a
minimum (T ′′ , T ′ )-separation.
30
Thus κ(βe◦ (v, v ′ )) = κ(γ◦ (t)), and thus γ◦ (t) is a minimum (T , T ′ )-separation as well.
It remains to prove that γ(t) is a leftmost minimum (T , T ′ )-separation. Let Y be the leftmost
minimum (T , T ′ )-separation. Then κ(Y ) = κ(γ◦ (t)) and Y ⊆ γ◦ (t).
y
It follows that Y is also the leftmost minimum (T , γ◦ (t))-separation, and thus γ(t) = Y .
Claim 2.
For all t, u ∈ V such that t ✁(0) u there is Y ⊆ U such that γ(u) ⊆ γ◦ (u) ⊆ Y ⊆ γ◦ (t)
and Y is a minimum (Tu , Tt )-separation.
Proof. By (TD.1), there is an edge (t′ , u′ ) on the path from t to u in T◦ such that Y := γ◦ (u′ ) =
e ′ , u′ ) is a minimum (Tu , Tt )-separation. Then γ(u) ⊆ γ◦ (u) ⊆ γ◦ (u′ ) = Y ⊆ γ◦ (t).
β(t
y
Claim 3.
For all (t, u) ∈ E(T (0) ), either γ(u) ⊆ γ(t) or γ(t) ∩ γ(u) = ∅.
Proof. Let (t, u) ∈ E(T (0) ). If t = r then we trivially have γ(u) ⊆ U = γ(t). Therefore, we assume
that t 6= r. Let t′ ∈ V such that t 6E(0) t′ and γ◦ (t) is a minimum (Tt , Tt′ )-separation and γ(t)
is a leftmost minimum (Tt , Tt′ )-separation. Such an t′ exists by Claim 1. Similarly, let u′ ∈ V
such that u 6E(0) u′ and γ◦ (u) is a minimum (Tu , Tu′ )-separation and γ(u) is a leftmost minimum
(Tu , Tu′ )-separation.
Suppose for contradiction that γ(u) 6⊆ γ(t) and γ(t) ∩ γ(u) 6= ∅, or equivalently, γ(u) 6⊆ γ(t).
Case 1: κ(γ(t)) ≤ κ(γ(u)).
Then either γ(t) ∈ Tu or γ(t) ∈ Tu .
Case 1a: γ(t) ∈ Tu .
If κ(γ(t) ∩ γ(u)) ≤ κ(γ(u)), then γ(t) ∩ γ(u) ∈ Tu and γ(t) ∩ γ(u) ∈ Tu′ . Thus γ(t) ∩ γ(u)
is a (Tu , Tu′ )-separation. As γ(t) ∩ γ(u) ⊂ γ(u) by our assumption γ(u) 6⊆ γ(t), this
contradicts γ(u) being a leftmost minimum (Tu , Tu′ )-separation. Thus κ(γ(t) ∩ γ(u)) >
κ(γ(u)).
By submodularity,
κ(γ(t) ∪ γ(u)) < κ(γ(t)) ≤ κ(γ◦ (t)).
Then γ(t) ∪ γ(u) ∈ Tt . Furthermore, γ(t) ∪ γ(u) ⊆ γ◦ (t) ∪ γ◦ (u) ⊆ γ◦ (t). Thus
γ(t) ∪ γ(u) ⊇ γ◦ (t) ∈ Tt′ , and γ(t) ∪ γ(u) is a (Tt , Tt′ )-separation. This contradicts
γ◦ (t) being a minimum (Tt , Tt′ )-separation.
Case 1b: γ(t) ∈ Tu .
If κ(γ(t) ∩ γ(u)) ≤ κ(γ(u)), then γ(t) ∩ γ(u) ∈ Tu and γ(t) ∩ γ(u) ∈ Tu′ . Thus γ(t) ∩ γ(u)
is a (Tu , Tu′ )-separation. As γ(t) ∩ γ(u) ⊂ γ(u) by our assumption γ(u) 6⊆ γ(t), this
contradicts γ(u) being a leftmost minimum (Tu , Tu′ )-separation. Thus κ(γ(t) ∩ γ(u)) >
κ(γ(u))
By posimodularity,
κ(γ(t) ∩ γ(u)) < κ(γ(t)).
This implies γ(t) ∩ γ(u) = γ(t) ∪ γ(u) ∈ Tt′ . By Claim 2, there exists a minimum
(Tu , Tt )-separation Y ⊇ γ(u). We have
γ(t) ∩ Y ∩ (γ(t) ∪ γ(u)) ⊆ γ(t) ∩ γ(u) ∩ (γ(t) ∪ γ(u)) = ∅.
Thus γ(t) ∩ γ(u) = γ(t) ∪ γ(u) ∈ Tt , and γ(t) ∩ γ(u) is a (Tt , Tt′ )-separation. As κ(γ(t) ∩
γ(u)) < κ(γ(t)), this contradicts γ(t) being a minimum (Tt , Tt′ )-separation.
31
Case 2: κ(γ(t)) > κ(γ(u)).
Then either γ(u) ∈ Tt or γ(u) ∈ Tt . As above, let Y ⊇ γ(u) be a minimum (Tu , Tt )-separation.
Then Y ∈ Tt , and as Y ∩ γ(u) = ∅, we have γ(u) ∈ Tt .
If κ(γ(t)∩ γ(u)) ≤ κ(γ(t)), then γ(t)∩ γ(u) is a (Tt , Tt′ )-separation. As γ(t)∩ γ(u) ⊂ γ(t), this
contradicts γ(t) being a leftmost minimum (Tt , Tt′ )-separation. Hence κ(γ(t)∩γ(u)) > κ(γ(t)).
By posimodularity,
κ(γ(t) ∩ γ(u)) < κ(γ(u)).
We have γ(t) ∩ γ(u) = γ(t) ∪ γ(u) ∈ Tu′ , and as γ(u) is a minimum (Tu , Tu′ )-separation, this
implies γ(t) ∩ γ(u) 6∈ Tu . Hence γ(t) ∪ γ(u) ∈ Tu .
Suppose for contradiction that κ(γ(t) ∩ γ(u)) ≤ κ(γ(u)). Then either γ(t) ∩ γ(u) ∈ Tu or
γ(t) ∩ γ(u) ∈ Tu , and as
(γ(t) ∪ γ(u)) ∩ γ(u) ∩ γ(t) ∩ γ(u) = γ(t) ∩ γ(u) ∩ γ(t) ∩ γ(u) = ∅,
we have γ(t) ∩ γ(u) ∈ Tu . Now we continue exactly as in Case 1a: w e have γ(t) ∩ γ(u) =
γ(t)∪γ(u) ∈ Tu′ , which contradicts γ(u) being a leftmost minimum (Tu , Tu′ )-separation. Thus
κ(γ(t) ∩ γ(u)) > κ(γ(u)).
By submodularity, κ(γ(t) ∪ γ(u)) < κ(γ(t)) ≤ κ(γ◦ (t)). Then γ(t) ∪ γ(u) ∈ Tt . Furthermore,
γ(t) ∪ γ(u) ⊆ γ◦ (t) ∪ γ◦ (u) ⊆ γ◦ (t). Thus γ(t) ∪ γ(u) ⊇ γ◦ (t) ∈ Tt′ , and γ(t) ∪ γ(u) is a
y
(Tt , Tt′ )-separation. This contradicts γ◦ (t) being a minimum (Tt , Tt′ )-separation.
To obtain a treelike decomposition, which is supposed to satisfy γ(u) ⊆ γ(t) whenever u is a
child of t, we need to restructure the tree, moving children u of a node t with γ(u) 6⊆ γ(t), and
thus γ(u) ∩ γ(t) = ∅ by Claim 3, upwards in the tree and attaching them as children to the first
ancestor s of t such that γ(u) ⊆ γ(s). To avoid inconsistencies, we do this by induction.
We shall define a sequence of directed trees T (i) , for i ranging from 0 to some h to be determined,
such that V (T (i) ) = V and the following conditions are satisfied for all i:
(i) if i ≥ 1, then the descendant order E(i) of T (i) is a subset of the descendant order E(i−1) of
T (i−1) , that is, if t E(i) u then t E(i−1) u;
(ii) for all (t, u) ∈ E(T (i) ), either γ(u) ⊆ γ(t) or γ(u) ∩ γ(t) = ∅.
We have already defined T (0) , and it clearly satisfies conditions (i) and (ii): condition (i) is void
for i = 0 and condition (ii) follows from Claim 3.
So let us assume we have defined T (i) for some i ≥ 0. Let us call a node u ∈ V \ {r} bad in T (i)
if γ(u) 6⊆ γ(t) for the parent t of u. If there are no bad nodes in T (i) , we let h := i and stop the
construction. Otherwise, let U (i) be the set of all E(i) -maximal bad nodes. Note that the nodes
u ∈ U (i) are mutually incomparable with respect to the order E(i) .
To define T (i+1) , for every u ∈ U (i) , we do the following:
• we delete the edge from u to its parent;
• we let su be the the last (that is, E(i) -maximal) node on the path from the root r to u such
that γ(u) ⊆ γ(su ) (such a node exists because γ(u) ⊆ U = γ(r));
• we add an edge from su to u.
32
Then T (i+1) satisfies condition (i), because
E(i+1) =E(i) \ (v, w) ∃u ∈ U (i) : su ✁(i) v ✁(i) u and u E(i) w .
To see that T (i+1) satisfies (ii), let (t, u) ∈ E(T (i+1) ). If u ∈ U (i) , then we have t = su and thus
γ(u) ⊆ γ(t). Otherwise (t, u) ∈ E(T (i) ), and we have γ(u) ⊆ γ(t) or γ(u)∩γ(t) = ∅ by the induction
hypothesis.
The construction terminates after at most height of T (0) many steps, because in each step the
maximum height of a bad node decreases.
We let T := T (h) .
Claim 4.
(T, γ) is a directed tree decomposition of κ.
Proof. It is immediate from the construction that T is a directed tree, that γ(u) ⊆ γ(t) for all
(t, u) ∈ E(T ), and that γ(r) = U for the root r. It remains to prove that for all t ∈ V and distinct
children u1 , u2 in T we have γ(u1 ) ∩ γ(u2 ) = ∅. Indeed, it follows from (i) that t E(0) u1 , u2 and
thus t E◦ u1 , u2 . Let t′ ∈ V (T◦ ) be E◦ -maximal such that t E◦ t′ E◦ u1 , u2 , and for i = 1, 2, let u′i
be the child of t′ in T◦ such that t′ ✁◦ u′i E◦ ui . Then
γ(ui ) ⊆ γ◦ (ui ) ⊆ γ◦ (u′i ) = βe◦ (t′ , u′i )
and βe◦ (t′ , u′1 ) ∩ βe◦ (t′ , u′2 ) = ∅. Thus γ(u1 ) ∩ γ(u2 ) = ∅.
y
We have defined τ and V = V (T ) such that τ is a bijection between T≤ℓ
max and V . It remains
to prove that (T, γ, τ ) satisfies (DTD.1) and (DTD.2). Recall that τ −1 (t) = Tt and τ −1 (u) = Tu .
To prove (DTD.2), let t ∈ V \ {r}. By Claim 1, there is a u ∈ V such that t 6E(0) u and γ(t) is
a leftmost minimum (Tt , Tu )-separation. As E is a subset of E(0) , we also have t 6E u.
To prove (DTD.1), let t, u ∈ V (T ) such that u 6E t. If t ✁ u then t ✁(0) u, and thus by Claim 2,
there is a Y ⊇ γ(u) that is a minimum (Tu , Tt )-separation. In the following, we assume that t 6E u.
Then γ(t) ∩ γ(u) = ∅, because (T, γ) is a directed tree decomposition. Moreover, γ(t) ∈ Tt and
γ(u) ∈ Tu by (DTD.2).
Let Y be a minimum (Tu , Tt )-separation. Suppose for contradiction that κ(Y ∪ γ(u)) < κ(Y ).
Then Y ∪ γ(u) ∈ Tu . Moreover, either Y ∪ γ(u) ∈ Tt or Y ∪ γ(u) ∈ Tt . As γ(u) ∩ γ(t) = ∅, we have
Y ∩ γ(t) ∩ (Y ∪ γ(u)) = ∅
and thus Y ∪ γ(u) ∈ Tt . Hence Y ∪ γ(u) is a (Tu , Tt )-separation of order less than κ(Y ). This
contradicts the minimality of Y .
Hence κ(Y ∪ γ(u)) ≥ κ(Y ) and therefore, by submodularity, κ(Y ∩ γ(u)) ≤ κ(γ(u)). Thus
Y ∩ γ(u) ∈ Tu , and Y ∩ γ(u) is a (Tu , γ◦ (u))-separation. By the minimality of γ(u), it follows that
γ(u) ⊆ Y ∩ γ(u) and thus γ(u) ⊆ Y .
6
Conclusions
Our main contribution is to make tangles, an important tool in structural graph theory, algorithmic.
The running time of all of our algorithms is nO(k) , where k is the order of the tangles and n
the size of the ground set. This running time is mainly caused by searching through all potential
bases for some separation. It is a very interesting open question if we can do the same with fixedparameter tractable algorithms. This would even be interesting for specific connectivity functions
such as the connectivity function κG and the cut-rank function ρG for graphs G or the connectivity
function of representable matroids. Possibly, this can be achieved by arguments building on Hlinený
and Oum’s fpt algorithm for computing branch width [11].
33
Even though the canonical decomposition theorem has useful applications as it is, specifically
in graph isomorphism testing, it would be even more useful if one actually understood it. However,
except for a few simple cases of low order, the parts β(t) of the decompositions remain mysterious,
for tangle nodes t and even more so for hub nodes t. For the connectivity function κG of graphs,
Carmesin et al. [2] obtained first results clarifying the structure of the decompositions.
References
[1] J. Carmesin, R. Diestel, M. Hamann, and F. Hundertmark. Canonical tree-decompositions of
finite graphs I. Existence and algorithms. ArXiv, arXiv:1305.4668v3 [math.CO], 2013.
[2] J. Carmesin, R. Diestel, M. Hamann, and F. Hundertmark. Canonical tree-decompositions of
finite graphs II. Essential parts. ArXiv, arXiv:1305.4909v2 [math.CO], 2013.
[3] J. Carmesin, R. Diestel, F. Hundertmark, and M. Stein. Connectivity and tree structure in
finite graphs. Combinatorica, 34(1):11–46, 2014.
[4] B. Courcelle and J. Engelfriet. Graph Structure and Monadic Second-Order Logic — A
Language-Theoretic Approach. Cambridge University Press, 2012.
[5] B. Courcelle and S.-I. Oum. Vertex-minors, monadic second-order logic, and a conjecture by
Seese. Journal of Combinatorial Theory, Series B, 97(1):91–126, 2007.
[6] E.D. Demaine, M.T. Hajiaghayi, and K. Kawarabayashi. Algorithmic graph minor theory: Decomposition, approximation, and coloring. In Proceedings of the 45th Annual IEEE Symposium
on Foundations of Computer Science, pages 637–646, 2005.
[7] J. Geelen, B. Gerards, and G. Whittle. Tangles, tree-decompositions and grids in matroids.
Journal of Combinatorial Theory, Series B, 99(4):657–667, 2009.
[8] M. Grohe, K. Kawarabayashi, and B. Reed. A simple algorithm for the graph minor decomposition – logic meets structural graph theory–. In Proceedings of the 24th Annual ACM-SIAM
Symposium on Discrete Algorithms, pages 414–431, 2013.
[9] M. Grohe and D. Marx. Structure theorem and isomorphism test for graphs with excluded
topological subgraphs. In Proceedings of the 44th ACM Symposium on Theory of Computing,
2012.
[10] M. Grohe and P. Schweitzer. Isomorphism testing for graphs of bounded rank width. ArXiv,
arXiv:1505.03737 [cs.DM], 2015. Conference version to appear in Proceedings FOCS 2015.
[11] P. Hlinený and S.-I. Oum. Finding branch-decompositions and rank-decompositions. SIAM
Journal on Computing, 38(3):1012–1032, 2008.
[12] F. Hundertmark. Profiles. An algebraic approach to combinatorial connectivity.
arXiv:1110.6207v1 [math.CO], 2011.
ArXiv,
[13] S. Iwata, L. Fleischer, and S. Fujishige. A combinatorial strongly polynomial algorithm for
minimizing submodular functions. Journal of the ACM, 48(4):761–777, 2001.
[14] K.-I. Kawarabayashi and P. Wollan. A simpler algorithm and shorter proof for the graph minor
decomposition. In Proceedings of the 43rd ACM Symposium on Theory of Computing, pages
451–458, 2011.
34
[15] S.-I. Oum. Rank-width and vertex-minors. Journal of Combinatorial Theory, Series B, 95:79–
100, 2005.
[16] S.-I. Oum and P. Seymour. Testing branch-width. Journal of Combinatorial Theory, Series
B, 97:385–393, 2007.
[17] S.-I. Oum and P.D. Seymour. Approximating clique-width and branch-width. Journal of
Combinatorial Theory, Series B, 96:514–528, 2006.
[18] S.-I. Oum and P.D. Seymour. Certifying large branch-width. In Proceedings of the 17th Annual
ACM-SIAM Symposium on Discrete Algorithms, pages 810–813, 2006.
[19] J. Oxley. Matroid Theory. Cambridge University Press, 2nd edition, 2011.
[20] B. Reed. Tree width and tangles: A new connectivity measure and some applications. In
R.A. Bailey, editor, Surveys in Combinatorics, volume 241 of LMS Lecture Note Series, pages
87–162. Cambridge University Press, 1997.
[21] N. Robertson and P.D. Seymour. Graph minors X. Obstructions to tree-decomposition. Journal
of Combinatorial Theory, Series B, 52:153–190, 1991.
[22] N. Robertson and P.D. Seymour. Graph minors XVI. Excluding a non-planar graph. Journal
of Combinatorial Theory, Series B, 77:1–27, 1999.
[23] A. Schrijver. A combinatorial algorithm minimizing submodular functions in strongly polynomial time. Journal of Combinatorial Theory, Series B, 80(2):346–355, 2000.
[24] D. Seese. The structure of models of decidable monadic theories of graphs. Annals of Pure
and Applied Logic, 53:169–195, 1991.
35
| 8cs.DS
|
arXiv:1801.07760v1 [math.AC] 23 Jan 2018
COMINIMAXNESS WITH RESPECT TO IDEALS OF
DIMENSION ONE
HAJAR ROSHAN-SHEKALGOURABI
Abstract. Let R be a commutative Noetherian ring, a be an ideal of R and
M be an R-module. It is shown that if ExtiR (R/a, M ) is minimax for all
i ≤ dim M , then the R-module ExtiR (N, M ) is minimax for all i ≥ 0 and for
any finitely generated R-module N with SuppR (N ) ⊆ V (a) and dim N ≤ 1.
As a consequence of this result we obtain that for any a-torsion R-module M
that ExtiR (R/a, M ) is minimax for all i ≤ dim M , all Bass numbers and all
Betti numbers of M are finite. This generalizes [8, Corollary 2.7]. Also, some
equivalent conditions for the cominimaxness of local cohomology modules with
respect to ideals of dimension at most one are given.
1. Introduction
Let R denote a commutative Noetherian ring with identity and a be an ideal of
R. For an R-module M , the ith local cohomology module of M with respect to a
is defined as
Hai (M ) ∼
ExtiR (R/an , M ).
= lim
−→
n∈N
For more details about the local cohomology, we refer the reader to [9].
In 1968, Grothendieck [12] conjectured that
for any ideal a of R and any finitely
generated R-module M , HomR R/a, Hai (M ) is a finitely generated R-module for
all i. One year later, by proving a counterexample, Hartshorne [13] showed that the
Grothendieck’s conjecture is not true in general even R is regular and introduced
the class of cofinite modules with respect to an ideal. He defined an R-module M
to be a-cofinite if SuppR (M ) ⊆ V (a) and ExtjR (R/a, M ) is finitely generated for
all j and posed the following question:
• For which rings R and ideals a is the module Hai (M ) a-cofinite for all i and
all finitely generated R-modules M ?
There are many papers that are devoted to study this question. For example,
see [13, 10, 11, 20, 6].
In [21], Zöschinger introduced the interesting class of minimax modules, modules
containing some finitely generated submodule such that the quotient module is
Artinian. As a generalization of the concept of a-cofinite modules, the concept
of a-cominimax modules was introduced in [4]. An R-module M is said to be acominimax if SuppR (M ) ⊆ V (a) and ExtjR (R/a, M ) is minimax for all j. Since the
concept of minimax modules is a natural generalization of the concept of finitely
generated modules, many authors studied the minimaxness of local cohomology
Date: January 25, 2018.
2010 Mathematics Subject Classification. 13D45, 13E05, 18E10.
Key words and phrases. Minimax modules, Cominimax modules, Krull dimension, Local cohomology modules.
1
2
ROSHAN
modules and answered the Hartshorn’s question in the class of minimax modules
(see [2, 3, 4, 15]).
As an important problem in commutative algebra is determining when the Bass
numbers and Betti numbers of a module is finite. In this direction, recently, Bahmanpour et al. in [8] proved that for any a-torsion R-module M that ExtiR (R/a, M )
is finitely generated for all i ≤ dim M , all Bass numbers µj (p, M ) and all Betti numbers βj (p, M ) of M are finite. In this paper, as a generalization of this result, we will
prove that the assertion in this result holds when we replace ”finitely generated”
by ”minimax”. More precisely, we shall show that:
Corollary 1.1. Let M be an R-module of dimension n such that SuppR (M ) ⊆ V (a)
and ExtiR (R/a, M ) is minimax for all i ≤ n. Then all Bass numbers µi (p, M ) and
all Betti numbers βi (p, M ) of M are finite.
Our main tools for proving this result is the following theorem which states
some conditions for the cominimaxness of local cohomology modules with respect
to ideals of dimension one.
Theorem 1.2. Let M be an R-module such that ExtiR (R/a, M ) is minimax for all
i ≤ dim M . Then the following assertions hold:
(1) The R-module Hbi (M ) is b-cominimax for all i ≥ 0 and for any ideal b ⊆ a
with dim R/b ≤ 1.
(2) The R-module ExtiR (N, M ) is minimax for all i ≥ 0 and for any finitely
generated R-module N with SuppR (N ) ⊆ V (a) and dim N ≤ 1.
The proof of Theorem 1.2 is given in Proposition 2.6 and Theorem 2.7.
Throughout the paper, we assume that R is a commutative Noetherian ring, a
is an ideal of R and V (a) is the set of all prime ideals of R containing a. For any
unexplained notation and terminology we refer the reader to [16].
2. MAIN RESULTS
Recall that a class of R-modules is a Serre subcategory of the category of Rmodules when it is closed under taking submodules, quotients and extensions. For
example, the classes of Noetherian modules, Artinian modules or minimax modules are Serre subcategories. As in standard notation, we let S stand for a Serre
subcategory of the category of R-modules. The following lemma which is needed
in the sequel, immediately follows from the definition of Ext and Tor functors.
Lemma 2.1. Let M be a finitely generated R-module and N ∈ S. Then ExtiR (M, N ) ∈
S and TorR
i (M, N ) ∈ S for all i ≥ 0.
Lemma 2.2. Suppose that M is a finitely generated R-module and N is an arbitrary
R-module. Let for some t ≥ 0, ExtiR (M, N ) ∈ S for all i ≤ t. Then ExtiR (L, N ) ∈ S
for all i ≤ t and any finitely generated R-module L with SuppR (L) ⊆ SuppR (M ).
Proof. See [2, Lemma 2.2].
Let us mention some elementary properties of the minimax modules that we
shall use.
Remark 2.3. The following statements hold:
(1) The class of minimax modules contains all finitely generated and all Artinian modules.
COMINIMAXNESS WITH RESPECT TO IDEALS OF DIMENSION ONE
3
(2) Let 0 → L → M → N → 0 be an exact sequence of R-modules. Then M
is minimax if and only if L and N are both minimax (see [5, Lemma 2.1]).
Thus any submodule and quotient of a minimax module is minimax.
(3) The set of associated primes of any minimax R-module is finite.
(4) Every zero-dimensional minimax R-module is Artinian.
(5) If R is a vector space, then every minimax R-module is finite length.
Lemma 2.4. Let a be an ideal of R, M be an R-module and n be a non-negative
i
integer such that ExtnR (R/a, M ) (resp. Extn+1
R (R/a, M )) is minimax. If Ha (M ) is
1
n
a-cominimax for all i < n, then HomR (R/a, Ha (M )) (resp. ExtR (R/a, Han (M )))
is minimax.
Proof. Since the class of minimax modules is a Serre subcategory, the assertion
follows from [1, Lemma 2.3].
The following lemma is well-known for a-cominimax modules.
Lemma 2.5. Let M be an a-torsion R-module such that dim M ≤ 1. Then M is
a-cominimax if and only if the R-modules HomR (R/a, M ) and Ext1R (R/a, M ) are
minimax.
Proof. See [14, Proposition 2.4].
The following proposition which states some conditions for the cominimaxness
of local cohomology modules with respect to ideals of dimension at most one, plays
a key role for proving the next theorem and the main result of this paper.
Proposition 2.6. Let M be an R-module of dimension n such that ExtjR (R/a, M )
is minimax for all j ≤ n. Then the R-module Hbi (M ) is b-cominimax for all i ≥ 0
and for any ideal a ⊆ b with dim R/b ≤ 1.
Proof. By Grothendieck’s Vanishing Theorem we only need to prove the assertion
for 0 ≤ i ≤ n. Let b be an arbitrary ideal of R containing a with dim R/b ≤ 1.
Then by assumption and Lemma 2.2, ExtjR (R/b, M ) is a minimax R-module for all
j ≤ n. We first prove the assertion for the case n = 0. Then by assumption, the
R-module
HomR (R/b, Γb(M )) = HomR (R/b, M )
is minimax and so is Artinian by Remark 2.3(4). Hence, Γb (M ) is Artinian by
virtue of Melkersson’s result [17, Theorem 1.3]. Now, the assertion follows from the
Grothendieck’s Vanishing Theorem and the fact that SuppR (Γb (M )) ⊆ V (b) and
the class of minimax modules contains all Artinian modules. Thus, it remains to
give the proof for the case n > 0. For this purpose, there are two cases to consider:
dim R/b = 0 or dim R/b = 1.
Case 1: If dim R/b = 0, then in the light of assumption and Remark 2.3(4),
HomR (R/b, M ) is Artinian. Hence by the argument in the case of n = 0, we may
conclude that Γb (M ) is b-cominimax. Now suppose, inductively, that 0 < i ≤ n
and the R-modules
Hb0 (M ), Hb1 (M ), · · · , Hbi−1 (M )
are b-cominimax. Since SuppR (Hbi (M )) ⊆ V (b) and the R-module ExtjR (R/b, M )
is minimax for all j ≤ n, we infer from Lemma 2.4 that HomR (R/b, Hbi (M )) is a
zero-dimensional minimax R-module and so is Artinian. Hence, Hbi (M ) is Artinian
by [17, Theorem 1.3], as desired.
4
ROSHAN
Case 2: Let dim R/b = 1. The proof is by induction on 0 ≤ i < n. Since
HomR (R/b, M/Γb(M )) = 0, it follows from the assumption and the exact sequence
0 → HomR (R/b, Γb(M )) → HomR (R/b, M ) → HomR (R/b, M/Γb(M ))
→ Ext1R (R/b, Γb(M )) → Ext1R (R/b, M )
that the R-modules HomR (R/b, Γb (M )) and Ext1R (R/b, Γb (M )) are minimax. Hence,
as dim Γb (M ) ≤ 1, the R-module Γb (M ) is b-cominimax by Lemma 2.5. Now suppose that the assertion holds for i − 1; we will prove it for i. By the inductive
hypotheses, the R-modules
Hb0 (M ), Hb1 (M ), · · · , Hbi−1 (M )
are b-cominimax. Since the R-modules ExtiR (R/b, M ) and Exti+1
R (R/b, M ) are
minimax, it follows from Lemma 2.4 that the R-modules
HomR (R/b, Hbi (M ) and Ext1R (R/b, Hbi (M ))
are minimax and so in view of Lemma 2.5 the R-module Hbi (M ) is b-cominimax,
for all i = 0, 1, · · · , n − 1. Since ExtnR (R/b, M ) is minimax, HomR (R/b, Hbn (M ))
is also minimax by Lemma 2.4. If there exists p ∈ SuppR (Hbn (M )) ⊆ V (b) with
dim R/p = 1, then it is easy to see that dim Mp ≤ n − 1 and so (Hbn (M ))p = 0 by
Grothendieck’s Vanishing Theorem, a contradiction. Therefore,
SuppR (Hbn (M )) ⊆ Max(R).
This implies that the R-module HomR (R/b, Hbn (M )) is Artinian by Remark 2.3(4).
Hence, Hbn (M ) is Artinian by [17, Theorem 1.3] and so is b-cominimax, as required.
Theorem 2.7. Let M be an R-module of dimension n such that ExtiR (R/a, M ) is
minimax for all i ≤ n. Then the R-module ExtiR (N, M ) is minimax for all i ≥ 0
and for any finitely generated R-module N with SuppR (N ) ⊆ V (a) and dim N ≤ 1.
Proof. Let N be a finitely generated R-module such that SuppR (N ) ⊆ V (a) and
dim N ≤ 1. Then, using [16, Theorem 6.4], there exist prime ideals p1 , · · · , pt of
R and a chain 0 = N0 ⊆ N1 ⊆ · · · ⊆ Nt = N of submodules of N such that
Nj /Nj−1 ∼
= R/pj for all j = 1, · · · , t. Since pj ∈ SuppR (N ), we deduce that
dim R/pj ≤ 1 and so in the light of Proposition 2.6, the R-module Hpi j (M ) is pj cominimax for all i ≥ 0 and for each j = 1, · · · , t. Thus, by [19, Corollary 3.10],
the R-module ExtiR (R/pj , M ) is minimax for all i ≥ 0 and for each j = 1, · · · , t.
Now, considering the exact sequences
0 → N1 →N2 → R/p2 → 0
0 → N2 →N3 → R/p3 → 0
..
.
0 → Nt−1 →Nt → R/pt → 0
we infer that ExtiR (N, M ) is minimax, as desired.
Now we are ready to state the main result of this paper. Recall that for each Rmodule M , all integers j ≥ 0 and all prime ideals p of R, the jth Bass number of M
with respect to p is defined as µj (p, M ) = dimk(p) ExtjRp (k(p), Mp ) and the jth Betti
COMINIMAXNESS WITH RESPECT TO IDEALS OF DIMENSION ONE
5
R
number of M with respect to p is defined as βj (p, M ) = dimk(p) Torj p (k(p), Mp ),
where k(p) := Rp /pRp . Recently, Bahmanpour et al. in [8, Corollary 2.7] proved
that for any a-torsion R-module M that ExtiR (R/a, M ) is finitely generated for all
i ≤ dim M , all Bass numbers µj (p, M ) and all Betti numbers βj (p, M ) of M are
finite. As an immediate consequence of Theorem 2.7 we obtain the next corollary
which is a generalization of [18, Theorem 1.9] and [8, Corollary 2.7] and shows that
the assertion in [8, Corollary 2.7] holds when we replace ”finitely generated” by
”minimax”.
Corollary 2.8. Let M be an R-module of dimension n such that SuppR (M ) ⊆ V (a)
and ExtiR (R/a, M ) is minimax for all i ≤ n. Then all Bass numbers µi (p, M ) and
all Betti numbers βi (p, M ) of M are finite.
Proof. Let p ∈ Spec(R) and k(p) = Rp /pRp be the residue field of Rp . If p ∈
/ V (a),
then Mp = 0 by assumption and there is nothing to prove. Otherwise, using Theorem 2.7 and letting N := Rp /pRp , we conclude that the Rp -module ExtiRp (k(p), Mp )
is minimax for all i ≥ 0. Since ExtiRp (k(p), Mp ) is also a k(p)-vector space, it must
be of finite length by Remark 2.3(5). Now the proof is completed by [19, Theorem
2.1].
Consequently, we get the following equivalent conditions for the cominimaxness
of local cohomology modules with respect to ideals of dimension at most one.
Corollary 2.9. Let M an R-module of dimension n and a be an ideal of R such
that dim R/a ≤ 1. Then the following conditions are equivalent:
(1) ExtiR (R/a, M ) is minimax for all i ≤ dim M ;
(2) Hai (M ) is a-cominimax for all i;
(3) ExtiR (R/a, M ) is minimax for all i;
(4) ExtiR (X, M ) is minimax for all i ≤ dim M and for any finitely generated
R-module X with SuppR (X) ⊆ V (a);
(5) ExtiR (X, M ) is minimax for all i ≤ dim M and for any finitely generated
R-module X with SuppR (X) = V (a);
(6) ExtiR (X, M ) is minimax for all i and for any finitely generated R-module
X with SuppR (X) ⊆ V (a);
(7) ExtiR (X, M ) is minimax for all i and for any finitely generated R-module
X with SuppR (X) = V (a).
Proof. The assertions follow from Proposition 2.6, Theorem 2.7 and Lemma 2.2.
Corollary 2.10. If (R, m) is a local ring and a be an ideal of R such that dim R/a =
1, then the following conditions are equivalent:
(1) Hai (M ) is a-cominimax for all i;
(2) µi (p, M ) is finite for all p ∈ V (a) and for all i ≤ dim M ;
(3) µi (p, M ) is finite for all p ∈ V (a) and for all i.
Proof. It yields from [7, Theorem 2.3] and Corollary 2.9.
References
[1] N. Abazari and K. Bahmanpour, Extension functors of local cohomology modules and serre
categories of modules, Taiwanese J. Math. 19 (2015), no. 1, 211–220.
[2] A. Abbasi and H. Roshan Shekalgourabi, Serre subcategory properties of generalized local
cohomology modules, Korean Ann. Math. 28 (2011), no. 1, 25–37.
6
ROSHAN
[3] A. Abbasi, H. Roshan Shekalgourabi, and D. Hassanzadeh-Lelekaami, Some results on the
local cohomology of minimax modules, Czechoslovak Math. J. 64 (2014), no. 139, 327–333.
[4] J. Azami, R. Naghipour, and B. Vakili, Finiteness properties of local cohomology modules for
a-minimax modules, Proc. Amer. Math. Soc. 137 (2009), no. 2, 439–448.
[5] K. Bahmanpour and R. Naghipour, On the cofiniteness of local cohomology modules, Proc.
Amer. Math. Soc. 136 (2008), no. 7, 2359–2363.
, Cofiniteness of local cohomology modules for ideals of small dimension, J. Algebra
[6]
321 (2009), 1997–2011.
[7] K. Bahmanpour, R. Naghipour, and M. Sedghi, On the finiteness of bass numbers of local
cohomology modules and cominimaxness, Houston J. Math. 40 (2014), no. 2, 319–337.
[8]
, Cofiniteness with respect to ideals of small dimensions, Algebr. Represent. Theory
18 (2015), no. 2, 369–379.
[9] M. P. Brodmann and R. Y. Sharp, Local cohomology: An algebraic introduction with geometric applications, Cambridge Studies in Advanced Mathematics 60, Cambridge University
Press, Cambridge, 1998.
[10] G. Chiriacescu, Cofiniteness of local cohomology modules, Bull. London Math. Soc. 32 (2000),
1–7.
[11] D. Delfino and T. Marley, Cofinite modules and local cohomology, J. Pure Appl. Algebra 121
(1997), no. 1, 45–52.
[12] A. Grothendieck, Cohomologie locale des faisceaux et theoremes de lefshetz locaux et globaux
(SGA2), 1968.
[13] R. Hartshorne, Affine duality and cofiniteness, Invent. Math. 9 (1969/1970), 145–164.
[14] Y. Irani, Cominimaxness with respect to ideals of dimension one, Bull. Korean Math. Soc.
54 (2017), no. 1, 289–298.
[15] A. Mafi, On the local cohomology of minimax modules, Bull. Korean Math. Soc. 48 (2011),
no. 6, 1125–1128.
[16] H. Matsumura, Commutative ring theory, Cambridge University Press, Cambridge, UK, 1986.
[17] L. Melkersson, On asymptotic stability for sets of prime ideals connected with the powers of
an ideal, Math. Proc. Camb. Phil. Soc. 107 (1990), 267–271.
, Properties of cofinite modules and application to local cohomology, Math. Proc.
[18]
Cambridge Philos. Soc. 125 (1999), 417–423.
[19]
, Modules cofinite with respect to an ideal, J. Algebra 285 (2005), 649–668.
[20] K. I. Yoshida, Cofiniteness of local cohomology modules for ideals of dimension one, Nagoya
Math. J. 147 (1997), 179–191.
[21] H. Zöschinger, Minimax-moduln, J. Algebra 102 (1986), 1–32.
Department of Basic Sciences, Arak University of Technology, P. O. Box 38135-1177,
Arak, Iran.
E-mail address: [email protected] and [email protected]
| 0math.AC
|
An efficient high-probability algorithm
for Linear Bandits
arXiv:1610.02072v2 [cs.DS] 13 Oct 2016
Gábor Braun1 and Sebastian Pokutta2
1 ISyE, Georgia
2 ISyE,
Institute of Technology, Atlanta, GA, USA. Email: [email protected]
Georgia Institute of Technology, Atlanta, GA, USA. Email: [email protected]
October 13, 2016
Abstract
For the linear bandit problem, we extend the analysis of algorithm CombEXP from Combes et al. [2015] to the
high-probability case against adaptive adversaries, allowing actions to come from an arbitrary polytope. √
We prove a
high-probability regret of O( T 2/3 ) for time horizon T. While this bound is weaker than the optimal O( T ) bound
achieved by GeometricHedge in Bartlett et al. [2008], CombEXP is computationally efficient, requiring only an efficient linear optimization oracle over the convex hull of the actions.
1 Introduction
We study sequential prediction problems with linear losses and bandit feedback against an adaptive adversary. At every
round t the forecaster chooses an action x t , and the adversary chooses a loss function Lt , and the forecaster suffers the
loss Lt ( x t ). The forecaster learns only the suffered loss after each round, while the adversary learns the forecaster’s
action xt . The forecaster’s aim is to minimize regret, which is the difference between the incurred loss and the loss of
the best single action in hindsight:
∑ Lt (xt ) − min ∑ Lt (x ).
t∈ T
x ∈ A t∈ T
In this work we focus on establishing regret bounds holding with high-probability with an efficient algorithm.
For algorithms with bandit feedback, exploration (occasionally playing random actions for learning) is a crucial
feature, however it does not have to be explicit as recently shown in Neu [2015], where exploration is achieved via
skewing loss estimators. One of the most studied regret minimization algorithm is EXP, which iteratively updates
the probabilities of each action via multiplication with factors exponential in its (estimated) loss. The variant EXP3
for multi-armed bandit problems first appeared in Auer et al. [2002], however optimal high-probability regret bounds
were first achieved in Dani and Hayes [2006]. The linear bandit setting is a generalization of the multi-armed bandit setting where, utilizing the linearity of losses, the goal is to improve the dependence on the number of actions
in the regret bound, which might be exponential in the dimension n. At the same time linear losses come naturally
into play when considering actions with a combinatorial structure, such as e.g., matchings, spanning trees, m-sets; see
Cesa-Bianchia and Lugosi [2012], Audibert et al. [2013] for an extensive discussion. For the linear bandit
√ setting, the
EXP-variant ComBand (Combinatorial Bandit) from Cesa-Bianchia and Lugosi [2012] has
optimal
O
(
T ) expected
√
regret, and in Bartlett et al. [2008] the modified version GeometricHedge achieves O( T ) regret with high probability. While these regret bounds practically do not depend on the number of actions, both maintain a distribution over
the (possibly exponentially large) action set A, which is infeasible in general due to the large data size, even though
ComBand is still efficient for many specific problems. Recently, a modification of the ComBand algorithm called
CombEXP (see Algorithm 1) was derived in Combes et al. [2015], which achieves general computational efficiency by
not maintaining a distribution of xt , but only the desired expectation x̂t of the distribution, and generating a new sparse
approximate distribution at every round.
In this work we provide a high-probability regret bound of O( T 2/3 ) for CombEXP against adaptive adversaries,
while generalizing it to general polytopes. The obtained bounds are any-time, i.e., the parameter choice is independent
1
of the time horizon T. Finally, our algorithm maintains computational efficiency given an efficient linear programming
oracle over the underlying polytope (the convex hull of actions). For comparison, we also show an O( T 2/3 ) regret in
the high-probability setting for the original ComBand.
The maximal matching problem is a good example where the linear programming oracle approach is useful, as it
has a polynomial time linear optimization algorithm Edmonds [1965], but no polynomial-size polyhedral description
Rothvoß [2014].
Related work
Our work is most closely related to the line of works on combinatorial bandit problems. The algorithm ComBand first
appeared in Cesa-Bianchia and Lugosi [2012], while GeometricHedge comes from Bartlett et al. [2008],
√ and CombEXP appeared in Combes et al. [2015]. Using interior point methods, an efficient algorithm with O( T ) expected
regret for linear bandit problems has been established in Abernethy et al. [2008].
For multiarmed bandit problems, the original version of EXP3 has high-probability regret √
Ω( T 2/3 ) against some
adaptive adversaries [Dani and Hayes, 2006, Theorem 1.2], however variants with optimal O( T ) regret exists, e.g.,
using accountants to control the exploration rate (see Dani and Hayes [2006]), or via the recent EXP3-IX with implicit
exploration (see Neu [2015]).
For convex loss functions, optimal high-probability regret bounds have been obtained in Hazan and Li [2016] with
running time being poly-exponential in the dimension, and in [Bubeck et al., 2016, Theorem 1] with polynomial running time provided the number of constraints of the underlying polytope is polynomial in the dimension. Optimal
regret bounds in expectation was first obtained in Bubeck and Eldan [2015]. However the case of convex loss does not
subsume the combinatorial/linear case, as with convex loss all inner points of the convex set are actions; with linear
losses the actions are limited to the vertices of the underlying polytope in most cases.
We refer the interested reader to the excellent survey of Bubeck and Cesa-Bianchi [2012] on bandit problems.
Contribution
Our main contribution is a high-probability regret bound for CombEXP from Combes et al. [2015] for adaptive adversaries over actions coming from arbitrary polytopes P ⊆ R n . Our algorithm, being a slight generalization of
CombExp, maintains computational efficiency. In particular, our contribution can be summarized as follows:
(i) High-probability bounds for an efficient algorithm. For CombEXP we establish a high-probability regret of
2
2n + 2
B + nB
T 2/3 ,
· ln
O
min{λ, 1}
δ
with probability 1 − δ, where B is the ℓ2 -diameter of P, and λ is a lower bound on the smallest eigenvalue of the
exploration covariance matrix, see Theorem 3.1 for the exact regret bound.
For comparison we show that the same method already provides a high-probability√regret bound of O( T 2/3 ) for
the original ComBand, albeit a suboptimal one as GeometricHedge achieves O( T ) regret.
(ii) Generalization of CombEXP and computational efficiency. We generalize CombEXP to actions arising from arbitrary polytopes contained in R n and to the case of adaptive adversaries. We maintain computational efficiency
of CombExp providing running times relative to a linear programming oracle over the underlying polytope P,
separating the complexity for learning from the complexity of linear optimization over P.
All our bounds are any-time, i.e., holding uniformly for all times T. In particular, our parameter choices are independent of T.
Outline
After a brief summary of the regret minimization framework in Section 2, we reanalyze CombEXP in Section 3. For
completeness we present a similar analysis for ComBand in Section 4.
We relegated various related materials to the the Appendix. In Section A we provide an any-time version of EXP
with time-varying parameters maintaining generalized distributions, defined by an arbitrary convex set in the positive
2
√
orthant, instead of the probability simplex. We prove an O( T ) regret bound in the full information case by standard
arguments, which forms the basis for our regret bounds for the bandit case. In Section B we recall concentration
inequalities that we use to establish high-probability bounds. Finally, in Sections C and D we provide (already known)
efficient algorithms for projection and distribution generation, which are key components in our algorithms. We include
those for completeness of exposition and to make parameters explicit.
2 Preliminaries
We will briefly recall the regret minimization framework to define our notation. In the sequential prediction problem
with linear losses, at every round t the forecaster chooses an action xt from a finite set A ⊆ R n and the adversary
⊺
chooses a loss vector Lt ∈ R n . The forecaster suffers the loss ℓt := Lt xt . The goal of the forecaster is to minimize
the regret
T
T
t =1
t =1
∑ L⊺t x.
∑ L⊺t xt − min
x∈ A
Against an oblivious adversary, who chooses the Lt independently of the forecaster’s actions, this is the extra loss
suffered by not playing the best single action in hindsight. However, this interpretation is clearly incorrect against an
adaptive adversary (the notion of policy regret from Arora et al. [2012] matches this interpretation). Nevertheless the
above notion of regret proved to be useful in many areas.
With bandit feedback the forecaster learns only the loss ℓt but not the actual loss vector Lt . An adaptive adversary
learns the forecaster’s action x t after round t, and can use it in later rounds to choose his actions.
We make various standard assumptions to bound the regret. The most important one is that the per round loss is
bounded, i.e., | L⊺t x | ≤ 1 for all x ∈ A. Under reasonably assumptions, this also implies that the set A of possible
actions A is bounded and we assume that k x k2 ≤ B and k x k1 ≤ B1 , with suitable positive numbers B, B1 . Clearly,
one can always choose B1 = nB, however we obtain finer bounds by keeping them separate. The bounds B1 and B
also serve as a proxy for the sparsity of the actions.
Following Cesa-Bianchia and Lugosi [2012] for ComBand, we shall use a fixed arbitrary distribution µ on A for
exploration, whose fitness for exploration is measured by a positive lower bound λ on the smallest eigenvalue of its
covariance matrix J:
J := E y∼µ [yy⊺ ] λI.
Here and below we denote by M N that N − M is a positive semi-definite matrix for symmetric matrices M
and N. When A is small then µ is typically the uniform distribution over A. For large A, common choices are the
uniform distribution on a barycentric spanner of A (see Hazan et al. [2014]), or the distribution on contact points of the
maximal volume ellipsoid contained in the convex hull P of A arising from John’s decomposition (John’s exploration;
see Dani et al. [2007]), transferred to A. In the latter two cases, J = I and λ = 1/n using the scalar product on
R n induced by the additional structure. John’s ellipsoid can be approximately estimated with a worse lower bound
λ = 1/n3/2 by Grötschel et al. [1993], however a constant factor approximation is NP-hard by Nemirovski [2006].
Recall that a barycentric spanner is a linear basis v1 , . . . , v n in P (the convex hull of A), such that every element of P
is a linear combination of the vi with coefficients from [−1, +1]. The basis v1 , . . . , v n is a C-approximate barycentric
spanner for some C > 1 if every element of P is a linear combination of the vi with coefficients from [−C, +C ]. A Capproximate barycentric spanners can be efficiently computed by O(n2 ln n/ ln C ) calls to a linear optimization oracle
over P by Awerbuch and Kleinberg [2004], which actually computes a spanner consisting of vertices of P. In this paper
we deliberately avoid using the scalar product induced by the structure to be able to directly use the bounds available
in the original space of the problem. Fortunately, the uniform distribution on an approximate barycentric spanner has
a close to optimal minimal eigenvalue even in the original space, see Lemma E.1, which allows us to preserve sparsity
of the original space. As such we assume that we have access to an exploration distribution over actions with sparse
support of size n, where n is the dimension of the vector space, from which we can efficiently sample. Note that
for specific problems exploration distributions with better minimal eigenvalue can be explicitly given; we refer the
interested reader to Cesa-Bianchia and Lugosi [2012] and follow-up work for a large set of such examples.
Let u := E y∼µ [y] denote the expectation of µ and let e denote the Euler constant. Instead of dealing directly
with A, it will be more convenient to use the convex hull P of A, then A contains the vertex set of P (and in many
applications the two are equal). We shall use the Kullback–Leibler divergence as Bregman divergence of the function
3
f ( x1 , . . . , x n ) = ∑ni=1 xi ln xi for projection:
n
KL( x, y) =
x
i =1
Rn
n
n
i =1
i =1
∑ xi ln yii − ∑ xi + ∑ yi .
Rn
we will use R n> a
:= { x ∈
In the following, for a vector a ∈
| xi > ai for all i ∈ [n]} to denote the a-positive
orthant. Moreover, a linear optimization oracle (or LP oracle) over a polytope P ⊆ R n finds for any linear objective
c ∈ R n a vertex x of P minimizing c⊺ x.
In all our bounds below, the O-notation only hides an absolute constant, i.e., all parameters of the algorithms are
explicit. However, in Section 1 the O-notation hides also other parameters, like the dimension n.
3 A high-probability regret bound for CombEXP
We provide an adaptation of CombEXP (Algorithm 1) with an O( T 2/3 ) regret with high probability against adaptive
adversaries, while maintaining computational efficiency. In a nutshell, EXP is run on the coordinates of the desired
expectation x̂t of xt , and a new distribution over vertices xt of P is generated in every round. In order to obtain
an efficient algorithm, we allow errors in the most resource-consuming components of the algorithm: the projection
step and the distribution generation. The accuracy of distribution generation is controlled by a parameter ε, and helps
maintaining a distribution with sparse support, to allow fast sampling and fast computation of the covariance matrix
Ct . The positive parameters ηt , γt control the learning rate and exploration rate of the algorithm. The role of the
shifting vector a ∈ R n is to avoid singularity issues with Kullback–Leibler divergence. Except for the shifting vector
a, these ideas already appeared in Combes et al. [2015].
The algorithm contains four resource-consuming steps: (1) projection (Line 10), (2) distribution generation (Line 3),
(3) sampling from the distribution, and (4) computing the covariance matrix. All the other steps are fast, depending
only polynomially on the dimension.
The major factor for the running time of sampling from the distribution (3), and computing the covariance matrix (4)
is the sparsity of the generated distribution, i.e., the number of possible outcomes. Sparse distributions (number of outcomes polynomial in the dimension) of sufficient accuracy can be efficiently generated by the decomposition algorithm
from Mirrokni et al. [2015], which we summarize as Algorithm 5 in Section D for the reader’s convenience. Common
choices of the exploration distribution µ are sparse, as discussed above, notwithstanding non-sparse distributions for µ
are also acceptable which have an efficient sampling method and a precomputed covariance matrix. Therefore we will
disregard the complexity of sampling and computation of the covariance matrix.
Finally, the projection step (Line 10) can be efficiently accomplished by the Frank–Wolfe algorithm (also called
conditional gradient), which we recall in Algorithm 4 in Section C. Note that if Algorithm 4 is used for the projection
step, it already provides a sparse linear decomposition of the desired expectation x̂t+1 with accuracy ε = 0, and
therefore makes a separate linear decomposition step unnecessary. Nevertheless it might be advantageous for specific
polytopes to use a specialized, more efficient projection algorithm and/or decomposition algorithm.
All in all, we measure complexity of only the most time-consuming tasks: projection and linear decomposition,
requiring the linear decomposition to be sparse. We report complexity of Algorithms 4 and 5 mentioned above in the
total number of linear optimization oracle calls over P. This relative complexity is often useful in applications where
fast linear programming oracles are available.
Now we are ready to state our main theorem on the regret and complexity of CombEXP.
Theorem 3.1 (High-probability regret bound for CombEXP for adaptive adversaries). For n ≥ 1 and with the choice
t−1/3
and
ηt := min{γ2t , γt λ}
2
Algorithm 1 achieves for any time T ≥ 1 the following regret: With probability at least 1 − δ, for any x ∈ P we have
r
!
T
n+2
k ak1 + B1
B( B + ε) 3
KL( a + x, a + x̂1 )
3
⊺
⊺
T 2/3
+2+
∑ ( Lt xt − Lt x ) ≤ 4 min{1, 2λT1/3} + B1 λ ln δ + (e − 2) λ
λ
4
t =1
B2 max{(k ak1 + B1 )/λ, ε} B1 max{1, B} B + ε √
2n + 2
+ O max 1,
,
, √
. (1)
T ln
λ
λ
δ
λ
γt :=
In particular, assuming α − ai ≤ zi ≤ β − ai for some 0 < α < β for all z ∈ P:
4
Algorithm 1 CombEXP
Require: polytope P ⊆ R n>− a , positive parameters ε, η1 ≥ η2 ≥ . . ., and 1/2 ≥ γ1 , γ2, . . .
Ensure: vertices xt of P as actions
1: x̂1 ∈ P arbitrary
2: for t = 1 to T do
3:
Find distribution pt with kE x ∼ pt [ x ] − x̂t k2 ≤ γt ε {approximate distribution}
4:
qt ← (1 − γt ) p t + γt µ
5:
Sample xt ∼ qt .
6:
Observe loss ℓt := L⊺t xt
7:
Ct ← E x ∼qt [ xx⊺ ]
8:
L̂t ← ℓt Ct−1 xt
9:
yt+1,i ← ( ai + x̂1,i )1−ηt+1/ηt ( ai + x̂t,i )ηt+1 /ηt exp(−ηt+1 L̂t,i ) − ai for all i ∈ [n]
10:
Find x̂t+1 ∈ P with KL( a + z, a + x̂t+1 ) ≤ KL( a + z, a + yt+1 ) + γt ηt+1 for all z ∈ P {approximate
projection}
11: end for
2
(i) Regret bound We have KL( a + x, a + x̂1 ) ≤ 4Bα so that the upper bound on the regret is proportional to T 2/3 .
With probability at least 1 − δ, for any x ∈ P we have
!
T
2 + kak
2
B
+
(
B
+
ε
)
2n
+
2
B
⊺
⊺
T 2/3 .
∑ ( Lt xt − Lt x ) ≤ O α + 1 min{λ, √λ} 1 · ln δ
t =1
(ii) Complexity. Using Algorithm 4 both for projection
and distribution generation (Lines 3 and 10) with ε = 0, the
algorithm makes altogether O
B4 β
α3 min{1,λ2 }
T 3 oracle calls to a linear optimization oracle over P.
Alternatively using a specialized projection algorithm in Line 10, and Algorithm 5 for distribution generation
in Line 3, then Algorithm 5 calls a linear optimization oracle over P at most O( B2 /ε2 ) T 5/3 times across all
rounds.
Obviously, the O( B2 /ε2 ) T 5/3 oracle calls in the last sentence does not contain the complexity of the specialized
projection algorithm.
Note that the bounds in Theorem 3.1 are any-time guarantees as the parameters of the algorithm do not depend on
the time horizon T. The constant factor in the regret bound can be slightly improved by a more sophisticated choice of
the γt and ηt , however, we preferred simple formulae for these parameters. Just as for EXP3, the choice of parameters
is different for the best expected regret and the best high-probability regret.
3.1 Proof of Theorem 3.1
In this section we will prove Theorem 3.1. We focus on the main regret bound, Equation (1), the other results easily
follow from it. See Propositions C.1 and D.1 for the complexity of Algorithms 4 and 5. The inequality KL( a + y, a +
2
x̂1 ) ≤ 4Bα is derived using ln z ≤ z − 1:
n
KL( a + y, a + x̂1 ) =
∑ (ai + yi ) ln
i =1
n
≤
∑ ( ai + yi )
i =1
n
n
ai + yi
− ∑ ( ai + yi ) + ∑ ( ai + x̂1,i )
ai + x̂1,i i=1
i =1
n ( y − x̂ )2
n
n
ky − x̂1 k22
ai + yi
4B2
1,i
− 1 − ∑ ( ai + yi ) + ∑ ( ai + x̂1,i ) = ∑ i
≤
≤
.
ai + x̂1,i
ai + x̂1,i
α
α
i =1
i =1
i =1
The proof of Equation (1) follows the standard approach, whereby we break-up the regret estimation into various
pieces, which we estimate separately:
T
T
∑ ( Lt x t − Lt x ) ≤ ∑
⊺
t =1
⊺
t =1
|
T
T
⊺
⊺
⊺
⊺
⊺
⊺
Lt xt − L̂t x̂t + ∑ ( L̂t x̂t − L̂t x ) + ∑ L̂t x − Lt x .
{z
Lemma 3.5
}
t =1
|
{z
}
Lemmas 3.3 and 3.4
5
t =1
|
{z
Lemma 3.6
}
(2)
given the history preLet E t [−] := E [− | x1 , L1 , . . . , x t−1 , Lt−1 , Lt ] denote the conditional expectation
operator
⊺
ceding round t and also the adversary’s action in round t. In particular, Ct = E t xt xt = (1 − γt ) Pt + γt J, with
Pt := E x ∼ pt [ xx⊺ ]. We first establish some basic bounds on quantities occurring in Algorithm 1.
Lemma 3.2 (Basic bounds). Let y ∈ P be arbitrary.
1
γt λ
B
k L̂t k2 ≤
γt λ
B
k L t k2 ≤
λ
kCt−1 k2 ≤
(3)
(4)
(5)
Proof. Equation (3) follows from Ct γt J γt λI. Inequality (4) follows via
| L̂t | = |ℓt · Ct−1 xt | ≤ |ℓt | · kCt−1 k2 · k xt k2 ≤
B
.
γt λ
Finally, (5) follows from the estimation
h
k L⊺t k2 = k L⊺t J J −1 k2 = E y∼µ L⊺t yy⊺ J −1
i
2
h
i
B
≤ Ey∼µ k L⊺t y · y⊺ J −1 k2 ≤ Ey∼µ k L⊺t yk2 kyk2 · k J −1 k2 ≤ .
λ
| {z } | {z } | {z }
≤1
≤B
≤1/λ
We now estimate the pieces of Equation (2). The following series of upper bounds are independent of the concrete
choice of the parameters γt , ηt . However, for the reader’s convenience in the last inequality of each estimation we
make the bound explicit by substituting the values for γt , ηt by the choices given in Theorem 3.1. We will tacitly use
the following inequality to estimate sums like ∑tT=1 γt :
( α +1
α +1
T
T
+α
≤ Tα+1 ,
−1 < α < 0
α
α
+
1
∑ t ≤ T α +1 − 1 + T α ≤ T α +1 + T α , α > 0
t =1
α +1
α +1
We first estimate the regret when using the loss estimators L̂t . For this we use a generalized variant of EXP (see
Lemma A.1), which works with arbitrary convex sets contained in the positive orthant.
Lemma 3.3.
T
∑ ( L̂t x̂t − L̂t x ) ≤
⊺
⊺
t =1
n
T
KL( a + x, a + x̂1 ) T −1
+ ∑ γt + (e − 2) ∑ ηt ∑ ( ai + x̂t,i ) L̂2t,i
ηT
t =1
t =1
i =1
n
T
KL( a + x, a + x̂1 ) T 2/3 3 2/3
η
≤4
T
+
(
e
−
2
)
+
∑ t ∑ (ai + x̂t,i ) L̂2t,i .
4
min{1, 2λT 1/3 }
t =1
i =1
(6)
Proof. This follows from Lemma A.1 with the L̂t as loss vectors and the a + x̂t as played actions. Note that a cancels
on the left hand side in ( L̂⊺t ( a + x̂t ) − L̂⊺t ( a + x )).
In a next step we estimate the last term of Equation (6).
Lemma 3.4. With probability at least 1 − δ
T
n
t =1
i =1
∑ ηt ∑ (ai + x̂t,i ) L̂2t,i
v
u T
ηt2
ηt
(k ak1 + B1 ) B2 u
1
t1
· ln
∑ γt +
∑
2
4
2
δ
λ
t =1
t =1 γ t
r
1
k ak1 + B1 3 2/3 (k ak1 + B1 ) B2 1
T
+
T ln .
≤
λ
4
2
δ
λ2
k ak1 + B1
≤
λ
T
6
(7)
Proof. This is a special case of the Azuma–Hoeffding inequality (recalled in Theorem B.1) using the bounds
n
n
B 2
(k ak1 + B1 ) B2
0 ≤ ∑ ( ai + x̂t,i ) L̂2t,i ≤ ∑ ( ai + x̂t,i )
≤
γt λ
( γt λ )2
i =1
i =1
and
Et
"
n
∑
i =1
( ai + x̂t,i ) L̂2t,i
#
= Et
"
n
∑ ( ai +
i =1
x̂t,i )ℓ2t e⊺i Ct−1 xt x⊺t Ct−1 ei
#
n
=
∑ (ai + x̂t,i )ℓ2t e⊺i Ct−1 Ct Ct−1 ei
i =1
n
≤
Next we bound the difference between the true loss
⊺
Lt x t
∑ (ai + x̂t,i )ei Ct−1 ei ≤
⊺
i =1
k ak1 + B1
.
γt λ
⊺
and the expected estimated loss L̂t x̂t .
Lemma 3.5. With probability at least 1 − δ
v
u
u
γt + t2
!
T
γ
1
(
3B
+
ε
)(
B
+
ε
)
t
T+
∑
∑ (1 − γt )2 ln δ
∑ L⊺t xt − L̂⊺t x̂t ≤
λ
t =1
t =1
t =1
!
1
B
1
B2 ε
p
ln
+
+3+
3
λ
δ
γ T (1 − γ T ) λ
B( B + ε) T
1
1
B
B2 ε
B+ε √
p
√
≤ 1+
ln
T
max
1,
ln
γ
+
+
3
+
+
O
max
1,
t
∑
λ
λ
δ
3 γ T (1 − γ T ) λ δ
λ
t =1
√
√
2
1
B ε
B+ε
2B
1
B( B + ε) 3 2/3
.
T + √ ( T 1/6 + T −1/6 ) ln + 3 +
+ O max 1, √
T
max 1, ln
≤ 1+
λ
4
δ
λ
δ
3 λ
λ
(8)
T
B( B + ε)
1+
λ
T
Proof. Let x̃t := E x ∼ pt [ x ] and xt := E t [ xt ] = (1 − γt ) x̃t + γt u. As also k x̂t − x̃t k2 ≤ γt ε, we have xt =
(1 − γt ) x̂t + γt v for v := u + 1−γtγt ( x̃t − x̂t ) with kvk2 ≤ B + ε(1 − γt ) ≤ B + ε. We consider the martingale
difference sequence
Xt := L⊺t xt − L̂⊺t x̂t − E t L⊺t xt − L̂⊺t x̂t = L⊺t xt − L̂⊺t x̂t − L⊺t xt + L⊺t x̂t = L⊺t xt − L̂⊺t x̂t + γt L⊺t ( x̂t − v).
⊺
Note that as x̃t x̃t = E x ∼ pt [ x ] E x ∼ pt [ x ]⊺ E x ∼ pt [ xx⊺ ] = Pt ≤ Ct /(1 − γt )
L̂⊺t x̃t
and
2
xt Ct−1 xt
B2
B2
≤
≤
,
1 − γt
γt λ (1 − γt )
γ T λ(1 − γ T )
⊺
= L̂⊺t x̃t x̃⊺t L̂t = ℓ2t x⊺t Ct−1 x̃t x̃⊺t Ct−1 xt ≤ x⊺t Ct−1 Pt Ct−1 xt ≤
| L̂⊺t x̂t − L̂⊺t x̃t | ≤ k L̂t k2 · k x̂t − x̃t k2 ≤
B2
B2 ε
γt ε =
,
γt λ
λ
hence
| Xt | ≤ L⊺t xt + L⊺t xt + L⊺t x̂t + L̂⊺t x̂t − L̂⊺t x̃t + L̂⊺t x̃t ≤ 3 +
and the variance of Xt is easily bounded by:
B
B2 ε
+p
,
λ
γ T (1 − γ T ) λ
i
h
i
h
Vart [ Xt ] ≤ E t ( L⊺t xt − L̂⊺t x̂t )2 = E t (ℓt · (1 − x⊺t Ct−1 x̂t ))2
i
h
i
h
≤ E t (1 − x⊺t Ct−1 x̂t )2 = Et 1 − 2x⊺t Ct−1 x̂t + x̂⊺t Ct−1 xt x⊺t Ct−1 x̂t
= 1 − 2xt ⊺ Ct−1 x̂t + x̂⊺t Ct−1 x̂t = 1 −
γ2t
γt (3B + ε)( B + ε)
1 − 2γt ⊺ −1
x
C
x
+
.
(v − 2xt )⊺ Ct−1 v ≤ 1 +
t
t
t
2
2
(1 − γt )
(1 − γt )
(1 − γt )2 λ
7
Hence Benett’s inequality (Theorem B.2, [Fan et al., 2012, (18)]) applied to the martingale difference sequence Xt
provides
!
T
1
B
1
B2 ε
⊺
⊺
⊺
∑ Lt xt − L̂t x̂t + γt Lt (x̂t − v) ≤ 3 pγ (1 − γ )λ + 3 + λ ln δ
T
T
t =1
v
!
u
u
1
(3B + ε)( B + ε) T
γt
t
ln .
+ 2 T+
∑
2
λ
δ
(1 − γt )
t =1
The claim follows by using | L⊺t ( x̂t − v)| ≤ 1 + B( B + ε)/λ.
Finally, we bound the difference between the true loss L⊺t x and the estimated loss L̂⊺t x for any point x ∈ P.
Lemma 3.6. For all 0 < δ < 1 with probability at least 1 − δ for every x ∈ R n simultaneously
v
u T
T
u2
2n
2n
1
1
k x k1 B
⊺
⊺
∑ L̂t x − Lt x ≤ 3 λ + γT λ ln δ + kx k1 t λ ∑ γt ln δ .
t =1
t =1
In particular, with probability at least 1 − δ, for all x ∈ P simultaneously
T
∑
t =1
⊺
L̂t x −
⊺
Lt x
2n
B
+ B1 T 2/3
≤ 1 B + 2T 1/3 ln
3λ
δ
r
4
1+
3T
r
2n
3
ln .
λ
δ
(9)
Remark 3.7. Restricting the statement for all x ≥ 0, the ln(2n/δ) can be replaced by ln(n/δ).
Proof. Let x = ±ei be a coordinate vector or its negation. Then
and
h
i
i
h
1
,
Vart L̂⊺t x − L⊺t x ≤ E t ( L̂⊺t x )2 ≤ E t x⊺ Ct−1 xt x⊺t Ct−1 x = x⊺ Ct−1 x ≤
γt λ
| L̂⊺t x − L⊺t x | ≤
B
1
+
.
λ
γt λ
Hence by Benett’s inequality (Theorem B.2, [Fan et al., 2012, (18)]) the claim follows for a fixed vector x = ±ei with
probability at least 1 − δ/(2n). Hence by the union bound, it holds for all x = ±ei simultaneously with probability
at least 1 − δ. Finally, the inequality for a general x follows by taking linear combinations with the absolute values of
the coefficients of x.
Summing up (6), (7), (8) (substituting δ/(2n + 2) for δ in the latter two) and (9) (substituting 2nδ/(2n + 2) for
δ), with probability at least 1 − δ yields (1) of Theorem 3.1.
4 A high-probability regret bound for ComBand
In this section we will show that ComBand of Cesa-Bianchia and Lugosi [2012] achieves a high-probability
regret
√
bound of O( T 2/3 ) without any modifications. While this is worse than the optimal regret of O( T ) obtained by
GeometricHedge in Bartlett et al. [2008], it shows that already Algorithm 2, the vanilla version of ComBand without
any correction terms suffices to achieve a high-probability regret bound.
Theorem 4.1. With the choice
ηt :=
γt λ
B2
and
8
γt :=
t−1/3
2
Algorithm 2 ComBand
Require: Losses Lt , action set A ⊆ R n , positive parameters η1 ≥ η2 ≥ . . ., 1/2 ≥ γ1 ≥ γ2 ≥ . . .
Ensure: actions xt ∈ A
1: for t = 1 to T do
−1 ⊺
2:
wt ( x ) ← ∑ti=
for all x
1 L̂i x
3:
Wt ← ∑ x wt ( x )
4:
pt ( x ) ← wt ( x )/Wt for all x
5:
qt ← (1 − γt ) p t + γt µ
6:
Sample xt ∼ qt .
⊺
7:
Observe loss ℓt := Lt xt .
8:
Ct ← E x ∼qt [ xx⊺ ]
9:
L̂t ← ℓt Ct−1 xt
10: end for
Algorithm 2 achieves regret
!
√ r
λ
N+2 √
3( e − 2) λ 3
B2
3B
N+2
2/3
√
T
+
O
n
ln
+n
+
+
1
+
ln
T
δ
2
λ
δ
4B2
B2
λ
!
r
N+2
λ
B
ln
+ n 2 T 2/3
≤O √
δ
B
λ
(10)
with probability at least 1 − δ for 0 < δ < 1.
Remark 4.2. Similar to Theorem 3.1, it is possible to change the ln(( N + 2)/δ) in the coefficient of T 2/3 to the
possibly much smaller ln((n + 2)/δ) with a suitable altering of the other constants. However, since an T 1/3 ln N
term will still remain in the regret bound, this does not seem to be a significant improvement.
We use the same notation as in Section 3.1 for CombEXP, which we recall here for the reader’s convenience. Let
E t [−] := E [− | x1 , L1 , . . . , x t−1, Lt−1 , Lt ] denote the conditional expectation operator given the history preceding
round t and also the adversary’s action in round t. Let x̃t := E x ∼ pt [ x ] and Pt := E x ∼ pt [ xx⊺ ] denote the expectation
and variance of distribution pt , respectively. Note that Ct = E t xt x⊺t = (1 − γt ) Pt + γt J.
Lemma 4.3 (Basic bounds). Let x, y1 , and y2 be arbitrary actions.
(i) Bounds on size
|y⊺1 Ct−1 y2 | ≤
B2
γt λ
(11)
| L̂⊺t x | ≤
B2
γt λ
(12)
i
h
E t x⊺t Ct−1 xt = n
(13)
(ii) Bounds on expectation
Proof. Equation (11) follows from the bounds ky1 k2 , ky2 k2 ≤ B and kCt−1 k2 ≤ 1/(γt λ), as Ct γt J γt λI.
Inequality (12) follows via
B2
| L̂⊺t x | = |ℓt · x⊺t Ct−1 x | ≤
.
γt λ
To prove (13), we use a trick using the trace function to compute the expectation:
h
i
i
h
E t x⊺t Ct−1 xt = E t Tr(Ct−1 xt x⊺t ) = Tr(Ct−1 Ct ) = Tr( I ) = n.
9
h
i
Remark 4.4. One can similarly prove E y∼ pt y⊺ Pt−1 y = n, but it will not be used in the following.
As in the case of CombEXP, the lemmas below are independent of the choice of the γt , ηt except for the last formula
in each lemma, where we particularize the bounds by substituting parameters.
First instead of the real regret, we estimate the regret computed using the estimators L̂t .
Lemma 4.5. With probability at least 1 − δ
T
∑ ( L̂⊺t x̃t − L̂⊺t x ) ≤
t =1
≤
v
u
t
B2 u 1 T
T
ln N
ηt
+
+ ( e − 2 ) n ∑
ηT
1
−
γt
λ
t =1
2B2 ln N 1/3
T
+ ( e − 2)
λ
ηt2
1
· ln
2 (1 − γ )2
2 t∑
δ
γ
t
=1 t
!
r
1
3λ 2/3
B
2T ln
.
n 2 (T
+ 2T 1/3 ) + √
δ
4B
λ
Proof. By Lemma A.1,
T
∑
t =1
h
i
T
ln N
⊺
⊺
L̂t x̃t − L̂t x ≤
+ (e − 2) ∑ ηt E y∼ pt ( L̂⊺t y)2 .
ηT
t =1
To estimate the last term, first note that
⊺
i
h
x C −1 x t
.
E y∼ pt ( L̂⊺t y)2 = E y∼ pt L̂⊺t yy⊺ L̂t = L̂⊺t Pt L̂t = ℓ2t xtT Ct−1 Pt Ct−1 xt ≤ t t
1 − γt
So far combining our estimates provides
T
⊺
⊺
∑ ( L̂t x̃t − L̂t x ) ≤
t =1
T
T
x ⊺ C −1 x t
x ⊺ C −1 x t
2B2 ln N 1/3
ln N
=
.
+ ∑ ηt t t
T
+ ( e − 2 ) ∑ ηt t t
ηT
1 − γt
λ
1 − γt
t =1
t =1
(14)
To estimate the last term on the right-hand side, we apply the Azuma–Hoeffding inequality using (11) and (13) for
bounding the summands and their expectation, which readily proves the lemma:
v
u T
⊺ −1
T
T
2
η2
xt Ct xt
ηt
B u
1
1
≤
n
+
η
∑ 1 − γt λ t 2 ∑ γ2 (1 −t γt )2 · ln δ .
∑ t 1 − γt
t =1
t =1 t
t =1
We turn our attention to the difference between the real loss vectors Lt and their estimators L̂t . We start by comparing the loss of the played action.
Lemma 4.6. With probability at least 1 − δ
T
∑
t =1
⊺
⊺
Lt xt − L̂t x̃t
v
!
u
3B2 T
1
B
γt
1
1 u
t
2+ p
≤ 2 ∑ γt +
ln + 2 T +
∑ (1 − γt )2 ln δ
3
δ
λ
γ
λ
(
1
−
γ
)
T
T
t =1
t =1
(15)
s
!
√
2
2B
3
1
9B 2/3
1
1
≤ T 2/3 +
2 + √ ( T 1/6 + T −1/6 ) ln + 2 T +
T
ln .
2
3
δ
λ
δ
λ
!
T
Proof. Let xt := E t [ xt ] = (1 − γt ) x̃t + γt u denote the conditional expectation of xt given the history before round
t and loss Lt . The statement is a special case of Benett’s inequality (see Theorem B.2) for the martingale
Xt := L⊺t xt − L̂⊺t x̃t − E t L⊺t xt − L̂⊺t x̃t = L⊺t xt − L̂⊺t x̃t − L⊺t xt + L⊺t x̃t = L⊺t xt − L̂⊺t x̃t + γt L⊺t ( x̃t − u).
Note that x̃t x˜t ⊺ Pt by Jensen’s inequality, therefore
L̂⊺t x̃t
2
= L̂⊺t x̃t x̃⊺t L̂t = ℓ2t x⊺t Ct−1 x̃t x̃⊺t Ct−1 xt ≤ x⊺t Ct−1 Pt Ct−1 xt ≤
10
x⊺t Ct−1 xt
B2
≤
,
1 − γt
γt λ (1 − γt )
hence
| Xt | ≤ 1 + p
B
γt λ (1 − γt )
and the variance of Xt is easily bounded by:
+ 2γt ≤ 2 + p
B
γ T λ(1 − γ T )
,
i
h
i
h
Vart [ Xt ] ≤ E t ( L⊺t xt − L̂⊺t x̃t )2 = E t (ℓt · (1 − x⊺t Ct−1 x̃t ))2
i
h
i
h
≤ E t (1 − x⊺t Ct−1 x̃t )2 = Et 1 − 2x⊺t Ct−1 x̃t + x̃⊺t Ct−1 xt x⊺t Ct−1 x̃t
= 1 − 2xt ⊺ Ct−1 x̃t + x̃⊺t Ct−1 x̃t = 1 −
γ2t
3γt B2
1 − 2γt ⊺ −1
⊺ −1
x
C
x
+
C
u
≤
1
+
u
−
2x
.
)
(
t
t
t
t
t
(1 − γt )2
(1 − γt )2
(1 − γt )2 λ
Benett’s inequality provides
T
∑
t =1
⊺
Lt x t
−
⊺
L̂t x̃t
+ γt L⊺t ( xt
1
− u) ≤
3
⊺
2+ p
B
γ T λ(1 − γ T )
!
v
u
1 u
ln + t2
δ
3B2
T+
λ
T
γt
∑ (1 − γt )2
t =1
!
1
ln .
δ
The claim follows by using | Lt ( x t − u)| ≤ 2.
Now we compare the losses Lt with their estimator L̂t for all fixed actions.
Lemma 4.7. For all 0 < δ < 1 with probability at least 1 − δ for every x ∈ A simultaneously
v
u
T
2
u 2B2 T 1
1
B
N
N
⊺
⊺
∑ L̂t x − Lt x ≤ 3 1 + γT λ ln δ + t λ ∑ γt · ln δ
t =1
t =1
r
r
√
2B2 1/3
N
3B 2/3
1
4
N
√
1+
ln +
T
≤
T
1+
ln .
3
λ
δ
3T
δ
λ
(16)
Proof. As customary for concentration inequalities, we start by a variance and size estimate:
h
i
i
h
⊺
B2
⊺
⊺
⊺
Vart L̂t x − Lt x ≤ E t ( L̂t x )2 ≤ E t x⊺ Ct−1 xt xt Ct−1 x = x⊺ Ct−1 x ≤
,
γt λ
and
| L̂⊺t x − L⊺t x | ≤ 1 +
⊺
B2
.
γt λ
⊺
Also note that L̂t x − Lt x is a martingale difference sequence. Hence by Benett’s inequality (see Theorem B.2) the
claim follows for a fixed action x with probability at least 1 − δ/N. Therefore by the union bound, it holds for all
x ∈ A simultaneously with probability at least 1 − δ.
Summing up (14), (15) (substituting δ/( N + 2) for δ), and (16) (substituting Nδ/( N + 2) for δ), we obtain (10)
with probability at least 1 − δ.
5 Concluding remarks
We would like
√ to mention that our method could be immediately strengthened to provide an optimal high-probability
regret of O( T ) using the correction term of GeometricHedge (see Bartlett et al. [2008]) and the identity
⊺ +
⊺
E n ∑ M̃i (n) X̃i2 (n) = E n X (n)⊺ M (n) M (n)⊺ Σ+
n −1 M̃ ( n) M̃( n) Σn −1 M ( n) M ( n) X ( n) ,
i ∈[d ]
11
√
used for establishing the O( T ) regret bound for the expected case under oblivious adversaries in [Combes et al., 2015,
supplementary material, proof of Theorem 6]. However, we were unable to verify this identity 1 , which is equivalent
to
2
E n ∑ M̃i (n) X̃i (n)
i ∈[d ]
2
= E n
∑
i ∈[d ]
M̃i (n) X̃i (n) ,
and as such we only claim the weaker bound of O( T 2/3 ). This is the only obstacle to combining
CombEXP with
√
GeometricHedge to obtain an efficient algorithm with optimal high-probability regret O( T ) for the adaptive case
using our method.
To put this into context, without the above identity also for the expected regret case under oblivious adversaries
we were only able to establish an O( T 2/3 ) regret bound, matching our high-probability regret bound for adaptive
adversaries.
References
J. Abernethy, E. Hazan, and A. Rakhlin. Competing in the dark: An efficient algorithm for bandit linear optimization. In Proceedings of the 21st Annual Conference on Learning Theory (COLT, pages 263–274, July 2008. URL
http://colt2008.cs.helsinki.fi/papers/123-Abernethy.pdf.
R. Arora, O. Dekel, and A. Tewari. Online bandit learning against an adaptive adversary: from regret to policy regret.
In J. Langford and J. Pineau, editors, In Proceedings of the 29th International Conference on Machine Learning,
pages 1503–1510, New York, NY, USA, July 2012. Omnipress. ISBN 978-1-4503-1285-1.
J.-Y. Audibert, S. Bubeck, and G. Lugosi. Regret in online combinatorial optimization. Mathematics of Operations
Research, 39(1):31–45, 2013.
P. Auer, N. Cesa-Bianchi, Y. Freund, and R. E. Schapire. The nonstochastic multiarmed bandit problem. Siam J.
Comput., 32(1):48–77, 2002.
B. Awerbuch and R. D. Kleinberg. Adaptive routing with end-to-end feedback: Distributed learning and geometric
approaches. In Proceedings of the thirty-sixth annual ACM symposium on Theory of computin, pages 45–53, New
York, NY, USA, June 2004. ISBN 1-58113-852-0. doi:10.1145/1007352.1007367.
P. L. Bartlett, V. Dani, T. Hayes, S. Kakade, A. Rakhlin, and A. Tewari. High-probability regret bounds for bandit online linear optimization. In 21th Annual Conference on Learning Theory (COLT 2008), July 2008. URL
http://eprints.qut.edu.au/45706/1/30-Bartlett.pdf.
S. Bubeck and N. Cesa-Bianchi. Regret analysis of stochastic and nonstochastic multi-armed bandit problems. Foundations and Trends in Machine Learning, 5(1):1–122, December 2012. doi:10.1561/2200000024.
S. Bubeck and R. Eldan. Multi-scale exploration of convex functions and bandit convex optimization. COLT 2016,
arXiv:1507.06580v1, 2015.
S. Bubeck, R. Eldan, and Y. T. Lee.
arXiv:1607.03084, July 2016.
Kernel-based methods for bandit convex optimization.
arXiv preprint
N. Cesa-Bianchia and G. Lugosi. Combinatorial bandits. Journal of Computer and System Sciences (Special Issue:
Cloud Computing 2011), 78(5):1404–1422, September 2012. doi:10.1016/j.jcss.2012.01.001.
R. Combes, M. S. Talebi Mazraeh Shahi, A. Proutiere, and M. Lelarge.
Combinatorial bandits revisited.
In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances
in Neural Information Processing Systems 28, pages 2116–2124. Curran Associates, Inc., 2015.
URL
http://papers.nips.cc/paper/5831-combinatorial-bandits-revisited.pdf.
1 As
of October 2016, we are discussing the matter with the authors of Combes et al. [2015].
12
V. Dani and T. P. Hayes. How to beat the adaptive multi-armed bandit. arXiv preprint, arXiv:cs/0602053, February
2006.
V. Dani, T. P. Hayes, and S. Kakade.
The price of bandit information for online optimization.
In Advances in Neural Information Processing Systems, volume 20, pages 345–352, 2007.
URL
http://machinelearning.wustl.edu/mlpapers/papers/NIPS2007_758.
J. Edmonds. Maximum matching and a polyhedron with 0,1-vertices. J. Res. Nat. Bur. Standards B, 69:125–130, 1965.
X. Fan, I. Grama, and Q. Liu. Hoeffding’s inequality for supermartingales. Stochastic Processes and their Applications,
122:3545–3559, 2012. doi:10.1016/j.spa.2012.06.009.
M. Grötschel, L. Lovász, and A. Schrijver. Geometric algorithms and combinatorial optimization, volume 2 of Algorithms and Combinatorics. Springer-Verlag, Berlin, second edition, 1993. ISBN 3-540-56740-2.
E. Hazan and Y. Li. An optimal algorithm for bandit convex optimization. arXiv preprint, arXiv:1603.04350, March
2016.
E. Hazan, Z. Karnin, and R. Meka. Volumetric spanners: an efficient exploration basis for learning. In JMLR: Workshop
and Conference Proceedings, volume 35, pages 1–15, 2014.
M. Jaggi. Revisiting Frank–Wolfe: Projection-free sparse convex optimization. In Proceedings of the 30th International
Conference on Machine Learning (ICML-13), pages 427–435, 2013.
S. Lacoste-Julien and M. Jaggi.
On the global linear convergence of Frank–Wolfe optimization variants.
In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 28, pages 496–504. Curran Associates, Inc., 2015.
URL
http://papers.nips.cc/paper/5925-on-the-global-linear-convergence-of-frank-wolfe-optimiz
C. H. Lim and S. J. Wright. Efficient Bregman projections onto the permutahedron and related polytopes. In Proceedings of the 19th International Conference on Artificial Intelligence and Statistics, pages 1205–1213, 2016. URL
http://www.jmlr.org/proceedings/papers/v51/lim16.html.
V. S. Mirrokni, R. P. Leme, A. Vladu, and S. C. wai Wong. Tight bounds for approximate Carathéodory and beyond.
arXiv preprint, arXiv:1512.08602, 2015.
A. Nemirovski. Efficient methods for large-scale convex optimization problems. Ekonomika i Matematicheskie Metody,
15, 1979.
A. Nemirovski. Advances in convex optimization: Conic programming. In Proceedings of the International Congress
of Mathematicians. EMS-European Mathematical Society Publishing House, 2006.
G. Neu.
Explore no more: Improved high-probability regret bounds for non-stochastic bandits.
In
C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 3168–3176. Curran Associates, Inc., 2015.
URL
http://papers.nips.cc/paper/5732-explore-no-more-improved-high-probability-regret-bounds
T. Rothvoß. The matching polytope has exponential extension complexity. Proceedings of STOC, pages 263–272,
2014.
13
A
Time-varying EXP algorithm with projections
Let R n>0 := x ∈ R n+ xi > 0 for all i ∈ [n] be the strictly positive orthant. In this section we provide a version of
EXP (see Algorithm 3) (1) that computes points in an arbitrary convex set P ⊆ R n>0 (as compared to distributions in
the probability simplex), and (2) that is any-time, i.e., the parameter choice is independent of T and the regret bounds
hold uniformly for any t ≤ T. We explicitly allow arbitrary dependence between the parameters ηt , input Lt , and the
points xt computed by the algorithm, to ease the use of the regret bound in applications.
Algorithm 3 EXP for convex sets contained in R n>0 with time varying parameters
Require: convex set P ⊆ R n>0 , start point x1 ∈ P, loss vectors Lt ∈ R n , positive parameters η1 ≥ η2 ≥ . . .,
satisfying ηt Lt,i ≥ −1 for i = 1, . . . , n
for t = 1 to T − 1 do
1− η /η η /η
x̃t+1,i ← x1,i t+1 t xt,it+1 t · e−ηt+1 Lt,i for all i ∈ [n]
Find xt+1 ∈ P such that KL(z, x t+1 ) ≤ KL(z, x̃t+1 ) + ε t for all z ∈ P. {approximate Bregman projection}
end for
Lemma A.1. Let P ⊆ R n>0 be a convex set and let ηt Lt,i ≥ −1 for all t and i, the vector xt computed by Algorithm 3
satisfy the following:
T
T
∑ Lt x t − ∑ Lt y ≤
⊺
⊺
t =1
t =1
n
T
KL(y, x1 ) T −1 ε t
+ ∑
+ (e − 2) ∑ ηt ∑ xt,i L2t,i .
ηT
η
t =1 t +1
t =1
i =1
Proof. The proof is an extension of the standard analysis of EXP, using the potential (KL(y, xt ) − KL(y, x1 ))/ηt to
measure progress:
n
n
n
x1,i
(17)
− ∑ x1,i + ∑ xt,i .
KL(y, xt ) − KL(y, x1 ) = ∑ yi ln
xt,i i=1
i =1
i =1
We compare this with the potential in the next round, first using x̃t+1 instead of xt+1 :
η
KL(y, x̃t+1 ) − KL(y, x1 ) = t+1
ηt
≤
ηt +1
ηt
n
x1,i
∑ yi ln xt,i
i =1
n
n
n
xt,i −ηt Lt,i ηt+1 /ηt
x1,i
∑ yi ln xt,i + ηt+1 ∑ yi Lt,i − ∑ x1,i + ∑ x1,i x1,i e
i =1
i =1
i =1
i =1
n
n
xt,i −ηt Lt,i
η
+ ηt+1 L⊺t y − ∑ x1,i + ∑ x1,i 1 + t+1
e
−1
ηt
x1,i
i =1
i =1
!
n
ηt +1
⊺
KL(y, x t ) − KL(y, x1 ) + ∑ xt,i (e−ηt Lt,i − 1) + ηt+1 Lt y,
=
ηt
i =1
n
removing the exponent ηt+1 /ηt using z a ≤ 1 + a(z − 1) for z > 0 and 0 < a < 1 (which is Jensen’s inequality for z a
as a function of a), and then plugging in (17). Rearranging and using the estimate e a ≤ 1 + a + (e − 2) a2 for a ≤ 1
with the choice a := −ηt L̂t,i ≤ 1 provides:
KL(y, x̃t+1 ) − KL(y, x1 ) KL(y, xt ) − KL(y, x1 )
1 n
xt,i e−ηt Lt,i − 1
− L⊺t y ≤
−
∑
ηt +1
ηt
ηt i =1
n
1 n
2
= − L⊺t xt + (e − 2)ηt ∑ xt,i L2t,i ,
−
η
L
+
(
e
−
2
)(
η
L
)
x
≤
t t,i
t t,i
t,i
∑
ηt i =1
i =1
Summing up for t = 1, . . . , T and rearranging leads to (using the value ηT +1 = ηT )
T
T
∑
t =1
L⊺t xt −
∑
t =1
T
L⊺t y − (e − 2) ∑ ηt
t =1
n
∑ xt,i L2t,i ≤
i =1
≤
T −1
∑
t =1
T −1
∑
t =1
KL(y, xt+1 ) − KL(y, x̃t+1 ) KL(y, x1 ) − KL(y, x̃ T +1)
+
ηt
ηT
KL(y, x1 )
εt
+
,
ηt +1
ηT
using KL(y, x̃ T +1) ≥ 0 and KL(y, xt+1 ) − KL(y, x̃t+1 ) ≤ ε t . The claim follows by rearranging.
14
B Concentration inequalities
We will use the following concentration inequalities.
Theorem B.1 (Azuma–Hoeffding inequality). For a martingale difference sequence Xt with at ≤ Xt ≤ bt almost
surely for constants at , bt , we have with probability at least 1 − δ
s
T
∑tT=1 (bt − at )2 ln(1/δ)
X
≤
.
t
∑
2
t =1
While the following inequality is stated only for b = 1 in [Fan et al., 2012, (18)] it easily generalizes via scaling to
arbitrary b > 0.
Theorem B.2 (Benett’s inequality [Fan et al., 2012, (18)]). For a supermartingale difference sequence Xt bounded
above by a positive constant Xt ≤ b, for any v ≥ 0 with probability at least 1 − δ:
T
T
∑ Vart [Xt ] ≥ v
∑ Xt ≤
or
t =1
t =1
C
b ln(1/δ)
+
3
q
2v ln(1/δ).
Projection for Kullback–Leibler divergence
We will now describe a generic, efficient, simple Frank–Wolfe algorithm for the projection step in Line 10 of Algorithm 1. We remark that there are many possibilities for improvements, such as, e.g., employing advanced variants
of the Frank–Wolfe algorithm (see e.g., Lacoste-Julien and Jaggi [2015]) or using customized algorithms for specific
polytopes. For example, in the case of the simplex P = { x ≥ 0 | ∑i xi = 1}, the projection of x is simply x/ ∑ni=1 xi
and for the the permutahedron there exist very fast, specialized projection methods (see e.g., Lim and Wright [2016]).
Algorithm 4 Projection for KL
Require: linear optimization oracle over a polytope P ⊆ [α, β]n , α > 0, upper bound B for the ℓ2 -diameter of P,
accuracy ε > 0, point x ∈ R n>0
Ensure: yK ∈ P with KL(z, yK ) ≤ KL(z, x ) + ε for all z ∈ P
y0 ∈ P
m
l any point
4B4 β
K ← α3 ε 2
for k = 1 to K do
s ∈ arg minz∈ P ∑ni=1 zi ln(yk−1,i /xi ) {Linear optimization oracle call}
yk ← ((k − 1)yk−1 + 2s)/(k + 1)
end for
return yK
Proposition C.1. Given a polytope P ⊆ [α, β]n with α > 0, an upper bound
for the ℓ2 -diameter of P, as well as an
B
accuracy ε > 0, Algorithm 4 computes an approximate projection with O
B4 β
α3 ε 2
oracle calls.
Proof. As the algorithm calls the oracle once per iteration, the bound on the number of oracle calls is immediate.
To prove the claimed accuracy of the returned point yK , note that the algorithm is the Frank–Wolfe algorithm for the
function f (z) := KL(z, x ). Recall that the gradient ∇ f (z) of f at z is given by (∇ f (z))i = ln(zi /xi ) and the Hessian
is a diagonal matrix ∇2 f (z) = diag(1/z1 , 1/z2 , . . . , 1/zn ). As 1/β ≤ 1/zi ≤ 1/α for z ∈ P, the function f is 1/αsmooth and 1/β-strongly convex on P in the ℓ2 -norm, and has curvature C f ≤ B2 /α. Let x ∗ := arg minz∈ P f (z),
i.e., the Bregman projection of x to P. By [Jaggi, 2013, Theorem 1], f (yK ) − f ( x ∗ ) ≤ 2C f /(K + 2), therefore by
strong convexity
1
2B2
kyK − x ∗ k22 ≤ KL(yK , x ) − KL( x ∗ , x ) ≤
.
2β
α ( K + 2)
15
Let z ∈ P be arbitrary. By the Pythagorean Theorem we have KL(z, x ∗ ) ≤ KL(z, x ) and thus
KL(z, yK ) − KL(z, x ) ≤ KL(z, y K ) − KL(z, x ∗ ) =
n
∑ zi ln
i =1
n
n
xi∗
− ∑ xi∗ + ∑ yK,i
yK,i i=1
i =1
n z −y
n
n
xi∗
B
K,i
≤ ∑ zi
− 1 − ∑ xi∗ + ∑ yK,i = ∑ i
( xi∗ − yK,i ) ≤ k x ∗ − yK k2
yK,i
yK,i
α
i =1
i =1
i =1
i =1
p
p q
2
2B β
B 2β
√
KL(y K , x ) − KL( x ∗ , x ) ≤
≤ ε.
≤
3/2
α
α
K+2
n
Plugging in K =
D
l
4B4 β
α3 ε 2
m
as set by the algorithm provides the result.
Linear decomposition
For the convenience of the reader, we briefly recall the decomposition algorithm (Algorithm 5) of Mirrokni et al. [2015]
that for a polytope P approximately decomposes any point x ∈ P into a convex combination of vertices of P, using
a linear optimization oracle over P. The algorithm uses Mirror Descent (see Nemirovski [1979]) to find a convex
combination.
Proposition D.1 ([Mirrokni et al., 2015, Theorem 3.5]). Given a polytope P with diameter at most 2D in ℓ2 -norm,
and a point x ∈ P, Algorithm 5 computes with O( D2 /ε2 ) calls to a linear optimization oracle over P a multiset x1 ,
. . . , xk of vertices for k = ⌈4D2 /ε2 ⌉ such that k∑ki=1 xi /k − x k2 ≤ ε.
Algorithm 5 Linear decomposition
Require: linear optimization oracle over polytope P, an inner point x ∈ P, precision ε
Ensure: vertices x1 , . . . , x k ∈ P such that k x − ∑i λi xi /kk2 ≤ ε
k ← ⌈4D2 /ε2 ⌉
η ← 4ε( p − 1)
y1 ← 0; z1 ← 0
for t = 1 to k do
⊺
Choose vertex xt ∈ arg miny∈ P yt y {Linear optimization oracle call}
z t +1 ← z t − η ( x − x t )
if kzt+1 k2 > 1 then
y t +1 ← z t +1 / k z t +1 k2
else
y t +1 ← z t +1
end if
end for
return x1 , . . . , xk
E Fitness of barycentric spanners for exploration
Let λmin (µ) denote the minimal eigenvalue of the covariance matrix E x ∼µ [ xx⊺ ] of a distribution µ. For exploration
one wishes to find a µ with a high minimal eigenvalue λmin (µ). Here we show that a uniform distribution on any
approximate barycentric spanner achieves within an O(n2 ) factor the best possible minimal eigenvalue using any scalar
product on R n . The free choice of scalar product and hence orthonormal basis allows preserving sparse representation
of a polytope P.
16
Lemma E.1. Let v1 , . . . , v n be a C-approximate barycentric spanner of a polytope P ⊆ R n . Then the uniform
distribution µv1 ,...,vn on the spanner satisfies
λmin (µv1 ,...,vn ) ≥
λmin (µ)
C 2 n2
for any distribution µ over P.
Proof. Using that the vi form a barycentric spanner, there are coefficients λ x,i for all x ∈ P satisfying
x = ∑ λ x,i vi ,
i
|λ x,i | ≤ C.
In particular, with α x := ∑ni=1 |λ x,i | ≤ Cn by Jensen’s inequality
xx⊺
n
n
i =1
i =1
∑ αx |λx,i |vi v⊺i C2 n ∑ vi v⊺i .
Hence E x ∼µ [ xx⊺ ] C2 n ∑ni=1 vi v⊺i = C2 n2 E x ∼µv1,...,vn [ xx⊺ ], from which the claim follows.
17
| 8cs.DS
|
iDroop: A dynamic droop controller to decouple power grid’s
steady-state and dynamic performance
arXiv:1612.05804v1 [math.OC] 17 Dec 2016
Enrique Mallada
Abstract— This paper presents a novel Dynam-i-c Droop
(iDroop) control mechanism to perform primary frequency
control with gird-connected inverters that improves the network dynamic performance. The work is motivated by the
dynamic degradation experienced by the power grid due to
the increase in asynchronous inverted-based generation. We
show that the widely suggested virtual inertia solution suffers
from unbounded noise amplification (infinite H2 norm) when
measurement noise is considered. This suggests that virtual
inertia could potentially further degrade the grid performance
once broadly deployed.
This motivates the proposed solution in this paper that
overcomes the limitations of virtual inertia controllers while
sharing the same advantages of traditional droop control. In
particular, our iDroop controllers are decentralized, rebalance
supply and demand, and provide power sharing. Furthermore,
our solution can improve the dynamic performance without
affecting the steady state solution. Our algorithm can be incrementally deployed and can be guaranteed to be stable using
a decentralized sufficient stability condition on the parameter
values. We illustrate several features of our solution using
numerical simulations.
I. I NTRODUCTION
Droop control has a long history in power system frequency control [1]. It is perhaps one of the simplest – and
yet very effective – decentralized mechanisms that achieve
synchronization and supply-demand balance in a network
with several generating resources running in parallel [2].
While its implementation may vary depending on the specific
device, its basic operational principle remains unchanged:
whenever frequency is above (below) the nominal value,
decrease (increase) power proportionally to the frequency
deviation. Thus it is usually referred as the primary layer of
the frequency control architecture [3]. Not surprisingly, its
many benefits have made droop control one of the features
of power system engineering that have successfully survived
decades of technological advances.
Unfortunately, the very principles that this mechanism
relies on are becoming less and less valid due to several reasons. Firstly, droop control relies on the fact that
demand is frequency dependent, yet with the increase of
power electronic based loads the aggregate load is becoming
less sensitive to frequency [4]. Secondly, while traditional
generation always provides some level of droop control,
renewable generation is insensitive to frequency fluctuations.
The remaining conventional generators are forced to handle
the whole burden of regulating the frequency. Thirdly, newer
types of generators have little or no inertia at all when
compared with traditional ones, which is slowly introducing
a dynamic degradation that concerns many utilities [5].
The author is with the Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218 USA (e-mail:
[email protected]). This work was supported by NSF CPS grant CNS
1544771, Johns Hopkins E2 SHI Seed Grant, and Johns Hopkins WSE
startup funds.
Recently, a large body of literature has been developed on
the design of distributed control mechanisms with the objective to synthetically generate frequency responsive generation
and demand that can overcome the diminishing participation
of these elements in primary frequency control [6]–[9]. In
particular, the proposed strategies seek to introduce a more
frequency responsive devices, either on the load side [6],
[10] or on the generation side [9], that not only implements
droop characteristics for primary frequency control, but also
guarantee higher level operational constraints such as restoring frequency to nominal value [11], preserving inter area
flow constraints, respecting thermal limits, and providing
efficiency [10]. While these solutions directly address the
loss of frequency responsiveness in the grid, they do not
explicitly address dynamic degradation.
Interestingly, droop control can in principle also provide
dynamic performance improvement by properly selecting the
droop coefficients [12]. For example, in an under-damped
power grid, decreasing the droop coefficient can reduce the
frequency Nadir (maximum frequency excursion) in the same
way increasing friction can reduce the overshoot of an underdamped spring-mass system. However this is not a practical
solution since it also requires the generator to take a larger
share of the supply-demand imbalance. As a result, the
efficiency of the steady-state resource allocation becomes
intrinsically coupled with the possible dynamic performance
improvement. It is the purpose of this work to eliminate this
coupling.
This paper proposes a novel dynam-i-c droop (iDroop)
control algorithm that is able to maintain all the desired
features of traditional droop control while providing enough
design flexibility to improve the dynamic performance. More
precisely, we present a control scheme –whose input is
frequency and output is power generation– that preserves the
same steady-state characteristics as traditional droop control,
maintains grid stability, and provides supply-demand balance. Our iDroop controller can be implemented by inverters
providing power from renewable sources or by intelligent
loads. Finally, we numerical demonstrate that one can use the
additional flexibility to improve the H2 -norm of the system.
Paper Organization: Section II describes the power network model as well as several inverter operational modes
used to interface with the grid. Section III uses dynamic and
steady state performance metrics to motivate the need for
a novel droop control solution. Section IV introduces the
proposed iDroop control and shows how our solution is able
to preserve the same steady state as droop control while providing enough flexibility to improve dynamic performance.
Moreover, we provide a decentralized sufficient condition
on the parameter values that guarantee the stability of our
controllers. We numerically illustrate the functionalities of
iDroop in Section V and conclude in Section VI.
II. N ETWORK M ODEL
We consider a power system composed by n buses denoted
by either i or j, i.e. i, j ∈ N := {1, . . . , n}. We use ij
to denote the transmission line that connects bus i and j,
and use E to denote the set of lines, i.e. ij ∈ E. Thus
the topology of the power system is described by the graph
G = (N, E). The admittance of the line ij is given by yij =
gij − ibij , where gij and bij denote the conductance and
susceptance of line ij, respectively. The state of the network
is described by the complex voltages (Vi )i∈N = (vi ejθi )i∈N
where vi and θi represent the voltage magnitude and phase
of bus i, respectively.
where Miv represents the virtual inertia.
C. Power Flow Model
We consider a Kron-reduced network model [21], where
constant impedance loads are implicitly included in the line
impedances of the reduced network. Thus every remaining
bus represents a grid generator. We further use the DC
network model which has been widely adopted for purpose
of designing frequency controllers for a long time [22], [23].
Therefore, the total electric power drained by the network
at bus i is
X
Pie = −
bij (θi − θj ),
(6)
j∈Ni
A. Generators and Loads
We model the dynamics of each conventional generator
using the standard swing equations [13], [14]. We denote the
frequency of each generator by ωi which evolves according
to
θ̇i = ωi ,
(1a)
1
Mi ω̇i = pin + qir − (Di + g )ωi − Pie ,
Ri
B. Inverters
In this paper we seek to develop a new control scheme
that can be implemented by inverters to improve the dynamic
performance of the power grid. Since the power electronics
of the inverters are significantly faster than the electromechanical dynamics of the generators, we assume that inverters
can statically update its power, i.e.
θ̇ = ω
(7a)
M ω̇ = pin + q r − (Rg −1 + D)ω − LB θ
(7b)
where M := diag(Mi , i ∈ N ), D := diag(Di , i ∈ N ),
Rg −1 = diag( R1g , i ∈ N ), ω := (ωi , i ∈ N ), pin =
i
(pin
i , i ∈ N ), LB is the bij -weighted Laplacian matrix
,
if i 6= j, ij ∈ E,
−b
P ij
(8)
(LB )ij =
b
,
if i = j,
ik
k∈Ni
0,
otherwise,
and q r := (qir , i ∈ N ) is given by
r,0
if i ∈ CP ,
qir,0 ,
1
r
if i ∈ DC,
qi = qi − Rir ωi ,
q r,0 − 1 ω − M v ω̇ , if i ∈ V I.
i i
i
Rr i
(9)
i
(2)
where ui is the command input. We assume that (2) represents the aggregate power of all the inverters connected at
bus i.
There are different operational modes in which inverters
can be interfaced with the power grid [15]–[18]. Here we
briefly review the most common ones:
Constant Power: This is the default operational mode in
today’s grid and amounts to setting
ui = qir,0 ,
D. Network Dynamics
Combining (1)-(6) we arrive to the following compact
description of the system dynamics
(1b)
where Mi denotes the aggregate generator inertia, Di is the
aggregate damping and frequency dependent load coefficient,
and Rig is the droop coefficient. pin
i denotes the net constant
power injection at bus i, qir is the controllable input power
injected by grid-connected inverters, and Pie denotes the net
electric power drawn by the grid.
qir = ui ,
where the set Ni denotes the set of neighboring buses
adjacent to bus i.
(3)
qir,0
where
is a constant parameter representing power generation set point.
Droop Control: This mode aims to change the power injection of the inverter to provide additional droop capabilities
by setting
1
(4)
ui = qir,0 − r ωi ,
Ri
where Rir is the droop coefficient.
Virtual Inertia: This operational mode has been recently
proposed [19], [20] as an alternative method to compensate
the loss of inertia and is given by
1
ui = qir,0 − r ωi − Miv ω̇i ,
(5)
Ri
The sets CP , DC and V I are the subsets of buses that
have inverters operating in constant power, droop control and
virtual inertia modes, respectively.
In the absence of higher layer controllers, such as automatic generation control [24], the system can synchronize
with a nontrivial frequency deviation from the nominal (ωi∗ 6=
0).1 Thus we refer to the vector (θ∗ (t) := θ∗ + ω ∗ t, ω ∗ , q r ∗ )
as a steady state solution of the system (7), with (7b) equal
d ∗
θ (t) = ω ∗ . Furthermore, using (7b) it is easy to
zero and dt
see that LB θ∗ (t) is constant, which implies that ω ∗ = 1n ω0∗ ,
with 1n ∈ Rn being the vector of all ones, and the scalar
ω0∗ given by
Pn
r,0
in
i=1 pi + qi
∗
ω0 = Pn
.
(10)
P
g −1
) + i∈DC∪V I Rir −1
i=1 (Di + Ri
In summary, the steady state solution of (7) and (9) is given
by (θ∗ (t), ω ∗ , q r ∗ ) = (θ∗ + 1n ω0∗ t, q r∗ ), where qir∗ = qir,0 , if
i ∈ CP , or qir∗ = qir,0 − Rir −1 ω0∗ , if i ∈ DC ∪ V I. Thus the
smallest invariant set that includes all possible steady states
is given by
E := {(θ+λ1n , ω, q r ) ∈ R3n: λ ∈ R, (7b) = 0, ωi = ω0∗ }.
(11)
1 We
assume here w.l.o.g. that the nominal frequency is 0
III. S TEADY S TATE AND DYNAMIC P ERFORMANCE
In this section we introduce two metrics, one for steady
state and another for dynamic performance, and illustrate
how the existing solutions for inverter control either cannot
produce any dynamic performance improvement, or the
performance improvement comes at the cost of steady state
deviation from the desired operational point.
∗
δqˆr is an optimal solution if and only if there exists a scalar
λ∗ such that
∗
αig δqˆig = λ∗ ∀i,
g∗
∗
αir δqˆir = λ∗ i ∈ DC ∪ V I,
(15)
∗
and i=1 δqˆi + δqˆir = ∆P .
To finalize the proof we just need to show that when Rir =
r
αi and Rig = αig , then the steady state deviations of droop
∗
control (δqig := − R1g ω0∗ and δqir ∗ := qir ∗ − qir,0 ) satisfy the
A. Steady State Performance
i
KKT conditions.
We now define the steady state performance metric used
® 1 ∗
− Rr ω0 , if i ∈ DC ∪ V I,
in this paper. Let δqig := − R1g ωi and δqir := qir − qir,0 be
r,0
∗
∗
r
r
i
i
By (9), δqi := qi − qi =
the power deviation of generators and inverters, respectively.
0,
otherwise.
After a disturbance these quantities need to be modified in
order to compensate a supply-demand imbalance, i.e.
Therefore, since by definition δq g ∗ = − R1g ω0∗ , then we have
i
n
that when Rir = αir and Rig = αig , (δq g ∗ , δq r ∗ ) satisfies (15)
X
X
δqir = ∆P,
δqig +
(12) for λ∗ = ω0∗ .
Finally, feasibility follows by definition of ∆P since
i=1
i∈DC∪V I
where ∆P denotes the power imbalance. We assign to both
conventional
and inverter-based generators a cost cgi (qig ) :=
g
αi
αri
g 2
r 2
r r
2 (qi ) and ci (qi ) := 2 (qi ) respectively. Thus given
a set deviations satisfying (12) the total system cost of
mitigating an imbalance of ∆P is given by
SS-Cost:
n
X
αg
i=1
X αr
i
i
(δqig )2 +
(δqir )2
2
2
(13)
i∈DC∪V I
One of the main attractive features of droop control is its
ability to share the supply-demand mismatch among different
resources. Recent works have shown that the steady state
allocation of droop control can be represented as the solution
of an optimization problem (see e.g. [6]).
In particular, if we define let δdi =: Di ωi denote the
frequency dependent demand deviation, then it can be shown
that the system (7) solves:
minimize
g
r
δqi ,δqi ,δdi
n
X
Rg (δq g )2
i
i=1
i
2
+
X Rr (δq r )2
(δdi )2
i
i
+
2Di
2
i∈DC∪V I
(14a)
subject to
n
X
i=1
r,0
g
pin
i +qi +δqi −δdi +
X
δqi = 0 (14b)
Pn
n
X
n
X
∗
−1
δqig + δqir ∗ = −(
Rig +
i=1
i=1
=
n
X
X
Rir −1 )ω0∗
i∈DC∪V I
r∗
∗
pin
i + qi − Di ω0 =: ∆P
i=1
∗
where the second steps follows from (10). Thus (δqig , δqir ∗ )
is feasible and satisfies the KKT conditions. Therefore,
∗
(δqig , δqir ∗ ) is an optimal allocation. Since the optimal so∗
∗
∗
lution is unique, we must have (δqig , δqir ∗ ) = (δqˆig , δqˆir ).
Remark 1: Theorem 1 illustrates the versatility of droop
control and how it can be used to optimally accommodate
supply-demand imbalances. However, it also highlights the
need to tune parameters that have a direct effect on the
network dynamics in order to achieve optimal steady state
performance.
Remark 2: The use of quadratic costs is standard in the
power system literature [29]. However, since this paper
focuses mostly on local analysis, the analysis can be extended to include nonlinear costs by substituting αiz with
∂ z
z∗
z ∗ where δq
denotes the equilibrium value
i
∂x ci (x)|x=δqi
of qiz (z refers to either g or r).
i∈DC∪V I
The proof of this claim is already standard in the community
(see e.g. [6], [7], [25]–[27]), we refer the reader to [6] for
a proof of a similar statement.
As a result of the steady state characteristic of droop
control, it is possible to optimally minimize the steady state
cost as it is summarized in the next theorem.
Theorem 1 (Droop Control Optimality): Let (θ∗ (t), ω ∗ ,
q r ∗ ) be the steady state solution of (7) where qir ∗ is given
∗
by (9). If Rir = αir and Rig = αig , then (δqig := − R1g ω0∗ ,
i
δqir ∗ := qir ∗ − qir,0 = − R1g ω0∗ ) is the unique allocation that
i
minimizes
state cost (13) subject to (12), where
Pnthe steady
r,0
in
∆P := i=1 pi + qi − Di ω0∗ .
Proof: We start by characterizing the optimal solution
of minimizing (13) subject to (12). Since the problem has a
strictly convex objective with linear constraints then there is
a unique solution that is characterized by the Karush-Kuhn∗
Tucker (KKT) conditions for optimality [28]. Thus, δqˆg and
B. Dynamic Performance
We now focus our attention on the dynamic performance
metric. In general, there are several metrics that can be
considered [30], and the change of droop control coefficients
can have direct effect in all of them. Here we focus H2 norm
of the system when the output is the frequency vector ω and
the system is being driven by stochastic white noise.
The main motivation of this particular choice of metric
is the need to characterize the effect of the increasing
generation volatility as well as the effect of the measurement
noise. We argue that while generator-based droop control can
be usually implemented without major measurement noise,
implementing droop control on inverters that are distributed
all over the transmission and distribution network needs to
account for these errors. Thus the use of H2 -norm seems
natural as it already accounts for stochastic inputs.
We assume that the net power injection of each bus is
given by Piin + ki1 wi1 (t), where w1 (t) = (wi1 (t), i ∈ N ) is a
vector of uncorrelated stochastic white noise with unit variance (E[w1T (τ )w1 (t)] = δ(t − τ )In ) that represents demand
fluctuations. Similarly, we assume that for the purposes of
implementing the droop control on the inverter the measured
frequency is given by ωi (t) + ki2 wi2 (t) where again w2 (t) =
(wi2 (t), i ∈ N ) is such that E[w2T (τ )w2 (t)] = δ(t − τ )In .
Finally, since to estimate ω̇i (t) one needs to obtain first
ωi , we assume that the measurement value of the frequency
d 2
wi , i ∈ N ).2
derivative is ω̇i + ki3 wi3 , where w3 = (wi3 := dt
Without loss of generality we make the following change
of variables
d
and w3 are not correlated processes (w3 6= dt
w2 ), one very
useful procedure to compute ||H||H2 (see [31]) is based on
using
||H||2H2 = tr(B T XB)
(22)
where X is the observability Grammian, i.e. X solves the
Lyapunov equation
AT X + XA = −C T C.
(23)
In the context of power systems the use of this methodology has been first used in [32], where the authors seek to
δθ(t) = θ(t) −
and δω(t) = ω(t) −
(16) compute the power losses incurred by the network in the proThus, defining the system output to be y(t) = δω(t) we can cess of resynchronizing generators after a disturbance. Since
combine (7) together with (9) to get the following MIMO then, several works have used similar metrics to evaluate
effect of controllers on the power system performance, see
system
e.g. [33], [34].
ï
òï ò ï
òï ò ï
ò"w1 #
˙
Remark 3: It is important to notice that in our case the
I 0 δθ
0 I δθ
0
0
0
w2 system is driven also by the derivative of the noise process
=
+
r −1
v
˙
δω
K1 −R
K2 −M K3 w
0 M̂ δω
−L −D̂
3
w2 . Thus, (22) can only be applied when K3 = 0. When
(17a) K3 6= 0, then (21) corresponds to a frequency weighted H2 ï ò
norm. More precisely, then noise process w3 is a frequency
δθ
y = [0 I]
(17b)
weighted process with weight function given by W (s) = sI,
δω
i.e. ŵ3 (s) = sŵ2 (s) where ŵk (s) is the Laplace Transform
where M v = diag(Miv ), Kx = diag(kix ) and wx = (wix ), of wk (t).
with x ∈ {1, 2, 3}, M̂ = M + M v , D̂ = D + R̂−1 and
The next theorem shows that droop controlled invertR̂−1 = Rg −1 + Rr −1 .
ers can indeed affect the performance by changing droop
Notice that for simplicity the system (17) implicitly as- parameters, and that inverters implementing virtual inertia
sumes that all the inverters in the system are operated using can drastically degrade the performance when measurement
the virtual inertia operation mode. However, it is possible noise is considered.
to model using (17) the droop controlled mode by setting
Theorem 2 (H2 -norm Computation): Assume
Miv = ki3 = 0. Moreover, if one wants to model the constant homogeneous parameter values, i.e. Mi = m, M v = mv ,
i
power mode one just needs to additionally set R1r = 0. To Di = d, Rg = rg , Rr = rr , k 1 = k1 , k 2 = k2 and k 3 = k3 .
i
i
i
i
i
i
simplify the discussion we will only consider two cases: 1) Let
All the inverters implement droop control (DC = N ); 2) All
r̂−1 = rg −1 + rr −1 , dˆ = d + r̂−1 ,
(24)
the inverters implement the virtual inertia control (V I = N ).
It will also be useful to write (17) in standard form
and let HDC and HV I denote the MIMO system (17) when
" #
all inverters implement droop control and virtual inertia,
ï ò
ï ò
ï
ò
w1
˙
δθ
δθ
δθ
respectively.
w
=A
+B
,
y
=
C
,
(18)
2
˙
δω
δω
δω
Then the squared H2 norm of HDC and HV I is given by
w3
where
n (k1 )2 + (k2 rr −1 )2
2
ï
ò
,
(25)
||HDC ||H2 =
0
I
2m(d + rg −1 + rr −1 )
A=
, C = [0 I] , and
(19)
−1
−1
−M̂ L −M̂ D̂
ò
ï
and
0
0
0
B=
.
(20)
−1
−1 −1
−1
v
M̂ K1 −M̂ R̂ K2 −M̂ M K3
||HV I ||2H2 = +∞,
(26)
Let H denote the LTI system (17). Then the square of the
H2 norm of (17) can be formally defined using
respectively.
Proof: The proof of this theorem is analogous to [32,
Dyn-Cost:
||H||2H2 = lim E[y T (t)y(t)]
(21) Lemmas 3.1 and 3.2]. We study the two cases (25) and (26)
t→∞
separately.
where y(t) is the output of (17) when the input w(t) =
Computing ||HDC ||2H2 : Notice that in this case M v = 0. In
(w1 (t), w2 (t), w3 (t)) is composed by a white noise process
2
with unit covariance (i.e., E[wk (τ )wkT (t)] = δ(t − τ )I for order to compute ||HDC ||H2 we first make the same change
of variable used in [32],
d
w2 ).
k ∈ {1, 2}) and its derivative (w3 = dt
The computation of the H2 norm has been widely studied
δθ = U θ0 and δω = U ω 0 ,
in modern control theory. In particular, in the case when w2
where U is the orthonormal transformation that diagonalizes
2 We use here the notation d w 2 to represent the frequency weighted noise
dt i
L, i.e. U T LB U = Γ where Γ = diag{λ1 = 0, . . . , λn } and
process with weight function given by wi (i2πf ) = i2πf , see Remark 3
U is assumed w.l.o.g. to be of the form U = [ √1n 1n | Un−1 ].
for more details.
ω0∗ t,
ω0∗ .
Thus, if we further transform y = U y 0 , w1 = U w10 and
w2 = U w20 , we can decouple (18) into n subsystems given
by
ò ñ 10 ô
ï 0ò ï
òï 0ò ï
0
1
0
0
θ
wi
θ̇i
= λi
0
0 + k1
k2
dˆ
ω
−
ω̇i0
−
−
r
wi2 ,
m
mr
m ï mò
HDC,i :
θ0
yi0 = [0 1] i0
ωi
(27)
where a simple computation using (22) and (23) shows that
(k1 )2 + (k2 rr −1 )2
||HDC,i ||2H2 =
.
2m(d + rg −1 + rr −1 )
P
Therefore, since ||HDC ||2H2 = ni=1 ||HDC,i ||2H2 we obtain
(25).
Computing ||HV I ||2H2 : To show that the norm ||HV I ||2H2 =
+∞ we will show that the transfer function of HV I has
nonzero feedthrough (σmax (H(i∞)) ≥ ε > 0).
To compute the transfer function we first notice that w3 =
ẇ2 which implies that we can model in the Laplace domain
ŵ3 (s) = sŵ2 (s).
Similar to the DC case, we can use a change of variable
to decouple the system into n different modes given by
0
ï 0ò
ï 0ò
ï 0ò
wi1
θ
θ
20
θ̇i
0
=Ai 0 +Bi wi ,
yi = Ci i0
ω
ω
ω̇i0
0
i
wi3
ò
ï
ò
ï
0
1
0
0
0
v
,
B
=
with
Ai =
ˆ
i
k1
k2
k3 m
d
− λm̂i − m̂
m̂ − m̂r r − m̂
Ci = [0 1]
where m̂ = m + mv .
We drop the subscript i and define ŵ0 (s) =
[ŵ10 (s) ŵ20 (s) ŵ30 (s)]T . Thus we can compute the transfer
function H(s) using
ŷ(s) = H(s)w0 (s) = C(sI − A)−1 B ŵ0 (s)
ô
ñ
dˆ
1
1
s + m̂
B ŵ0 (s)
C
=
λ
∆(s)
− m̂
s
1 λ
=
− m̂ s B ŵ0 (s)
∆(s)
0
s k1
k2
mv
=
ŵ (s)
− m̂r
− k3m̂
r
m̂
∆(s)
" 0 #
ŵ1 (s)
s k1
v
k2
k3 m
ŵ20 (s)
=
−
−
r
m̂r
m̂
∆(s) m̂
ŵ30 (s)
î
Ä
äó ïŵ0 (s)ò
s
k1
k3 mv
k2
1
=
−
s
+
m̂
m̂
m̂r r
ˆ
ŵ20 (s)
s2 + d s + λ
m̂
m̂
ˆ
d
λ
where ∆(s) = s2 + m̂
s + m̂
and in the last step we used the
0
0
relationship ŵ3 (s) = sŵ2 (s). It follows that when f → +∞,
mv
H(i2πf ) → 0 − k3m̂
,
which implies that ||HV I ||2H2 = +∞.
Theorem 2 provides an interesting insight on how the
different controllers described in Section II affect the system
performance and illustrates the effect of measurement errors
on it. To understand the performance changes, we provide
the H2 norm of the swing equations (HSW ) without any
additional control (rr = mv = k2 = k3 = 0)
n
||HSW ||2H2 =
(k1 )2 .
(28)
2m(d + rg −1 )
Thus it is easy to see that adding droop control introduces
a larger damping (d + rg −1 + rr −1 ) > d + rg −1 . However,
there is an intrinsic tradeoff since if rr is small enough, then
the noise amplification takes over and increases the norm.
But perhaps the most interesting result from this theorem is
(26), which implies that the massive use of virtual inertia in
the network can amplify the stochasticity of the system and
introduce huge volatility in the frequency fluctuations.
C. The Need for a Better Solution
The analysis provided in sections III-A and III-B shows
that none of the existing solutions can simultaneously achieve
efficient steady state operation while improving dynamic
performance.
On the one hand, while Theorem 1 shows indeed that
droop control can be used to optimally allocate resources
by carefully setting the parameters Rir = αir and Rig = αig ,
this tie between control parameters and economic efficiency
makes it impossible to further improve the dynamic performance without incurring on an additional steady state cost
(13). Therefore, if one wants to operate the system in an
efficient steady state, then Rir cannot be used to improve the
dynamic performance.
On the other hand, while the use of virtual inertia has been
suggested to be a viable solution to improve the dynamic
performance without losing steady state efficiency, Theorem
2 shows that this solution cannot be widely adopted since
it will induce a large noise amplification that can hinder the
secure operation of the power grid.
As a result, all the existing solutions for inverter control
cannot provide a dynamic performance improvement without
sacrificing either steady-state or dynamic performance.
IV. DYNAM - I - C D ROOP C ONTROL ( I D ROOP )
We now introduce our iDroop control. The main underlying idea on the design of the proposed controller is to
leverage the flexibility that virtual inertia controllers provide
while controlling the noise amplification using a filtering
stage.
Let ID ⊂ N be the set of network buses that implement
iDroop. Then we propose to control the power injected to
the grid by the inverter using
iDroop:
qi = qir,0
ä
Ä + xi
ẋi = δi − R1r ωi − xi − νi ω̇i
(29)
i
A few comments are in order. Firstly, the parameter νi
place a role similar to Miv in the virtual inertia controller.
However, since this term is integrated in (29), it is not longer
interpreted as a virtual inertia. Secondly, the parameters δi
and νi can be independently tuned to reduce the noise introduced by the frequency (ki2 wi2 ) or the frequency derivative
(ki3 wi3 = ki3 ẇi2 ). This can be easily seen when these noise
processes are introduced in (29) giving
Å
ã
1
δi k 2
ẋi = δi − r ωi − xi − νi ω̇i − ri wi2 − νi ki3 ẇi2 . (30)
Ri
Ri
Finally, while the stability (in the absence of noise) of (7)
with (9) is trivially guaranteed, the additional integration
stage requires a more detailed analysis.
We do not provide here an explicit formula for
||HiDroop ||H2 in this paper and leave it for future research.
Instead we will numerically illustrate in the next section the
effect of νi and δi on this norm and how the performance of
iDroop compares with the one of droop controlled inverters.
In the rest of this section we show that indeed our controllers
are able preserve the steady state behavior for arbitrary
parameter values δi and νi , and characterize a sufficient
condition for asymptotic stability.
A. Steady State Optimality
We now show that our iDroop controllers provide the same
steady state properties as traditional droop control. We do
this by showing that iDroop achieves the minimum SS-Cost
(13).
Theorem 3 (iDroop Optimality): Let (θ∗ (t), ω ∗, q r ∗ ) be
the steady state solution of (7) where qir is given by (29).
Then the steady state solution of (7) and (29) is given by
ω∗
θ(t)∗= θ∗ +1n ω0∗ t, ωi∗ = ω0∗ ∀i and qi∗ −qir,0 = x∗i = − 0r .
Ri
Moreover, if
Rir
=
αir
and
Rig
=
αig ,
then
∗
δqig
:=
− R1g ω0∗ ,
i
δqir ∗ := qir ∗ − qir,0 is the unique allocation that minimizes
the
state cost (13) subject to (12), where ∆P =
Pn steady
in
p
+
qir,0 − Di ω0∗ .
i=1 i
Proof: The proof of this theorem relies on Theorem 1.
We will show that iDroop has a steady state behavior that
is identical to the standard droop control when DC ∪ V I =
ID. Thus, we can use then use Theorem 1 to show that
indeed iDroop preserves the optimality characteristics of the
traditional droop control.
We now characterize the steady state behavior of (7) and
(29). Similarly to (7) and (9), we can use (7b) to show that
ω̇ = 0 only if θ(t) = θ∗ (t) := θ∗ + 1n ω0∗ t. Using (29)
with ẋ = 0 and ω̇ = 0, we get that Rir x∗i = −ωi∗ = −ω0∗ .
Therefore, the steady state deviation of the inverter i is given
by
1
δqir ∗ = qi∗ − qir,0 = x∗i = − r ω0∗ .
Ri
Since the steady state deviation of conventional generators
∗
is by definition δqig = − R1g ω0∗ , then we have obtained the
i
same allocation described in Theorem 1. It follows then that
g
when αir = Rir and αi = Rig , the steady state solution of
(29) is an allocation that minimizes (13) subject to (12).
B. Stability Analysis
We now show that the controllers described in (29) can
preserve the stability of the network.
We use the same change of variable used in (16) together
with δx(t) = x(t) − x∗ . Therefore (7) and (29) become
˙ = δω
δθ
(31a)
˙ = −(D + Rg −1 )δω − LB δθ + δx
M δω
˙ = −Kδ (Rr −1 δω + δx) − Kν δω
˙
δx
where Kν = diag(νi , i ∈ N ).
3
(31b)
(31c)
It is easy to see
3 To simplify notation we assume here that every bus iDroop. However, the
results can be generalized for any combination of iDroop with the controllers
in (9).
that now the steady state solutions of Theorem 3 become
(δθ∗ , δω ∗ , δx∗ ) = (1n α, 0, 0) for any α ∈ R. Thus the set
of equilibria of (31) is given by
Ê = {(δθ, δω, δx) : δω = δx = 0, δθ = 1n α, α ∈ R}.
(32)
Theorem 4 (Asymptotic Convergence): Whenever the following condition holds,
νi
νi Rir −1
g −1
)
+
>
0
and
(D
+R
> 0,
i
i
δi (νi +Rir −1 )
νi + Rir −1
(33)
the iDroop control (29) converges asymptotically to the set
of equilibria Ê described in (32).
Proof: We will show that under the conditions of the
theorem, the following Lyapunov function decreases along
trajectories:
1
1
V (δθ, δω, δx) = δθTLB δθ + δω T M δω
2
2
(34)
1
+ (δx + Kν δω)T T (δx + Kν δω).
2
where T ∈ Rn×n is a positive definite diagonal matrix to
be defined later (T 0 and T = diag(v), with v ∈ Rn ).
The Lyapunov function V is inspired on [35] where a similar
derivative term is used to damp oscillations.
The change of V along the trajectories of (31) is given by
T
T
˙
˙
˙
˙
V̇ = δθT LB δθ+δω
M δω+(δx+K
ν δω) T (δx+Kν δω)
(35)
= δθT LB δω + δω T (−(D + Rg −1 )δω − LB δθ + δx)
+ (δx + Kν δω)T T Kδ (−Rr −1 δω − δx)
(36)
= − δω T (D+Rg −1 +Kν T Kδ Rr −1 ) δω−δxT T Kν δx
+ δω T (I − Kν T Kδ −T Kδ Rr −1 )δx
(37)
where (36) follows from (31), and (37) from rearranging the
terms.
We now choose T so that the cross term in (37) is zero.
Since T is assumed to be a diagonal matrix, then we get
0 = I − Kν T Kδ −T Kδ Rr −1
0 = I − T Kδ (Kν + R
T = Kδ
−1
(Kν + R
r −1
)
r −1 −1
)
⇐⇒
⇐⇒
(38)
Therefore, using (38), it follows that
V̇ = − δω T (D+Rg −1+Kν Rr −1 (Kν +Rr −1 )−1 )δω
− δxT Kν Kδ−1 (Kν +Rr −1 )−1 δx
(39)
Thus, it follows from (33) that V̇ ≤ 0 and we can
now apply LaSalle’s Invariance Principle [36] to show that
(δθ(t), δω(t), δx(t)) converges to the largest invariant set
M ⊂ {V̇ ≡ 0}. Using (39) we can show that V̇ ≡ 0 implies
that δω(t) ≡ 0 and δx(t) ≡ 0, which implies in turn (through
˙ ≡ 0. Therefore, it must follow that δθ(t) ≡
(31a)) that δθ
δθ∗ . Finally, using (31b) we get 0 ≡ LB δθ(t) ≡ LB δθ∗
which, since the graph is connected, implies that θ∗ = 1n α
for some α ∈ R. Therefore, every trajectory converges to
the set of equilibria Ê given in (32), i.e. M ⊂ Ê.
we introduce a step change ∆P30 = −0.5p.u. in the power
injection at bus 30 (where generator 10 is located). Fig. 2
shows the evolution of the frequency deviations for the four
cases: (a) CP, (b) DC, (c) VI and (d) iDroop. It can be seen
that both VI (Fig. 2-c) and iDroop (Fig. 2-d) can reduce the
Nadir (minimum frequency achieved), while achieving the
the same steady state as the DC mode (Fig. 2-b).
9
37
10
26
30
25
29
28
2
38
27
18
1
8
24
17
16
3
15
1
6
35
14
4
12
6
19
7
31
23
13
11
2
20
10
32
9
(b) Droop Control
0.3
0.2
0.2
0.1
0.1
0
0
-0.1
-0.1
36
34
8
(a) Constant Power
0.3
22
δqir p.u.
5
21
7
33
5
0
4
5
10
15
(c) Virtual Inertia
0.3
0
5
Fig. 1: IEEE 39 Bus System (New England)
0.2
0.2
0.1
0.1
0
0
-0.1
-0.1
0
5
10
15
0
5
t
V. N UMERICAL I LLUSTRATIONS
10
15
iDroop
Droop Control
60
||HDC ||2H2
60
40
20
0
40
20
0
8
0.2
6
0.15
4
δ
(a) Constant Power
t
However, the way this is achieved is in the case of VI
and iDroop is completely different. This can be seen in Fig.
3 where we show the power deviation experienced by each
individual inverter. In particular, the step change in the power
injection introduces a discontinuity in ω̇30 . This is shown
in Fig. 3-c where the inverter at bus 30 has a step change
in power. iDroop on the other hand is able to perform a
similar task using less peak power and with a more desirable
behavior. This drastic change on the power experienced by
the VI also illustrates why it is expected that this solution
will produce noise amplification.
0.1
2
ωi rad/s
15
Fig. 3: Power deviation of inverter
||HiDroop ||2H2
In this section we numerically illustrate some of the
features of our iDroop control using the IEEE 39 bus
system shown in Fig. 1. The network parameters as well as
the stationary starting point were obtained from the Power
System Toolbox (PST) [37] dataset.
Before building the dynamic model, we perform the Kron
reduction for every load bus. Thus the final system has 10
buses that correspond to each generator in the nework. The
inertia parameters of each generators are also obtained from
the PST dataset. Throughout the time domain simulations
we assume that the aggregate generator damping and load
frequency sensitivity parameter is Di = 0.1. We also set the
droop coefficient of each generator to Rig = 15.
On each bus we add an additional inverter-based generator
and vary their operational mode using either one of the
three modes described in (9) (CP=constant power, DC=droop
control and VI=virtual inertia), or the iDroop mode. The
droop coefficient is set to Rir = 15 in all the cases. For
illustration purposes we select parameters so that the VI
mode and iDroop have a similar frequency transient behavior.
In particular, we choose Miv = 0.15, δi = 6, νi = 0.9.
8
0.2
6
0.15
4
0.1
2
0.05
ν
δ
0.05
ν
(b) Droop Control
0
0
-0.2
-0.2
-0.4
-0.4
-0.6
-0.6
-0.8
-0.8
-1
Fig. 4: H2 -norm comparison between iDroop and Droop
Control
-1
0
5
10
15
0
5
(c) Synthetic Inertia
ωi rad/s
10
(d) iDroop
0.3
3
δqir p.u.
39
10
15
10
15
(d) iDroop
0
0
-0.2
-0.2
-0.4
-0.4
-0.6
-0.6
-0.8
-0.8
-1
-1
0
5
10
t
15
0
5
t
Fig. 5: Frequency fluctuations when δi = 6 and νi = 0.01
Fig. 2: Frequency deviations after perturbation
We initialize the system in steady state and at time t = 1
Finally, since the VI mode has unbounded H2 -norm,
we show in Fig. 4 the H2 -norm of the system when the
inverters are using only the DC mode and iDroop on a high
measurement noise regime (ki1 = 0.1, ki2 = ki3 = 5). We can
see that DC does not vary with δ and ν (as expected) while
iDroop does. However, more importantly, iDroop is able to
achieve better performance than DC despite using frequency
derivative measurements. This is also illustrated in Figure 5
where we simulate the frequency fluctuations for the values
of δi = 6 and νi = 0.01.
VI. C ONCLUDING R EMARKS
This paper studies the intrinsic trade off between steady
state and dynamic performance of inverter-based droop control and several of its variants. We show that while the
standard droop control can improve the dynamic performance, it can only achieve it by losing steady state efficiency.
Moreover, our analysis also shows that the popular alternative
of adding virtual inertia is very sensitive to measurement
noise and can increase the frequency variance unboundedly.
To solve these issues we propose a new control scheme
(iDroop) that is able to tune dynamic performance without
altering steady state efficiency. We characterize the set of
parameter values that can guarantee the stability of the
steady state solution and illustrate its behavior numerically.
In particular, we show that iDroop is able to reduce the Nadir
and variance when compared with Constant Power (CP) and
Droop Control (DC) modes.
R EFERENCES
[1] H. Ziebolz and A. R. Co, “Control means for power generating
systems,” Patent, Dec., 1947.
[2] P. Almeras and N. B. P. Picte, “Regulation of an assembly of electric
generating units,” Jul. 1951.
[3] J. Machowski, J. Bialek, and D. J. Bumby, Power System Dynamics,
ser. Stability and Control. John Wiley & Sons, Aug. 2011.
[4] A. J. Wood, B. F. Wollenberg, and G. B. Sheble, “Power generation,
operation and control,” John Wiley&Sons, 1996.
[5] B. J. Kirby, “Frequency Regulation Basics and Trends,” Oak Ridge
National Laboratory, Tech. Rep., Jan. 2005.
[6] C. Zhao, U. Topcu, N. Li, and S. H. Low, “Design and Stability of
Load-Side Primary Frequency Control in Power Systems,” Automatic
Control, IEEE Transactions on, vol. 59, no. 5, pp. 1177–1189, 2014.
[7] X. Zhang and A. Papachristodoulou, “A real-time control framework
for smart power networks: Design methodology and stability,” Automatica, vol. 58, pp. 43–50, Aug. 2015.
[8] F. Dörfler, J. Simpson-Porco, and F. Bullo, “Breaking the Hierarchy:
Distributed Control & Economic Optimality in Microgrids,” arXiv.org,
Jan. 2014.
[9] N. Li, L. Chen, C. Zhao, and S. H. Low, “Connecting automatic
generation control and economic dispatch from an optimization view,”
in Proceedings of the American Control Conference, Massachusetts
Inst. of Technology, Cambridge, United States. IEEE, Jan. 2014, pp.
735–740.
[10] E. Mallada, C. Zhao, and S. H. Low, “Optimal load-side control for
frequency regulation in smart grids,” arXiv.org, Oct. 2014.
[11] E. Mallada and S. H. Low, “Distributed frequency-preserving optimal
load control,” in IFAC World Congress, 2014, pp. 5411–5418.
[12] A. E. Motter, S. A. Myers, M. Anghel, and T. Nishikawa, “Spontaneous synchrony in power-grid networks,” Nature Physics, vol. 9,
no. 3, pp. 191–197, Feb. 2013.
[13] D. W. C. Shen and J. S. Packer, “Analysis of Hunting Phenomena
in Power Systems by Means of Electrical Analogues,” Proceedings of
the IEE - Part II: Power Engineering, vol. 101, no. 79, pp. 21–34,
Feb. 1954.
[14] A. R. Bergen and D. J. Hill, “A Structure Preserving Model for
Power System Stability Analysis,” Power Apparatus and Systems,
IEEE Transactions on, vol. PAS-100, no. 1, pp. 25–35, 1981.
[15] C. T. Lee, R. P. Jiang, and P. T. Cheng, “A grid synchronization
method for droop-controlled distributed energy resource converters,”
Ieee Transactions on Industry Applications, vol. 49, no. 2, pp. 954–
962, 2013.
[16] M. C. Chandorkar, D. M. Divan, and R. Adapa, “Control of Parallel
Connected Inverters in Standalone Ac Supply-Systems,” Ieee Transactions on Industry Applications, vol. 29, no. 1, pp. 136–143, 1993.
[17] J. W. Simpson-Porco, F. Dörfler, and F. Bulbo, “Synchronization and
power sharing for droop-controlled inverters in islanded microgrids,”
Automatica, vol. 49, no. 9, pp. 2603–2611, Sep. 2013.
[18] J. Liu, Y. Miura, and T. Ise, “Comparison of Dynamic Characteristics Between Virtual Synchronous Generator and Droop Control
in Inverter-Based Distributed Generators,” Power Electronics, IEEE
Transactions on, vol. 31, no. 5, pp. 3600–3611, May 2016.
[19] H. P. Beck and R. Hesse, “Virtual synchronous machine,” in 2007 9th
International Conference on Electrical Power Quality and Utilisation,
EPQU. TU Clausthal, Clausthal-Zellerfeld, Germany, Dec. 2007.
[20] J. Driesen and K. Visscher, “Virtual synchronous generators,” in IEEE
Power and Energy Society 2008 General Meeting: Conversion and
Delivery of Electrical Energy in the 21st Century, PES. IEEE, New
York, United States, Sep. 2008.
[21] P. Varaiya, F. F. Wu, and R.-L. Chen, “Direct Methods for Transient
Stability Analysis of Power Systems: Recent Results,” Proceedings of
the Ieee, vol. 73, no. 12, pp. 1703–1715, Jan. 1985.
[22] G. Quazza, “Automatic control in electric power systems,” Automatica,
vol. 6, no. 1, pp. 123–150, Jan. 1970.
[23] ——, Criteria for equitable participation of areas in tie-line power
and frequency control of an interconnected power system.
Automazione e Strumentazione, 1966.
[24] F. deMello, R. Mills, and W. B’Rells, “Automatic Generation Control
Part II-Digital Control Techniques,” Power Apparatus and Systems,
IEEE Transactions on, vol. PAS-92, no. 2, pp. 716–724, 1973.
[25] C. Zhao, E. Mallada, and S. H. Low, “Distributed generator and
load-side secondary frequency control in power networks,” 2015 49th
Annual Conference on Information Sciences and Systems (CISS), pp.
1–6, 2015.
[26] X. Zhang, N. Li, and A. Papachristodoulou, “Achieving real-time
economic dispatch in power networks via a saddle point design
approach,” in Power & Energy Society General Meeting, 2015 IEEE.
IEEE, 2015, pp. 1–5.
[27] A. Kasis, E. Devane, and I. Lestas, “Primary frequency regulation with
load-side participation: stability and optimality,” arXiv.org, 2016.
[28] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge
University Press, Mar. 2004.
[29] D. S. Kirschen and G. Strbac, Fundamentals of Power System Economics, ser. Kirschen/Power System Economics. Chichester, UK:
John Wiley & Sons, Oct. 2004.
[30] N. W. Miller, M. Shao, and S. Venkataraman, “Frequency Response
Study ,” General Electric, Tech. Rep., Nov. 2011.
[31] J. C. Doyle, K. Glover, P. P. Khargonekar, and B. A. Francis, “Statespace solutions to standard H2 and H∞ control problems,” Automatic
Control, IEEE Transactions on, vol. 34, no. 8, pp. 831–847, Aug.
1989.
[32] E. Tegling, B. Bamieh, and D. F. Gayme, “The Price of Synchrony:
Evaluating the Resistive Losses in Synchronizing Power Networks,”
IEEE Transactions on Control of Network Systems, vol. 2, no. 3, pp.
254–266, 2015.
[33] B. K. Poolla, S. Bolognani, and F. Dörfler, “Optimal Placement of
Virtual Inertia in Power Grids,” Oct. 2015.
[34] E. Tegling, M. Andreasson, J. W. Simpson-Porco, and H. Sandberg,
“Improving performance of droop-controlled microgrids through distributed PI-control,” Jan. 2016.
[35] F. Paganini and E. Mallada, “A Unified Approach to Congestion
Control and Node-Based Multipath Routing,” Networking, IEEE/ACM
Transactions on, vol. 17, no. 5, pp. 1413–1426, 2009.
[36] H. K. Khalil, Nonlinear systems, 3rd ed. Prentice Hall, 2002.
[37] J. H. Chow and K. W. Cheung, “A toolbox for power system dynamics
and control engineering education and research,” Power Systems, IEEE
Transactions on, vol. 7, no. 4, pp. 1559–1564, 1992.
| 3cs.SY
|
Active Orthogonal Matching Pursuit for Sparse Subspace
Clustering
arXiv:1708.04764v1 [cs.LG] 16 Aug 2017
Yanxi Chen, Gen Li, and Yuantao Gu∗
submitted June 20, 2017, accepted August 9, 2017, for publication in
IEEE Signal Processing Letters
Abstract
Sparse Subspace Clustering (SSC) is a state-of-the-art method for clustering highdimensional data points lying in a union of low-dimensional subspaces. However, while
`1 optimization-based SSC algorithms suffer from high computational complexity, other
variants of SSC, such as Orthogonal Matching Pursuit-based SSC (OMP-SSC), lose clustering accuracy in pursuit of improving time efficiency. In this letter, we propose a novel
Active OMP-SSC, which improves clustering accuracy of OMP-SSC by adaptively updating data points and randomly dropping data points in the OMP process, while still
enjoying the low computational complexity of greedy pursuit algorithms. We provide
heuristic analysis of our approach, and explain how these two active steps achieve a
better tradeoff between connectivity and separation. Numerical results on both synthetic data and real-world data validate our analyses and show the advantages of the
proposed active algorithm.
Keywords: Sparse subspace clustering, orthogonal matching pursuit, active algorithm, subspace detection property, connectivity
1
Introduction
In a big-data era, unsupervised learning plays a significant role in analyzing numerous
unlabeled data. In many applications, such as motion segmentation and face clustering,
∗
The authors are with the Department of Electronic Engineering and Tsinghua National Laboratory for
Information Science and Technology (TNList), Tsinghua University, Beijing 100084, China. The corresponding author of this work is Y. Gu (E-mail: [email protected]).
1
high-dimensional data samples are drawn approximately from a union of low dimensional
subspaces. Subspace clustering (SC) [1] refers to the problem of clustering these data
points into their original subspaces, and various methods have been developed to solve
this problem [1–16]. Among them, Sparse Subspace Clustering (SSC), the first one to
introduce sparse representation into SC problem, is a state-of-the-art method, with elegant
formulation and theoretical guarantees to work under weak conditions, e.g., at the presence
of noise [17] and outliers [18]. SSC algorithm solves SC problem by first finding a sparse
self-representation for each data point, and then applying spectral clustering [19] to the
similarity matrix. However, it is inefficient to solve an optimization problem for sparse
representation (as the original SSC algorithm does) when analyzing large-scale data. One
way to speed up SSC is replacing `1 optimization with greedy pursuit, e.g., Orthogonal
Matching Pursuit (OMP) [20], for sparse self representation [21–23], which improves the
time efficiency of SSC by several orders of magnitude for large-scale problem. Another way
is to compress data points (i.e., to reduce dimension) before applying SSC algorithm, leading
to Compressed Subspace Clustering algorithm [24–26], whose performance is guaranteed by
works in random projection and Restricted Isometry Property (RIP) of subspaces [27, 28].
One motivation of this letter is the drawbacks of existing SSC algorithms. Traditional `1
optimization-based SSC (`1 -SSC) suffers from high computational complexity, while OMPSSC loses clustering accuracy. Our goal is to develop a fast SSC algorithm with high
clustering accuracy. We are also motivated by the insight that the final clustering accuracy
depends on two major properties of the similarity matrix: Subspace Detection Property
(SDP) [18], i.e., “no false connection”, among points from different subspaces, and connectivity among points from the same subspace. While there is rich literature on theoretical
analysis of SDP, much fewer works concern about the connectivity problem [29]. [30] introduces a post-processing procedure, and [31] modifies the optimization target as a weighted
sum of `1 norm and nuclear norm of the representation matrix in order to tradeoff between
separation and connectivity.
In this letter, we proposed a novel algorithm, active OMP-SSC (A-OMP-SSC), which
introduces two active-style processes, including adaptively updating and randomly dropping
data points. We analyze how the proposed algorithm obtains a better tradeoff between
connectivity and SDP, thus improving clustering accuracy while preserving advantages of
greedy methods in time efficiency. Numerical results validate our analyses and show the
2
advantages of our algorithm.
2
Preliminary
Given data matrix X = [x1 , x2 , · · · , xN ] ∈ RD×N , where D is data dimension and N is
total number of data points, the aim of SC is to cluster these data points into their original
subspaces. In the preprocessing, all points are `2 normalized, i.e., kxi k2 = 1, ∀i.
2.1
SSC and OMP-SSC
The core idea behind SSC is that, each data point can be approximately represented as a
sparse linear combination of other points, where the nonzero entries correspond to those
from the same subspace. The first step of SSC is to find a sparse representation for each
point by solving an optimization problem of
min kci k1 +
ci
λ
kxi − Xci k22
2
s.t.
cii = 0,
(1)
where ci denotes the representation coefficient vector for xi and λ is a balance parameter.
The second step is to apply spectral clustering to the similarity matrix A = |C| + |C|T and
get the clustering labels.
OMP-SSC is a faster version of SSC. It finds the sparse representation by OMP instead.
Given data point xi and dictionary composed of all other data points, OMP iteratively finds
an atom that has the largest absolute inner product with residual, adds this atom to the
neighbor set, projects xi onto the span of its current neighbors, and updates residual, until
the iteration number reaches a certain value or the norm of residual is small enough.
2.2
Geometric Analysis
Most literature focuses theoretical analyses of SSC on SDP.
Definition 1. [18] Subspace Detection Property (SDP) holds if and only if it holds that
for all i, ci is non-trivial (i.e., not all-zeros), and cij 6= 0 only when xi and xj lie in the
same subspace.
A geometric framework for theoretical analysis of SDP was first proposed in [18], and
has been thoroughly studied since then. In comparison, only a few works analyze the
3
connectivity problem of SSC. It is shown in [29] that, when subspace dimension exceeds
three, there is no guarantee that data points from the same subspace will form a connected
component even in the noiseless case.
3
Active OMP-SSC
3.1
Intuition
As mentioned before, the final clustering accuracy of SSC depends on both connectivity
and SDP of the similarity matrix obtained by the self-representation step. Actually, SDP
holding true is not necessary for correct clustering results due to the robustness of spectral
clustering. On the other hand, spectral clustering could fail if the connectivity of similarity
matrix is so weak that data points from the same subspace are separated. The ideal case is
that both connectivity and SDP are strong, which is unrealistic in practice at the presence of
noise. Therefore, the problem of increasing clustering accuracy becomes to tradeoff between
connectivity and SDP.
3.2
Algorithm
The proposed A-OMP-SSC algorithm operates iteratively on all data points. For xi , we
first find its sparse representation ci on the current dictionary, indexed by a set D, using
OMP,
ci = arg min kxi − Xĉi k2 ,
ĉi
s.t. kci k0 = d and
cij = 0, ∀j ∈ ({1, · · · , N }\D) ∪ {i},
(2)
where d denotes the iteration number of OMP. Since the goal of OMP here is not calculating
exactly the representation coefficients, but finding a few neighbors reliably, the iteration
number d is not very large, which will be specified later. We then calculate the representation
residual
ri = xi − Xci .
(3)
After the sparse representation step, we update xi by adding an offset term to it and keeping
it `2 normalized
x0i = (xi + bri )/kxi + bri k2 ,
4
(4)
Algorithm 1 The Proposed A-OMP-SSC
Input: Data points X ∈ RD×N , dictionary index set D = {1, · · · , N }, OMP iteration
number d, modifier parameter b > 0, dropping probability p ∈ (0, 1), cluster number k.
for i = 1, · · · , N do
1:
2:
Find ci using OMP (with d iterations) by (2).
3:
Calcuate residual ri by (3).
4:
Update data point: xi ← x0i by (4).
5:
Update dictionary: D ← D\{i} with probability p.
6:
end for
7:
Apply spectral clustering (with number of clusters k given) to similarity matrix A.
Output: Clustering labels.
where b > 0 denotes a modifier parameter. Then, with certain probability p we drop this
point from the future representation, i.e., removing index i from the dictionary index set.
The process above is repeated sequentially for all data points. Finally, we apply spectral
clustering to the similarity matrix, as done in SSC algorithms. Notice that OMP-SSC is a
special case of A-OMP-SSC, with modifier parameter b = 0 and dropping probability p = 0.
The detailed procedure of the proposed algorithm is included in Algorithm 1.
3.3
Discussion
3.3.1
Sequentially updating data points
From the construction of matrix A comes the intuition that, a two-direction connection
doesn’t contribute to connectivity. If xi chooses xj (j > i) as its neighbor, we should
encourage xj to choose the data points other than xi . By the updating operation in (4),
we push xi further away from its neighbors, as illustrated in Fig. 1.
This will be verified in an ideal case. Suppose (3) is a correct decomposition of xi ,
where x̄i = Xci lies in its ideal subspace and ri in its orthogonal subspace. Notice that the
projection operation in OMP ensures that xT
i x̄i is nonnegative. After the updating of (4)
we have
T
x0i x̄i =
x̄T
(x̄i + (b + 1)ri )T x̄i
T
i x̄i
=
≤ x̄T
i x̄i = xi x̄i .
kxi + bri k2
kxi + bri k2
5
(5)
o
x0i =
7i
x
ri
xi
xj
xi +bri
jjxi +bri jj
xi + bri
Figure 1: A visualization that updating data point xi may push it further away from its
neighbor xj . Without loss of generality, we suppose xT
i xj > 0.
The inequality in (5) comes from the fact that, when b ≥ 0,
kxi + bri k22 = kx̄i + (b + 1)ri k22 = kx̄i k22 + (b + 1)2 kri k22
≥ kx̄i + ri k22 = 1,
(6)
and the equality holds when b = 0.
From (6), kxi + bri k2 is symmetric about b0 = −1. In practice, since signal-to-noise
ratio (SNR) of residual is generally lower than that of the data point itself, a positive b
introduces less noise than the symmetric negative one in the updating step because it has
a smaller absolute value, which is preferred. The choice of b reflects a tradeoff between
improving connectivity and avoiding worsening SNR too much.
3.3.2
Randomly dropping data points
Dropping xi is more radical than updating it, as it is now impossible for xj , j > i to choose
xi as its neighbor. However, dropping data points could be risky since it reduces the data
density, which is undesirable in SSC as it worsens SDP.
3.3.3
OMP iteration number
Intuitively, in OMP process, a new edge improves connectivity if it is a true connection, but
weakens SDP otherwise. For OMP, once an atom is added to the neighbor set at a certain
iteration, it cannot be removed. Therefore, as iteration number d increases, connectivity
increases but SDP decreases, as pointed out in [23]. Thus, it is not the best to choose d
as the subspace dimension, as done in [22]. Choosing a smaller OMP iteration number can
improve clustering accuracy and linearly decrease time consumption.
6
Another advantage of choosing a small OMP iteration number in our active algorithm is
that, it allows more freedom for a correct choice of sparse representation coefficient satisfying
SDP as ci has fewer non-zero entries. For the SC problem, the essential information from
data is not the data points themselves, but their relationship instead. This inspires the
updating and dropping steps in A-OMP-SSC, which seek to further exploit this relationship
by changing the distribution of data points adaptively and thus changing the choice of
representation coefficients.
3.3.4
When it benefits
A-OMP-SSC works well when data density is not too small. This comes from the fact
that, the updating step changes the distribution of points, and the dropping step gradually
decreases data density. When the starting data density is relatively high, there are still
many near neighbors for later points, so SDP would not badly decrease.
3.3.5
Computational complexity
In OMP-SSC, each iteration requires N inner products with complexity O(D). Thus, for
N points, each d iterations, the total complexity of self-representation step is O(N 2 dD).
For A-OMP-SSC, only an additional O(D) updating step is required for each point, which
is neglegible. Moreover, the dropping step gradually decreases the number of points during
the process, which improves time efficiency.
4
Numerical Experiments
In this section, we conduct numerical experiments to validate our analysis and illustrate
the advantages of active OMP-SSC algorithm1 .
4.1
Synthetic Experiments
We first conduct experiments to examine the effects of modifier parameter b, dropping probability p, and OMP iteration number d, respectively. Next, we compare `1 -SSC, OMP-SSC,
and A-OMP-SSC algorithms, with respect to clustering error rate, connectivity, and SDP.
1
The MATLAB codes for the proposed methods and all experiments are available at http://gu.ee.
tsinghua.edu.cn/codes/A_OMP_SSC.zip.
7
1
1
0.6
0.4
0.2
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.06
0.04
0.02
0.01
0.005
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.06
0.04
0.02
0.01
0.005
0.8
noise level
noise level
0.8
0.5
0.45
0.6
0.4
0.2
0
0
-4 -3 -2 -1
0
1
2
3
4
5
b
0
0.2
0.4
0.6
0.8
1
p
Figure 2: Clustering error rate versus noise level and parameter b (left) or p (right).
We also validate that A-OMP-SSC shows advantages over OMP-SSC when data density is
not too small. Finally, the superior time efficiency of the proposed algorithm is verified.
The generated synthetic samples are randomly permuted as we meet in practice.
4.1.1
Modifier parameter b
We randomly generate 3 independent linear subspaces in 40-dimensional ambient space,
each of which is of dimension 6 and has 45 data samples. The noise level, i.e., additive
Gaussian noise strength of each sample, varies from 0 to 1. We set p = 0 and d = 3.
Clustering results with different choices of modifier parameter b are demonstrated in Fig. 2
(left), where each result is the average of 100 independent trials. According to Fig. 2 (left),
we read that clustering error rate is the highest at b ≈ −1 and lowest at 1. As b increases
from −1, clustering error rate first drops, leveraged by improvement in connectivity, but
then rises due to a decrease in SNR and thus a reduction in SDP. (Due to limited space,
results on connectivity and SDP are not shown here.) Also notice that a positive b is better
than a negative one symmetric about −1 thanks to higher SNR.
4.1.2
Dropping probability p
The data is generated the same way as in the previous experiment. We set b = 0 and
d = 3. Clustering results with different choices of dropping probability p are demonstrated
in Fig. 2 (right), which shows that clustering error rate reaches a minimum at p ≈ 0.8. A
large p up to 1 is a good choice in this experiment, because the subspace dimension and
OMP iteration number are small while the number of data points per subspace is relatively
large. Therefore, the decreasing data density would not be a big trouble.
8
clustering error rate
0.3
subspace dim. = 12
24
36
0.25
0.2
0.15
0.1
0.05
0
0
0.2
0.4
0.6
0.8
1
OMP iteration number / subspace dimension
Figure 3: Effects of OMP iteration number on clustering error rate for OMP-SSC (triangle
markers) and A-OMP-SSC (star markers).
4.1.3
OMP iteration number d
We randomly generate 3 independent linear subspaces in R120 , each of which has 60 data
samples. For A-OMP-SSC, b = 1, p = 0.3. This experiment is conducted under three
conditions, with subspace dimension set to 12, 24, 36, and the corresponding noise level 0.9,
0.7, 0.5. Effects of OMP iteration number d are demonstrated in Fig. 3. As d increases from
1 to subspace dimension, clustering error rate first drops owing to increasing connectivity,
but then rises due to decreasing SDP, just as we discussed. A reasonable choice of d is
approximately one over three of subspace dimension. Notice that A-OMP-SSC doesn’t
perform well when subspace dimension is 36, which is a result of the small data density, as
we will discuss later.
4.1.4
Performance comparison
We take the second smallest eigenvalue of the normalized Laplacian of a cluster as a metric
of connectivity (with range [0, 1]); the larger it is, the stronger connectivity is [22]. We
also use the percentage of points satisfying SDP as a metric of SDP (with range [0, 100]);
the larger it is, the stronger SDP is. The data is generated the same way as in the first
experiment. Based on previous results, we set b = 1, p = 0.8, and d = 3. A performance
comparison of three algorithms is demonstrated in Fig. 4. Results show that A-OMP-SSC
outperforms OMP-SSC in clustering accuracy, thanks to great improvement in connectivity,
despite minor loss of SDP. A-OMP-SSC also outperforms `1 -SSC at slightly high noise level.
This is because the number of chosen neighbors in the solution of `1 -SSC is generally around
or larger than subspace dimension, therefore SDP percentage quickly decreases to zero as
9
0.4
0.3
0.2
0.1
0
0
0.2
0.4
0.6
0.8
0.5
100
0.4
80
0.3
60
0.2
40
0.1
20
0
1
SDP percentage
L1-SSC
OMP-SSC
A-OMP-SSC
connectivity
clustering error rate
0.5
0
0
0.2
noise level
0.4
0.6
0.8
1
noise level
Figure 4: Comparison between `1 -SSC, OMP-SSC and A-OMP-SSC. Left: clustering error
rate; right: connectivity (left axis, blue solid lines) and SDP percentages (right axis, red
dashed lines).
4
noise level = 0.2
0.6
running time (second)
clustering error rate
0.4
0.3
0.2
0.1
0
0
50
100
150
200
samples per subspace
L1-SSC
OMP-SSC
A-OMP-SSC
3
2
1
0
0
50
100
150
200
samples per subspace
Figure 5: Algorithm comparisons under various problem scales. Left: clustering error rate;
right: running time (in seconds).
noise level increases. This result demonstrates the robustness of A-OMP-SSC to white
Gaussian noise.
4.1.5
Data density
We use the same parameter set as the performance comparison experiment, except that
the number of samples per subspace varies from 10 to 200. A comparison of clustering
error rate between SSC algorithms under certain noise levels is demonstrated in Fig. 5
(left). OMP-SSC performs better when data density is small, while A-OMP-SSC possesses
an increasing advantage as data density rises, which validates our previous analysis on the
success condition of A-OMP-SSC.
We also compare the running time of SSC algorithms under the same simulation setting.
10
Table 1: EYaleB clustering error rate in percentage.
# clusters
2
3
5
8
13
21
38
OMP-SSC
1.11
2.64
5.52
8.43
14.4
19.8
28.7
A-OMP-SSC
1.10
2.23
4.11
6.94
12.2
16.8
22.3
The result is shown in Fig. 5 (right). Despite the disadvantage in clustering accuracy in
some scenarios, a major advantage of OMP-SSC is that it is much faster than traditional
`1 -SSC, which is consistent with the results from existing literature [22, 23]. Furthermore,
A-OMP-SSC is even faster than OMP-SSC, because the operation of updating data points
doesn’t change the overall computational complexity, while the operation of dropping data
points gradually reduces data density.
4.2
Face Clustering
We test our algorithm on EYaleB, a dataset of human-face pictures from 38 persons, each
with around 64 photos under various illumination. In our experiment, the pictures are
downsampled to 48 × 42. It is known that human-face images of a subject can be approximated by a 9-dimensional subspace [2]. We set b = 0.5, p = 0.2 and d = 3. A comparison of
OMP-SSC and A-OMP-SSC on clustering error rate is demonstrated in Table 1. The result
is the average of 200 trials. For each trial, given the number of clusters k, we randomly
choose k subjects, mix all images of these subjects together, and permutate them before applying clustering algorithms. The result shows that A-OMP-SSC outperforms its reference
on human face clustering problem, especially in the scenario of multiple classes. Notice that
the state-of-the-art clustering error rate of `1 -SSC in the case of all 38 subjects is 31.0% [22],
while its running time is prohibitive. This highlights the fact that A-OMP-SSC enjoys both
superior time efficiency and higher clustering accuracy in such large-scale problems.
5
Conclusion
We propose a fast and accurate A-OMP-SSC algorithm, which obtains a better tradeoff
between connectivity and SDP via adaptively updating and randomly dropping data points.
We heuristically explain the intuitions behind our algorithm and analyze the condition where
it works. Numerical results validate our analyses, and demonstrate the advantages of our
11
algorithm over OMP-SSC and `1 -SSC. The proposed active mechanism may benefit many
other SC algorithms.
References
[1] R. Vidal, “Subspace clustering,” Signal Processing Magazine, IEEE, vol. 28, no. 2,
pp. 52–68, 2011.
[2] E. Elhamifar and R. Vidal, “Sparse subspace clustering: Algorithm, theory, and applications,” IEEE transactions on pattern analysis and machine intelligence, vol. 35,
no. 11, pp. 2765–2781, 2013.
[3] G. Liu, Z. Lin, S. Yan, J. Sun, Y. Yu, and Y. Ma, “Robust recovery of subspace
structures by low-rank representation,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 35, no. 1, pp. 171–184, 2013.
[4] R. Vidal and P. Favaro, “Low rank subspace clustering (lrsc),” Pattern Recognition
Letters, vol. 43, pp. 47–61, 2014.
[5] C.-Y. Lu, H. Min, Z.-Q. Zhao, L. Zhu, D.-S. Huang, and S. Yan, “Robust and efficient
subspace segmentation via least squares regression,” Computer Vision–ECCV 2012,
pp. 347–360, 2012.
[6] C. Lu, J. Feng, Z. Lin, and S. Yan, “Correlation adaptive subspace segmentation by
trace lasso,” in Proceedings of the IEEE International Conference on Computer Vision,
pp. 1345–1352, 2013.
[7] C. Lu, J. Tang, M. Lin, L. Lin, S. Yan, and Z. Lin, “Correntropy induced l2 graph for
robust subspace clustering,” in Proceedings of the IEEE International Conference on
Computer Vision, pp. 1801–1808, 2013.
[8] R. Heckel and H. Bölcskei, “Robust subspace clustering via thresholding,” IEEE Transactions on Information Theory, vol. 61, no. 11, pp. 6320–6342, 2015.
[9] Y. Zhang, Z. Sun, R. He, and T. Tan, “Robust subspace clustering via half-quadratic
minimization,” in Proceedings of the IEEE International Conference on Computer Vision, pp. 3096–3103, 2013.
12
[10] D. Park, C. Caramanis, and S. Sanghavi, “Greedy subspace clustering,” in Advances
in Neural Information Processing Systems, pp. 2753–2761, 2014.
[11] C. Yang, D. Robinson, and R. Vidal, “Sparse subspace clustering with missing entries,”
in Proceedings of the 32nd International Conference on Machine Learning (ICML-15),
pp. 2463–2472, 2015.
[12] C. You, C.-G. Li, D. P. Robinson, and R. Vidal, “Oracle based active set algorithm
for scalable elastic net subspace clustering,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pp. 3928–3937, 2016.
[13] X. Peng, Z. Yu, Z. Yi, and H. Tang, “Constructing the l2-graph for robust subspace
learning and subspace clustering,” IEEE transactions on cybernetics, vol. 47, no. 4,
pp. 1053–1066, 2017.
[14] H. Hu, Z. Lin, J. Feng, and J. Zhou, “Smooth representation clustering,” in Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3834–3841,
2014.
[15] C.-G. Li, C. You, and R. Vidal, “Structured sparse subspace clustering: A joint affinity
learning and subspace clustering framework,” IEEE Transactions on Image Processing,
2017.
[16] C. Peng, Z. Kang, M. Yang, and Q. Cheng, “Feature selection embedded subspace
clustering,” IEEE Signal Processing Letters, vol. 23, no. 7, pp. 1018–1022, 2016.
[17] Y.-X. Wang and H. Xu, “Noisy sparse subspace clustering,” in Proc. Int. Conf. Mach.
Learn., pp. 1–8, 2013.
[18] M. Soltanolkotabi and E. J. Candes, “A geometric analysis of subspace clustering with
outliers,” The Annals of Statistics, vol. 40, no. 4, pp. 2195–2238, 2012.
[19] U. Von Luxburg, “A tutorial on spectral clustering,” Statistics and computing, vol. 17,
no. 4, pp. 395–416, 2007.
[20] J. A. Tropp and A. C. Gilbert, “Signal recovery from random measurements via orthogonal matching pursuit,” IEEE Transactions on information theory, vol. 53, no. 12,
pp. 4655–4666, 2007.
13
[21] E. L. Dyer, A. C. Sankaranarayanan, and R. G. Baraniuk, “Greedy feature selection for
subspace clustering.,” Journal of Machine Learning Research, vol. 14, no. 1, pp. 2487–
2517, 2013.
[22] C. You, D. Robinson, and R. Vidal, “Scalable sparse subspace clustering by orthogonal
matching pursuit,” in Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 3918–3927, 2016.
[23] M. Tschannen and H. Bölcskei, “Noisy subspace clustering via matching pursuits,”
arXiv preprint arXiv:1612.03450, 2016.
[24] X. Mao and Y. Gu, “Compressed subspace clustering: A case study,” in Signal and
Information Processing (GlobalSIP), 2014 IEEE Global Conference on, pp. 453–457,
IEEE, 2014.
[25] R. Heckel, M. Tschannen, and H. Bölcskei, “Dimensionality-reduced subspace clustering,” Information and Inference: A Journal of the IMA, p. iaw021, 2017.
[26] Y. Wang, Y.-X. Wang, and A. Singh, “A theoretical analysis of noisy sparse subspace
clustering on dimensionality-reduced data,” arXiv preprint arXiv:1610.07650, 2016.
[27] G. Li and Y. Gu, “Distance-preserving property of random projection for subspaces,” in
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
2017.
[28] G. Li and Y. Gu, “Restricted isometry property of gaussian random projection for
finite set of subspaces,” arXiv preprint arXiv:1704.02109, 2017.
[29] B. Nasihatkon and R. Hartley, “Graph connectivity in sparse subspace clustering,”
in Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on,
pp. 2137–2144, IEEE, 2011.
[30] Y. Wang, Y.-X. Wang, and A. Singh, “Graph connectivity in noisy sparse subspace
clustering,” CoRR abs/1504.01046, 2016.
[31] Y.-X. Wang, H. Xu, and C. Leng, “Provable subspace clustering: When lrr meets ssc,”
in Advances in Neural Information Processing Systems, pp. 64–72, 2013.
14
| 7cs.IT
|
1
Graph-Based Blind Image Deblurring From a
Single Photograph
arXiv:1802.07929v1 [cs.CV] 22 Feb 2018
Yuanchao Bai, Student Member, IEEE, Gene Cheung, Senior Member, IEEE, Xianming Liu, Member, IEEE,
Wen Gao, Fellow, IEEE
Abstract—Blind image deblurring, i.e., deblurring without
knowledge of the blur kernel, is a highly ill-posed problem. The
problem can be solved in two parts: i) estimate a blur kernel
from the blurry image, and ii) given estimated blur kernel, deconvolve blurry input to restore the target image. In this paper,
we propose a graph-based blind image deblurring algorithm by
interpreting an image patch as a signal on a weighted graph.
Specifically, we first argue that a skeleton image—a proxy that
retains the strong gradients of the target but smooths out the
details—can be used to accurately estimate the blur kernel
and has a unique bi-modal edge weight distribution. Then, we
design a reweighted graph total variation (RGTV) prior that can
efficiently promote a bi-modal edge weight distribution given a
blurry patch. Further, to analyze RGTV in the graph frequency
domain, we introduce a new weight function to represent RGTV
as a graph l1 -Laplacian regularizer. This leads to a graph spectral
filtering interpretation of the prior with desirable properties,
including robustness to noise and blur, strong piecewise smooth
(PWS) filtering and sharpness promotion. Minimizing a blind
image deblurring objective with RGTV results in a non-convex
non-differentiable optimization problem. We leverage the new
graph spectral interpretation for RGTV to design an efficient
algorithm that solves for the skeleton image and the blur
kernel alternately. Specifically for Gaussian blur, we propose a
further speedup strategy for blind Gaussian deblurring using
accelerated graph spectral filtering. Finally, with the computed
blur kernel, recent non-blind image deblurring algorithms can
be applied to restore the target image. Experimental results
demonstrate that our algorithm successfully restores latent sharp
images and outperforms state-of-the-art methods quantitatively
and qualitatively.
Index Terms—Blind image deblurring, graph signal processing,
non-convex optimization
I
I. I NTRODUCTION
Mage blur is a common image degradation, which is caused
by out-of-focus photography or motions between objects
and camera during exposure time. The blur process is usually
modeled as
b = x ⊗ k + n,
(1)
where b is the observed blurry image, x is the latent sharp
image, k is the blur kernel, n is the noise and ⊗ is the
convolution operator. As an inverse problem, image deblurring
is to recover the latent sharp image x from the blurry image
Yuanchao Bai and Wen Gao are with the School of Electronics Engineering
and Computer Science, Peking University, Beijing, 100871, China. (e-mail:
{yuanchao.bai, wgao}@pku.edu.cn).
Gene Cheung is with the National Institute of Informatics, Tokyo, 1018430,
Japan (e-mail: [email protected]).
Xianming Liu is with the School of Computer Science and Technology, Harbin Institute of Technology, Harbin, 150001, China (e-mail: [email protected]).
b. There are two categories of image deblurring depending
on whether the blur kernel k is known, i.e., non-blind image
deblurring [1] and blind image deblurring [2], [3]. We focus
on the blind image deblurring problem, where both the latent
image x and the blur kernel k are unknown and must be
restored given only the blurry image b. It is a highly ill-posed
problem, since the feasible solution of the problem is not only
unstable but also non-unique.
To overcome the ill-posedness, for blind image deblurring,
it is important to design a prior that promotes image sharpness and penalizes blurriness. However, conventional gradientbased priors of natural images tend to fail [3], because they
usually favor blurry images with mostly low frequencies in
the Fourier domain. Recently, many sophisticated image priors
are proposed to deal with this problem, for example, l0 -norm
based prior [4], low-rank prior [5] and dark channel prior [6].
Besides these priors, with the advance of graph signal processing (GSP) [7], graph-based priors have been designed for
different image applications [8]–[11]. By modeling pixels as
nodes with weighted edges that reflect inter-pixel similarities,
images can be interpreted as signals on graphs. In this paper,
we explore the relationship between graph and image blur, and
propose a graph-based prior for blind image deblurring.
Specifically, instead of directly computing the natural image, we argue that a skeleton image—a piecewise smooth
(PWS) proxy that retains the strong gradients of the target
image but smooths out the details—is sufficient to estimate
the blur kernel. We observe that, unlike blurry patches, the
edge weights of a graph for the skeleton image patch have a
unique bi-modal distribution. We thus propose a reweighted
graph total variation (RGTV) prior to promote the desirable
bi-modal distribution given a blurry patch. We juxtapose
and analyze the advantages of RGTV against previous graph
smoothness priors, such as graph total variation (GTV) [12]–
[15] and the graph Laplacian regularizer [9] in the nodal
domain. To analyze RGTV in graph frequency domain, we
define a new graph weight function so that RGTV can be
expressed as a graph l1 -Laplacian regularizer, similar in form
to previous graph l2 -Laplacian regularizer [9]. Doing so means
that RGTV can be interpreted as a low-pass graph spectral
filter with desirable properties such as robustness to noise /
blur and strong PWS filtering.
Based on our graph spectral interpretation of RGTV, we
design an efficient algorithm that solves for the skeleton image
and the blur kernel alternately. Moreover, specifically for
Gaussian blur, we propose a further speedup strategy for blind
Gaussian deblurring using accelerated graph spectral filtering
2
[16]. Finally, with the estimated blur kernel k, we de-convolve
the blurry image using a non-blind deblurring method, like [4],
[17], [18]. The contributions of this paper are summarized as
follows:
1) We propose a graph-based image prior RGTV that
promotes a bi-modal weight distribution to reconstruct
a skeleton patch from a blurry observation, so that a
suitable blur kernel can be simply derived thereafter.
2) We introduce a graph weight function so that GTV
/ RGTV can be expressed as a graph l1 -Laplacian
regularizer. The prior can then be interpreted as a lowpass graph filter with desirable spectral properties. To
the best of our knowledge, we are the first to provide a
graph frequency interpretation of GTV.
3) Based on our spectral interpretation of RGTV, we design
an efficient algorithm to solve the non-convex nondifferentiable optimization problem alternately, where
the two sub-problems to solve for the skeleton image
and the blur kernel have closed-form solutions.
4) Specifically for Gaussian blur, we propose a further
speedup strategy for blind Gaussian deblurring using
accelerated graph spectral filters [16].
Experiments demonstrate that the proposed algorithm is competitive or even better than the state-of-the-art methods with
lower complexity.
The outline of the paper is organized as follows. We review
the related works on image deblurring and GSP in Sec. II. We
introduce necessary GSP definitions and skeleton image in
Sec. III. RGTV prior and its analysis in both nodal and graph
frequency domains are presented in Sec. IV. Blind deblurring
algorithm is proposed in Sec. V. Experiments and conclusions
are in Sec. VI and Sec. VII, respectively.
II. R ELATED W ORK
A. Image Deblurring
Depending on whether the blur kernel is known in advance
or not, image deblurring is divided into two categories.
1) Non-blind image deblurring: Non-blind image deblurring [1] is to recover latent sharp image with known blur
kernel, which has been studied for decades. It is an ill-posed
problem since the inverse process is unstable to noise; even a
small amount of noise will lead to severe distortions. Classic
methods to solve this problem include Wiener filter [19]
and Richardson-Lucy algorithm [20], [21]. In recent years,
non-blind image deblurring is modeled as an optimization
problem. Many famous image priors, such as total variation
(TV) [22] and sparse priors [17], [23], have been introduced
for regularization to deal with the ill-posedness in image
restoration.
2) Blind image deblurring: Blind image deblurring [2], [3]
is to recover a latent sharp image without knowledge of the
underlying blur kernel. Conventional gradient-based priors of
natural images tend to fail [3], because they usually favor
blurry images with mostly low frequencies in the Fourier
domain. With the progress of regularization and optimization,
more sophisticated priors have been introduced to solve this
problem, such as framelet based prior [24], l0 -norm based prior
[4], [25], sparse-coding based prior [26], low-rank prior [5]
and dark channel prior [6], etc. However, each has its own
shortcomings. Framelet based prior [24] relies on manually
crafted wavelet functions, which are not general enough to
handle heterogeneous blurring scenarios. l0 -norm based prior
[4], [25] is combinatorial in nature (thus non-convex), and
its convex relaxation to l1 -norm requires sensitive parameter
tuning for optimal performance. Sparse-coding based prior
[26] assumes statistical similarity between the training set and
the target image, which may not be true in practice. Low-rank
prior [5] requires solving costly singular value decomposition
(SVD), which has complexity O(N 3 ) in general. Dark channel
prior [6] is a comprehensive prior combining l0 -norm and nonlinear dark channel computation, thus also suffers from high
complexity.
Instead of the above mentioned priors, we investigate the
potential of graph-based priors for the blind image deblurring
problem. Our proposed graph-based prior RGTV is highly
data-adaptive and can be efficiently solved using fast graph
spectral filters via a novel interpretation of RGTV as a graph
l1 -Laplacian regularizer.
B. Graph Based Image Prior
GSP [7] is an emerging field to study signals on irregular
data kernels described by graphs. By modeling pixels as
nodes with weighted edges that reflect inter-pixel similarities,
images (or image patches) can be interpreted as graph signals.
For image restoration, graph based image priors, such as
graph Laplacian regularizer [9] and GTV [12]–[15], have been
designed for different inverse problems.
1) Graph Laplacian Prior: In [8], Hu et al. designed a
scheme to soft-decode a JPEG-compressed PWS image by
optimizing the desired graph signal and the similarity graph
in a unified framework. In [10], Liu et al. soft-decoded JPEGcompressed natural images by using a combination of three
priors, including a new graph smoothness prior called Left
Eigenvectors of Random walk Graph Laplacian (LERaG),
a compact dictionary trained by sparse representation and
Laplacian distribution of discrete cosine transform coefficients.
Pang et al. [9] analyzed graph Laplacian regularization in the
continuous domain and provided insights for image denoising.
With a doubly-stochastic graph Laplacian, Kheradmand et al.
[11] developed a framwork for non-blind image deblurring.
2) Graph Total Variation Prior: GTV is one of the pDirichlet energy prior, which enjoys both desirable PWSpreserving properties and convexity. In [12], Elmoataz et at.
analyzed the discrete p-Dirichlet energy in image and manifold
processing. In [13], Hidane et al. employed GTV for nonlinear multi-layered representation of graph signals. In [14],
Couprie et al. proposed a dual constrained GTV regularization
on graphs. In [15], Berger et al. used GTV to recover a smooth
graph signal from noisy samples taken on a subset of graph
nodes.
In this paper, we propose a novel RGTV prior to solve the
blind image deblurring problem. The proposed method is a
non-trivial extension of our recent work [27]: i) we provide
analysis of GTV and RGTV in the graph spectral domain,
3
the first to do so in the literature; ii) based on our spectral
interpretation of RGTV, we design an efficient alternating iterative algorithm to solve the non-convex optimization problem;
and iii) specifically for Gaussian blur, we propose a speedup
strategy for blind Gaussian deblurring using accelerated graph
spectral filtering [16].
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 1. Illustrations of different kinds of images. (a) a true natural
image. (b) a blurry image. (c) a skeleton image. (d), (e) and (f) are
patches in red squares of (a), (b) and (c), respectively.
III. GSP D EFINITIONS AND S KELETON I MAGE
where W is an adjacency matrix of size N ×N , xi and xj are
the intensity values at pixels i and j of the image x, and σ is a
parameter. 0 ≤ wij ≤ 1 and the larger wij is, the more similar
the nodes i and j are to each other. Unlike the bilateral filter
[28], we do not include Euclidean distance in the edge weight
definition in (2), so that a denser graph would lead to stronger
filtering among pixels of similar intensities. (Experimentally,
graph density is chosen as a tradeoff between filtering strength
and computation complexity.)
Given the adjacency matrix W, a combinatorial graph
Laplacian matrix L is a symmetric matrix defined as:
L , diag(W1) − W
(3)
where 1 is a vector of all 1’s. diag(·) is an operator constructing a square diagonal matrix with the elements of input vector
on the main diagonal.
Given (2) and (3), L is a real symmetric matrix, so there is
an orthogonal matrix U that diagonalizes L via the Spectral
Theorem,
L = UΛUT
(4)
where Λ is a diagonal matrix containing eigenvalues λk , k ∈
{1, . . . , N }. Each column uk in U is an eigenvector corresponding to λk . Given wi,j is non-negative from (2), L is a
positive semi-definite (PSD) matrix. Hence λk ≥ 0 for each
k and xT Lx ≥ 0 for arbitrary graph signal x. In the GSP
literature [7], the non-negative eigenvalues λk are interpreted
as graph frequencies and corresponding eigenvectors in U as
graph frequency components. Together, they define the graph
spectrum for graph G.
B. Skeleton Image and its Bi-modal Weight Distribution
We define a skeleton image—a PWS version of the target
image—as a proxy for the blind image deblurring problem.
The skeleton image retains the strong gradients in a natural
image but smooths out the minor details, which is similar to
a structure extracted image [29] or an edge-aware smoothed
image [30]. An illustrative example is shown in Fig. 1. Both
Fraction
We first define GSP concepts needed in our work. A graph
G(V, E, W) is a triple consisting of a finite set V of N nodes
(image pixels) and a finite set E ⊂ V × V of M edges. Each
edge (i, j) ∈ E is undirected with a corresponding weight wij
which measures the similarity between nodes i and j. Here
we compute the weights using a Gaussian kernel [7]:
kxi − xj k2
,
(2)
[W]i,j = wi,j = exp −
σ2
0.1
Fraction
0.05
0.025
0
0.05
0
0
100
d=|xi-xj|
(a)
200
0.2
Fraction
A. Definitions in GSP
0
100
d=|xi-xj|
0.1
200
(b)
0
0
100
d=|xi-xj|
200
(c)
Fig. 2. Edge weight distribution around image edges. (a) a true natural patch.
(b) a blurry patch. (c) a skeleton patch.
the target natural image and its skeleton image are sharper
than the blurry image in the middle.
In order to differentiate between blurriness and sharpness
in a pixel patch in a more mathematically rigorous manner,
we further seek statistical descriptions of these patches. As
an illustrative example, we construct a fully connected graph
for each of three representative local patches (highlighted by
red in Fig. 1) and examine its edge weight distribution, where
edge weight wi,j is computed using (2). Fig. 2 shows the edge
weight distributions (histograms) of the representative patches
in Fig. 1d–1f. x-axis is the discrete inter-pixel difference d =
|xi − xj | for edge weight wi,j ; edge weight wi,j in (2) is a
monotonically decreasing function of d. y-axis shows fractions
of weights given d for different image patches. We make the
following key observation from the histograms:
Both the target natural patch and its skeleton version
have bi-modal distributions of edge weights, while
the blurred patch does not, due to low-pass filtering
during the blur process.
Bi-modal distribution means that the inter-pixel differences
in an image patch are either very small or very large, i.e., the
patch is PWS. To be demonstrated experimentally later, the
PWS skeleton patch is just as valuable as the target natural
image in terms of computing an appropriate blur kernel, but
the skeleton patch can be more easily reconstructed from a
blurry patch than the natural patch.
Similar observations can be made on sparser graphs also.
Given the desirable statistical property of the skeleton image
patch, we next design a signal prior to promote a bi-modal
distribution of edge weights given an observed blurry patch.
IV. G RAPH - BASED I MAGE P RIOR AND A NALYSIS
A. Reweighted Graph Total Variation Prior
We propose a reweighted graph total variation (RGTV)
prior (regularizer) to promote the aforementioned bi-modal
edge weight distribution in a target pixel patch. We first define
4
B. Comparisons with Graph Laplacian Prior
1
Graph Laplacian
Graph TV
Reweighted Graph Laplacian
Reweighted Graph TV
0.04
0.5
For comparison, we also analyze commonly used graph
Laplacian regularizers. The graph Laplacian regularizer [9] is
expressed as
Value
Value
0.03
Graph Laplacian
Graph TV
Reweighted Graph Laplacian
Reweighted Graph TV
0.02
0
0.01
0
-0.5
0
0.1
0.2
0.3
0.4
0.5
0
0.1
d=|xi-xj|
0.2
0.3
0.4
0.5
xT Lx =
i=1 j=1
d=|xi-xj|
(a)
(b)
Fig. 3. Curves of regularizers and their corresponding first-derivatives for
each (i, j) pair. d is normalized to [0, 1]. wi,j = 0.1 for graph Laplacian
and graph TV. σ = 0.1 for reweighted graph Laplacian and reweighted graph
TV.
the gradient operator of a graph signal x. The gradient of node
i ∈ V is defined as ∇i x ∈ RN and its j-th element is:
(∇i x)j , xj − xi ,
(5)
The graph total variation (GTV) [12]–[15] is defined as
kxkGT V =
=
X
k diag(Wi,· )∇i xk1
i∈V
N
N X
X
wi,j |xj − xi |,
(6)
where Wi,· is the i-th row of the adjacency matrix W. GTV
initializes W using for example (2) and keeps it fixed, and
hence does not promote bi-modal distribution of edge weights.
Specifically, since (6) is separable, we can analyze the behavior
of GTV using a single node pair (i, j) separately like a twonode graph. With d = |xi − xj | and fixed wi,j , the regularizer
for pair (i, j) is wi,j d, which is a linear function of d with
slope wi,j . The curve of wi,j d has only one minimum at d = 0,
as shown in Fig. 3a. Minimizing (6) only pushes d towards 0,
i.e., smoothing the image x.
Instead of using fixed W, we extend the conventional
graph TV to RGTV, where the graph weights W(x) are also
functions of x,
X
kxkRGT V =
k diag(Wi,· (x))∇i xk1
i∈V
N X
N
X
i=1 j=1
wi,j (xi , xj )|xj − xi |,
(7)
where Wi,· (x) is the i-th row of W(x) and wi,j (xi , xj ) is the
(i, j) element of W(x). This extension makes a fundamental
difference, because the regularizer for pair (i, j) now becomes
2
wi,j (xi , xj )|xj − xi | = exp(−d2 /σ√
) · d. The curve of this
regularizer has one maximum at σ/ 2 and two minima at 0
and +∞, as shown
if d
√ in Fig. 3a. Minimizing (7) reduces d √
is smaller than σ/ 2 and amplifies d if d is larger than σ/ 2.
Thus, RGTV regularizer can effectively promote the desirable
bi-modal edge weight distribution of sharp images.
Using the aforementioned RGTV prior, we propose an
optimization function for blind image deblurring in Sec. V.
wi,j (xj − xi )2
(8)
Like GTV, graph Laplacian initializes L and keeps it fixed, and
hence does not promote the desirable bi-modal edge weight
distribution. With d = |xi − xj | and fixed wi,j , the prior for
each node pair (i, j) is wi,j d2 , which is a quadratic function
of d with coefficient wi,j . The curve of wi,j d2 has only one
minimum at d = 0, as shown in Fig. 3a. Minimizing (8) only
pushes d to 0, i.e., smoothing the image x.
We extend the conventional graph Laplacian to the reweighted graph Laplacian. Similar to RGTV, we define the reweighted graph Laplacian as
xT L(x)x =
N X
N
X
i=1 j=1
i=1 j=1
=
N X
N
X
wi,j (xi , xj ) · (xj − xi )2 ,
(9)
where Laplacian matrix L(x) is a function of x and
wi,j (xi , xj ) is the same as the definition in (7). Then, the
regularizer for pair (i, j) becomes wi,j (xi , xj )(xj − xi )2 =
exp(−d2 /σ 2 ) · d2 . The curve has one maximum at d = σ
and two minima at 0 and +∞, as shown in Fig. 3a. Thus the
reweighted graph Laplacian also promotes the desirable bimodal edge weight distribution, which explains its effectiveness in previous works on restoration of PWS images using
this prior [8], [9], [31].
However, it has one significant drawback. Taking the first
derivative of its function results in exp(−d2 /σ 2 ) · 2d(1 −
d2 /σ 2 ), as shown in Fig. 3b. limd→0 exp(−d2 /σ 2 ) · 2d(1 −
d2 /σ 2 ) = 0, which means that the promotion of bi-modal edge
weight distribution tends to slow down significantly when d
is close to 0 in practice.
Different from reweighted graph Laplacian, the first derivative of the cost function of RGTV is exp(−d2 /σ 2 ) · (1 −
2d2 /σ 2 ), as shown in Fig. 3b. limd→0 exp(−d2 /σ 2 ) · (1 −
2d2 /σ 2 ) = 1, which means that RGTV can effectively promote a bi-modal edge weight distribution even as d approaches 0. As an illustration, Fig. 4 compares the performance
of reweighted graph Laplacian and the proposed RGTV in
promoting bi-modal distribution from a blurry image experimentally. We observe that RGTV restores a better skeleton
image with sharp edges than Reweighted graph Laplacian.
C. Spectral Analysis of GTV and RGTV
As a signal smoothness prior defined with respect to the
graph [7], GTV generalizes the well-known TV notion [22]
and its edge-preserving property to the graph signal domain.
However, the usage of l1 -norm means that there is no natural
graph spectral interpretation—i.e., promotion of certain low
graph frequencies—like the graph Laplacian regularizer.
In this paper, we introduce a spectral interpretation of GTV
by rewriting it as a novel l1 -Laplacian operator on a graph,
inspired by [12]. Based on the spectral interpretation of GTV,
5
adjacency matrix Γ with the new weight function (12) and
rewrite (11) in matrix form for GTV as
LΓ , diag(Γ1) − Γ.
(a)
(b)
We call LΓ the l1 -Laplacian matrix. LΓ is a real symmetric
PSD matrix. With LΓ , we are able to analyze the spectrum of
GTV like (4),
(c)
Fig. 4. (a) Image blurred by Gaussian blur, σb = 1.5. (b) Promoting bimodal distribution by reweighted graph Laplacian. (c) Promoting bi-modal
distribution by
√ RGTV. We set σ = 0.1 for reweighted graph Laplacian and
σ = 0.1 × 2 for RGTV to ensure the maximum of both priors are at
d = 0.1.
we examine the graph spectrum of RGTV iteratively as the
edge weights are updated across iterations. The desirable
properties of RGTV become evident when carefully examining
the low frequencies of its corresponding graph spectrum and
its gradual transformation. Moreover, the new graph spectral
interpretation of RGTV leads to an efficient algorithm for
the non-convex and non-differentiable blind image deblurring
problem, and an accelerated graph spectral filtering implementation specifically for Gaussian blur (to be discussed in
Sec. V).
Since a graph spectrum is defined with respect to a variation
operator like the graph Laplacian, towards a spectral interpretation for GTV, we first define a new Laplacian operator for
GTV. The conventional Laplacian operator on graph signal x
can be deduced from the derivative of (8),
∂xT Lx
i
= 2 · (Lx)i = c ·
N
X
j=1
wi,j · (xi − xj ),
N
N
X
X
=c·
wi,j xi −
wi,j xj
j=1
(10)
j=1
where c is a coefficient derived from the derivative apart from
the Laplacian operator.
Similarly, we take the sub-derivative of (6), as GTV is nondifferentiable. By sub-differentiating and applying an upperbound function to (6) near zero (detailed derivation is included
in Appendix A), we get:
(∂kxkGT V )i = c0 ·
N
X
j=1
γi,j · (xi − xj ),
N
N
X
X
= c0 ·
γi,j xi −
γi,j xj
j=1
(11)
j=1
where c0 is a coefficient similar to c in (10) and
wi,j
γi,j =
max{|xj − xi |, }
(13)
(12)
where is introduced as a small constant for numerical stability around 0. We see from the equation that when |xj −xi | < ,
γi,j = (1/)wi,j , which is upper-bounded by 1/. In our
experiments, we fix at 0.01.
Considering γi,j as a new graph weight defined by (12),
(11) is in the same form as (10). Hence we can define a new
LΓ = UΓ ΛΓ UTΓ
(14)
where ΛΓ is a diagonal matrix containing graph frequencies of
GTV, and UΓ contains corresponding graph frequency components as columns. min {diag(ΛΓ)} is 0 and max {diag(ΛΓ )}
is upper-bounded by maxi 2d i , where di is the degree of
node i, as proven in Appendix B. The bounds of diag(ΛΓ )
are used in later stability analysis of our algorithm in Sec. V.
For clarity of presentation, in the following we denote by
LW the conventional graph Laplacian matrix computed from
adjacency matrix W, and denote by LΓ the l1 -Laplacian
matrix computed from adjacency matrix Γ.
Fig. 5 and Fig. 6 show experiment results for a simple
one-dimensional PWS signal, demonstrating that the spectrum
of GTV has better PWS properties than conventional graph
Laplacian. First, the PWS property of the second eigenvector
(lowest AC frequency component) of GTV is more robust than
the graph Laplacian regularizer when noise and/or blur are
applied to the ideal signal, as shown in Fig. 5d, 5e and 5f.
Second, we claim that GTV is a stronger PWS-preserving
filter than the graph Laplacian regularizer; we show this
by examining the relative eigenvalues λk /λ2 . Specifically,
consider the known graph-spectral filter from a standard MAP
formulation [9], where the nodal domain filter x∗ = (I + µ ·
L{W,Γ} )−1 y can be expressed in the graph frequency domain
as:
!
1
∗
UT{W,Γ} y,
(15)
x = U{W,Γ} diag
{W,Γ}
1 + µ · λk
where {W, Γ} means either the conventional graph Laplacian
operator or our proposed l1 -Laplacian operator is employed.
µ is a parameter trading off the importance of the fidelity
term and the graph-signal smoothness prior in the original
MAP formulation, and k ∈ {1, . . . , N }. In Fig. 6, we observe
{W,Γ}
{W,Γ}
{W,Γ}
{W,Γ}
= 0 and λ2
/λ2
= 1 for both
that λ1
/λ2
Γ
Γ
GTV and graph Laplacian, but λk /λ2 of GTV for k > 2 is
W
much larger than λW
k /λ2 of the graph Laplacian regularizer,
meaning that GTV penalizes high graph frequencies more
severely than the graph Laplacian regularizer.
Spectral Analysis of RGTV: Based on the spectral analysis
of GTV above, we further analyze the behaviour of the
spectrum of RGTV. RGTV inherits the desirable spectral
properties—robustness to noise and blur, and strong PWS
filtering—from GTV and promotes bi-modal edge weight
distribution at the same time. Since the adjacency matrix
W(x) of RGTV is a function of signal x, there is no fixed
spectrum for RGTV. Instead, we first initialize weights for
RGTV and compute its spectrum like GTV, and then we
update the weights and examine the gradual transformation
of spectrum iteratively. Considering the graph-spectral filter
6
1.5
1.5
1.5
1
1
1
0.5
0.5
0.5
0
0
0
-0.5
0
20
40
60
80
100
-0.5
0
20
40
(a)
60
80
100
-0.5
0.15
0
0
0
-0.05
-0.05
60
80
100
-0.15
100
V. B LIND I MAGE D EBLURRING A LGORITHM
GTV
GL
Using the blur image model in (1), we pose the blind image
deblurring problem as an optimization as follows using our
proposed RGTV prior:
-0.1
-0.1
40
80
0.05
-0.05
-0.1
60
0.1
0.05
20
40
0.15
GTV
GL
0.1
0.05
0
20
(c)
0.15
GTV
GL
0.1
-0.15
0
(b)
-0.15
0
20
40
(d)
60
80
0
100
20
40
(e)
60
80
100
(f)
Fig. 5. Illustrative experiments of graph spectrum on 1D graph signals. (a)
an ideal PWS signal. (b) a PWS signal with Gaussian noise, σn = 0.02. (c) a
PWS signal blurred by a Gaussian blur, σb = 1. (d), (e) and (f) are the lowest
AC frequency components of GTV and GL (graph Laplacian) corresponding
to (a), (b) and (c). The graphs are constructed as a 4-neighbour adjacency
matrix with weight parameter σ = 0.3.
×104
λk/λ2
λk/λ2
GTV
GL
GTV
GL
150
300
4
2
200
1
2
3
4
5
0
6
Index k of eigenvalues
100
50
100
0
-2
200
400
GTV
GL
6
λk/λ2
8
1
2
3
4
5
0
6
1
2
(a)
3
4
5
6
Index k of eigenvalues
Index k of eigenvalues
(b)
(c)
Fig. 6. Relative eigenvalues. (a), (b) and (c) represent the curves of relative
eigenvalue λk /λ2 of GTV and GL (graph Laplacian) with respect to k. Each
plot corresponds to the signal in Fig. 5a, Fig. 5b and Fig. 5c, respectively.
1.5
0.15
2500
iter1
iter2
iter3
0.1
1
λk/λ2
0
1000
500
-0.05
0
0
-0.1
-500
-0.5
0
20
40
(a)
60
80
100
-0.15
iter1
iter2
iter3
2000
1500
0.05
0.5
0
20
40
(b)
60
80
100
1
2
3
4
5
6
Index k of eigenvalues
1
x̂, k̂ = arg min kx ⊗ k − bk22 + βkxkRGT V + µkkk22 (17)
2
x,k
where the first term is the data fidelity term, and the remaining
two terms are regularization terms for variables x and k,
respectively. β and µ are two corresponding parameters.
The optimization (17) is non-convex and non-differentiable,
which is challenging to solve. Here we apply a coarse-to-fine
strategy [32] to solve (17), in order to make the solver robust
even for large blur kernels. In the coarse-to-fine strategy, we
construct an image pyramid by down-sampling the blurry image and perform blind image deblurring scale-by-scale. In each
scale, we estimate k̂ and x̂ alternatingly, and then up-sample
k̂ as the initial value for the finer scale. The optimization
algorithm in each scale is sketched in Algorithm 1.
Algorithm 1 Blind Deblurring Algorithm in Each Scale
Input: Blurry image b and kernel size h × h.
1: Initialize k̂ with delta function or the result from coarser scale.
2: while not converge do
Compute x̂ by solving (18).
Compute k̂ by solving (21).
β ← β/1.1.
endwhile
Output: Estimated blur kernel k̂ and skeleton image x̂.
(c)
Fig. 7. Illustrative experiments of graph spectrum of RGTV on a 1D graph
signal. (a) a PWS signal blurred by a Gaussian blur σb = 1 with Gaussian
noise σn = 0.0001. (b) is the second eigenvectors of RGTV in each iteration.
(c) represents the curves of relative eigenvalue λk /λ2 with respect to k in
each iteration. The graphs are constructed as a 4-neighbour adjacency matrix
with weight parameter σ = 0.3.
again, the ideal RGTV filter satisfies (I + µ · LΓ (x∗ )) x∗ = y.
We initialize x(0) = y and perform the following filter (16)
iteratively until convergence,
1
(n)
(n)T
(n+1)
x
= UΓ diag
UΓ y
(16)
Γ
(n)
1 + µ · λk (x )
(n)
From both the analysis on the graph nodal and graph
spectral domain, we can theoretically conclude that RGTV
is an effective prior for the blind image deblurring, and more
effective than previous graph smoothness priors, such as graph
Laplacian regularizer [9]–[11] and conventional GTV [13]–
[15].
where UΓ = UΓ (x(n) ).
A one-dimensional illustrative experiment is shown in
Fig. 7. The first AC graph frequency component of RGTV
becomes sharper (Fig. 7b), and the low-pass graph filtering
strength becomes larger (Fig. 7c) with iteration, which demonstrates the edge-sharpness promotion property of RGTV from
the graph frequency perspective.
The minimizer x̂ is our PWS proxy—the skeleton image—
in order to estimate a good blur kernel k̂. To restore the natural
sharp image given estimated blur kernel k̂, we can use recent
non-blind image deblurring algorithms to deblur the blurry
image b such as [4], [17], [18].
A. Skeleton Image Restoration
Given k̂, optimization (17) to solve x becomes:
1
x̂ = arg min kx ⊗ k̂ − bk22 + βkxkRGT V
2
x
(18)
RGTV is a non-differentiable and non-convex prior, where
the edge weights are functions of x. To solve (18), we
leverage on the spectral analysis in Sec. IV-C and employ an
alternating scheme with the proposed l1 -Laplacian of GTV to
approximate RGTV; i.e., we first optimize x with initialized
LΓ , then we update LΓ with LΓ (x̂) and optimize x again. The
alternating algorithm runs iteratively until convergence as the
solution to (18). The steps of solving (18) is summarized in
Algorithm 2.
7
Fixing LΓ and k̂ to solve for x, the problem becomes a
non-blind image deblurring problem with a graph Laplacian
regularizer:
1
x̂ = arg min kx ⊗ k̂ − bk22 + β · xT LΓ x
2
x
(19)
As (19) is a quadratic convex optimization function, it is
equivalent to solving the following system of linear equations,
(K̂T K̂ + 2β · LΓ )x̂ = K̂T b
(20)
where K̂ is a block circulant with circulant blocks (BCCB)
matrix that is the matrix representation of convolving with
k̂. The matrix K̂T K̂ + 2β · LΓ is a real symmetric positivedefinite matrix, proved in Appendix C. Further, one can verify
if (20) is well-conditioned numerically via a Power Method
[33] to check its condition number1 : running Power Method
twice to compute the maximum and minimum eigenvalues of
K̂T K̂+2β·LΓ and checking the condition number λmax /λmin .
We found that the matrix typically has small condition number
in our experiments. In the rare case when the condition number
is large, for stability we can add an iterative refinement term
I and solve (20) iteratively, details described in pg.146 [34].
Since the left-hand-side matrix is sparse, positive definite and symmetric, we can solve (20) efficiently using the
Conjugate Gradient (CG) method [35]. In practice, we can
implement the K̂x as 2D convolution and accelerate it with
Fast Fourier Transform (FFT), and implement LΓ x as locally
graph filter, instead of matrix computation.
Algorithm 2 Solving (18)
Input: Blurry image b and estimated kernel k̂.
1: Initialize LΓ as an unweighted graph Laplacian.
2: while not converge do
Update x̂ by solving (20) with CG.
Update LΓ = LΓ (x̂) using (12) and (13).
endwhile
Output: Restored skeleton image x̂.
B. Blur Kernel Estimation
To solve k given x̂, we make a slight modification by
solving k in the gradient domain to avoid artifacts [36], [37].
The optimization (17) becomes:
1
k̂ = arg min k∇x̂ ⊗ k − ∇bk22 + µkkk22
2
k
(21)
where ∇ is the gradient operator. (21) is a quadratic convex
function and has a closed-form solver like deconvolution. We
accelerate the solver via FFT [36]. After obtaining k̂, we
threshold
P the negative elements to zeros and normalize k̂ to
ensure i k̂i = 1.
The rationale for successful kernel estimation with skeleton
image x̂ is that (21) is an over-determined function. Because
the kernel k is much smaller than the image x, the skeleton
image x̂ with restored sharp edges is enough for kernel
estimation.
1 https://en.wikipedia.org/wiki/Condition
number#Matrices
C. Acceleration for Specific Gaussian Blur Deblurring
Gaussian blur is a widely-assumed blur type in image
restoration applications, such as out-of-focus deblurring or
image super-resolution [38]–[40]. Assuming we know a priori
that the blur type is Gaussian or close to Gaussian, we propose
an acceleration for blind Gaussian blur deblurring.
In Algorithm 1, solving (18) with general blur kernel k̂ takes
most of the running time. Under the assumption of Gaussian
blur, we replace k̂ (or K̂) with graph filter I + a · LΓ , where
LΓ is first initialized as an unweighted graph Laplacian. The
filter I + a · LΓ with a < 0 is a smoothing process, which can
be considered as an approximation of Gaussian blur. To set a
suitable initial value for parameter a, we manually blur several
sharp images with Gaussian blurs σb ∈ (0, 2], as σb ∈ (0, 2] are
commonly used in practice. We learn the optimal a = −0.07
from the sharp and blurred image pairs using the least square
method,
a = arg min k(I + a · LΓ )X − Yk22 .
a
(22)
where matrix X = [x1 , x2 , ..., xn ] represents n sharp images,
matrix Y = [y1 , y2 , ..., yn ] represents corresponding blurred
images.
With I + a · LΓ , the skeleton image restoration function (18)
is modified to (23),
1
x̂ = arg min k(I + a · LΓ )x − bk22 + βkxkRGT V . (23)
2
x
The advantage of (23) is that I + a · LΓ and graph Laplacian
xT LΓ x in (19) now share the same graph frequency bases. The
closed-form solution (20) now becomes:
g(LΓ )
b
x̂ =
g 2 (LΓ ) + 2β · LΓ
g(ΛΓ )
= UΓ
UTΓ b
(24)
g 2 (ΛΓ ) + 2β · ΛΓ
where g(X) = I + a · X. (24) is a polynomial graph filter to
signal b and can be implemented with an accelerated Lanczos
method [16], which is faster than solving (20) with CG for this
specific problem. Lanczos method computes an orthonormal
basis VZ = [v1 , ..., vZ ] of the Krylov subspace KZ (LΓ , b) =
span{b, LΓ b, ..., LZ−1
b} and the corresponding symmetric
Γ
scalar tridiagonal matrix HZ :
α 1 β2
β2 α2 β3
.
∗
.
.
β3 α3
VZ LΓ VZ = HZ =
(25)
.
.
..
.. β
M
βM αM
The approximation of x̂ with order Z Lanczos method is:
x̂ = f (LΓ )b ≈ kbk2 VZ f (HZ )e1 := fZ ,
(26)
a = arg min k(I + a · LΓ )x̂ − bk22 .
(27)
where e1 ∈ RZ is the first unit vector. f (HZ ) is inexpensive
given Z N .
We update x̂, LΓ = LΓ (x̂) and parameter a using (27)
iteratively until convergence,
a
8
Algorithm 3 Accelerated Blind Gaussian Blur Deblurring
Input: Blurry image b and kernel size h × h.
1: Initialize LΓ as an unweighted graph Laplacian.
Initialize blur with I + a · LΓ smoothing.
2: Computing x̂ by solving (23):
while not converge do
Update x̂ using Lanczos method (25) and (26).
Update LΓ = LΓ (x̂) using (12) and (13).
Update a using (27).
endwhile
3: Compute k̂ by solving (21).
Output: Estimated blur kernel k̂ and skeleton image x̂.
Afterwards, a satisfactory skeleton image x̂ can be restored
and then we can compute the blur kernel k̂ using (21), as
shown in Algorithm 3.
VI. E XPERIMENTS AND D ISCUSSIONS
In this section, we set up comprehensive experiments to
verify the effectiveness of our proposed algorithms in solving
various blind image deblurring problems. The proposed algorithms are evaluated on three kinds of blurred cases, including
artificial blurred database, real motion blurred images, and
Gaussian blurred images. For all cases, we compare the
performance of the proposed algorithms with the best existing
blind image deblurring algorithms in the literature. We first evaluate the performance of Algorithm 1 on the artificial blurred
database [26], which is a widely used database appropriate for
both qualitative and quantitative assessment. Further, we apply
Algorithm 1 on real motion blurred images to demonstrate its
practicality. In the third part, we consider specific Gaussian
blur introduced by manual blurring or image scaling. We
evaluate the accelerated Algorithm 3 in this part. All the
experiments are implemented on the Matlab 2015a platform
with i7-4765T CPU.
We tune the parameters of the proposed Algorithm 1 based
on the artificial blurred database [26] and find them satisfactory for almost all the cases. The down-sampling factor for
coarse-to-fine strategy is set to log2 3. We construct a fourneighbour adjacency graph on the image as a trade-off between
performance and computation complexity. In (2), σ = 0.1. In
(17), β = 0.01 and µ = 0.05. The same parameters are used
for Algorithm 3. In blind image deblurring, the kernel size is
unknown and is an important parameter. For fair comparisons,
we set the same kernel size for all the algorithms to estimate
the blur kernel in each case. Then, we use the same non-blind
image deblurring algorithm to reconstruct sharp images with
estimated blur kernels. The detail settings are described in each
experiment.
A. Artificial Blurred Database
We evaluate the qualitative and quantitative performance of
the proposed algorithm on a large database introduced by Sun
et al. [26]. The database consists of 640 gray-scale blurry
images (typically 1024×768), which were made by convolving
80 high-resolution sharp images with 8 blur kernels from Levin
et al. [3] and were added 1% white Gaussian noise.
TABLE I
Q UANTITATIVE COMPARISONS OF ALL METHODS OVER THE ENTIRE
DATABASE .
Blind Deblurring
Method
Cho et al. [43]
Krishnan et al. [42]
Levin et al. [41]
Cho & Lee [36]
Xu & Jia [37]
Sun et al. [26]
Michaeli & Irani [44]
Lai et al. [45]
Ours + [18]
Pan et al. [6]
Ours + [4]
Mean error
ratio
28.1
11.7
6.6
8.7
3.6
2.5
2.6
2.1
2.0
1.6
1.5
Worst error
ratio
165.0
133.2
40.9
111.1
65.3
30.5
9.3
17.9
9.2
8.8
5.1
Success rate
(r ≤ 5)
11.7%
24.8%
46.7%
65.5%
85.8%
93.4%
95.9%
97.3%
99.7%
99.1%
99.8%
We compare with nine best existing blind image deblurring
algorithms, i.e., Levin et al. [41], Krishnan et al. [42], Cho
et al. [43], Cho & Lee [36], Xu & Jia [37], Sun et al. [26],
Michaeli & Irani [44], Lai et al. [45] and Pan et al. [6]. The
results of the first eight algorithms, except Pan et al. [6], are
offered by Sun et al. [26] or their authors on their websites.
These algorithms estimate blur kernels and then use non-blind
deblurring algorithm [18] to restore the latent sharp images as
the final step. In the experiments, kernel size is set to 51×51
by all these algorithms. The code of Pan et al. [6] is provided
online, so we run their code on the database. To fairly compare
with Pan et al. [6], we do not modify their codes but use the
same setting as theirs to run the proposed algorithm again, i.e.,
kernel size = 51 × 51 and use sophisticated [4] as the final
non-blind deblurring algorithm.
Considering the upper-bound performance of blind image
deblurring is non-blind image deblurring with ground-truth
blur kernel, we measure the quantitative performances of all
the algorithms with error ratio, introduced by Levin et al. [3],
r=
kx − xk̂ k2
,
kx − xk k2
(28)
where xk̂ and xk are the images restored by estimated kernel
k̂ and ground-truth kernel k, respectively. x is the groundtruth sharp image. r ≥ 1 and the smaller, the better. Same as
Michaeli & Irani [44], we assume that r ≤ 5 is the threshold
to decide the success of deblurred results.
Fig. 8 reports the fractions of images that can be restored
under different error ratios. Our algorithm and [6], [26], [44],
[45] are close when r ≤ 2, but ours increases faster and
quickly becomes the best when r > 2. We also introduce
three statistical measures of error ratios, i.e., the mean and
worst error ratio, and success rate, in Table. I. The proposed
algorithm is superior to the competing algorithms in all three
measures. Apart from quantitative assessments, Fig. 9 shows
some qualitative examples of challenging cases. The proposed
algorithm achieves more robust results on these cases, compared with the competing algorithms.
B. Real Motion Blurred Images
Besides artificial blurred database, we further apply the
proposed algorithm on real motion blurred images. For real
9
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 9. Three Deblurring Examples from Artificial Database. (a) Blurry Image. (b) Sun et al. [26]. (c) Michaeli & Irani [44]. (d) Lai et al. [45]. (e) Pan et
al. [6]. (f) The proposed algorithm. The images are better viewed in full size on computer screen.
better than Krishnan et al. [42], Levin et al. [41] and Michaeli
& Irani [44], and is competitive against Pan et al. [6] in
experiments, as shown in Fig. 10–13.
1
0.9
Fractions of Images
0.8
0.7
Michaeli & Irani
Sun et al.
Xu & Jia
Levin et al.
Krishnan et al.
Cho & Lee
Cho et al.
Lai et al.
Ours+[16]
Pan et al.
Ours+[3]
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1
2
3
4
5
6
7
8
9
10
Error Ratio
Fig. 8. Cumulative distribution of error ratios on artificial blurry database.
Each algorithm estimates blur kernels and uses [18] (solid line) or [4] (dash
line) as the final step non-blind deblurring.
images, the blurred cases are usually more complicated, including different depth of field (Fig. 10), mixture of human and
scenes (Fig. 11), more complex and severer motions (Fig. 12,
13), etc. We compare our algorithm with recent methods, of
which the implementations are available, i.e., Krishnan et al.
[42], Levin et al. [41], Michaeli & Irani [44] and Pan et al. [6].
In each experiment, all the algorithms are applied to estimate the blur kernel and then we use sophisticated nonblind image deblurring algorithm in [4] to restore latent sharp
images. The proposed algorithm can robustly estimate blur
kernels and results in fewer artifacts in the restored images.
The visual comparisons show that our algorithm is obviously
C. Gaussian Blurred Images
In Sec. V-C, we propose a specific acceleration for blind
Gaussian blur deblurring in Algorithm 3. The experiments
are set up by considering both the manual Gaussian blur and
the unknown blur introduced by image scaling. In Fig. 14,
the ground-truth image is blurred by Gaussian blur with
σb = 1.85. In Fig. 15, the unknown blur is introduced before
2× bilinear downsampling using “imresize” Matlab function.
We apply Michaeli & Irani [44], Pan et al. [6], Algorithm 1
and Algorithm 3 to estimate the blur kernels. Afterwards, we
use the classic hyper-Laplacian based non-blind deblurring
algorithm [17] to restore the latent sharp image. The restored
results, PSNRs and running time are reported in Fig. 14–15
and Table. II, respectively.
Algorithm 3 can estimate a blur kernel that is similar to
the ground-truth kernel, and the restored images are obviously sharpened without artifacts, as shown in Fig. 14–15.
Further, for the same or slightly better PSNR performance,
the accelerated Algorithm 3 is significantly faster than the
competing algorithms in Table. II. We note that the kernel
restored by Algorithm 1 is the most similar to the groundtruth kernel. Overall, both the proposed algorithms outperform
the competing methods in the specific Gaussian deblurring
experiments.
VII. C ONCLUSION
The proposed RGTV is an effective prior to promote image
sharpness and penalize blurriness. In this paper, we analyze its
10
(a)
(c)
(f)
Fig. 10. Real Blind Motion Deblurring Example. Image size: 618 × 464,
kernel size: 69 × 69. (a) Blurry image. (b) Krishnan et al. [42]. (c) Levin
et al. [41]. (d) Michaeli & Irani [44]. (e) Pan et al. [6]. (f) The proposed
Algorithm 1. The images are better viewed in full size on computer screen.
TABLE II
Q UANTITATIVE PERFORMANCE OF ALGORITHM 3 COMPARED WITH
ALGORITHM 1 ON SPECIFIC G AUSSIAN - LIKE BLURRY IMAGES .
Fig. 14
Fig. 15
(b)
(c)
(d)
(e)
(f)
(d)
(e)
Examples
(a)
(b)
Images
Blurred
Micheali & Irani [44]
Pan et al. [6]
Alg. 1
Alg. 3
Up-sampled
Micheali & Irani [44]
Pan et al. [6]
Alg. 1
Alg. 3
PSNR(dB)
26.7
30.6
30.4
30.8
30.8
32.1
32.6
33.0
33.9
33.8
Time(s)
1150.0
421.3
30.0
7.8
1142.0
428.3
30.0
7.9
advantages in both nodal and graph frequency domain. In the
nodal domain, RGTV can promote the bi-modal weight distribution of sharp images from blurry observations. In the graph
frequency domain, RGTV enjoys desirable spectral properties,
including robustness to noise and blur, strong PWS-preserving
filtering and sharpness promotion. We design a robust blind
image deblurring algorithm using RGTV. Experimental results
demonstrate that the proposed algorithm can deal with various
blind image deblurring scenarios, and the reconstructed sharp
results are visually and numerically better than the state-ofthe-art methods.
The proposed algorithm with RGTV mainly focuses on
uniform blur, i.e., blur follows the convolution model (1)
in the data fidelity term of (17). Leveraging our theoretical
Fig. 11. Real Blind Motion Deblurring Example. Image size: 800 × 532,
kernel size: 69 × 69. (a) Blurry image. (b) Krishnan et al. [42]. (c) Levin
et al. [41]. (d) Michaeli & Irani [44]. (e) Pan et al. [6]. (f) The proposed
Algorithm 1. The images are better viewed in full size on computer screen.
analysis on RGTV in both nodal and graph spectral domains,
one can easily employ RGTV to solve non-uniform deblurring
problems also, as long as there is a more sophisticated nonuniform blur model substituting for the current convolution
model. In the future, we will extend our algorithm to solve
non-uniform camera motion deblurring problems, by combining RGTV with the existing camera geometric model [46]–
[48]. Moreover, as there are many fast-developing imaging
technologies, such as light field imaging and multi-spectral
imaging, we also would like to use RGTV for blind deblurring
in these advanced imaging systems.
A PPENDIX A
D ERIVATION OF WEIGHT FUNCTION OF GTV
As GTV is non-differentiable, we take the sub-derivative of
GTV, resulting in the following sub-differential:
( w
i,j
· (xi − xj ), xi 6= xj ;
(∂kxkGT V )i,j = |xi −xj |
(29)
[−1, 1],
x i = xj .
Because the denominator of the weight term in the subdifferential (29), |xi − xj |, goes to 0 as xi approaches xj , for
numerical stability we approximate (29) with the following
when |xi − xj | is less than a small ; i.e.,
( w
i,j
−xj | · (xi − xj ), |xi − xj | ≥ ;
(∂kxkGT V )i,j = |xwii,j
· (xi − xj ),
|xi − xj | < .
(30)
11
(a)
(c)
(a)
(b)
(c)
(d)
(e)
(f)
(b)
(d)
Fig. 13. Real Blind Motion Deblurring Example. Image size: 1229 × 825,
kernel size: 95 × 95. (a) Blurry image. (b) Krishnan et al. [42]. (c) Levin
et al. [41]. (d) Michaeli & Irani [44]. (e) Pan et al. [6]. (f) The proposed
Algorithm 1. The images are better viewed in full size on computer screen.
(e)
(f)
Fig. 12. Real Blind Motion Deblurring Example. Image size: 972 × 966,
kernel size: 69 × 69. (a) Blurry image. (b) Krishnan et al. [42]. (c) Levin
et al. [41]. (d) Michaeli & Irani [44]. (e) Pan et al. [6]. (f) The proposed
Algorithm 1. The images are better viewed in full size on computer screen.
where is a small constant close to 0. (30) is equivalent to
earlier (11) and (12). (30) can actually be interpreted as the
derivative of an upper-bound function of GTV:
(
wi,j |xi − xj |,
|xi − xj | ≥ ;
(kxkGT V )i,j = wi,j
wi,j
2
2 (xi − xj ) + 2 , |xi − xj | < .
(31)
where (kxkGT V )i,j is the same as GTV (kxkGT V )i,j except
when |xi − xj | < , when (kxkGT V )i,j has a quadratic
w
behavior with minimum at i,j
2 . The l1 adjacency matrix Γ
we use for GTV spectral analysis is based on this upper-bound
weight function (31), which is reasonable, since (31) satisfies
lim→0 kxkGT V = kxkGT V .
A PPENDIX B
B OUNDEDNESS OF EIGENVALUES OF LΓ
Lower bounded: Because adjacency matrix Γ is a nonnegative matrix and the definition of graph Laplacian matrix
(a)
(c)
(b)
(d)
(e)
(f)
Fig. 14. Unknown Gaussian Blur Example. Image size: 768 × 512, kernel
size: 7×7, σb = 1.85. (a) Blurry image and ground-truth kernel. (b) Restored
image by the proposed Algorithm 3. (c) Restored kernel by Michaeli & Irani
[44]. (d) Restored kernel by Pan et al. [6]. (e) Restored kernel by the proposed
Algorithm 1. (f) Restored kernel by the proposed Algorithm 3.
LΓ (13), LΓ is a PSD matrix. Therefore, 0 is the lowest
eigenvalue of LΓ with 1 its eigenvector.
Upper bounded: We prove the upper boundedness based on
Gershgorin Circle Theorem.
Theorem 1 (Gershgorin Circle Theorem). Let A be a n × n
complex matrix P
with entries ai,j . For i ∈ {1, ..., n}, Ri is
defined as Ri = j6=i |ai,j |. Every eigenvalue of A lies within
12
(a)
For image blur caused by out-of-focus
or motion, each
P
element of blur kernel ki ≥ 0 and
i ki = 1. Therefore,
K̂ is a non-negative matrix and K̂ · 1 = 1, i.e., λ = 1
is an eigenvalue of matrix K̂ and corresponding eigenvector
is a constant vector 1. This property is intuitive; blurring
a constant signal results in the signal itself under suitable
boundary conditions. Therefore, 1 6∈ S1 and S1 ∩ S2 = ∅.
S1 ∩ S2 = ∅ means that the quadratic forms of two positive
semi-definite matrices, i.e., xT (K̂T K̂)x and xT LΓ x, cannot
be 0 at the same time. Therefore, K̂T K̂ + 2β · LΓ (β > 0) is
a positive definite matrix.
(b)
R EFERENCES
(c)
(d)
(e)
(f)
Fig. 15. Unknown Blur Introduced Before Image Scaling. Image size:
768 × 512. (a) 2× up-sampled low-resolution image. (b) Restored image
by the proposed Algorithm 3. (c) Restored kernel by Michaeli & Irani [44].
(d) Restored kernel by Pan et al. [6]. (e) Restored kernel by the proposed
Algorithm 1. (f) Restored kernel by the proposed Algorithm 3.
at least one of the Gershgorin discs D(ai,i , Ri ), i.e.,
|λ − ai,i | ≤ Ri
(32)
P
di maxj (wi,j )
≤ di
For matrix A = LΓ , ai,i = j6=i γi,j ≤
P
P
d maxj (wi,j )
and Ri = j6=i |ai,j | = j6=i γi,j ≤ i
≤ di . di is
the degree of node i. Therefore, we have
2di
),
Since di ≤ n, λ is upper-bounded for fixed .
max λi ≤ max(ai,i + Ri ) ≤ max(
i
i
i
(33)
A PPENDIX C
P ROOF OF P OSITIVE D EFINITENESS
Proof : We prove xT (K̂T K̂ + 2β · LΓ )x > 0 (β > 0), for
∀x ∈ Rn and x 6= 0. Here we set β = 0.5 without loss of
generality.
xT (K̂T K̂ + LΓ )x = xT (K̂T K̂)x + xT LΓ x,
X
= kK̂xk22 +
γi,j (xi − xj )2
(34)
i,j
P
In (34), kK̂xk22 ≥ 0. i,j γi,j (xi − xj )2 ≥ 0, for γi,j ≥ 0
given the weight definition (12). Therefore, xT (K̂T K̂ + LΓ )x
is a positive semi-definite matrix.
We further prove xT (K̂T K̂ + LΓ )x is strictly positive. We
define two sets:
S1 = {x|xT (K̂T K̂)x = 0, x ∈ Rn , x 6= 0}
T
n
S2 = {x|x LΓ x = 0, x ∈ R , x 6= 0}
(35)
(36)
As both K̂T K̂ and LΓ are already positive semi-definite, we
only need to prove S1 ∩ S2 = ∅.
LΓ is a combinatorial graph Laplacian of a connected graph;
by definition it must have one constant eigenvector corresponding to eigenvalue 0 and the multiplicity of eigenvalue 0 is 1.
Therefore, S2 = {1}, where 1 is a constant vector with all
elements 1.
[1] P. C. Hansen, J. G. Nagy, and D. P. O’Leary, Deblurring images :
matrices, spectra, and filtering.
SIAM, Society for Industrial and
Applied Mathematics, 2006.
[2] M. S. C. Almeida and L. B. Almeida, “Blind and semi-blind deblurring
of natural images,” IEEE Transactions on Image Processing, vol. 19,
no. 1, pp. 36–52, 2010.
[3] A. Levin, Y. Weiss, F. Durand, and W. T. Freeman., “Understanding
blind deconvolution algorithms,” IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 33, no. 12, pp. 2354–2367, Dec 2011.
[4] J. Pan, Z. Hu, Z. Su, and M. H. Yang, “L0-regularized intensity and
gradient prior for deblurring text images and beyond,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PP, no. 99, pp.
1–1, 2016.
[5] W. Ren, X. Cao, J. Pan, X. Guo, W. Zuo, and M. H. Yang, “Image
deblurring via enhanced low-rank prior,” IEEE Transactions on Image
Processing, vol. 25, no. 7, pp. 3426–3437, 2016.
[6] J. Pan, D. Sun, H. Pfister, and M.-H. Yang, “Blind image deblurring
using dark channel prior,” in Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition, June 2016.
[7] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, “The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular
domains,” in IEEE Signal Processing Magazine, vol. 30, no.3, May
2013, pp. 83–98.
[8] W. Hu, G. Cheung, and M. Kazui, “Graph-based dequantization of
block-compressed piecewise smooth images,” IEEE Signal Processing
Letters, vol. 23, no. 2, pp. 242–246, Feb 2016.
[9] J. Pang and G. Cheung, “Graph Laplacian regularization for inverse
imaging: Analysis in the continuous domain,” in IEEE Transactions on
Image Processing, vol. 26, no.4, April 2017, pp. 1770–1785.
[10] X. Liu, G. Cheung, X. Wu, and D. Zhao, “Random walk graph
laplacian-based smoothness prior for soft decoding of jpeg images,”
IEEE Transactions on Image Processing, vol. 26, no. 2, pp. 509–524,
Feb 2017.
[11] A. Kheradmand and P. Milanfar, “A general framework for regularized,
similarity-based image restoration,” IEEE Transactions on Image Processing, vol. 23, no. 12, pp. 5136–5151, 2014.
[12] A. Elmoataz, O. Lezoray, and S. Bougleux, “Nonlocal discrete regularization on weighted graphs: A framework for image and manifold
processing,” IEEE Transactions on Image Processing, vol. 17, no. 7,
pp. 1047–1060, 2008.
[13] M. Hidane, O. Lzoray, and A. Elmoataz, “Nonlinear multilayered
representation of graph-signals,” Journal of Mathematical Imaging and
Vision, vol. 45, no. 2, pp. 114–137, 2013.
[14] C. Couprie, L. Grady, L. Najman, J.-C. Pesquet, and H. Talbot, “Dual
constrained tv-based regularization on graphs,” SIAM Journal on Imaging Sciences, vol. 6, no. 3, pp. 1246–1273, 2013.
[15] P. Berger, G. Hannak, and G. Matz, “Graph signal recovery via primaldual algorithms for total variation minimization,” IEEE Journal of
Selected Topics in Signal Processing, vol. 11, no. 6, pp. 842–855, Sept
2017.
[16] A. Susnjara, N. Perraudin, D. Kressner, and P. Vandergheynst, “Accelerated filtering on graphs using lanczos method,” arXiv preprint
arXiv:1509.04537, 2015.
[17] D. Krishnan and R. Fergus, “Fast image deconvolution using hyperlaplacian priors,” in Proceedings of Neural Information Processing
Systems, 2009, Conference Proceedings, pp. 1033–1041.
13
[18] D. Zoran and Y. Weiss, “From learning models of natural image patches
to whole image restoration,” in Proceedings of IEEE International
Conference on Computer Vision, 2011, Conference Proceedings, pp.
479–486.
[19] N. Wiener, Extrapolation, interpolation, and smoothing of stationary
time series, with engineering applications. Technology Press of the
Massachusetts Institute of Technology, 1950.
[20] W. H. Richardson, “Bayesian-based iterative method of image restoration,” Journal Of the Optical Society Of America, vol. 62, no. 1, 1972.
[21] L. B. Lucy, “An iterative technique for the rectification of observed
distributions,” Astronomical Journal, vol. 79, no. 6, pp. 745–754, Jun.
1974.
[22] L. I. Rudin, S. Osher, and E. Fatemi, “Nonlinear total variation based
noise removal algorithms,” Physica D: Nonlinear Phenomena, vol. 60,
no. 1-4, pp. 259–268, 1992.
[23] A. Levin, R. Fergus, F. Durand, and W. T. Freeman, “Image and depth
from a conventional camera with a coded aperture,” Acm Transactions
on Graphics, vol. 26, no. 3, 2007.
[24] J.-F. Cai, H. Ji, C. Liu, and Z. Shen, “Framelet-based blind motion deblurring from a single image,” IEEE Transactions on Image Processing,
vol. 21, no. 2, pp. 562–572, 2012.
[25] L. Xu, S. Zheng, and J. Jia, “Unnatural l0 sparse representation for
natural image deblurring,” in Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2013, Conference Proceedings,
pp. 1107–1114.
[26] L. Sun, S. Cho, J. Wang, and J. Hays, “Edge-based blur kernel estimation
using patch priors,” in Proceedings of IEEE International Conference
on Computational Photography, 2013.
[27] Y. Bai, G. Cheung, X. Liu, and W. Gao, “Blind image deblurring via
reweighted graph total variation,” in Proceedings of IEEE International
Conference on Acoustics, Speech and Signal Processing, 2018, Conference Proceedings.
[28] C. Tomasi and R. Manduchi, “Bilateral filtering for gray and color
images,” in Proceedings of the IEEE International Conference on
Computer Vision, Bombay, India, 1998.
[29] L. Xu, Q. Yan, Y. Xia, and J. Jia, “Structure extraction from texture via
relative total variation,” ACM Trans. Graph., vol. 31, no. 6, pp. 1–10,
2012.
[30] L. Xu, C. Lu, Y. Xu, and J. Jia, “Image smoothing via l0 gradient
minimization,” ACM Trans. Graph., vol. 30, no. 6, pp. 1–12, 2011.
[31] W. Hu, X. Li, G. Cheung, and O. Au, “Depth map denoising using graphbased transform and group sparsity,” in IEEE International Workshop on
Multimedia Signal Processing, Pula, Italy, October 2013.
[32] R. Fergus, B. Singh, A. Hertzmann, S. T. Roweis, and W. Freeman,
“Removing camera shake from a single photograph,” ACM Transactions
on Graphics, vol. 25, pp. 787–794, 2006.
[33] J. H. Wilkinson, The algebraic eigenvalue problem. Clarendon Press
Oxford, 1965, vol. 87.
[34] N. Parikh and S. Boyd, “Proximal algorithms,” Foundations and Trends
in Optimization, vol. 1, no. 3, pp. 127–239, 2014.
[35] S. Boyd and L. Vandenberghe, Convex optimization.
Cambridge
university press, 2004.
[36] S. Cho and S. Lee, “Fast motion deblurring,” ACM Transactions on
Graphics, vol. 28, no. 5, p. article no. 145, 2009.
[37] L. Xu and J. Jia, “Two-phase kernel estimation for robust motion deblurring,” in Proceedings of European Conference on Computer Vision,
Berlin, Heidelberg, 2010, pp. 157–170.
[38] P. Sung Cheol, P. Min Kyu, and K. Moon Gi, “Super-resolution image
reconstruction: a technical overview,” IEEE Signal Processing Magazine,
vol. 20, no. 3, pp. 21–36, 2003.
[39] S. Farsiu, M. D. Robinson, M. Elad, and P. Milanfar, “Fast and robust
multiframe super resolution,” IEEE Transactions on Image Processing,
vol. 13, no. 10, pp. 1327–1344, 2004.
[40] Q. Shan, Z. Li, J. Jia, and C. K. Tang, “Fast image/video upsampling,”
Acm Transactions on Graphics, vol. 27, no. 5, pp. 32–39, 2008.
[41] A. Levin, Y. Weiss, F. Durand, and W. T. Freeman, “Efficient marginal
likelihood optimization in blind deconvolution,” in Proceedings of IEEE
Conference on Computer Vision and Pattern Recognition, Washington,
DC, USA, 2011, pp. 2657–2664.
[42] D. Krishnan, T. Tay, and R. Fergus, “Blind deconvolution using a
normalized sparsity measure,” in Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition, Washington, DC, USA, 2011,
pp. 233–240.
[43] T. S. Cho, S. Paris, B. K. P. Horn, and W. T. Freeman, “Blur kernel estimation using the radon transform,” in Proceedings of IEEE Conference
on Computer Vision and Pattern Recognition, June 2011, pp. 241–248.
[44] T. Michaeli and M. Irani, “Blind deblurring using internal patch recurrence,” in Proceedings of European Conference on Computer Vision,
Cham, 2014, pp. 783–798.
[45] L. Wei-Sheng, D. Jian-Jiun, L. Yen-Yu, and C. Yung-Yu, “Blur kernel
estimation using normalized color-line priors,” in 2015 IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2015, Conference
Proceedings, pp. 64–72.
[46] M. Hirsch, S. Sra, B. Scholkopf, and S. Harmeling, “Efficient filter flow
for space-variant multiframe blind deconvolution,” in 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition,
2010, Conference Proceedings, pp. 607–614.
[47] M. Hirsch, C. J. Schuler, S. Harmeling, and B. Scholkopf, “Fast removal
of non-uniform camera shake,” in 2011 International Conference on
Computer Vision, 2011, Conference Proceedings, pp. 463–470.
[48] O. Whyte, J. Sivic, A. Zisserman, and J. Ponce, “Non-uniform deblurring
for shaken images,” International Journal of Computer Vision, vol. 98,
no. 2, pp. 168–186, 2012.
| 1cs.CV
|
arXiv:1611.06342v1 [cs.LG] 19 Nov 2016
Quantized neural network design under
weight capacity constraint
Sungho Shin, Kyuyeon Hwang, and Wonyong Sung
Department of Electrical and Computer Engineering
Seoul National University
Seoul, 08826 Korea
[email protected], [email protected], [email protected]
Abstract
The complexity of deep neural network algorithms for hardware implementation
can be lowered either by scaling the number of units or reducing the word-length of
weights. Both approaches, however, can accompany the performance degradation
although many types of research are conducted to relieve this problem. Thus,
it is an important question which one, between the network size scaling and
the weight quantization, is more effective for hardware optimization. For this
study, the performances of fully-connected deep neural networks (FCDNNs) and
convolutional neural networks (CNNs) are evaluated while changing the network
complexity and the word-length of weights. Based on these experiments, we
present the effective compression ratio (ECR) to guide the trade-off between the
network size and the precision of weights when the hardware resource is limited.
1
Introduction
Deep neural networks (DNNs) begin to find many real-time applications, such as speech recognition,
autonomous driving, gesture recognition, and robotic control (Sak et al., 2015; Chen et al., 2015;
Shin & Sung, 2016; Corradini et al., 2015). Recent works show that the precision required for
implementing fully-connected deep neural networks (FCDNNs), convolutional neural networks
(CNNs) or recurrent neural networks (RNNs) needs not be very high, especially when the quantized
networks are trained again to learn the effects of lowered precision. In the fixed-point optimization
examples shown in Hwang & Sung (2014), neural networks with ternary weights showed quite good
performance which was close to that of floating-point arithmetic. However, the performance of DNNs
usually degrades when the weights are represented using a very low precision. Thus, we have a
question whether it might be a better option to reduce the network size, instead of severely quantizing
the weights, for efficient implementations.
In this work, we compare the performance of FCDNNs and CNNs under two constraints for hardware
implementation, one is reducing the network size and the other is lowering the precision of the
weights. We conduct the experiments with FCDNNs for phoneme recognition and CNNs for image
classification. To control the network complexity, the number of units in each hidden layer is varied
in the FCDNN. For the CNN, the number of feature maps for each layer is changed. The retraining
based quantization algorithm is used for fixed-point optimization of weights (Hwang & Sung, 2014).
Based on the experiments, we propose a metric called the effective compression ratio (ECR) that
compares the complexity of floating-point and fixed-point networks showing the same performance.
This analysis intends to provide a guideline to network size and word-length determination for
efficient hardware implementation of deep neural networks (DNN).
2
Related Work
Fixed-point design of DNNs with ternary weights show quite good performances that are very close
to the floating-point results (Hwang & Sung, 2014; Anwar et al., 2015a; Shin et al., 2016). The
ternary weight based FCDNN is used for VLSI implementations, by which the algorithms can operate
with only on-chip memory consuming very low power (Kim et al., 2014). The CNN is implemented
by XNOR-bitcounting operations (Rastegari et al., 2016). Binary weight based deep neural network
design is also studied (Courbariaux et al., 2015). Pruned floating-point weights are utilized for
efficient GPU-based implementations, where small valued or less important weights are forced to
zero to reduce the number of arithmetic operations and the memory space for weight storage (Yu
et al., 2012; Han et al., 2015; Anwar et al., 2015b).
Most of the above works are experimented using large size neural networks. However, mobile or
embedded portable devices have limited resources, and thus small size fixed-point networks showing
good performances are very needed.
3
Fixed-Point FCDNN and CNN Design
This section explains the design of FCDNN and CNN with varying network complexity and weight
precision.
3.1
FCDNN and CNN Design
In this work, we examine an FCDNN for phoneme recognition and a CNN for image classification.
The reference DNN has four hidden layers. Each of the hidden layers has Nh units; the value of Nh
is altered to control the complexity of the network. We conduct experiments with Nh value of 32,
64, 128, 256, 512, and 1024. The input layer of the network has 1,353 units to accept 11 frames
of a Fourier-transform-based filter-bank with 40 coefficients (+energy) distributed on a mel-scale,
together with their first and second temporal derivatives. The output layer consists of 61 softmax
units which correspond to 61 target phoneme labels (Mohamed et al., 2012). Phoneme recognition
experiments were performed on the TIMIT corpus.
The CNN used is for CIFAR-10 dataset (Krizhevsky & Hinton, 2009). It contains a training set of
50,000 and a test set of 10,000 images. We divided the training set to 40,000 images for training and
10,000 images for validation. The reference CNN has 3 convolution and max-pooling layers, a fully
connected hidden layer with 64 units, and the output with 10 softmax units. We control the number
of feature maps in each convolution layer. The reference size has 32-32-64 feature maps with a 5
by 5 kernel size as used in Krizhevskey (2014). To know the effects of network size variation, the
number of feature maps is reduced or increased. The configurations of the feature maps used for the
experiments are 8-8-16, 16-16-32, 32-32-64, 64-64-128, 96-96-192, and 128-128-256. Note that the
fully connected layer in the CNN is not changed.
3.2
Fixed-Point Optimization of DNNs
Reducing the word-length of weights brings several advantages in hardware based implementation of
neural networks. First, it lowers the arithmetic precision, and thereby reduces the number of gates
needed for multipliers. Second, the size of memory for storing weights is minimized, which would
be a big advantage when keeping them on a chip, instead of external DRAM or NAND flash memory.
Note that FCDNNs demand a very large number of weights. Third, reduced arithmetic precision or
minimization of off-chip memory accesses leads to low power consumption.
The fixed-point DNN algorithm design consists of three steps: floating-point training, direct quantization, and retraining of weights. Refer to Hwang & Sung (2014) for the details.
4
Analysis of Quantization Effects
The fixed-point performance of the FCDNN is shown in Figure 1a, where the number of hidden
units in each layer varies. The performance of direct 2 bits (ternary levels), direct 3 bits (7-levels),
retrain-based 2 bits, and retrain-based 3 bits are compared with the floating-point results. We can
2
find that the performance gap between the floating-point and the retrain-based fixed-point networks
converges very fast as the network size grows. Direct quantization does not show good results at any
network size. In this figure, the performance of the floating-point network almost saturates when the
network size is about 1024. Note that the TIMIT corpus that is used for training has only 3 hours
of data. Thus, the network with 1024 hidden units can be considered in the ‘training-data limited
region’. Here, the gap between the floating-point and the fixed-point networks almost vanishes when
the network is in the ‘training-data limited region’. However, when the network size is limited,
such as 32, 64, 128, or 256, there exists some performance gap between the floating-point and
highly quantized networks even if retraining on the quantized networks is performed. The similar
Phone error rate (%)
2-bit direct
3-bit direct
2-bit retrain
80
3-bit retrain
floating
60
40
32
64
128
256
512
Miss classification rate (%)
100
80
floating result
60
3bit direct
2bit retrain
3bit retrain
40
20
1,024
Size of the network
2 bit direct
16
32
64
128
192 256
Size of the network
(b) CNN
(a) FCDNN
Figure 1: Comparison of retrain-based and direct quantization. All weights are quantized with ternary
and 7-level weights. Note that each FCDNN has four hidden layers. In the figure (b), x-axis label
“16, 32, 64, 128, 192, 256” represents the number of feature map is “8-8-16, 16-16-32, 32-32-64,
64-64-128, 96-96-192, 128-128-256”.
experiments are conducted for the CNN with varying feature map sizes, and the results are shown in
Figure 1b. The configurations of the feature maps used for the experiments are 8-8-16, 16-16-32,
32-32-64, 64-64-128, 96-96-192, and 128-128-256. The size of the fully connected layer is not
changed. In this figure, the floating-point and the fixed-point performances also converge very fast as
the number of feature maps increases. The floating-point performance saturates when the feature
map size is 128-128-256, and the gap is less than 1% when comparing the floating-point and the
retrain-based 2-bit networks. This suggests that a fairly high-performance feature extraction can be
designed even using very low-precision weights if the number of feature maps can be increased.
5
Efficient DNN Design with Hardware Constraints
As the number of quantization levels decreases, the memory space needed is reduced at the cost of
sacrificing the accuracy. Therefore, there can be a trade-off between the network size reduction and
aggressive quantization. Figure 2a shows the framewise phoneme error rate on TIMIT corpus while
varying the layer size of FCDNNs with a various number of quantization bits from 2 to 8 bits. Note
that the network has four hidden layers containing the same number of units.
In this section, we propose a guideline for finding the optimal bit-widths when the desired accuracy
or the network size is given. Note that we assume 2n − 1 quantization levels are represented by n
bits (i.e. 2 bits are required for representing a ternary weight). For simplicity, all layers are quantized
with the same number of quantization levels. Based on this observation, we introduce a metric called
the effective compression ratio (ECR), which is defined as follows:
Size of a floating-point network
ECR =
(1)
Size of a quantized network showing the same performance
Figure 2b describes how to compare the hardware efficiency of floating-point and fixed-point networks.
In this figure, we assume the target performance of 32.87% which can be obtained using a floatingpoint DNN with the network size of 256. This graph shows that the target performance can be
obtained with the network size of N 1 when 3 bits weights are used, and that of N 2 when 2 bits
3
40
2-bit retrain
3-bit retrain
45
Phone error rate (%)
Phone error rate (%)
50
4-bit retrain
5-bit retrain
6-bit retrain
40
7-bit retrain
8-bit retrain
floating
35
2-bit retrain
3-bit retrain
floating
35
30
30
64
128
256
512
256
1,024
N1
N2
512
Size of the network
Size of the network
(a)
(b)
Figure 2: (a) Framewise phone error rate of phoneme recognition FCDNNs with respect to the size of
the networks for weights with quantization. (b) Obtaining an effective number of parameters for the
uncompressed network.
2-bit retrain
ECR
8
3-bit retrain
4-bit retrain
5-bit retrain
6
6-bit retrain
7-bit retrain
8-bit retrain
4
32
64
128
256
512
Size of the network
Figure 3: Effective compression ratio (ECR) of fixed-point networks with respect to the layer size of
reference floating-point networks.
quantization is employed. Thus, the compression obtained by 3-bit quantization can be roughly
figured as (32 × 2562 )/(3 × N 12 ). If N 1 is very close to 256, there can be about 10 (= 32/3) times
compression. But, if N 1 is 512, the compression drops to only about 2.5 because the number of
parameters of FCDNN is proportional to the square of the network size.
The ECRs for various network sizes and quantization bits are shown in Figure 3. The figure illustrates
that the 2-bit or (maybe 1-bit) quantization can lead to the best compression ratio when the target
performance is low, which means a high phoneme error rate in this example. However, for designing
a fairly high-performance network, increasing the network size with severe quantization does not
yield hardware efficient networks. The optimum number of bits for obtaining the performance
corresponding to that of 512 sizes floating-point DNNs is 4 or 5 bits. Further reducing the wordlength demands very increased size networks, and as a result, the total number of bits increases.
6
Conclusion
Hardware efficient deep neural networks can be designed either by lowering the number of units in
each layer or reducing the number of bits for weight quantization. We evaluate the performance of
fixed-point deep neural networks and analyze the trade-off between the complexity and the precision
of the weights. This study shows that low-performance hardware efficient DNNs can be designed
with severely quantized weights. In the low-performance region, the DNN performance increases
very rapidly as the network size grows. Thus, it is possible to compensate for the quantization effects
by slightly increasing the network size. However, for a high-performance DNN design, compensation
of quantization effects by increasing the network size is difficult, and thus severe quantization does
not lead to efficient hardware design. The effective compression ratio is given for a DNN design
when the network size and the precision vary.
4
Acknowledgments
This work was supported in part by the Brain Korea 21 Plus Project and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIP) (No. 2015R1A2A1A10056051).
References
Anwar, Sajid, Hwang, Kyuyeon, and Sung, Wonyong. Fixed point optimization of deep convolutional
neural networks for object recognition. In Acoustics, Speech and Signal Processing (ICASSP),
2015 IEEE International Conference on, pp. 1131–1135. IEEE, 2015a.
Anwar, Sajid, Hwang, Kyuyeon, and Sung, Wonyong. Structured pruning of deep convolutional
neural networks. arXiv preprint arXiv:1512.08571, 2015b.
Chen, Chenyi, Seff, Ari, Kornhauser, Alain, and Xiao, Jianxiong. Deepdriving: Learning affordance
for direct perception in autonomous driving. In Proceedings of the IEEE International Conference
on Computer Vision, pp. 2722–2730, 2015.
Corradini, Maria Letizia, Giantomassi, Andrea, Ippoliti, Gianluca, Longhi, Sauro, and Orlando,
Giuseppe. Robust control of robot arms via quasi sliding modes and neural networks. In Advances
and Applications in Sliding Mode Control systems, pp. 79–105. Springer, 2015.
Courbariaux, Matthieu, Bengio, Yoshua, and David, Jean-Pierre. Binaryconnect: Training deep
neural networks with binary weights during propagations. In Advances in Neural Information
Processing Systems, pp. 3105–3113, 2015.
Han, Song, Mao, Huizi, and Dally, William J. Deep compression: Compressing deep neural network
with pruning, trained quantization and huffman coding. CoRR, abs/1510.00149, 2, 2015.
Hwang, Kyuyeon and Sung, Wonyong. Fixed-point feedforward deep neural network design using
weights +1, 0, and -1. In Signal Processing Systems (SiPS), 2014 IEEE Workshop on, pp. 1–6.
IEEE, 2014.
Kim, Jonghong, Hwang, Kyuyeon, and Sung, Wonyong. X1000 real-time phoneme recognition VLSI
using feed-forward deep neural networks. In Acoustics, Speech and Signal Processing (ICASSP),
2014 IEEE International Conference on, pp. 7510–7514. IEEE, 2014.
Krizhevskey, A. CUDA-convnet, 2014.
Krizhevsky, Alex and Hinton, Geoffrey. Learning multiple layers of features from tiny images. 2009.
Mohamed, Abdel-rahman, Dahl, George E, and Hinton, Geoffrey. Acoustic modeling using deep
belief networks. Audio, Speech, and Language Processing, IEEE Transactions on, 20(1):14–22,
2012.
Rastegari, Mohammad, Ordonez, Vicente, Redmon, Joseph, and Farhadi, Ali. XNOR-Net: Imagenet
classification using binary convolutional neural networks. arXiv preprint arXiv:1603.05279, 2016.
Sak, Haşim, Senior, Andrew, Rao, Kanishka, and Beaufays, Françoise. Fast and accurate recurrent
neural network acoustic models for speech recognition. arXiv preprint arXiv:1507.06947, 2015.
Shin, Sungho and Sung, Wonyong. Dynamic hand gesture recognition for wearable devices with low
complexity recurrent neural networks. In Circuits and Systems (ISCAS), 2016 IEEE International
Symposium on, pp. 2274–2277. IEEE, 2016.
Shin, Sungho, Hwang, Kyuyeon, and Sung, Wonyong. Fixed-point performance analysis of recurrent
neural networks. In 2016 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), pp. 976–980. IEEE, 2016.
Yu, Dong, Seide, Frank, Li, Gang, and Deng, Li. Exploiting sparseness in deep neural networks for
large vocabulary speech recognition. In Acoustics, Speech and Signal Processing (ICASSP), 2012
IEEE International Conference on, pp. 4409–4412. IEEE, 2012.
5
| 9cs.NE
|
Asymptotic prime divisors over complete intersection
rings
arXiv:1403.6972v2 [math.AC] 11 Nov 2016
Dipankar Ghosh
Tony J. Puthenpurakal
27th March, 2014
Abstract
Let A be a local complete intersection ring. Let M, N be two finitely generated
A-modules and I an ideal of A. We prove that
[[
i>0 n>0
AssA ExtiA (M, N/I n N )
is a finite set. Moreover, we prove that there exist i0 , n0 > 0 such that for all
i > i0 and n > n0 , we have
2i0
n0
n
AssA Ext2i
A (M, N/I N ) = AssA ExtA (M, N/I N ) ,
2i0 +1
(M, N/I n0 N ) .
(M, N/I n N ) = AssA ExtA
AssA Ext2i+1
A
We also prove the analogous results for complete intersection rings which arise
in algebraic geometry. Further, we prove that the complexity cxA (M, N/I n N )
is constant for all sufficiently large n.
1
Introduction
Let A be a commutative Noetherian ring, I an ideal of A, and M a finitely generated Amodule. Brodmann [2] proved that the set of associated prime ideals AssA (M/I n M) is
independent of n for all sufficiently large n. Thereafter, L. Melkersson and P. Schenzel
generalized Brodmann’s result in [6, Theorem 1] by proving that
n
AssA TorA
i (M, A/I )
1
2010 Mathematics Subject Classification. Primary 13H10, 13D07; Secondary 13A02, 13A15.
Key words and phrases. Complete intersection rings; Graded rings and modules; Associate primes;
Ext and Tor; Eisenbud operators; Support variety.
2
1
is independent of n for all large n and for a fixed i > 0.
Later, D. Katz and E. West proved the above result in a more general way [5, 3.5];
if N is a finitely generated A-module, then for a fixed i > 0, the sets
n
AssA TorA
i (M, N/I N)
and
AssA ExtiA (M, N/I n N)
are stable for all large n. So, in particular, for a fixed i > 0,
[
AssA ExtiA (M, N/I n N)
n>0
is a finite set. In this context, the second author [7, page 368] raised a question about
what happens if we vary i (> 0) also? More precisely,
[[
(†)
is the set
AssA ExtiA (M, N/I n N) finite?
i>0 n>0
The motivation for the question (†) came from the following two questions. They
were raised by W. Vasconcelos [8, 3.5] and Melkersson and Schenzel [6, page 936]
respectively.
[
(1)
Is the set
AssA ExtiA (M, A) finite?
(2)
Is the set
i>0
[[
i>0 n>1
n
AssA TorA
i (M, A/I ) finite?
Recently, the second author [7, Theorem 5.1] proved that if A is a local complete
L
intersection ring and N =
n>0 Nn is a finitely generated graded module over the
Rees ring R(I), then
[[
AssA ExtiA (M, Nn )
i>0 n>0
is a finite set. Moreover, he proved that there exist i0 , n0 > 0 such that
2i0
AssA Ext2i
(M,
N
)
=
Ass
Ext
(M,
N
)
,
n
A
n
0
A
A
2i0 +1
2i+1
AssA ExtA
(M, Nn ) = AssA ExtA
(M, Nn0 )
for all i > i0 and n > n0 . In particular, if N is a finitely generated A-module, then N
L
L
can be taken as n>0 (I n N) or n>0 (I n N/I n+1 N). In the present study, we prove that
the question (†) has an affirmative answer for a local complete intersection ring. We
also analyze the stability of the sets of associated prime ideals which occurs periodically
after a certain stage.
Let (A, m) be a local complete intersection ring. Let M, N be two finitely generated
A-modules and I an ideal of A. The complexity of the pair (M, N) is defined to be
the number
µ(ExtnA (M, N))
<∞ ,
cxA (M, N) = inf b ∈ N lim sup
nb−1
n→∞
2
where µ(D) denotes the minimal number of generators of a finitely generated A-module
D. In [7, Theorem 7.1], the second author proved that cxA (M, I j N) is constant for all
j ≫ 0. In Section 6, we prove that
(††)
cxA (M, N/I j N) is constant for all j ≫ 0.
The rest of the paper is organized as follows. In Section 2, we give some graded
module structures which we use in order to prove our main results. The finiteness
results on asymptotic primes are proved in Section 3; while the stability results are
shown in Section 4. Then, in Section 5, we prove the analogous results on associated
primes for complete intersection rings which arise in algebraic geometry. Finally, in
Section 6, we prove (††).
2
Module structure
In this section, we give the graded module structures which we are going to use in
order to prove our main results.
Let Q be a commutative Noetherian ring and f = f1 , . . . , fc a Q-regular sequence.
Set A := Q/(f). Let M and D be two A-modules, and let M be finitely generated.
2.1. Let F : · · · → Fn → · · · → F1 → F0 → 0 be a projective resolution of M by
finitely generated free A-modules. Let tj : F(+2) → F, 1 6 j 6 c be the Eisenbud
operators defined by f = f1 , . . . , fc (see [3, Section 1]). By [3, 1.4], the maps tj are
determined uniquely up to homotopy. In particular, they induce well-defined maps
tj : ExtiA (M, D) −→ Exti+2
A (M, D) for all i and 1 6 j 6 c.
By [3, 1.5], the maps tj (j = 1, . . . , c) commute up to homotopy. Thus
Ext⋆A (M, D) :=
M
ExtiA (M, D)
i>0
turns into a graded T := A[t1 , . . . , tc ]-module, where T is the graded polynomial ring
over A in the cohomology operators tj defined by f with deg(tj ) = 2 for all 1 6 j 6 c.
Furthermore, these structures depend only on f, are natural in both module arguments
and commute with the connecting maps induced by short exact sequences.
2.2. If projdimQ (M) is finite and D is a finitely generated A-module, then Gulliksen
[4, 3.1] proved that Ext⋆A (M, D) is a finitely generated graded T -module.
L
2.3. Let I be an ideal of A. Let R(I) = n>0 I n X n be the Rees ring associated to I.
L
We consider R(I) as a subring of the polynomial ring A[X]. Let N = n>0 Nn be a
3
graded R(I)-module. Let u ∈ R(I) be a homogeneous element of degree s. Consider
the A-linear maps given by multiplication with u:
u
Nn −→ Nn+s
for all n.
By applying HomA (F, −) on the above maps and using the naturality of the operators
tj , we have the following commutative diagram of complexes:
tj
HomA (F, Nn )
//
HomA (F(+2), Nn )
u
tj
HomA (F, Nn+s )
//
u
HomA (F(+2), Nn+s ).
Now, taking cohomology, we obtain the following commutative diagram of A-modules:
tj
ExtiA (M, Nn )
//
Exti+2
A (M, Nn )
u
ExtiA (M, Nn+s )
tj
//
u
Exti+2
A (M, Nn+s )
for all i, n and 1 6 j 6 c. Thus
E (N ) :=
MM
ExtiA (M, Nn )
i>0 n>0
turns into a bigraded S := R(I)[t1 , . . . , tc ]-module, where we set deg(tj ) = (0, 2) for
all 1 6 j 6 c and deg(uX s ) = (s, 0) for all u ∈ I s , s > 0.
L
2.4. Suppose N is a finitely generated A-module. Set L := n>0 (N/I n+1 N). Note
L
n
that R(I, N) =
n>0 I N and N[X] = N ⊗A A[X] are graded modules over R(I)
and A[X] respectively. Since R(I) is a graded subring of A[X], we set that N[X]
is a graded R(I)-module. Therefore L is a graded R(I)-module, where the graded
structure is induced by the sequence
0 −→ R(I, N) −→ N[X] −→ L(−1) −→ 0.
Therefore, by the observations made in Section 2.3,
MM
E (L) =
ExtiA (M, N/I n+1 N)
i>0 n>0
is a bigraded module over S = R(I)[t1 , . . . , tc ].
Let us recall the following result from [7, 1.1].
Theorem 2.5. Let Q be a Noetherian ring of finite Krull dimension, and let f =
f1 , . . . , fc be a Q-regular sequence. Set A := Q/(f). Let M be a finitely generated
L
A-module with projdimQ (M) finite. Let I be an ideal of A, and let N = n>0 Nn be
a finitely generated R(I)-module. Then
MM
E (N ) :=
ExtiA (M, Nn )
i>0 n>0
is a finitely generated bigraded S = R(I)[t1 , . . . , tc ]-module.
4
We first prove the main results of this article for a ring A which is of the form
Q/(f), where Q is a regular local ring and f = f1 , . . . , fc is a Q-regular sequence. Then
we deduce the main results for a local complete intersection ring with the help of the
following well-known lemma:
b the completion of A with
Lemma 2.6. Let (A, m) be a Noetherian local ring and A
respect to m. Let D be an A-module. Then
b
AssA (D) = {q ∩ A : q ∈ AssAb(D ⊗A A)}.
3
Asymptotic associated primes: Finiteness
In this section, we prove the announced finiteness result for the set of associated prime
ideals of the family of Ext-modules ExtiA (M, N/I n N), (i, n > 0), where M and N are
finitely generated modules over a local complete intersection ring A and I ⊆ A is an
ideal (see Corollary 3.2).
Theorem 3.1. Let Q be a Noetherian ring of finite Krull dimension, and let f =
f1 , . . . , fc be a Q-regular sequence. Set A := Q/(f). Let M and N be two finitely
generated A-modules with projdimQ (M) finite, and let I be an ideal of A. Then
[[
AssA ExtiA (M, N/I n N)
is a finite set.
i>0 n>0
Proof. For a fixed n > 0, consider the short exact sequence of A-modules:
0 −→ I n N/I n+1 N −→ N/I n+1 N −→ N/I n N −→ 0.
L
n+1
Taking direct sum over n > 0 and setting L :=
N), we obtain the
n>0 (N/I
following short exact sequence of graded R(I)-modules:
0 −→ grI (N) −→ L −→ L(−1) −→ 0,
which induces an exact sequence of graded R(I)-modules for each i > 0:
ExtiA (M, grI (N)) −→ ExtiA (M, L) −→ ExtiA (M, L(−1)).
Taking direct sum over i > 0 and using the naturality of the cohomology operators tj ,
we get the following exact sequence of bigraded S = R(I)[t1 , . . . , tc ]-modules:
M
M
M
I nN
Φ
Ψ
i
−→
Vi,n −→
Vi,n−1,
ExtA M, n+1
I N
i,n>0
i,n>0
i,n>0
where Vi,n := ExtiA (M, N/I n+1 N) for each i > 0 and n > −1. Let
M
U=
Ui,n := Image(Φ).
i,n>0
5
Then, for each i, n > 0, considering the exact sequence of A-modules:
0 → Ui,n → Vi,n → Vi,n−1,
we have
AssA (Vi,n ) ⊆ AssA (Ui,n ) ∪ AssA (Vi,n−1)
⊆ AssA (Ui,n ) ∪ AssA (Ui,n−1 ) ∪ AssA (Vi,n−2 )
..
.
[
⊆
AssA (Ui,j ) [as AssA (Vi,−1 ) = φ for each i > 0].
06j6n
Taking union over i, n > 0, we obtain
[
AssA (Vi,n ) ⊆
i,n>0
[
AssA (Ui,n ).
(3.1.1)
i,n>0
Since grI (N) is a finitely generated graded R(I)-module, by Theorem 2.5,
M
I nN
i
ExtA M, n+1
I N
i,n>0
is a finitely generated bigraded S -module, and hence U is a finitely generated bigraded
S -module. Therefore, by [9, Lemma 3.2],
[
AssA (Ui,n )
is a finite set.
(3.1.2)
i,n>0
Now the result follows from (3.1.1) and (3.1.2).
An immediate corollary is the following:
Corollary 3.2. Let (A, m) be a local complete intersection ring. Let M, N be two
finitely generated A-modules and I an ideal of A. Then
[[
AssA ExtiA (M, N/I n N)
i>0 n>0
is a finite set.
b = Q/(f), where Q is a regular
Proof. Since A is a local complete intersection ring, A
local ring and f = f1 , . . . , fc is a Q-regular sequence.
b we have that
Applying Theorem 3.1 for the ring A,
[
[
i
n
n b
i
c
b
b
b
AssAb ExtA (M, N/I N) ⊗A A =
AssAb ExtAb M, N /(I A) N
i,n>0
i,n>0
is a finite set, and hence the result follows from Lemma 2.6.
6
4
Asymptotic associated primes: Stability
In the present section, we analyze the asymptotic behaviour of the sets of associated
prime ideals of Ext-modules ExtiA (M, N/I n N), (i, n > 0), where M and N are finitely
generated modules over a local complete intersection ring A and I ⊆ A is an ideal (see
Corollary 4.3). We first prove the following theorem:
Theorem 4.1. Let Q be a Noetherian ring of finite Krull dimension, and let f =
f1 , . . . , fc be a Q-regular sequence. Set A := Q/(f). Let M and N be two finitely
generated A-modules with projdimQ (M) finite, and let I be an ideal of A. Then there
exist i0 , n0 > 0 such that for all i > i0 and n > n0 , we have
2i
n0
0
AssA ExtA
(M, N/I n N) = AssA Ext2i
(M,
N/I
N)
,
A
2i0 +1
n
n0
AssA Ext2i+1
(M,
N/I
N)
=
Ass
Ext
(M,
N/I
N)
.
A
A
A
To prove this result, we assume the following lemma which we prove at the end of
this section.
Lemma 4.2. Let (Q, n) be a Noetherian local ring with residue field k, and let f =
f1 , . . . , fc be a Q-regular sequence. Set A := Q/(f). Let M and N be two finitely
generated A-modules with projdimQ (M) finite, and let I be an ideal of A. Then
n
λA HomA k, Ext2i
and λA HomA k, Ext2i+1
(M, N/I n N)
A (M, N/I N)
A
are given by polynomials in i, n with rational coefficients for all sufficiently large (i, n).
Proof of Theorem 4.1. Applying Theorem 3.1, we may assume that
[[
AssA ExtiA (M, N/I n N) = {p1 , p2 , . . . , pl }.
i>0 n>0
L
Set Vi,n := ExtiA (M, N/I n N) for each i, n > 0, and V := i,n>0 Vi,n .
We first prove that there exist some i′ , n′ > 0 such that
AssA (V2i,n ) = AssA (V2i′ ,n′ ) for all i > i′ and n > n′ .
(4.2.1)
To prove the claim (4.2.1), it is enough to prove that for each pj , where 1 6 j 6 l,
there exist some ij0 , nj0 > 0 such that exactly one of the following alternatives must
hold:
either pj ∈ AssA (V2i,n ) for all i > ij0 and n > nj0 ;
or pj ∈
/ AssA (V2i,n ) for all i > ij0 and n > nj0 .
Localizing at pj , and replacing Apj by A and pj Apj by m, it is now enough to prove
that there exist some i′ , n′ > 0 such that
either m ∈ AssA (V2i,n ) for all i > i′ and n > n′ ;
(4.2.2)
or m ∈
/ AssA (V2i,n ) for all i > i′ and n > n′ .
(4.2.3)
7
But from Lemma 4.2, λA (HomA (k, V2i,n )) is given by a polynomial in i, n with rational
coefficients for all sufficiently large (i, n), and hence there exist i′ , n′ > 0 such that
either
HomA (k, V2i,n ) 6= 0 for all i > i′ and n > n′ ;
or
HomA (k, V2i,n ) = 0 for all i > i′ and n > n′ ,
which is equivalent to that either (4.2.2) is true, or (4.2.3) is true.
Applying a similar procedure as in the even case, we see that there exist i′′ , n′′ > 0
such that
AssA (V2i+1,n ) = AssA (V2i′′ +1,n′′ ) for all i > i′′ and n > n′′ .
Now (i0 , n0 ) := max{(i′ , n′ ), (i′′ , n′′ )} satisfies the required result of the theorem.
An immediate corollary of the Theorem 4.1 is the following:
Corollary 4.3. Let (A, m) be a local complete intersection ring. Let M and N be two
finitely generated A-modules, and let I be an ideal of A. Then there exist i0 , n0 > 0
such that for all i > i0 and n > n0 , we have
2i0
n
n0
AssA Ext2i
A (M, N/I N) = AssA ExtA (M, N/I N) ,
2i0 +1
AssA Ext2i+1
(M, N/I n N) = AssA ExtA
(M, N/I n0 N) .
A
b = Q/(f), where Q is a regular local ring and f = f1 , . . . , fc is a
Proof. Assume A
b we see that there
Q-regular sequence. Then, by applying Theorem 4.1 for the ring A,
exist i0 , n0 > 0 such that for all i > i0 and n > n0 , we have
2i0
2i
n
n0
b
b
AssAb ExtA (M, N/I N) ⊗A A = AssAb ExtA (M, N/I N) ⊗A A ,
2i0 +1
n
n0
b
b
AssAb Ext2i+1
(M,
N/I
N)
⊗
A
=
Ass
Ext
(M,
N/I
N)
⊗
A
.
b
A
A
A
A
A
The result now follows from Lemma 2.6.
We now give
Proof of Lemma 4.2. For each fixed n > 0, consider the short exact sequence of Amodules:
0 −→ I n N −→ N −→ N/I n N −→ 0,
which induces an exact sequence of A-modules for each i, n:
n
ExtiA (M, I n N) → ExtiA (M, N) → ExtiA (M, N/I n N) → Exti+1
A (M, I N).
Taking direct sum over i, n and using the naturality of the cohomology operators tj ,
we obtain an exact sequence of bigraded S = R(I)[t1 , . . . , tc ]-modules:
U −→ T −→ V −→ U(0, 1),
8
where
U=
M
Ui,n :=
i,n>0
T =
M
V =
ExtiA (M, I n N),
i,n>0
Ti,n :=
i,n>0
M
M
M
ExtiA (M, N),
i,n>0
Vi,n :=
i,n>0
M
ExtiA (M, N/I n N),
and
i,n>0
U(0, 1) is same as U but the grading is twisted by (0, 1). Setting
X := Image(U → T ), Y := Image(T → V ) and Z := Image(V → U(0, 1)),
we have the following commutative diagram of exact sequences of bigraded S -modules:
U ❍❍
❍❍
##
0
:: X
✉✉
✉✉
// T
// U(0, 1)
//
;; V ●●
✈;; ●●●●
✇
●
✈
✇
♦
●## ♦♦♦ 77
✈✈
##
✇✇
:: Z ❖❖❖❖
:: Y ■■
■■
■■ ✈✈✈
■■ ✈✈✈
❖❖❖
$$ ✈
$$ ✈
''
0,
0
0
which gives the following two short exact sequences of bigraded S -modules:
0 → X → T → Y → 0 and 0 → Y → V → Z → 0.
Now applying HomA (k, −) to these short exact sequences, we get the following exact
sequences of bigraded S -modules:
0 −→ HomA (k, X) −→ HomA (k, T ) −→ HomA (k, Y ) −→ C −→ 0,
(4.3.1)
0 −→ HomA (k, Y ) −→ HomA (k, V ) −→ D −→ 0,
(4.3.2)
where C := Image HomA (k, Y ) −→ Ext1A (k, X) ,
D := Image (HomA (k, V ) −→ HomA (k, Z)).
L
i
n
By virtue of Theorem 2.5, U =
i,n>0 ExtA (M, I N) is a finitely generated bigraded S -module, and hence X = Image(U → T ) is so. Therefore HomA (k, X) and
Ext1A (k, X) are finitely generated bigraded S -modules. Being an S -submodule of
Ext1A (k, X), C is also a finitely generated bigraded S -module. Since HomA (k, X)
and Ext1A (k, X) are annihilated by the maximal ideal of A, HomA (k, Xi,n ) and Ci,n
both are finitely generated k-modules, and hence they have finite length as A-modules
for each i, n > 0. Therefore, by applying the Hilbert-Serre Theorem to the bigraded
S -modules HomA (k, X) and C, we obtain
X
λA (HomA (k, Xi,n ))z i w n =
X
λA (Ci,n )z i w n =
i,n>0
i,n>0
PX (z, w)
,
(1 − z 2 )c (1 − w)r
PC (z, w)
(1 − z 2 )c (1 − w)r
9
(4.3.3)
(4.3.4)
for some polynomials PX (z, w) and PC (z, w) over Z.
L
L
i
For a fixed n > 0,
i>0 Ti,n =
i>0 ExtA (M, N) is a finitely generated graded
L
A[t1 , . . . , tc ]-module, and hence HomA k, i>0 Ti,n is also so. By similar argument as
before, λA (HomA (k, Ti,n )) < ∞ for each i > 0. Therefore, again by the Hilbert-Serre
Theorem, we have for each n > 0:
X
λA (HomA (k, Ti,n ))z i =
i>0
PS (z)
(1 − z 2 )c
for some polynomial PS (z) ∈ Z[z]. Multiplying both sides of the above equation by
w n , and then taking sums over n > 0, we get
X
λA (HomA (k, Ti,n ))z i w n =
i,n>0
PS (z)
PT (z, w)
=
,
c
2
(1 − z ) (1 − w)
(1 − z 2 )c (1 − w)r
(4.3.5)
where PT (z, w) := PS (z)(1 − w)r−1 ∈ Z[z, w].
Now considering (4.3.1), we have an exact sequence of A-modules:
0 −→ HomA (k, Xi,n ) −→ HomA (k, Ti,n ) −→ HomA (k, Yi,n ) −→ Ci,n −→ 0
for each i, n > 0. So, the additivity of length functions gives
λA (HomA (k, Xi,n )) − λA (HomA (k, Ti,n )) + λA (HomA (k, Yi,n )) − λA (Ci,n ) = 0
for each i, n > 0. Multiplying the above equation by z i w n , then taking sum over
i, n > 0, and using (4.3.3), (4.3.4) and (4.3.5), we obtain
X
λA (HomA (k, Yi,n ))z i w n =
i,n>0
PY (z, w)
,
(1 − z 2 )c (1 − w)r
(4.3.6)
where PY (z, w) := −PX (z, w) + PT (z, w) + PC (z, w) ∈ Z[z, w].
Recall that by Theorem 2.5, U is a finitely generated bigraded S -module. Now
observe that Z is a bigraded submodule of U(0, 1). Therefore Z is a finitely generated
bigraded S -module, hence HomA (k, Z) is so, and hence
D = Image(HomA (k, V ) −→ HomA (k, Z))
is also so. Observe that λA (Di,n ) < ∞ for each i, n > 0. Therefore, once again by
applying the Hilbert-Serre Theorem, we have
X
λA (Di,n )z i w n =
i,n>0
PD (z, w)
(1 − z 2 )c (1 − w)r
for some polynomial PD (z, w) over Z.
Now considering (4.3.2), we have an exact sequence of A-modules:
0 −→ HomA (k, Yi,n ) −→ HomA (k, Vi,n ) −→ Di,n −→ 0
10
(4.3.7)
for each i, n > 0, which gives
λA (HomA (k, Yi,n)) − λA (HomA (k, Vi,n )) + λA (Di,n ) = 0.
Multiplying the above equation by z i w n , then taking sum over i, n > 0, and using
(4.3.6) and (4.3.7), we obtain
X
λA (HomA (k, Vi,n ))z i w n =
i,n>0
PV (z, w)
,
(1 − z 2 )c (1 − w)r
where PV (z, w) := PY (z, w) + PD (z, w) ∈ Z[z, w]. Therefore it follows that
λA (HomA (k, V2i,n ))
and
λA (HomA (k, V2i+1,n ))
are given by polynomials in i, n with rational coefficients for all sufficiently large (i, n),
which completes the proof of the lemma.
5
Asymptotic associated primes: The geometric
case
Let V be an affine or projective variety over an algebraically closed field K. Let A be
the coordinate ring of V . Then V is said to be a locally complete intersection variety
if all its local rings are complete intersection. Thus
• in the affine case, Ap is a local complete intersection ring for every p ∈ Spec(A).
• in the projective case, A(p) is a local complete intersection ring for every p ∈
Proj(A). Recall that A(p) is the degree zero part of the graded ring S −1 A, where
S is the collection of all homogeneous elements in A r p.
In this section, we prove the results analogous to Theorems 3.1 and 4.1 for the
coordinate rings of locally complete intersection varieties. In the affine case, we prove
the following general result:
Theorem 5.1. Let A = Q/a, where Q is a regular ring of finite Krull dimension
and a ⊆ Q is an ideal so that aq ⊆ Qq is generated by a Qq -regular sequence for each
q ∈ Var(a). Let M and N be two finitely generated A-modules, and let I be an ideal of
A. Then the set
[[
AssA ExtiA (M, N/I n N)
is finite.
i>0 n>0
Moreover, there exist i0 , n0 > 0 such that for all i > i0 and n > n0 , we have
2i0
n
n0
AssA Ext2i
A (M, N/I N) = AssA ExtA (M, N/I N) ,
2i0 +1
AssA Ext2i+1
(M, N/I n N) = AssA ExtA
(M, N/I n0 N) .
A
11
Proof. For x ∈ A, set D(x) = {p ∈ Spec(A) : x ∈
/ p}. As in [7, Proof of Theorem 6.1],
we have
Spec(A) = D(g1 ) ∪ · · · ∪ D(gm ) for some g1 , . . . , gm ∈ A
such that the localization Agj by {gj l : l > 0} has the form Qj /aj for some regular
ring Qj of finite Krull dimension and some ideal aj of Qj generated by a Qj -regular
sequence. Clearly, for any A-module E, we obtain
[
(5.1.1)
AssA (E) = {q ∩ A : q ∈ AssAgj (Egj ) for some j = 1, . . . , m}.
Since localization Agj is flat over A, we have
ExtiA (M, N/I n N)
gj
= ExtiAgj Mgj , Ngj /(IAgj )n Ngj
(5.1.2)
for all i, n > 0 and j = 1, . . . , m. Therefore, from (5.1.1) and (5.1.2), it is enough to
prove the result for the ring Agj = Qj /aj for each j. Note that Qj is a regular ring of
finite Krull dimension, and hence projdimQj (Mgj ) is finite. Therefore the result now
follows by applying the Theorems 3.1 and 4.1 to each Agj = Qj /aj .
Now we prove the analogous result to Theorem 5.1 in the projective case. Let us
fix the following hypothesis:
Hypothesis 5.2. Let K be a field not necessarily algebraically closed, and let Q =
K[X0 , X1 , . . . , Xr ] be a polynomial ring over K, where deg(Xi ) = 1 for all i. Let a be
a homogeneous ideal of Q. Set A := Q/a. Suppose A(p) is a complete intersection ring
for every p ∈ Proj(A).
Let m be the unique maximal homogeneous ideal of A. If E is a graded A-module,
then define the set of relevant associated primes of E as
∗
AssA (E) := AssA (E) r {m}.
Theorem 5.3. With the Hypothesis 5.2, let M and N be two finitely generated graded
A-modules, and let I be a homogeneous ideal of A. Then the set
[[
∗
AssA ExtiA (M, N/I n N)
is finite.
i>0 n>0
Moreover, there exist i0 , n0 > 0 such that for all i > i0 and n > n0 , we have
∗
2i0
n
n0
AssA Ext2i
A (M, N/I N) = AssA ExtA (M, N/I N) ,
2i0 +1
∗
AssA Ext2i+1
(M, N/I n N) = ∗ AssA ExtA
(M, N/I n0 N) .
A
∗
Proof. For a homogeneous element x ∈ A, set ∗ D(x) = {p ∈ Proj(A) : x ∈
/ p}. As in
[7, Proof of Theorem 6.3], we have
Proj(A) = ∗ D(g1 ) ∪ · · · ∪ ∗ D(gm ) for some homogeneous g1 , . . . , gm ∈ A
12
such that Agj = Qj /aj for some regular ring Qj of finite Krull dimension and some
homogeneous ideal aj of Qj generated by a Qj -regular sequence. Clearly, for any graded
A-module E, we obtain
[
∗
AssA (E) = {q ∩ A : q ∈ AssAgj (Egj ) for some j = 1, . . . , m}.
Similarly, as in the proof of Theorem 5.1, the result follows by applying the Theorems 3.1 and 4.1 to each Agj = Qj /aj .
6
Support varieties
Let (A, m) be a local complete intersection ring of codimension c. Let M and N be
two finitely generated A-modules. The complexity of the pair (M, N) is defined to be
the number
µ(ExtnA (M, N))
cxA (M, N) = inf b ∈ N lim sup
<∞ ,
nb−1
n→∞
where µ(D) denotes the minimal number of generators of a finitely generated A-module
D.
In this section, we prove the following theorem:
Theorem 6.1. Let (A, m) be a local complete intersection ring. Let M and N be two
finitely generated A-modules, and let I be an ideal of A. Then
cxA (M, N/I j N)
is constant for all j ≫ 0.
As in [7, Section 7.2], we may assume that our local complete intersection A
(1) is complete. So A = Q/(f1 , . . . , fc ), where (Q, n) is a regular local ring and
f = f1 , . . . , fc ∈ n2 is a Q-regular sequence.
(2) has an algebraically closed residue field k.
Let U and V be two finitely generated A-modules. Let
M
Ext⋆A (U, V ) :=
ExtiA (U, V )
i>0
be the total Ext module of U, V over the ring A[t1 , . . . , tc ] of cohomology operators tj
defined by f, where deg(tj ) = 2 for all j = 1, . . . , c. Set
C(U, V ) := Ext⋆A (U, V ) ⊗A k.
Since Ext⋆A (U, V ) is a finitely generated graded A[t1 , . . . , tc ]-module, C(U, V ) is a finitely
generated graded module over
T := A[t1 , . . . , tc ] ⊗A k = k[t1 , . . . , tc ].
13
Define the support variety V (U, V ) of U, V as the zero set in k c of the annihilator
of C(U, V ) in T , that is
V (U, V ) := {(b1 , . . . , bc ) ∈ k c : P (b1 , . . . , bc ) = 0 ∀ P ∈ annT (C(U, V ))} ∪ {0}.
By [1, 2.4(2)], we have
(@)
cxA (U, V ) = dim(V (U, V )) = dimT (C(U, V )).
Before proving the Theorem 6.1, we give the following well-known lemma:
L
Lemma 6.2. Let R =
n>0 Rn be a Noetherian standard N-graded ring, and let
L
M = n>0 Mn be a finitely generated N-graded R-module. Then there exists j0 > 0
such that
dimR0 (Mj ) = dimR0 (Mj0 ) for all j > j0 .
Proof. Since M is a finitely generated N-graded module over a Noetherian standard
N-graded ring R, there exists j0 > 0 such that
annR0 (Mj ) = annR0 (Mj0 ) for all j > j0 ,
and hence dimR0 (Mj ) = dimR0 (Mj0 ) for all j > j0 .
6.3. Since grI (N) =
Theorem 2.5,
M
L
j>0 I
j
N/I j+1 N is a finitely generated graded R(I)-module, by
Ext⋆A (M, I j N/I j+1N) =
j>0
MM
ExtiA (M, I j N/I j+1 N)
i>0 j>0
is a finitely generated bigraded R(I)[t1 , . . . , tc ]-module, and hence
M
Ext⋆A (M, I j N/I j+1 N) ⊗A k
j>0
is a finitely generated graded R(I)[t1 , . . . , tc ] ⊗A k = F (I)[t1 , . . . , tc ]-module, where
F (I) is the fiber cone of I which is a finitely generated k-algebra. Writing
F (I)[t1 , . . . , tc ] = k[x1 , . . . , xm ][t1 , . . . , tc ] = T [x1 , . . . , xm ],
and using Lemma 6.2, we have
dimT Ext⋆A (M, I j N/I j+1 N) ⊗A k
is constant for all j ≫ 0,
and hence cxA (M, I j N/I j+1 N) is constant for all j ≫ 0, by (@).
Now we give
14
Proof of Theorem 6.1. Fix j > 0. Consider the short exact sequence of A-modules
0 −→ I j N/I j+1 N −→ N/I j+1 N −→ N/I j N −→ 0,
which induces the following exact sequence of A-modules for each i:
j
Exti−1
A (M, N/I N) −→
ExtiA (M, I j N/I j+1N) −→ ExtiA (M, N/I j+1 N) −→ ExtiA (M, N/I j N) −→
j
j+1
Exti+1
N).
A (M, I N/I
Taking direct sum over i and setting
M
M
ExtiA (M, I j N/I j+1 N) and Vj :=
ExtiA (M, N/I j N),
Uj :=
i>0
i>0
we obtain an exact sequence of A[t1 , . . . , tc ]-modules:
ϕ1
ϕ2
ϕ3
ϕ4
Vj (−1) −→ Uj −→ Vj+1 −→ Vj −→ Uj (1).
Set Zj := Image(ϕ1 ), Xj := Image(ϕ2 ) and Yj := Image(ϕ3 ). Thus we have the
following commutative diagram of exact sequences:
Vj (−1)
0
// Uj
// U (1)
// Vj+1
//
<< Vj ❏❏❏
❋❋
88 j
▲▲▲
::
❍❍❍
③
②<<
♣
✉
❏
❋
♣
③
②
%%
♣
❍$$
▲&&
""
②
③
✉✉
X
Y
Z
(1)
Z
j
j
j
j
7
7
9
9
❑❑❑
●●
❍❍
❖❖❖
●● rrr88
♣♣♣
❍❍ ✈✈✈::
❑❑%% sssss
❖❖''
## r
## ✈
♣♣♣
0
0
0
0.
Now consider the short exact sequences of A[t1 , . . . , tc ]-modules:
0 → Xj → Vj+1 → Yj → 0 and 0 → Yj → Vj → Zj (1) → 0.
Tensoring these sequences with k over A, we get the following exact sequences of
T = k[t1 , . . . , tc ]-modules:
Φj
Xj ⊗A k −→Vj+1 ⊗A k −→ Yj ⊗A k −→ 0,
Ψj
Yj ⊗A k −→Vj ⊗A k −→ Zj (1) ⊗A k −→ 0.
Now for each j > 0, set Xj′ := Image(Φj ) and Yj′ := Image(Ψj ) to get the following
short exact sequences of T -modules:
0 −→ Xj′ −→Vj+1 ⊗A k −→ Yj ⊗A k −→ 0,
(6.3.1)
0 −→ Yj′ −→Vj ⊗A k −→ Zj (1) ⊗A k −→ 0.
(6.3.2)
L
From Section 6.3,
j>0 Uj is a finitely generated graded R(I)[t1 , . . . , tc ]-module,
L
L
and hence its submodule
j>0 Zj is also so. Therefore
j>0 (Zj ⊗A k) is a finitely
generated graded module over
R(I)[t1 , . . . , tc ] ⊗A k = F (I)[t1 , . . . , tc ] = T [x1 , . . . , xm ].
15
Therefore, by Lemma 6.2, dimT (Zj ⊗A k) = z for all sufficiently large j, where z is some
constant. Now considering the short exact sequences (6.3.1) and (6.3.2), we obtain that
dimT (Vj+1 ⊗A k) = max{dimT (Xj′ ), dimT (Yj ⊗A k)},
dimT (Vj ⊗A k) = max{dimT (Yj′ ), z} > z
(6.3.3)
(6.3.4)
for all sufficiently large j, say j > j0 .
Note that dimT (Vj ⊗A k) = cxA (M, N/I j N) for all j > 0. Thus it is enough to
prove that the stability of dimT (Vj ⊗A k) holds for all sufficiently large j.
If dimT (Vj ⊗A k) = z for all j > j0 , then we are done. Otherwise there exists some
j > j0 such that dimT (Vj ⊗A k) > z, and hence for this j we have
dimT (Vj ⊗A k) = dimT (Yj′ ) 6 dimT (Yj ⊗A k) 6 dimT (Vj+1 ⊗A k).
First equality above occurs from (6.3.4), second inequality occurs because Yj′ is a
quotient module of Yj ⊗A k, and the last inequality occurs from (6.3.3).
Note that dimT (Vj+1 ⊗A k) > z. So, by applying a similar procedure, we have
dimT (Vj+1 ⊗A k) 6 dimT (Vj+2 ⊗A k).
In this way, we have a bounded non-decreasing sequence
dimT (Vj ⊗A k) 6 dimT (Vj+1 ⊗A k) 6 dimT (Vj+2 ⊗A k) 6 · · · 6 dim(T ) < ∞,
which eventually stabilize somewhere, and hence the required stability holds.
Acknowledgements
We thank the anonymous reviewer for his/her careful reading of our manuscript and
many valuable comments. The first author would like to thank NBHM, DAE, Govt.
of India for providing financial support for this study.
References
[1] L. L. Avramov and R.-O. Buchweitz. Support varieties and cohomology over
complete intersections. Invent. Math. 142 (2000), 285-318.
[2] M. Brodmann. Asymptotic stability of Ass(M/I n M). Proc. Amer. Math. Soc.
74 (1979), 16-18.
[3] D. Eisenbud. Homological algebra on a complete intersection, with an application to group representations. Trans. Amer. Math. Soc. 260 (1980), 35-64.
16
[4] T. H. Gulliksen. A change of ring theorem with applications to Poincaré series
and intersection multiplicity. Math. Scand. 34 (1974), 167-183.
[5] D. Katz and E. West. A linear function associated to asymptotic prime divisors. Proc. Amer. Math. Soc. 132 (2004), 1589-1597.
[6] L. Melkersson and P. Schenzel. Asymptotic prime ideals related to derived
functors. Proc. Amer. Math. Soc. 117 (1993), 935-938.
[7] T. J. Puthenpurakal. On the finite generation of a family of Ext modules.
Pacific J. Math. 266 (2013), 367-389.
[8] W. V. Vasconcelos. Cohomological degrees of graded modules. pp. 345-392 in
Six lectures on commutative algebra (Bellaterra, 1996), edited by J. Elias et
al., Progr. Math., 166, Birkhäuser, Basel, 1998.
[9] E. West. Primes associated to multigraded modules. J. Algebra 271 (2004),
427-453.
Department of Mathematics
Indian Institute of Technology Bombay
Powai
Mumbai 400076
India
E-mail: [email protected]
[email protected]
17
| 0math.AC
|
Generalizing Bottleneck Problems
arXiv:1802.05861v1 [cs.IT] 16 Feb 2018
Hsiang Hsu∗ , Shahab Asoodeh†, Salman Salamatian‡, and Flavio P. Calmon∗
∗ Harvard University, {hsianghsu, fcalmon}@g.harvard.edu, † University of Chicago, [email protected],
‡ Massachusetts Institute of Technology, [email protected]
Abstract—Given a pair of random variables (X, Y ) ∼ PXY
and two convex functions f1 and f2 , we introduce two bottleneck functionals as the lower and upper boundaries of
the two-dimensional convex set that consists of the pairs
(If1 (W ; X), If2 (W ; Y )), where If denotes f -information and W
varies over the set of all discrete random variables satisfying
the Markov condition W → X → Y . Applying Witsenhausen
and Wyner’s approach, we provide an algorithm for computing
boundaries of this set for f1 , f2 , and discrete PXY , . In the
binary symmetric case, we fully characterize the set when (i)
f1 (t) = f2 (t) = t log t, (ii) f1 (t) = f2 (t) = t2 − 1, and (iii) f1
and f2 are both ℓβ norm function for β > 1. We then argue that
upper and lower boundaries in (i) correspond to Mrs. Gerber’s
Lemma and its inverse (which we call Mr. Gerber’s Lemma),
in (ii) correspond to estimation-theoretic variants of Information
Bottleneck and Privacy Funnel, and in (iii) correspond to Arimoto
Information Bottleneck and Privacy Funnel.
I. I NTRODUCTION
Few information-theoretic constructs have captured the attention of machine learning researchers and practitioners as
the Information Bottleneck (IB) [1]. Given two correlated
random variables X and Y with joint distribution PXY , the
goal of the IB is to determine a mapping PW |X that produces
a new representation W of X such that (i) W → X →
Y and (ii) I(W ; Y ) is maximized (information preserved)
while minimizing I(W ; X) (compression). This tradeoff can
be quantified by the Lagrangian functional B(PXY , λ) ,
max I(W ; Y )−λI(W ; X). The IB has proved useful in many
PW |X
machine learning problems, such as clustering [2] and natural
language processing [3]. More recently, the IB framework
has been used to analyze the training process of deep neural
networks [4], [5].
In an inverse context, the Privacy Funnel (PF), introduced
in [6], seeks to determine a mapping PW |X that minimizes
I(W ; Y ) (privacy leakage) while assuring I(W ; X) ≥ x (revealing useful information). Analogously, the PF can be solved
by considering the functional F(PXY , λ) , min I(W ; Y ) +
PW |X
λI(W ; X). The privacy funnel (and its variants) has shown to
be useful in information-theoretic privacy [6], [7].
The choice of mutual information in both the IB and the PF
frameworks does not seem to carry any specific “operational”
significance. It does, however, have a desirable practical consequence: it leads to self-consistent equations [1, Eq. 28] that
can be solved iteratively in the IB case. In fact, this property is
unique to mutual information among many other information
metrics [8]. Nevertheless, at least in theory, one can replace the
mutual information with a broader family of measures based
on f -divergences1.
In this paper, we consider a wider class of bottleneck
problems which includes the IB and the PF. We define f information between two random variables X and Y as
If (X; Y ) , Df (PXY kPX PY ), and introduce the following
bottleneck functional
Bf1 ,f2 (PXY , x) ,
max
W →X→Y
If2 (W ; Y ) s.t.If1 (W ; X) ≤ x,
(1)
and the funnel functional
Ff1 ,f2 (PXY , x) ,
If2 (W ; Y ) s.t.If1 (W ; X) ≥ x,
(2)
where f1 and f2 are convex functions. Different incarnations
of f -information have already appeared, e.g., T -information
in [9] for f (t) = |t − 1|. These metrics possess “operational”
significance that are arguably more useful in statistical learning
and privacy applications than mutual information. For instance,
total variation and Hellinger distance play important roles
in hypothesis testing [10] and χ2 -divergence in estimation
problems [6]. Formulations (1) and (2) for a broader class of
divergences can be potentially useful to emerging applications
of information theory in machine learning.
Computing B and F reduces to characterizing the upper and
lower boundaries, respectively, of the two-dimensional set
o
n
(3)
If1 (W ; X), If2 (W ; Y ) : W → X → Y .
min
W →X→Y
It is worth mentioning that studying (3) is at the heart
of the strong data processing inequalities [11] as well
as fundamental limits of privacy [7]. Witsenhausen et al.
[12]
investigated the lower boundaryo of a related set
n
(H(X|W ), H(Y |W )) : W → X → Y , where H(·) is the
entropy function. In particular, they proposed an algorithm for
analytically computing H(Y |W ) based on a dual formulation.
When X is binary and PY |X is a binary symmetric channel
(BSC), the lower bound of H(Y |W ) corresponds to the wellknown Mrs. Gerber’s Lemma [13]. Related convex techniques
have also been used to characterize some network information
theoretic regions [14].
We generalize the approach in [12] to study boundaries of
(3) for a broader class of f -information metrics, characterizing properties of new bottleneck problems of the form (1)
and (2). In particular, we investigate the estimation-theoretic
1 Given
two probability distributions P ≪ Q and a convex function
fi :
h
dP
) .
(0, ∞) → R with f (1) = 0, f -divergences is Df (P kQ) , EQ f ( dQ
variants of information bottleneck and privacy funnel using
χ2 -divergence, which we call Estimation Bottleneck and Estimation Privacy Funnel, respectively. In the binary symmetric
case, the upper boundary corresponds to the inverse of Mrs.
Gerber’s Lemma, which we call Mr. Gerber’s Lemma. We
further extend these lemmas for Arimoto conditional entropy
[15].
This paper is organized as follows. Section II introduces the
geometry of bottleneck problems. In Section III, we formulate
variational bottleneck problems and explore their use, and
provide further applications on information inequalities in
Section IV. Proofs of the results are available in [16].
II. G EOMETRIC P ROPERTIES
A. Notation
Let X and Y be two random variables having joint distribution PXY with supports X = [m] , {1, · · · , m} and
Y = [n], respectively. We denote by PX = q ∈ ∆m the
marginal probability vector with
P entries [PX (1), · · · , PX (m)],
where ∆m , {x ∈ Rm : m
i=1 xi = 1, xi ≥ 0} is a mdimensional simplex. We denote by T ∈ Rn×m the stochastic matrix whose entries are the channel transformation
PY |X , i.e. [T]i,j = PY |X (i|j); thus, PY = Tq ∈ ∆n .
For a discrete
random variable W with support
W, let
pw = PX|W (1|w), PX|W (2|w), . . . , PX|W (m|w) , and let
the marginal of W be PW (w) = αw . We denote by hm
thePentropy function, i.e. hm : ∆m → R with hm (q) =
− i∈[m] qi log qi and 0 log 0 , 0. Finally, we denote the
convex hull of a set A by convA, and the boundary of a set
by ∂A.
Let f : ∆m → R and g : ∆n → R be continuous and
bounded mappings over simplices of dimension m and n,
respectively. We study the set (3) by first considering a more
general context, and then specialize it to different information
metrics in following sections. We consider the tuple
(4)
E f (pw ) , E g (Tpw ) ,
P
where E [f (pw )] =
w∈W αw f (pw ), and E [g (Tpw )] =
P
α
g
(Tp
).
Recall
that X, Y , and W form the
w
w∈W w
Markov chain W → X → Y . Therefore, we are interested
in the following set for a fixed channel T:
, {(q, E [f (pw )] , E [g (Tpw )]) |
X
X
pw ∈ ∆m ,
αw pw = q,
αw = 1}. (5)
w∈W
∆
LT (q, x)
= inf {y| (q, x, y) ∈ C (T)} ,
UT (q, x)
∆
= sup {y| (q, x, y) ∈ C (T)} .
(6)
(7)
Under appropriate conditions on x (depending on the choice
of f ), {y| (q, x, y) ∈ C (T)} is non-empty, and hence the
compactness of C (T) allows one to replace infimum and
supremum in (6) and (7) with minimum and maximum,
respectively. Moreover, it follows from the convexity of C (T)
that LT (q, ·) is convex and UT (q, ·) is concave.
C. Dual Formulations
Since C(T) is a convex set, its upper and lower boundaries
are equivalently represented by its supporting hyperplanes. We
use the dual approach introduced in [12] to evaluate LT (q, ·)
and UT (q, ·). For a given λ, define the conjugate function
L∗T (q, λ) , min {−λx + y| (q, x, y) ∈ C (T)} .
(8)
Note that the graph of LT (q, ·) is the lower boundary of C (T).
It then follows that the point (x, y) that achieves the minimum
in (8) lies on the lower boundary of C (T) with supporting line
of slope λ, and hence corresponds to a point (x, LT (p, x)).
We now turn our attention to evaluating L∗T (q, ·). Let
∆
Sλ (T) = {(p, y − λx) | (p, x, y) ∈ S (T)} .
(9)
We observe that Sλ (T) is the graph of the function φ (·, λ)
on ∆m given by
φ (p, λ) = g (Tp) − λf (p) , p ∈ ∆m .
(10)
Since the mapping y − λx preserves convexity, we have
B. Geometry of Bottleneck Problems
C (T)
Let the upper and lower boundaries of C be denoted by UT
and LT , respectively, i.e., we have
w∈W
Moreover, we define S (T) , {(p, f (p) , g (Tp)) |p ∈ ∆m }.
The next lemma is a direct generalization of [12, Lemma 2.1].
Lemma 1. C (T) is convex and compact with C (T) =
convS (T). In addition, all points in C (T) can be written
as a convex combination of at most m + 1 points of S (T); in
other words, |W| ≤ m + 1.
Cλ (T) , convSλ (T) = {(q, y − λx) | (q, x, y) ∈ C (T)}
(11)
as the convex hull of the graph of φ (·, λ). Thus, L∗T (q, λ)
is given by the lower convex envelope of φ (·, λ) at q. The
same would go, mutatis mutandis, for UT (q, x): its conjugate
∗
function UT
(q, ·), defined as
∗
UT
(q, λ) , max {−λx + y| (q, x, y) ∈ C (T)}
coincides with the upper concave envelope of φ (·, λ) at q.
These properties leads to a procedure for characterizing
LT (q, ·) and UT (q, ·). We illustrate LT (q, ·) in details and
UT (q, ·) will follow by using concave envelope instead. For
L∗T (q, λ) = z ∗ , there are two scenarios:
1) Trivial case: If (q, z ∗ ) is in both Sλ and Cλ , then z ∗ =
g (Tq) − λf (q). In this case, (q, x, LT (q, x)) simply
reduces to (q, f (q) , g (Tq)), and the optimal W has
PW (w) = 1 for some w, independent of X.
2) Non-trivial case: If z ∗ 6= φ (q, λ), then (q, z ∗ ) ∈ Cλ
is the convex combination of points (pi , φ (pi , λ)) ∈
Sλ , with
Pk weights αi where i ∈ [k] for some k ≥ 2,
and i=1 αi = 1. Then (q, x, LT (q, x)) is given by
Pk
i=1 αi (pi , f (pi ) , g (Tpi )). Moreover, an optimal W
is attained by PW (i) = αi and pw = pi , i ∈ [k].
Algorithm 1 Computing (x, LT (q, x)) at slope λ
Input: λ, q
Output: (x, LT (q, x))
1: Compute φ (p, λ), p ∈ ∆m
2: L∗
T (p, λ) ← convex envelope of φ (p, λ)
3: if L∗
T (q, λ) = φ (q, λ) then
4:
return (f (q) , g (Tq))
5: else
n
k
k
6:
(αi , pi )i=1 ← ai , φ−1 (bi , λ) i=1 |bi ∈ φ (·, λ) ,
o
Pk
∗
i=1 ai bi = LT (q, λ)
P
Pk
k
7:
return
α
f
(p
)
,
α
g
(Tp
)
i
i
i=1 i
i=1 i
Pm+1
Hence, the points on the graph of LT (q, ·) can be obtained
by only considering the points of φ(·, λ) which differs from
its convex envelope L∗T (·, λ) since those are exactly the points
where W is not induced from the trivial case. Algorithm 1
summarizes our previous discussions.
Observe that for a given PX we have
D. Matched Channels
The geometry of bottleneck problems leads to intriguing properties of pw . Our previous discussions reveal that
the points {pi }ki=1 used to form the convex envelope
of φ (·, λ) are special: they determine a channel PX|W
such that for any distribution PW , the resulting value of
(q, E [f (pw )] , E [g (Tpw )]) is on the boundary of C (T) with
supporting line of slope λ. In this case, we say that the points
k
{pi }i=1 form a matched channel for T and f , g.
Definition 1 (Matched Channel). For a fixed channel PY |X
and f , g, we say that PX|W is matched to PY |X if there exists
PW such that |W| ≥ 2 and
(E [f (pw )] , E [g (Tpw )]) = (x, LT (q, x)) .
(12)
Using an elementary result in convex geometry (see
Lemma 2 in Appendix.), we immediately have the following
theorem.
Theorem 1. Let PX|W =w = pw be a matched channel for
PY |X . Then for any PW , we have
E [f (pw )] , E [g (Tpw )] = (x, LT (q, x)) .
(13)
Proof. See the Appendix.
From Theorem 1, we know that for any distribution PX ,
matched channels PX|W are entirely determined by the points
on the curve φ (·, λ) whose convex combinations lead to the
convex envelope of φ at PX . It implies that as long as φ (·, λ)
meets its convex envelope at PX , small perturbation around
PX does not change the matched channels PX|W but simply
change the weight αw . Thus, optimal mappings PW |X are
surprisingly robust to small errors in estimation of PX , which
could potentially give pragmatic advantages when applying
bottleneck problems to real data. However, if PX changes, we
can recover the matched channels by first solving αi via q =
i=1 αi pw (i), where pw (i) = PX|W =i , and then applying
Bayes’ rules.
Note that the properties above only hold when f and g do
not depend on PX . Specifically, matched channels do not exist
for the cases studied in Section III-C and III-D.
III. G ENERALIZING B OTTLENECK P ROBLEMS
In this section, we demonstrate how the tools developed in
Section II can be applied to new bottleneck problems of the
form (1) and (2). We then revisit the IB and PF, and also study
their estimation-theoretic variants.
Consider the Markov chain W → X → Y . Our goal is to
describe the achievable pairs of f -divergences
(Df1 (PW X kPW PX ) , Df2 (PW Y kPW PY )) .
Df1 (PW X kPW PX )
"
#
X
X
PX|W (x|w)
=
PW (w)
PX (x)f1
PX (x)
w∈W
x∈X
X
=
PW (w)Df1 PX|W =w kPX ,
(14)
(15)
(16)
(17)
w∈W
and hence If1 (W ; X) can be expressed as
X
If1 (W ; X) =
αw f (pw ) = E f (pw ) ,
(18)
w∈W
for some function f . Similarly, define Df2 PY |W kPY =
g(PY |W ), we have
X
If2 (W ; Y ) =
αw g(Tpw ) = E g (Tpw ) .
(19)
w∈W
Hence the corresponding set {(q, If1 (W ; X), If2 (W ; Y ))} for
varying W has the same form as C(T). Letting
φ (p, λ) = Df2 (TpkPY ) − λDf1 (pkPX ) ,
(20)
we can thus apply Algorithm 1 to characterize Bf1 ,f2 (PXY , ·)
and Ff1 ,f2 (PXY , ·).
Next, we show that how the usual IB and PF fit in our
formulation, and study their estimation-theoretic counterparts.
We note, however, that the previous analysis does not require
f1 = f2 .
A. Information Bottleneck
Assuming f1 (t) = f2 (t) = t log t in the bottleneck
functional (1), we have Df1 (PW X kPW PX ) = I (W ; X) and
Df2 (PW Y k PW PY ) = I (W ; Y ). Thus, the set of points
{(x, UT (q, x)) |0 ≤ x ≤ H (X)} corresponds to the set of
solutions of the IB problem.
It is worth mentioning that the same geometric approach
can also be applied directly to entropy functions which also
leads to the IB formulation. In fact, this is exactly the setting
studied in [12]. Specifically, choosing f1 = hm and f2 = hn ,
the set of points
{(hm (q) − x, hn (Tq) − LT (q, x)) |0 ≤ x ≤ hm (q)}
(21)
0.25
χ
XY
χ
X
0.35
Iβ W Y
( ; )
0.20
0.15
0.10
Assuming f1 (t) = f2 (t) = t log t in the funnel functional
(2), the set {(x, LT (q, x)) |0 ≤ x ≤ H (X)} corresponds to
the set of solutions of the privacy funnel, introduced in [6].
Equivalently, using the entropy function as in Section III-A,
the set of points
β
=5
= 10
0.25
β
0.20
=2
β
=1
0.15
0.10
0.05
0.00
β
0.30
χ
B. Privacy Funnel
0.40
B 2 (P , x)
F 2 (P Y, x)
0.30
2(W; Y)
also corresponds to the set of solutions of the IB. The IB is
closely related to strong data processing inequalities. See [11,
Proposition 2] for more details in the case of the BSC (see
also Fig. 1 (right)).
0.05
0.0
0.2
0.4
2
χ (W; X)
0.6
0.8
0.00
0.0
0.2
0.4
0.6
I (W; X)
0.8
1.0
β
{(hm (q) − x, hn (Tq) − UT (q, x)) |0 ≤ x ≤ hm (q)}
(22)
also corresponds to the set of solutions, which follows from
the fact that UT (q, ·) is monotonically non-decreasing; see
Fig. 1 (right).
Fig. 1. T follows BSC with crossover probability δ = 0.1 and P {X = 1} =
q = 0.1. Left: The estimation bottleneck and privacy
funnel. Right: Set of
achievable pairs of Arimoto mutual information Iβ (W ; X), Iβ (W ; Y ) for
BSC with δ = 0.2 and q = 0.4. Note that when β = 1, the upper and lower
boundaries correspond to the IB and the PF.
C. Estimation Bottleneck
where the privacy is measured in terms of MMSE. The
practical significance of (25) is justified as follows. Suppose
Y represents private data (e.g. political preferences) and X
(e.g. movie rating) is correlated with Y . The main objective,
formulated by (25), is to construct a privacy-assuring mapping
PX|W such that the information disclosed about Y by W is
minimized, thus minimizing privacy leakage, while preserving
the estimation efficiency that W provides about X. Similarly,
the solutions of the estimation privacy funnel (25) correspond
to the set of points {(x, LT (q, x)) |0 ≤ x ≤ m − 1} with
f1 (t) = f2 (t) = t2 − 1; see Fig. 1 (left).
One can move away from the usual IB and define new
bottleneck problems by considering different functions f1 and
f2 . For instance, if f1 (t) = f2 (t) = t2 − 1, then the
corresponding f -information, called χ2 -information, is defined
as
PW,X (W, X)
∆
2
− 1,
If1 (W ; X) = χ (W ; X) = E
PW (W ) PX (X)
(23)
We simplify the notation in (1) for χ2 -information as
Bχ2 (PXY , x) ,
max
W →X→Y
χ2 (W ; Y ) s.t.χ2 (W ; X) ≤ x, (24)
The reason to specifically study χ2 - information are twofold.
χ2 (X; Y ) =
Pd First, it has been shown in [6] that
th
i=1 λi (X; Y ), where λi (X; Y ) is the i principal inertia
component (PIC) of PX,Y and d = min {|X | , |Y|} − 1.
Moreover, if the PICs between X and Y are large, then
the minimum mean square error (MMSE) mmse (X|Y ) of
estimating X given Y will be small [6, Theorem 1], thereby
making reliable estimations. Hence, if the goal of an estimation
problem is to minimize mmse (Y |W ), we can equivalently
consider maximizing χ2 (Y ; W ). Second, following the spirit
of the IB, we also add the constraint χ2 (W ; X) ≤ x for
the new representation W , as χ2 -divergence serves as sharp
bounds for any f -divergence [17].
Due to the above connection between Bχ2 (PXY , x)
and estimation problems, we call Bχ2 (PXY , x) Estimation Bottleneck (EB) problem. Clearly, the set of points
{(x, UT (q, x)) |0 ≤ x ≤ m − 1} corresponds to the set of
solutions of (24); see Fig. 1 (left). The bound x ≤ m − 1
comes from the PIC analysis [6].
D. Estimation Privacy Funnel
Motivated by the connection between χ2 -information and
estimation problems mentioned in Section III-C, we propose
Fχ2 (PXY , x) ,
min
W →X→Y
χ2 (W ; Y ) s.t.χ2 (W ; X) ≥ x, (25)
IV. M RS .
AND
M R . G ERBER ’ S L EMMAS
The study of upper and lower boundaries of achievable mutual information pairs are essential in multi-user information
theory [13]. In the binary symmetric case, we not only rephrase
Mrs. Gerber’s lemma [13], but also derive its counterpart for
the PF. Furthermore, we discuss analogous results to Mrs. and
Mr. Gerber’s lemmas for Arimoto conditional entropy.
A. Mr. Gerber’s Lemma
We apply the duality argument, developed in Section II-C,
for LT and UT to characterize the IB and the PF the binary
symmetric case. In particular, let PY |X be the BSC with
crossover probability δ and q = Pr (X = 1) ≤ 12 . For
a ∈ [0, 1], denote ā = 1 − a. We denote by hb (q) the binary
entropy function h2 ([q, q̄]).
Let f1 = f2 = h2 . It was shown in [12] that
(26)
LT (q, x) = hb δ ⋆ h−1
b (x) , ∀x ∈ [0, hb (q)] ,
where h−1
: [0, 1] → [0, 12 ] is the inverse function of hb (·)
b
and a ⋆ b , (1 − a) b + (1 − b) a, for a, b ∈ [0, 1]. Eq. (26)
is well-known as Mrs. Gerber’s Lemma (MGL). In this case,
the matched channel is also a BSC with crossover probability
h−1
b (x). Using the approach outlined in Section II, we derive
a counterpart result for the upper boundary UT , and call it Mr.
Gerber’s Lemma.
Theorem 2 (Mr. Gerber’s Lemma). For 0 ≤ q ≤ 21 , we have
q
UT (q, x) = αhb δ ⋆
+ ᾱhb (δ) ,
(27)
z
where x = αhb qz and z = max (α, 2q), α ∈ [0, 1].
Proof. See the Appendix.
In summary, in the binary symmetric case, the set of
solutions for the IB follows from Mrs. Gerber’s Lemma (26)
and is given by {(hb (q) − x, hb (q ⋆ δ) − LT (q, x))}, and the
set of solutions for the PF follows from Mr. Gerber’s Lemma
(27) and is given by {(hb (q) − x, hb (q ⋆ δ) − UT (q, x))}.
The upper and lower boundaries of the achievable pairs
{I (W ; X) , I (W ; Y ) : W → X → Y } are therefore characterized by Mrs. and Mr. Gerber’s Lemmas, respectively.
B. Achievable Pairs of Arimoto Conditional Entropy
Beside χ2 -divergence and the entropy functions, one can
choose ℓβ -norm k·kβ for f and g in (4), which results
in Arimoto’s version of conditional Rényi entropy (Arimoto
conditional entropy) [15] of order β > 1:
Hβ (X|W ) ,
X
β
log
αw kpw kβ .
1−β
(28)
w∈W
When β = 1, we define H1 (X|W ) = H(X|W ). Hence,
the set of achievable Arimoto conditional entropy pairs
(Hβ (X|W ) , Hβ (Y |W )) can be obtained by the nonlinear
mapping:
β
β
(x, y) 7→
log x,
log y , (x, y) ∈ C (T) . (29)
1−β
1−β
With (28) at hand, Arimoto mutual information [15] of order
β > 1 can be defined as Iβ (X; W ) , Hβ (X) − Hβ (X|W ),
where Hβ (X) is the Rényi entropy of order β. Arimoto
conditional entropy has been proven useful in approximating
the minimum error probability of Bayesian M -ary hypothesis
testing [15].
C. Arimoto’s Mr. and Mrs. Gerber’s Lemmas
Due to the importance of Arimoto conditional entropy [15],
we study the extensions of Mr. and Mrs. Gerber’s Lemmas for
Arimoto conditional entropy, naming them Arimoto’s Mr. and
Mrs. Gerber’s Lemmas respectively.
n
o
Let Kβ (X|W ) = exp 1−β
H
(X|W
)
, and also
β
o β
n
1−β
Kβ (X) = exp β Hβ (X) . Since Hβ (X|W ) ≤ Hβ (X)
o
n
x
is strictly
for β > 1 and the mapping x 7→ exp 1−β
β
decreasing, we have Kβ (X|W ) ≥ Kβ (X). Define LT (q, x)
and UT (q, x) respectively as the minimum and maximum of
Kβ (Y |W ) when Kβ (X|W ) = x for Kβ (X) ≤ x ≤ 1. For
simplicity, denote Kβ (q) = Kβ (X) if X ∼ Bernoulli (q).
In this case, following section II-C, we have φ (p, λ) =
Kβ (δ ⋆ p) − λKβ (p), which leads to the following theorem.
Theorem 3 (Arimoto’s Mrs. Gerber’s Lemma). For 0 ≤ q ≤
∆
1/2, let L(β) = {(x, LT (q, x)) |Kβ (q) ≤ x ≤ 1} and for β >
1, we have
L(β) = {(Kβ (p) , Kβ (p ⋆ δ)) |0 ≤ p ≤ q}
In particular,
Hβ (X|W ) ≥
β
1−β
β
1−β
log y
=
min
W →X→Y
(β)
log x for (x, y) ∈ L
(30)
Hβ (Y |W ) s.t.
.
Proof. See the Appendix.
Analogous to this theorem, we also obtain the following
generalization of Mr. Gerber’s Lemma.
Theorem 4 (Arimoto’s Mr. Gerber’s Lemma). For 0 ≤ q ≤
∆
1/2, let U (β) = {(x, UT (q, x)) |Kβ (q) ≤ x ≤ 1} and for β >
1, we have
q
q
n
o
, αKβ
U (β) =
ᾱ + αKβ
⋆ δ + ᾱKβ (δ) ,
z
z
(31)
where α ∈ [0, 1] , z = max {α, 2q}. In particular, we have
β
β
max Hβ (Y |W ) s.t. Hβ (X|W ) ≤ 1−β
log x
1−β log y =
W →X→Y
for (x, y) ∈ U (β) .
Consequently, for β > 1, Arimoto’s Mrs. and Mr.
Gerber’s Lemmas jointly characterize the achievable sets
{Iβ (W ; X), Iβ (W ; Y ) : W → X → Y }; see Fig. 1 (right).
V. F INAL R EMARKS
In this paper, we study the geometric structure behind
bottleneck problems, and generalize the IB and PF to a broader
class of f -divergences. In particular, we consider estimationtheoretic variants of the IB and PF. Moreover, we show how
bottleneck problems can be used to calculate the counterpart
of Mrs. Gerber’s lemma (called Mr. Gerber’s Lemma), and
derive versions of Mrs. and Mr. Gerber’s lemmas for Arimoto
conditional entropy. These results can be potentially useful for
new applications of information theory in machine learning.
A PPENDIX
A. Lemma 2
Lemma 2 ( [18]). Let A be a connected and non-empty subset
of Rn , and B = convA. Assume there exists x ∈ ∂B such
m
that xP∈
/ A, then there exists {xi }i=1 , where mP≤ n with
x = i∈[m] αi xi , xi ∈ ∂A, 0 < αi < 1, and i αi = 1.
Furthermore, conv {xi }m
i=1 ⊆ ∂B.
B. Proof of Theorem 1 (Matched Channel)
Recall that LT (q, ·) is determined parametrically in λ by
the points where φ (·, λ) does not match its convex envelope
L∗T (·, λ). Thus, the columns of the channel transformation
matrix of a matched channel correspond to extreme points
PX|W (·|i) = pi where φ (·, λ) matches L∗T (·,P
λ). However,
there exists αi where the convex combination i∈[k] αi pi =
q corresponds to a point φ (q, λ) 6= L∗T (q, λ). Using lemma 2,
any non-trivial convex combination of pi will result in a point
q which is on the convex envelope of φ (·, λ) and determines
a corresponding point on the curve LT (q, ·).
C. Proof of Theorem 2 (Mr. Gerber’s Lemma)
Take f = g = h2 , we have φ (p, λ) = hb (p ⋆ δ) −
2
λhb (p). For λ ≥ (1 − 2δ) , φ (·, λ) is convex in p, and
2
∗
UT (q, λ) = hb (δ). For 0 ≤ λ < (1 − 2δ) , φ (·, λ) is
1
concave in a region centered at p = 2 , where it reaches a
local maximum. Consequently, if φ 21 , λ < hb (δ), the upper
convex envelope of Sλ is the linear combination of (0, φ (0, λ))
1
and (1,φ (1, λ)) and UT (q, x) = hb (δ). Assuming
1 p ≤ 2 , if
1
φ 2 , λ > hb (δ), then there exists pλ ∈ 0, 2 such that
∗
for p ≤ pλ , (p, UT
(q, λ)) ∈ Cλ is a convex combination
of
(0, φ (0, λ)) and (pλ , φ (pλ , λ)). Finally, if φ 12 , λ = hb (δ),
then anypoint on the upper
convex envelope of Sλ also lies
in conv φ (0, λ) , φ 21 , λ , φ (1, λ) .
Hence, assuming p ≤ 21 , the distribution PW,X that achieves
UT (q, x) will be of two cases:
1) P r (X = 1|W = 0) = 0, P r (X = 1|W = 1) = αp with
P r (W = 1) = α, 2p ≤ α ≤ 1.
2) W
assuming
values
in
{0, 1, 2}
with
P r (X = 1|W = 0) = 0, P r (X = 1|W = 1) = 1,
P r (X = 1|W = 2) = 21 , with P r (W = 0) = 1−p− α2 ,
P r (W = 1) = p − α2 , P r (W = 2) = α and
0 ≤ α ≤ 2p.
Rearranging (1) and (2), the result in (27) follows.
D. Proof of Theorem 3 (Arimoto’s Mr. Gerber’s Lemma)
2
Since φ (·, λ) is convex for λ ≤ (1 − 2δ) . For λ >
2
(1 − 2δ) , φ′′ (·, λ) is negative on an interval [pλ , p¯λ ], symmetric at p = 21 , and is positive elsewhere with local maximum
at p = 21 . By symmetry, the lower convex envelope of the
graph φ (·, λ) is obtained by replacing p = pλ if p ∈ [pλ , p¯λ ].
Therefore, for a given q ≤ 21 , if pλ ≥ q, then (q, LT (q, λ))
is a convex combination of (pλ , φ (pλ , λ)) and (p¯λ , φ (p¯λ , λ)).
Hence, we have LT (q, x) = Kβ (p ⋆ δ) for x = Kβ (p) and
0 ≤ p ≤ q.
R EFERENCES
[1] N. Tishby, F. C. Pereira, and W. Bialek, “The information bottleneck
method,” in Proc. of IEEE Allerton, 2000.
[2] N. Tishby and N. Slonim, “Data clustering by markovian relaxation and
the information bottleneck method,” in Proc. of NIPS, 2001.
[3] N. Slonim and N. Tishby, “Document clustering using word clusters via
the information bottleneck method,” in Proc. of ACM SIGIR, 2000.
[4] N. Tishby and N. Zaslavsky, “Deep learning and the information
bottleneck principle,” in Proc. of IEEE ITW, 2015.
[5] R. Shwartz-Ziv and N. Tishby, “Opening the black box of deep neural
networks via information,” arXiv preprint arXiv:1703.00810, 2017.
[6] F. P. Calmon, A. Makhdoumi, M. Médard, M. Varia, M. Christiansen,
and K. R. Duffy, “Principal inertia components and applications,”
vol. 63, no. 9, pp. 5011–5038, 2017.
[7] F. P. Calmon, A. Makhdoumi, and M. Médard, “Fundamental limits of
perfect privacy,” in Proc. of IEEE ISIT, 2015.
[8] P. Harremoës and N. Tishby, “The information bottleneck revisited or
how to choose a good distortion measure,” in Proc. of IEEE ISIT, 2007.
[9] Y. Polyanskiy and Y. Wu, “Dissipation of information in channels with
input constraints,” vol. 62, no. 1, pp. 35–55, Jan 2016.
[10] ——, “Lecture notes on information theory,” Lecture Notes for ECE563
(UIUC), vol. 6, pp. 2012–2016, 2014.
[11] F. P. Calmon, Y. Polyanskiy, and Y. Wu, “Strong data processing
inequalities in power-constrained gaussian channels,” in Proc. of IEEE
ISIT, 2015.
[12] H. Witsenhausen and A. Wyner, “A conditional entropy bound for a pair
of discrete random variables,” vol. 21, no. 5, pp. 493–501, 1975.
[13] A. El Gamal and Y.-H. Kim, Network information theory. Cambridge
university press, 2011.
[14] C. Nair, “Upper concave envelopes and auxiliary random variables,” Int.
J. Adv. Eng. Sci. Appl. Math., vol. 5, no. 1, pp. 12–20, 2013.
[15] I. Sason and S. Verdú, “Arimoto-Rényi conditional entropy and bayesian
m-ary hypothesis testing,” arXiv preprint arXiv:1701.01974, 2017.
[16] H. Hsu, S. Asoodeh, S. Salamatian, and F. P. Calmon, “Generalizing
bottleneck problems - extended version.” [Online]. Available:
https://github.com/HsiangHsu/ISIT-18-Extended-Version
[17] A. Makur and L. Zheng, “Bounds between contraction coefficients,” in
Proc. of IEEE Allerton, 2015.
[18] H. G. Eggleston, Convexity. Wiley Online Library, 1966.
| 7cs.IT
|
1
Optimal Power Allocation and Secrecy Sum Rate
in Two-Way Untrusted Relaying
Ali Kuhestani1, Phee Lep Yeoh2, and Abbas Mohammadi1
Engineering Department, Amirkabir University of Technology, Tehran, Iran.
2 School of Electrical and Information Engineering, The University of Sydney, NSW, Australia.
arXiv:1709.01330v1 [cs.CR] 5 Sep 2017
1 Electrical
Abstract—In this paper, we examine the secrecy performance
of two-way relaying between a multiple antenna base station
(BS) and a single antenna mobile user (MU) in the presence
of a multiple antenna friendly jammer (FJ). We consider the
untrusted relaying scenario where an amplify-and-forward relay
is both a necessary helper and a potential eavesdropper. To
maximize the instantaneous secrecy sum rate, we derive new
closed-form solutions for the optimal power allocation (OPA)
between the BS and MU under the scenario of relaying with
friendly jamming (WFJ). Based on the OPA solution, new closedform expressions are derived for the ergodic secrecy sum rate
(ESSR) with Rayleigh fading channel. Furthermore, we explicitly
determine the high signal-to-noise ratio slope and power offset of
the ESSR to highlight the benefits of friendly jamming. Numerical
examples are provided to demonstrate the impact of the FJ’s
location and number of antennas on the secrecy performance.
I. I NTRODUCTION
R
ELAYING is a proven approach to improve energy
efficiency, extend coverage and increase the throughput
of wireless communication networks. Recently, the benefits
of relaying have been viewed from the viewpoint of wireless
physical-layer security (PLS) [1]. A key area of interest is
the untrusted relaying scenario where the source-to-destination
transmission is assisted by a relay which may also be a
potential eavesdropper [2]. This scenario occurs in largescale wireless systems such as heterogeneous networks and
device-to-device (D2D) communications, where confidential
messages are often retransmitted by intermediate nodes.
Secure transmission utilizing an untrusted relay was first
studied in [3], where an achievable secrecy rate was derived. In
[4], it was found that introducing a friendly jammer (FJ) could
result in a positive secrecy rate for a one-way untrusted relay
link with no direct source-destination transmission. Indeed,
many recent papers on untrusted relay communications have
focussed on the one-way relaying scenario [5]– [7]. Recently,
several works have considered the more interesting scenario
of two-way untrusted relaying [8]-[10] where physical-layer
network coding can provide security enhancement since the
relay receives a superimposed signal from the two sources
instead of each individual signal [11].
Considering two-way untrusted relaying, [8] proposed a
game-theoretic power control scheme between the two single
antenna sources and multiple FJs. In the proposed scheme, the
optimal jamming power was derived for the special case where
the FJ is very close to the amplify-and-forward (AF) relay. In
[9], multiple antennas were considered at the sources and a
joint optimization of transmit covariance matrices and relay
selection was proposed for AF relaying without FJs. In [10],
an iterative algorithm was proposed to jointly optimize the
multiple antenna sources and AF relay beamformers such that
the instantaneous secrecy sum rate without FJs is maximized.
In this paper, we investigate the PLS of a two-way untrusted
AF relaying system, where a multiple antenna base station
(BS) exchanges confidential messages with a single antenna
mobile user (MU) in the presence of a multiple antenna
FJ. The relay is considered to be both a necessary helper
and a potential curious eavesdropper. For this system, we
analyze the secrecy performance under the scenario of relaying
with friendly jamming (WFJ). In particular, we formulate the
optimal power allocation (OPA) between the BS and MU that
maximizes the instantaneous secrecy sum rate of two-way
untrusted relaying. Based on this, we derive a new closedform solution for the OPA in the large-scale multiple antenna
(LSMA) regime for the BS and FJ. According to our OPA
solution, the ergodic secrecy sum rate (ESSR) of the optimized
system is derived for Rayleigh fading channel. We further
characterize the high signal-to-noise ratio (SNR) slope and the
high SNR power offset of the ESSR which explicitly captures
the impact of the distance dependent channel gains and the
number of BS and FJ antennas on the ESSR. Finally, numerical
results are provided to reveal the secrecy performance advantage of the OPA scheme compared with equal power allocation
and without friendly jamming. We highlight the significant
ESSR advantage of employing a FJ and the impacts of the FJ
location and number of BS and FJ antennas on the ESSR.
II. S YSTEM M ODEL
We consider a two-way relay network as depicted in Fig. 1
where a BS and MU exchange information via an untrusted
AF relay. Depending on the required secrecy sum rate for the
BS and MU, a FJ may be employed to transmit noise-like
jamming signals to the relay. The BS and FJ are equipped
with NBS and NFJ antennas, respectively, while the relay and
MU are equipped with a single antenna. Similar to [8], we
consider that the FJ is an external power seller that may be
activated to boost the secrecy sum rate of the BS and MU.
The two-way relaying transmission is performed in two
phases. In Phase 1, shown with solid lines, both the BS and
MU transmit their information to the relay. Simultaneously, the
FJ can be activated to transmit jamming signals to the relay.
In Phase 2, shown with dashed lines, the relay broadcasts a
combined version of the received signals to the BS and MU.
Based on knowledge of its own signal and the FJ’s jamming
2
respectively, where nBS and nMU are the AWGN at the BS
and MU, respectively.
Based on (3) and (4), and using (2), the signal-tointerference-plus-noise ratios (SINRs) at the BS (after performing MRC) and at the MU are given by
First Phase
Second Phase
(1 − λ)γbr γmr
,
(1 + λ)γbr + (1 − λ)γmr + γfr + ǫ
λγbr γmr
,
γMU =
λγbr + (2 − λ)γmr + γfr + ǫ
γBS =
User
(5)
(6)
respectively, where ǫ = 1. Based on (1), the SINR at the
untrusted relay is given by
Base Station
λγbr + (1 − λ)γmr
,
γfr + ǫ
γR =
Relay
Fig. 1. Two-way relaying between a multiple antenna base station (BS) and
a mobile user (MU) via an untrusted amplify-and-forward (AF) relay in the
presence of a multiple antenna friendly jammer (FJ).
signal, the BS and MU extracts the information signal from
its opposite counterpart. To utilize the multiple antennas, we
apply maximum ratio transmission (MRT) at the BS and FJ
in Phase 1, and maximum ratio combining (MRC) at the BS
in Phase 2.
We consider a total transmit power budget for the BS and
MU of P with power allocation factor λ ∈ (0, 1) such that
the transmit powers at the BS and MU are λP and (1 − λ)P ,
respectively. For simplicity, the transmit power at the relay and
the FJ are set to P . The complex Gaussian channel vectors
from the BS to relay, MU to relay and FJ to relay are denoted
by hbr ∼ CN (0NBS ×1 , µbr INBS ×1 ), hmr ∼ CN (0, µmr ) and
hfr ∼ CN (0NFJ ×1 , µfr INFJ ×1 ), respectively, where 0 is the
zero matrix, I is the identity matrix, and µbr , µmr , µfr are
the channel gains based on the distance-dependent path loss
for each antenna branch. We define γbr = ρkhbr k2 , γmr =
ρ|hmr |2 and γfr = ρkhfr k2 where ρ = σP2 and the additive
n
white Gaussian noise (AWGN) at all nodes are zero-mean with
variance σn2 . Also γ br = ρµbr , γ mr = ρµmr and γ fr = ρµfr .
Let x1 and x2 denote the transmitted scalar symbols from
the BS and MU, respectively. In Phase 1, the received signal
at the relay is expressed as
√
p
√
yR = λP khbr kx1+ (1 − λ)P hmr x2+ P khfr kxfj +nR , (1)
where xfj and nr represent the jamming signal and the AWGN
at the relay, respectively.
In Phase 2, the relay amplifies its received signal in (1) by
an amplification factor of
s
P
G=
, (2)
2
2
λP khbr k + (1 − λ)P |hmr |2 + P khfr k + σn2
and transmits a broadcast signal to the BS and MU given by
xR = GyR . After canceling the self interference and jamming
signals in yR , the corresponding received signals at the BS
and MU are given by
p
yBS = (1 − λ)P Ghbr hmr x2 + Ghbr nR + nBS , (3)
√
(4)
yMU = λP Gkhbr khmr x1 + Ghmr nR + nMU ,
(7)
where we assume that the relay performs multiuser decoding to
estimate the signals from the BS and MU. We adopt the wellknown SINR approximation of ǫ = 0 and note that from a
security perspective this corresponds to the maximum intercept
probability for the relay SINR in (7). From (5)–(7), we can
conclude that the jamming signal decreases all the SINR terms
but as observed in the next section, it has a more dominant
impact on the relay SINR.
The instantaneous secrecy sum rate of the two-way untrusted relaying system is given by [9] Rs = [IBS + IMU −
IR ]+ , where IK = 21 log2 (1 + γK ) is the achievable rate at
source K ∈ {BS , MU}, IR = 21 log2 (1 + γR ) is the information leaked to the untrusted relay, and [x]+ = max{0, x}. As
such, thehinstantaneous
i secrecy sum rate can be re-expressed
as Rs =
1
2
log2 Φ(λ)
+
Φ(λ) =
, where
(1 + γBS )(1 + γMU )
.
1 + γR
(8)
III. O PTIMAL P OWER A LLOCATION FOR S ECRECY S UM
R ATE M AXIMIZATION
In this section, we derive the OPA at the BS and MU that
maximizes the instantaneous secrecy sum rate. Indeed, we
need to consider the following optimization problem
λ⋆ = arg max Φ(λ), s.t. 0 < λ < 1
(9)
Note that Φ(λ) is a quasi-concave function of λ in the feasible
set. As such, by taking the derivative of φ(λ) w.r.t. λ and
setting it to zero, the OPA when γbr ≫ γmr is derived as
p
2
2
⋆ −2γmr−γfr+γmr 2γmr+3γmr γfr−2γmr+γfr−γfr
λ =
. (10)
γbr γmr
To obtain further insights from (10), we consider the case
where γfr ≫ γmr and γmr ≫ 1 which results in
γfr
.
(11)
λ⋆ =
γbr
This OPA applies to the case of high SNRs in both the relayto-jammer link and the relay-to-MU link. Note that γfr ≫
γmr occurs when the number of antennas at the FJ NFJ is
significantly large, or the relay is located much closer to the FJ
compared to the MU. It is also worth of noting that while we
have considered LSMA at both the BS and the FJ in deriving
(11), NBS should be larger than NFJ to ensure that λ < 1.
3
IV. E RGODIC S ECRECY S UM R ATE A NALYSIS
In this section, we apply the OPA solution (11) to derive new
closed-form expressions for the ESSR of two-way untrusted
relaying. Recall that the ESSR can be expressed as [9]
n
o
o
1 h n
Rs =
E ln(1 + γBS ) + E ln(1 + γMU )
2 ln 2 |
{z
} |
{z
}
I1
I2
n
oi
(12)
− E ln(1 + γR ) .
{z
}
|
I3
As such, in the following subsections, we proceed to evaluate
the ESSR terms of I1 , I2 and I3 .
Substituting λ⋆ in (11) into (5)–(7), the SINRs at the BS,
MU and relay for γbr ≫ γmr and γmr ≫ 1 yields
γBS = γmr
1−
1+
γfr
γbr
2γfr
γbr
γMU =
,
γmr
2
and
γR = 1.
(13)
′
T
+1 TX
+1
X
√
p −4b b x
bp e p q ,
Ei(−x) ≈ −4 2πa1 a2
(14)
p=1 q=1
= 2(T1+1) , a2 =
cot(θp−1(q−1) )−cot(θp(q) )
, and .065
θp(q) −θp−1(q−1)
′
1
2(T ′ +1) ,
=
and bp(q)
< θ1 < ... < θT +1(T ′ +1) .
T and T are positive integers that control the accuracy of the
approximation.
Using (13) and (14), I1 is obtained as (see Appendix A)
n
o
I1 = E ln(1 + γBS )
(4bp bq −1)(2− 3 )
√
̺
Γ(NFJ + NBS ) γ fr NBS
γ mr
= 4 2πa1 a2 e
Γ(NFJ )Γ(NBS ) γ br
′
T
+1
T
+1
h
X Xp
(4bp bq − 1)(3 − ̺3 )
×
bp − A1 Ei
γ mr
p=1 q=1
+
NFJ +N
XBS −1
Ai+1 −
i=1
+
e
(4bp bq −1)(3− 3 )
̺
−
γ mr
3
3− ̺
(γ
1
(i−1)!
∂
i−1
h
mr
)i
=
̺
where
NBS −1
u
(1 − 4bp qp )i Ei(−
i−1
X
(1 − 4bp bq )k (3 − ̺3 )k i
1 +
(u−1)
3
(4bp bq −1)(3− ̺
)
)
γ mr
i!
i(i − 1)...(i − k)
k=0
γ fr
γ br
i
and
Based on (13), in the high SNR regime with ρ → ∞, we
have ln(1 + γBS ) ≈ ln(γBS ) and ln(1 + γMU ) ≈ ln(γMU ). As
such, we can evaluate I1 and I2 as follows
n
o
n
n
γfr o
γfr o
I1 = E ln(γmr ) + E ln(1 −
) − E ln(1 + 2
) ,
γ
γ
|
{z
} |
{z br } |
{z br }
I1,1
I1,2
I1,3
n
o
n
o
I2 = E ln(γMU ) = E ln(γmr ) − ln 2,
{z
}
|
(20)
,
ANFJ +NBS +1−i
(15)
=
NFJ −1
and finally, using γR in (13), I3 is given by
n
o
I3 = E ln(1 + γR ) = ln 2.
where the values of I1,1 , I1,2 and I1,3 are (see Appendix B)
I1,1 = ln(γ mr ) − C,
(21)
1
Γ(NFJ + NBS ) γ fr NBS
×
I1,2 =
Γ(NFJ )Γ(NBS ) γ br
(−1)NBS +1
NFJX
+NBS h
̺−1
̺−1
Bi C1 ln(
× B1 Li2 (
)+
)
̺
̺
i=2
+
i−1
i
X
Cj
,
(1 − ̺)1−j − (−̺)1−j
1−j
j=2
(22)
and
Γ(NFJ + NBS ) γ fr NBS
Γ(NFJ )Γ(NBS ) γ br
h
2̺
2̺
2̺ − 2 i
× D1 ln(3) ln(
) + Li2 (
) − Li2 (
)
2̺ − 3
2̺ − 3
2̺ − 3
NFJX
+NBS i−1
ln(3)
2 Di h
ln(3)
−
+
+
i−1
i
−
1
(2̺)
(2̺
− 3)i−1
i=2
i
Bj
,
(23)
(2̺)1−j − (2̺ − 2)1−j
+
1−j
I1,3 =
(NFJ −1)(NFJ −2)....(NFJ +1−i)
(̺ −
(i−1)!
(−1)j−1 ̺−j
NFJ −i
1)
, Ci−j
=
, DNFJ +NBS +1−i
=
(j−1)!
(NFJ −1)(NFJ −2)....(NFJ +1−i)
NFJ −i
(1
−
̺)
and
P(i−1)!
∞ xk
Li2 (x) = k=1 k2 is the Dilogarithm function [13].
where BNFJ +NBS +1−i =
|u= 1̺ . According to γMU
in (13) and using [11, Eq. (4.331.2)], I2 can be obtained as
n
o
2
2
,
(16)
I2 = E ln(1 + γMU ) = −e− γ mr Ei −
γ mr
∂ui−1
where S∞ is the high SNR slope in bits/s/Hz/ (3 dB) and L∞
is the high SNR power offset in 3 dB units, which are defined
respectively as
Rs∞
R∞
and L∞ = lim log2 ρ − s . (19)
S∞ = lim
ρ→∞
ρ→∞ log2 ρ
S∞
I1,1
To derive the ESSR term of I1 , we use Lemma 3 in
[15] where the exponential integral function Ei(−x) [13] is
expressed as a closed-form solution as
where a1
by applying the general asymptotic form given by [12]
(18)
Rs∞ = S∞ log2 ρ − L∞ ,
Therefore, our closed-form expression for the asymptotic
ESSR of the optimized WFJ network is given by
1
Rs∞ =
2I1,1 + I1,2 − I1,3 − 2 ln 2 .
(24)
2 ln 2
Using (19), the high SNR slope and the high SNR power offset
are given by
S∞ = 1
(17)
We can further evaluate the asymptotic ESSR when ρ → ∞
and
(25)
I1,3
C
I1,2
+
+
+ 1. (26)
L∞ = − log2 µmr −
2 ln 2 2 ln 2 ln 2
To characterize the impact of NFJ and NBS on the ESSR
4
14
10
9
Exact, WFJ
OPA (10), WFJ
EPA, WFJ
OPA,WoFJ
FJ
12
(24)
N FJ=4, Eq. (28)
FJ
N =8, Eq. (28)
(24)
FJ
Ergodic Secrecy Sum Rate (bits/s/Hz)
8
Ergodic Secrecy Sum Rate (bits/s/Hz)
Exact,
Exact NFJ=4
NFJ=4,
=4, Eq.
Eq. (19)−(21)
(15)−(17)
N
7
6
5
4
3
10
NFJ=16, Eq. (28)
(24)
8
8
6
4
6
22 24 26 28
4
2
2
1
0
0
5
10
15
20
ρ [dB]
25
30
35
0
0
40
Fig. 2. ESSR versus transmit SNR, where NBS = 64, NFJ = 1, µbr =
µmr = 1, µfr = 4.
5
10
15
20
ρ [dB]
25
30
35
40
Fig. 3. ESSR versus transmit SNR for different number of antennas at the
FJ, where NBS = 256, µbr = µmr = 1, µfr = 4.
4.8
As such, the ESSR power offset is derived as
L∞ = − log2 µmr +
3NFJ γ fr
C
+
+ 1.
(NBS − 1)γ br ln 2 ln 2
(28)
Expression (28) shows that the power offset of WFJ depends
on both NFJ and NBS . As expected, we see that increasing the
number of BS antennas NBS decreases the ESSR power offset
which corresponds to an increase in the ESSR. Interestingly,
our results reveal that increasing the number of FJ antennas
NFJ actually increases the ESSR power offset which corresponds to a decrease in the ESSR.
4.6
Ergodic Secrecy Sum Rate (bits/s/Hz)
power offset in (26), we consider γbr ≫ γfr and approximate
2γfr
fr
fr
fr
ln(1 − γγbr
) ≈ − γγbr
and ln(1 + 2γ
γbr ) ≈ γbr . Since γfr and
γbr are independent, using Lemma 2.9 in [14] results in
n o n 1 o
NFJ γ fr
=−
I1,2 ≈ −E γfr E
and
γbr
(NBS − 1)γ br
2NFJ γ fr
.
(27)
I1,3 ≈
(NBS − 1)γ br
4.4
4.2
4
3.8
NBS=64, d=0.5
3.6
NBS=64, d=1
N
=256, d=0.5
BS
3.4
0
NBS=256, d=1
2
4
6
8
Number of FJ antennas (NFJ)
10
Fig. 4. Effect of distance between the FJ and relay on the ESSR for different
number of FJ antennas NFJ , where µbr = µmr = 1, n = 2, ρ = 20 dB.
The dashed lines correspond to the asymptotic SNR analysis from Eq. (24).
V. N UMERICAL E XAMPLES
This section provides numerical examples to verify the
accuracy of the derived ESSR expressions. We compare our
LSMA-based ESSR performance with the exact ESSR where
the OPA is numerically evaluated for finite numbers of BS and
FJ antennas using the bisection method. We also evaluate the
equal power allocation (EPA) between the BS and MU (i.e.,
λ = 0.5) and the scenario without friendly jamming (WoFJ)
where the power is distributed optimally between the users.
To verify the accuracy of the derived LSMA-based ESSR
expressions in (15)–(17) and (24), we conduct the following
simulations in Figs. 2 and 3 [8]: For simplicity and without
loss of generality, we assume that the BS, the MU, the relay
and the FJ are located at (−1, 0), (1, 0), (0, 0) and (0.3, 0.4),
respectively. Moreover, the path loss factor is α = 2.
Fig. 2 shows the ESSR versus ρ in dB for NBS = 64
and NFJ = 1. As can be observed from the figure: 1) The
ESSR using the OPA solution in (10) is sufficiently tight at
medium and high transmit SNRs, 2) The SNR gap between
the optimized network and EPA is approximately 4.3 dB
compared to WFJ when the ESSR is 5 bits/s/Hz, and 3) The
optimized WFJ significantly improves the ESSR performance
compared to the optimized WoFJ. For example, for a target
ESSR of 5 bits/s/Hz, the SNR advantage of WFJ over WoFJ
is approximately 18 dB which reveals the clear advantage of
using a FJ to establish an energy-efficient secure network.
Fig. 3 depicts the ESSR versus ρ in dB for NBS = 256 and
different values of NFJ = 4, 8, 16. We set T = T ′ = 20. As
can be seen from the figure, the Monte Carlo simulation of the
exact ESSR matches precisely with the ESSR expression in
(15)–(17) and is well-approximated by the asymptotic ESSR
expression in (24) in the high SNR regime. Furthermore, we
note that the ESSR decreases with increasing NFJ which can
be explained by the fact that the high-SNR power offset L∞
that was derived in (28) increases with increasing NFJ .
5
In Fig. 4, we examine the effect of the FJ location and
the number of FJ and BS antennas on the ESSR. The figure
highlights that for a given FJ location and number of BS
antennas, there is an optimal number of FJ antennas that
maximizes the ESSR. For example, when NBS = 64 and
d = 0.5, the ESSR is maximized when NFJ = 2. For a fixed FJ
location, we see that NFJ must increase with increasing NBS
to achieve a higher ESSR. This is because the received SINR
at the relay increases with increasing NBS with MRT at the
BS. By employing a FJ with more antennas, the information
leakage to the relay can be reduced.
VI. C ONCLUSION
We examined the OPA and ESSR of a two-way relaying
network including a BS, a MU, an untrusted relay and a FJ. We
considered the BS and FJ are equipped with multiple antennas
denoted as NBS and NFJ , respectively, and derived the OPA
between the BS and the MU such that the instantaneous
secrecy sum rate of the network is maximized. Based on
the OPA solution, new closed-form ESSR expressions were
derived for the optimized WFJ network in the LSMA regime.
Our asymptotic ESSR expressions highlighted that the ESSR
of WFJ depends on the number of antennas NBS and NFJ .
Numerical results revealed that exploiting a FJ improves the
ESSR significantly compared to without friendly jamming.
A PPENDIX A
To evaluate I1 , we first derive the cumulative distribution
∆ fr
function (cdf) of γfb = γγbr
as follows
Z ∞ Z βx
∂
fγfr (α)fγbr (β) dαdβ
fγfb (x) =
∂x 0
0
Γ(NFJ + NBS ) γ fr NBS
xNFJ −1
=
×
×
, (29)
γ
Γ(NFJ )Γ(NBS )
γ br
(x + fr )NFJ +NBS
γ br
where the last expression follows from Leibniz integral rule
and substituting the probability density function (pdf) of γfr
and γbr , and using [13, Eq. (3.381.4)]. Due to the fact that
γmr > 1 and γfb < 1, the cdf of γBS in (5) is obtained as
γ(1 + 2γ ) o
n
n
o
fb
FγBS (γ) = Pr γBS < γ = Eγfb Fγmr
1 − γfb
n − γ 1+2γfb o
(a)
γ mr
1−γfb
= 1 − Eγfb e
γ NBS
2γ Γ(N
(b)
FJ + NBS )
fr
×
= 1 − e γ mr
Γ(NFJ )Γ(NBS )
γ br
Z 1
3γ
1
z NFJ−1
dz,
(30)
e− γ mr × 1−z
(z + γγ fr )NFJ +NBS
0
br
where (a) and (b) follow from substituting the cdf of γmr and
the pdf of γfb , respectively. By using a change of variables
1
u = 1−z
and then applying the partial fraction decomposition
from [13, Eq. (2.102)], the cdf of γBS is formed. By using the
obtained cdf and [13, Eq. (3.352.4)], I1 is expressed as
o
n
Γ(NFJ + NBS ) γ fr NBS
=−
E ln 1 + γBS
Γ(NFJ )Γ(NBS ) γ br
3u−2
NFJX
+NBS Z ∞ e γ mr Ei(− 3u−2 )
γ mr
Ai
×
du.
(31)
1 i
(u
−
)
1
̺
i=1
By using a change of variables v = 3u−2
γ mr , the result in (14)
along with [11, Eq. (3.352.2)] and [11, Eq. (3.351.4)], the
result in (15) can be achieved after simple manipulations.
A PPENDIX B
Using [11, Eq. (4.352.2)], I1,1 is obtained as (21). Furthermore, using the pdf of γfb in (29), I1,2 is derived as
n
o Γ(N + N ) γ NBS
FJ
BS
fr
E ln(1 − γfb ) =
Γ(NFJ )Γ(NBS ) γ br
NFJX
+NBS Z 1
1
ln(u)
Bi
×
du.(32)
i
(−1)NBS +1 i=1
0 (u − ̺)
Using [13, Eq. (2.727.1)] and the partial fraction
decomposition of [13, Eq. (2.101)], I1,2 is computed as
shown in (22). The third part I1,3 can be calculated in the
same way as I1,2 and the final expression can be found in (23).
Acknowledgements
The authors deeply thank Prof. Lajos Hanzo for useful
discussions and constructive comments to improve the paper.
R EFERENCES
[1] A. Yenner and S. Ulukus, “Wireless physical-layer security: lessons
learned from information theory,” Proc. IEEE, vol. 103, no. 10, pp.
1814-1825, Sep. 2015.
[2] X. Chen, D. W. K. Ng, W. Gerstacker, and H-H. Chen, “A survey on
multiple-antenna techniques for physical layer security,” IEEE Commun.
Surveys Tuts., doi: 10.1109/COMST.2016.2633387.
[3] Y. Oohama, “Coding for relay channels with confidential messages,” in
Proc. Inf. Theory Workshop, 2001, pp. 87-89.
[4] X. He and A. Yener, “Two-hop secure communication using an untrusted
relay: A case for cooperative jamming,” in Proc. IEEE Globecom, New
Orleans, LA, Dec. 2008, pp. 1-5.
[5] L. Sun, T. Zhang, Y. Li, and H. Niu, “Performance study of two-hop
amplify-and-forward systems with untrustworthy relay nodes,” IEEE
Trans. Veh. Technol., vol. 61, no. 8, pp. 3801-3807, Oct. 2012.
[6] A. Kuhestani, A. Mohammadi and M. Noori, “Optimal power allocation
to improve secrecy performance of non-regenerative cooperative systems
using an untrusted relay,” IET Commun., vol. 10, no. 8, pp. 962-968,
May 2016.
[7] A. Kuhestani, A. Mohammadi, and M. Masoudi, “Joint optimal power
allocation and relay selection to establish secure transmission in uplink
transmission of untrusted relays network”, IET Network., vol. 5, no. 2,
pp. 30-36, Mar. 2016.
[8] R. Zhang, L. Song, Z. Han, and B. Jiao, “Physical layer security for
two-way untrusted relaying with friendly jammers,” IEEE Trans. Veh.
Technol., vol. 61, no. 8, pp. 3693-3704, Oct. 2012.
[9] J. Huang, A. L. Swindlehurst, “Joint transmit design and node selection
for one-way and two-way untrusted relay channels,” Asilomar Conference on Signals, Systems and Computers, pp. 1555-1559, Nov. 2013.
[10] J. Mo, M. Tao, Y. Liu, and R. Wang, “Secure beamforming for MIMO
two-way communications with an untrusted relay,” IEEE Trans. Signal
Process., vol. 62, no. 9, May 2014.
[11] K. P. Roshandeh, A. Kuhestani, M. Ardakani, and C. Tellambura,
“Ergodic sum rate analysis and efficient power allocation for a massive
MIMO two-way relay network,” IET Commun., vol. 11, no. 2, pp. 211217, Jan. 2017.
[12] A. Lozano, A. Tulino, and S. Verdu, “High-SNR power offset in
multiantenna communication,” IEEE Trans. Inf. Theory, vol. 51, no. 12,
pp. 4134-4151, Dec. 2005.
[13] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
Products, 7th ed. New York: Academic, 2007.
[14] A. M. Tulino and S. Verdu, “Random matrix theory and wireless communications,” Foundations and Trends in Communications and Information
Theory, vol. 1, no. 1, pp. 1-182, Jun. 2004.
[15] A. Alkheir and M. Ibnkahla, “An accurate approximation of the exponential integral function using a sum of exponentials,” IEEE Commun.
Lett., vol. 17, no. 7, pp. 1364-1367, Jul. 2013.
| 7cs.IT
|
arXiv:1508.06216v1 [cs.DS] 25 Aug 2015
Cardinality Estimation Meets Good-Turing
Reuven Cohen Liran Katzir Aviv Yehezkel
Department of Computer Science
Technion
Haifa 32000, Israel
August 26, 2015
Abstract
Cardinality estimation algorithms receive a stream of elements whose order might
be arbitrary, with possible repetitions, and return the number of distinct elements.
Such algorithms usually seek to minimize the required storage and processing at the
price of inaccuracy in their output. Real-world applications of these algorithms are
required to process large volumes of monitored data, making it impractical to collect
and analyze the entire input stream. In such cases, it is common practice to sample
and process only a small part of the stream elements. This paper presents and analyzes a generic algorithm for combining every cardinality estimation algorithm with
a sampling process. We show that the proposed sampling algorithm does not affect
the estimator’s asymptotic unbiasedness, and we analyze the sampling effect on the
estimator’s variance.
1
Introduction
Consider a very long stream of elements x1 , x2 , x3 , . . . , with repetitions. Finding the number
n of distinct elements is a well-known problem with numerous applications. The elements
might represent IP addresses of packets passing through a router [21, 24, 33], elements in a
large database [29], motifs in a DNA sequence [27], or nodes of RFID/sensor networks [36].
One can easily find the exact value of n by comparing the value of a newly encountered
element, xi , to every (stored) value encountered so far. If the value of xi has not been seen
before, it is stored as well. After all of the elements are treated, the stored elements are
counted. This simple approach does not scale if storage is limited, or if the computation
performed for each element xi should be minimized. In these cases, the following cardinality
estimation problem should be solved:
The cardinality estimation problem
Instance: A stream of elements x1 , x2 , x3 , . . . with repetitions, and an integer m. Let n be
the number of different elements, namely n = |{x1 , x2 , x3 , . . .}|, and let these elements
be {e1 , e2 , . . . , en }.
Objective: Find an estimate n
b of n using only m storage units, where m ≪ n.
As an application example, x1 , x2 , x3 , . . . could be IP packets received by a server. Each
packet belongs to one of n IP flows e1 , e2 , . . . , en , and the cardinality n represents the number
of active flows. By monitoring the number of distinct flows during every time period, a router
can estimate the network load imposed on the end server and detect anomalies. For example,
it can detect DDoS attacks on the server when the number of flows significantly increases
during a short time interval [17, 24].
Several algorithms have been proposed for the cardinality estimation problem [9, 10, 19,
28, 32, 33], all of which were designed to work on the entire stream, namely, without sampling.
However, real-world applications are required to process large volumes of monitored data,
making it impractical to collect and process the entire stream. For example, this is the case
for IP packets received over a high-speed link, because a 100 Gbps link creates a 1 TB log
file in less than 1.5 minutes. In such cases, only a small part of the stream is sampled and
processed [12, 14].
In this paper we present and analyze a generic algorithm that adds a sampling process
into every cardinality estimation procedure. The proposed algorithm consists of two steps:
(a) cardinality estimation of the sampled stream using any known cardinality estimator; (b)
estimation of the sampling ratio. We show that the proposed algorithm does not affect the
original estimator’s asymptotic bias (accuracy), and we analyze the algorithm’s effect on the
estimator’s variance (precision).
A naive approach to solving the cardinality estimation problem is to estimate the cardinality of the sampled stream and view it as an estimation for the cardinality of the whole
(unsampled) stream. However, this approach yields poor results because it ignores the probability of elements that do not appear in the sample. For example, we simulated a stream
of n = 10, 000 distinct elements whose frequency in the stream follows uniform distribution
∼ U(102 , 104 ). We then sampled 0.1% of the stream and used the HyperLogLog algorithm
[19] with m = 200 storage units to estimate the cardinality of the sample. We repeated
this test 200 times, each on a different stream of 10, 000 distinct elements, and averaged
the results. We found that the mean estimated cardinality
is E[b
n] ≈ 9, 100, which means
a bias of 9%, and that the relative variance is Var nnb ≈ 0.0552. In contrast, our proposed
algorithm computed a mean estimated cardinality
of E[b
n] ≈ 9, 900, namely a bias of only
1%, and a relative variance of only Var nnb ≈ 0.0118.
The rest of this paper is organized as follows. Section 2 discusses previous work. Section
3 presents our first algorithm (Algorithm 1) for combining the sampling process with a
generic cardinality estimation procedure. In addition, this section presents an analysis of
the asymptotic bias and variance of Algorithm 1. Section 4 presents our enhanced algorithm
(Algorithm 2), which uses subsampling in order to reduce the memory cost of Algorithm 1.
This section also presents an analysis of the asymptotic bias and variance of Algorithm 2.
Section 5 presents simulation results that validate our analysis in Sections 3 and 4. Finally,
Section 6 concludes the paper.
2
Related Work
Several works address the cardinality estimation problem [9, 10, 19, 28, 32, 33] and propose
statistical algorithms for solving it. These algorithms are efficient because they make only
one pass on the data stream, and because they use a fixed and small amount of storage.
The common approach is to use a random hash function that maps each element ej into a
low-dimensional data sketch h(ej ), which can be viewed as a random variable. The hash
function guarantees that h(ej ) is identical for all the appearances of ej . Thus, the existence
of duplicates, i.e., multiple appearances of the same element, does not affect the value of the
extreme order statistics. Let h be a hash function and h(xi ) denote the hash value of xi .
Then, an order statistics estimator or a bit pattern estimator can be used to estimate the
value of n. An order statistics estimator keeps the smallest (or largest) m hash values. These
values are then used to estimate the cardinality [6, 10, 28, 30, 32]. A bit pattern estimator
keeps the highest position of the leftmost (or rightmost) “1” bit in the binary representation
of the hash values in order to estimate the cardinality [9, 19].
Real-world applications of cardinality estimation algorithms are required to process large
volumes of monitored data, making it impractical to collect and analyze the entire input
stream. In such cases, it is common practice to sample and process only a small part of the
stream elements. For example, routers use sampling techniques to achieve scalability. The
industry standard for packet sampling is sFlow [1], short for “sampled flow”. Using a defined
sampling rate N, an average of 1 out of N packets is randomly sampled. The flow samples
are then sent as sFlow datagrams to a central monitoring server, which analyzes the network
traffic.
Although sampling techniques provide greater scalability, they also make it more difficult
to infer the characteristics of the original stream. One of the first works addressing inference
from samples is the Good-Turing frequency estimation, a statistical technique for estimating
the probability of encountering a hitherto unseen element in a stream, given a set of past
samples. For a recent paper on the Good-Turing technique, see [22].
Several other works have addressed the problem of inference from samples. For example,
the detection of heavy hitters, elements that appear many times in the stream, is studied in
[5]. The authors propose to keep track of the volume of data that has not been sampled.
Then, a new element is skipped only when its effect on the estimation will “not be too large.”
The case where the elements are packets has also been addressed. In such cases, the heavy
hitters are called elephants. The accuracy of detecting elephant flows is studied in [34] and
[35]. The authors use Bayes’ theorem for determining the threshold of sampled packets,
which indicates whether or not a flow is an elephant in the entire stream.
Other works have dealt with exploiting protocol-level information of sampled packets in
order to obtain accurate estimations of the size of flows in the network. For example, in [15]
the authors present a TCP-specific method whose estimate is based on the TCP SYN flag in
the sampled packets. Another method, which uses TCP sequence numbers, is presented in
[37]. These methods can also be used to estimate the cardinality of the flows in the network,
i.e., the number of active flows. However, both methods are limited to TCP flows. In this
paper we present a generic algorithm that does not make any assumptions regarding the
type of the input elements.
Related to the cardinality estimation problem is the problem of finding a uniform sample
of the distinct values in the stream. Such a sample can be used for a variety of database
management applications, such as query optimization, query monitoring, query progress
indication and query execution time prediction [4, 7, 8]. Additional applications of the
uniform sample pertain to approximate query answering, such as estimating the mean, the
variance, and the quantiles over the distinct values of the query [2, 3, 26]. Several algorithms
provide a uniform sample of the stream; for example, the authors of [25] show how to find
such a sample in a single data pass. Several variations of this work are also proposed in
[11, 20, 23]. However, all the discussed approaches require scanning the entire input stream,
which is usually impractical. In this paper we present a generic algorithm that does not
require a full data pass over the input stream.
The above works consider uniform packet sampling, where each packet is sampled with a
fixed probability. Previous works have also dealt with size-dependent flow sampling, where
packets are sampled with different probability, according to their flow size. The first works
on size-dependent flow sampling study the problem of deciding which records in a given set
of flow records should be discarded when storage constraints allow only a small fraction to
be kept [13, 14, 16]. The sampling decision in these works is made off-line: a flow is first
received and only then discarded or stored. In [31], the on-line version of this problem is
studied. In this version, upon receiving a packet, the algorithm needs to determine whether
to keep it. The authors develop a new packet sampling method that samples each packet with
probability f (b
s), where f is a decreasing function of the estimated size of the corresponding
flow when the packet is received, and the size of the flow is estimated using a small sketch
that stores the approximate sizes of all flows.
3
3.1
Cardinality Estimation with Sampling
Preliminaries: Good-Turing Frequency Estimation
The Good-Turing frequency estimation technique is useful in many language-related tasks
where one needs to determine the probability that a word will appear in a document.
Let X = {x1 , x2 , x3 , . . .} be a stream of elements, and let E be the set of all different
elements E = {e1 , e2 , . . . , en }, such that xi ∈ E. Suppose that we want to estimate the
probability π(ej ) that a randomly chosen element from X is ej . A naive approach is to
#(e )
choose a sample Y = {y1 , y2, . . . , yl } of l elements from X, and then to let π(ej ) = l j ,
where #(ej ) denotes the number of appearances of ej in Y . However, this approach is
inaccurate, because for each element ej that does not appear in Y even once (an “unseen
element”), #(ej ) = 0, and therefore π(ej ) = 0.
Let Ei = {ej |#(ej ) = i} be the set of elements that appear i times in the sample Y .
P
|
Thus,
|Ei | · i = l. The Good-Turing frequency estimation claims that Pbi = (i + 1) |Ei+1
is
l
a consistent estimator for the probability Pi that an element of X appears in the sample i
times.
c0 = |E1 | /l. In other words,
For the special case of P0 , we get from Good-Turing that P
the hidden mass P0 can be estimated by the relative frequency of the elements that appear
exactly once in the sample Y . For example, if 1/10 of the elements in Y appear only once
in Y , then approximately 1/10 of the elements in X do not appear in Y at all (i.e., they are
unseen elements).
3.2
The Proposed Algorithm
We now show how to use Good-Turing in order to combine a sampling process with a
generic cardinality estimation procedure, referred to as Procedure 1. As before, let X =
{x1 , x2 , x3 , . . .} be the entire stream of elements, and let Y = {y1 , y2 , . . . , yl } be the sampled
stream. Assume that the sampling rate is P , namely, 1/P of the elements of X are sampled
into Y . Let n and ns be the number of distinct elements in X and Y respectively. The
algorithm receives the sampled stream Y as an input and returns an estimate for n. The
algorithm consists of two steps: (a) estimating ns using Procedure 1 (any procedure, such
as in [9, 19, 28, 32]); (b) estimating n/ns , the factor by which to multiply the cardinality ns
of the sampled stream in order to estimate the cardinality n of the full stream.
To estimate ns in step (a), Procedure 1 is invoked using m storage units. To estimate
n/ns in step (b), we note that P0 = (n − ns )/n and thus 1/(1 − P0 ) = n/ns . Therefore, the
problem of estimating n/ns is reduced to estimating the probability P0 of unseen elements.
c0 = |E1 | /l is a consistent estimator for P0 . Thus,
As indicated above, by Good-Turing, P
we only need to find the number |E1 | of elements that appear exactly once in the sampled
stream Y . To compute the value of |E1 | precisely, one should keep track of all the elements in
Y and ignore each previously encountered element. This is done by Algorithm 1 below using
O(l) storage units. We later show (Algorithm 2 in Section 4) that the number of storage
units can be reduced by estimating the value of |E1 | /l.
Algorithm 1
(cardinality estimation with sampling)
(a) Estimate the number ns of distinct elements in the sample Y by invoking a cardinality
estimation procedure (Procedure 1) on this sample using m storage units.
(b) Determine the ratio n/ns by computing
computed precisely and l is known.
1
c0 ,
1−P
c0 = |E1 | /l. The value of |E1 | is
where P
[s as an estimator for the cardinality of the entire stream X.
(c) Return n
b = nbs · n/n
3.3
Analysis of Algorithm 1
In this section we analyze the asymptotic bias and variance of Algorithm 1, assuming that
the HyperLogLog algorithm [19] is used as Procedure 1. This
nb algorithm is the best known
cardinality estimator and it has a relative variance of Var n ≈ 1.08/m, where m is the
number of used storage units. Our main result is Theorem 1, where we prove that the
sampling does not affect the estimator’s asymptotic unbiasedness, and we show the effect of
the sampling rate P on the estimator’s variance.
We start with three preliminary lemmas. The first lemma shows how to compute the
probability distribution of a random variable that is a product of two normally distributed
random variables whose covariance is 0:
Lemma 1 (Product distribution)
Let X and Y be two random variables satisfying X → N (µx , σx2 ) and Y → N µy , σy2 , such
that Cov [X, Y ] = 0. Then, the product X · Y asymptotically satisfies the following:
X · Y → N µx µy , µ2y σx2 + µ2x σy2 .
A proof is given in [38].
The next lemma, known as the Delta Method, can be used to compute the probability distribution for a function of an asymptotically normal estimator using the estimator’s
variance:
Lemma 2 (Delta Method)
√
Let θm be sequence of random variables satisfying m(θm − θ) → N (0, σ 2 ), where θ and σ 2
are finite valued constants. Then, for every function g for which g ′ (θ) exists and g ′(θ) 6= 0,
the following holds:
√
2
m(g(θm ) − g(θ)) → N 0, σ 2 g ′(θ) .
A proof is given in [38].
The last lemma states a normal limit law for the estimation of |E1 | /l, where |E1 | and l
are as described in Section 3.1:
Lemma 3 (Random Sample’s Coverage)
1
2
\
|E
1 | /l → N |E1 | /l, l ((|E1 | + 2 |E2 |)/l − (|E1 | /l) ) .
A proof is given in [18].
We are now ready to start our analysis. Our first lemma summarizes the distribution of
P0 :
Lemma 4
1
c
P0 → N P0 , l P0 (1 − P0 ) + P1 , where l is the sample size.
Proof:
For the expectation, the following holds
h i
c0 = E[|E1 | /l] = |E1 | /l.
E P
c0 in Algorithm 1, and the second is because
The first equality is due to the definition of P
|E1 | and l are constants. By Good-Turing we get that |E1 | /l → P0 .
For the variance, the following holds
h i
c0 = Var [|E1 | /l] = 1/l · ((|E1 | + 2 |E2 |)/l − (|E1 | /l)2 ).
Var P
c0 in Algorithm 1. The second equality is due
The first equality is due to the definition of P
2
to Lemma 3. Finally,
due to Good-Turing we get that 1/l · ((|E1 | + 2 |E2 |)/l − (|E1 | /l) ) →
1
P0 (1 − P0 ) + P1 .
l
As shown in [19], when sampling
used, Procedure 1 estimates n with mean value n
is2not
n2
n
and variance m , namely, n
b → N n, m . The following theorem states the asymptotic bias
and variance of Algorithm 1 for P < 1.
Theorem 1
2
2
0 )+P1
Algorithm 1 estimates n with mean value n and variance nl P0 (1−P
+ nm , namely, n
b→
2
(1−P
)
0
2
2
0 )+P1
+ nm , where l is the sample size, and m is the storage size used for
N n, nl P0 (1−P
(1−P0 )2
estimating ns . In addition, P0 and P1 satisfy:
P
1. E[P0 ] = n1 ni=1 e−P ·fi .
P
2. E[P1 ] = Pn ni=1 fi · e−P ·fi
where fi is the frequency of element ei in X.
Proof:
Applying the Delta Method (Lemma 2) on 1 − P0 yields that
1
1
1 P0 (1 − P0 ) + P1
→N
.
,
c0
1 − P0 l
(1 − P0 )4
1−P
According to [9]:
n2s
.
n
bs → N ns ,
m
Next we show that
Cov ns ,
1
1−P0
(1)
(2)
and ns have zero covariance:
1
n
= Cov ns ,
1 − P0
n
s
n
n
= E Cov ns ,
| ns + Cov E[ns | ns ] , E
| ns
ns
ns
n
= 0 + Cov ns ,
ns
n
n
− E[ns ] E
= E ns ·
ns
ns
n
= E[n] − ns ·
ns
= n − n = 0.
The first equality is due to the P0 definition. The second equality is due to the law of total
covariance. The third equality is because ns and n/ns are independent when ns is known.
The fourth equality is due to the covariance definition. The fifth and sixth equalities are due
to the expectation definition and algebraic manipulations.
Applying the distribution product property (Lemma 1) for Eqs. (1) and (2) yields that:
n
bs
n2s P0 (1 − P0 ) + P1 n2s
1
n
b=
→ N n,
.
+
c0
l
(1 − P0 )4
m (1 − P0 )2
1−P
Finally, substituting ns = n · (1 − P0 ) yields that:
n2 P0 (1 − P0 ) + P1 n2
n
b → N n,
.
+
l
(1 − P0 )2
m
The resulting asymptotic variance depends on both P0 and P1 , which are determined
according to the sampling rate P and fi , the frequency of each distinct element in the
stream. Thus, the final part of the proof is to compute their expectation. For P0 we get
that:
n
n
n
n
1X
1X
1 X −1 P ·fi
1 X −P ·fi
E[P0 ] =
(1 − P )fi =
((1 − P )1/P )P ·fi =
(e )
=
e
.
n i=1
n i=1
n i=1
n i=1
The first equality is due to the expectation and P0 definitions. The second and the last
equalities are due to algebraic manipulations. The third equality is due to the known limit
result where (1 − x)1/x → e−1 when x → 0 (in our case P → 0).
For P1 we get that:
n
n
n
1X
1X
P X
fi −1
1/P P ·(fi −1)
E[P1 ] =
fi · P (1 − P )
=
fi · P ((1 − P ) )
=
fi · e−P ·fi .
n i=1
n i=1
n i=1
The first equality is due to the expectation and P1 definitions. The second and third equalities
are due to algebraic manipulations and the same known limit result noted above.
4
Reducing the Computational Cost of Algorithm 1
4.1
Algorithm 2 with Subsampling
Algorithm 1 computes |E1 | precisely. To this end, it uses O(l) storage units, which is linear
in the sample size. We now show how to reduce this cost by approximating the value of |E1 |
using a subsample U of the sample Y (see Figure 1).
Algorithm 2
(cardinality estimation with sampling and subsampling)
Same as Algorithm 1, except that in step (b) the ratio |E1 | /l is estimated by invoking Procedure 2 using only u ≪ l storage units.
Procedure 2:
1. Uniformly subsample u elements from the sampled stream Y . Let this subsample be
U.
X (the original stream)
Y (a sample of X)
U
(a sample of Y)
Figure 1: The relationship between X, Y and U
2. Compute (precisely) the number |U1 | of elements that appear only once in U.
c0 = |U1 | /u.
3. Return P
The intuition behind Algorithm 2 is that the cheap operation of Algorithm 1, estimating
ns , is performed on the whole sample Y , whose length is l, while the expensive operation,
computing the number of elements that appear only once (|E1 |), is performed on a small
subsample U of length u, where u ≪ l.
Uniform subsampling (step (1) in Procedure 2) can be implemented using one-pass reservoir sampling [39], as follows. First, initialize U with the first u elements of Y , namely,
y1 , y2 , . . . , yu , and sort them in decreasing order of their hash values. When a new element
is sampled into Y , its hash value is compared to the current maximal hash value of the
elements in U. If the hash value of the new element is smaller than the current maximal
hash value of U, the new value is stored in U instead of the element with the maximal hash
value. After all of the elements are treated and the sample Y is created, U is a uniform
subsample of length u.
We now analyze the running time complexity of Algorithm 2. Both steps (a) and (b)
are performed using a simple pass over the sample Y , and require O(1) operations per
sampled element. Thus, these steps require O(l) operations. Step (b) requires additional
O(u) operations for each insertion of an element into U. On the average, there are O(log l)
such insertions. The total complexity is thus O(l + u · log l) = O(l), which is similar to that
of Algorithm 1. However, the main advantage of Algorithm 2 over Algorithm 1 is that it
requires only m + u storage units, while Algorithm 1 requires m + l storage units, where
u ≪ l.
Next, we analyze the asymptotic bias and variance of Algorithm 2, assuming that the
HyperLogLog algorithm [19] is used as Procedure 1. Then we generalize the analysis for any
cardinality estimation procedure.
4.2
Analysis of Algorithm 2
Our main result is Theorem 2, which proves that the subsampling does not affect the asymptotic unbiasedness of the estimator and analyzes the effect of the sampling rate P on the
estimator’s variance, with respect to the storage sizes m and u.
P Let Zi be the set of elements that appear exactly i times in the subsample U; thus,
|Zi | · i = u and Z1 is the set of elements that appear only once in U. |Z1 | can be written
using indicator variables as:
|Z1 | =
u
X
Ij ,
where
j=1
1 if the j’th element in U has a single appearance in the subsample
0 otherwise.
\
Consider the estimator |E
1 | /l for |Z1 | /u. By definition, the variable |Z1 | follows a hypergeometric distribution, which can be relaxed to a binomial distribution if u ≪ l [38]. Thus,
due to binomial distribution properties, the expectation is
h
i
\
E |E
|
/l
|
|E
|
= E[|Z1 | /u] = E[Ij ] = |E1 | /l,
(3)
1
1
Ij =
and the variance is
h
i
\
Var |E
|
/l
|
|E
|
= Var [|Z1 | /u] = 1/u · Var [Ij ] = 1/u · |E1 | /l · (1 − |E1 | /l).
1
1
(4)
The following lemma summarizes the distribution of P0 :
Lemma 5
c0 → N P0 , 1 2P0 (1 − P0 ) + P1 .
P
u
Proof:
For the expectation, the following holds
h i
h
i
h h
ii
\
\
c0 = E |E
E P
|
/l
=
E
E
|E
|
/l
|
|E
|
= E[|E1 | /l] = |E1 | /l.
1
1
1
The first equality is due to Procedure 2. The second equality is due to the law of total
expectation. The third equality is due to Eq. 3. The fourth equality is due to Lemma 3.
By Good-Turing we get that |E1 | /l → P0 . For the variance, the following holds:
h i
h
i
\
c
Var P0 = Var |E1 | /l
h h
ii
h
h
ii
\
\
= Var E |E
|
/l
|
|E
|
+
E
Var
|E
|
/l
|
|E
|
1
1
1
1
= 1/u · ((|E1 | + 2 |E2 |)/l − (|E1 | /l)2 ) + 1/u · |E1 | /l · (1 − |E1 | /l)
= 2/u · ((|E1 | + |E2 |)/l − (|E1 | /l)2 ).
The first equality is due to Procedure 2. The second equality is due to the law of total
variance. The third equality is due to Eq. 4 and Lemma 3. The fourth equality is due to
algebraic manipulations.
By Good-Turing we get that 2/u · ((|E1 | + |E2 |)/l − (|E1 | /l)2 ) → u1 2P0 (1 − P0 ) + P1 .
The following theorem states the asymptotic bias and variance of Algorithm 2 for P < 1.
Theorem 2
2
2
(1−P0 )+P1
Algorithm 2 estimates n with mean value n and variance nu 2P0(1−P
+ nm , namely,
2
0)
2
2
(1−P0 )+P1
+ nm . In addition, P0 and P1 can be estimated as described in
n
b → N n, nu 2P0(1−P
2
0)
Theorem 1.
Proof:
Applying the Delta Method (see Section 3.3) on 1 − P0 yields that:
1
1 2P0 (1 − P0 ) + P1
1
.
,
→N
c0
1 − P0 u
(1 − P0 )4
1−P
(5)
According to [9]:
n2s
n
bs → N ns ,
.
(6)
m
i
h
1
= 0 (see Section 3.3); applying the distribution product property
Recall that Cov ns , 1−P
0
(see Section 3.3) for Eqs. (5) and (6) yields that:
n
bs
n2s 2P0 (1 − P0 ) + P1 n2s
1
n
b=
→ N n,
.
+
c0
u
(1 − P0 )4
m (1 − P0 )2
1−P
Finally, substituting ns = n · (1 − P0 ) yields that:
n2 2P0 (1 − P0 ) + P1 n2
n
b → N n,
.
+
u
(1 − P0 )2
m
The resulting asymptotic variance depends on both P0 and P1 , which are determined
according to the sampling rate P and fi , the frequency of each distinct element in the
stream, as was described in Section 3.3.
The analysis above assumes that the HyperLogLog algorithm [19] is used as Procedure
1. Recall that the asymptotic relative efficiency (ARE) of cardinality estimator n
b is defined
2
1
as the ratio ARE = nm · Var[b
.
For
example,
the
ARE
of
bottom-m
sketches
[28]
is 1.00,
n]
and the ARE of the maximal-term sketch in [9] is 0.93. The following theorem generalizes
Theorem 2 for any cardinality estimation procedure.
Theorem 3
2
(1−P0 )+P1
1 n2
+
, namely,
Algorithm 2 estimates n with mean value n and variance nu 2P0(1−P
2
)
ARE
m
0
2
(1−P0 )+P1
1 n2
, where ARE is the asymptotic relative efficiency of
n
b → N n, nu 2P0(1−P
+ ARE
2
m
0)
Procedure 1. In addition, P0 and P1 can be estimated as described in Theorem 1.
The proof is identical to that of Theorem 2.
5
Simulation Results
In this section we validate our analysis for the asymptotic bias and variance of Algorithm
1 and Algorithm 2, as stated in Theorems 1 and 2 respectively. We implement both algorithms using the HyperLogLog [19] as Procedure 1, and simulate a stream of n distinct
elements. Each distinct element ej appears fj times in the original (unsampled) stream.
These frequencies are determined according to the following models:
1. Uniform distribution: The frequency of the elements is uniformly distributed between
100 and 10, 000; i.e., fj ∼ U(102 , 104 ).
2. Pareto distribution: The frequency of the elements follows the heavy-tailed rule with
shape parameter α and scale parameter s = 500; i.e., the frequency probability function
is p(fj ) = αsα f −α−1 , where α > 0 and fj ≥ s > 0. The scale parameter s represents
the smallest possible frequency.
Pareto distribution has several unique properties. In particular, if α ≤ 2, it has infinite
variance, and if α ≤ 1, it has infinite mean. As α decreases, a larger portion of the probability
mass is in the tail of the distribution, and it is therefore useful when a small percentage of
the population controls the majority of the measured quantity.
Table 1 presents the simulation results for Algorithm 1 using uniformly distributed frequencies. The number of distinct elements is n = 10, 000. Thus, the expected length of
the original stream X is 10, 000 · 100+10,000
= 50.5 · 106 . We examine two sampling rates:
2
P = 1/100 (Table 1(a)) and P = 1/1000 (Table 1(b)). We use different m values, and for
every m average the results over 200 different runs. In each table row we present, for every
m, the bias and the variance. The bias column is only from the simulations and it is always
very close to 0, as proven in our analysis. For the variance we have two values: one from the
analysis (Theorem 1) and one from the simulations.
The results in Table 1 show very good agreement between the simulation results and our
analysis. First, as already said, the bias values are all very close to 0. Second, the simulation
variance is always very close to the analyzed variance.
variance
analysis simulation
50 0.0023 0.0200
0.0191
100 0.0134 0.0100
0.0116
150 0.0094 0.0067
0.0057
m
bias
(a) P = 1/100
variance
analysis simulation
50 0.0141 0.0209
0.0174
100 0.0094 0.0114
0.0099
150 0.0036 0.0096
0.0087
m
bias
(b) P = 1/1000
Table 1: Simulation results for Algorithm 1 using uniformly distributed frequencies
Next, we consider Algorithm 2 and seek to validate Theorem 2. Table 2 presents the
simulation results for uniform distribution of the frequencies. The total storage budget is
200 units, which are partitioned between m and u. The number of distinct elements is
n = 10, 000. We examine again two sampling rates: P = 1/100 and P = 1/1000. Table
3 presents results for the Pareto distribution of the frequencies, with α = 1.1, n = 10, 000,
P = 1/100, and a total storage budget of 2, 000 units. The results are averaged again over
200 runs, and the variance from the analysis is determined according to Theorem 2.
m
u
bias
10 190 0.0439
50 150 0.0025
100 100 0.0029
150 50 0.0037
190 10 0.0058
variance
analysis simulation
0.1000
0.1149
0.0200
0.0217
0.0101
0.0121
0.0068
0.0075
0.0060
0.0054
(a) P = 1/100
m
u
bias
10 190 0.0093
50 150 0.0184
100 100 0.0114
150 50 0.0060
190 10 0.0142
variance
analysis simulation
0.1000
0.1081
0.0200
0.0199
0.0101
0.0118
0.0068
0.0059
0.0058
0.0053
(b) P = 1/1000
Table 2: Simulation results for Algorithm 2 using uniform distribution and m + u = 200
storage units
In both tables we see again that the bias is indeed practically 0 and that the variance of
the algorithm as found by the simulations is very close to the variance found by our analysis.
These results are very consistent, for both frequency distributions, both sampling rates, and
all m and u values. As expected, when m + u increases (more storage is used), the variance
decreases.
variance
analysis simulation
50 1950 0.00005 0.0200
0.0217
100 1900 0.0189
0.0100
0.0104
500 1500 0.0011
0.0020
0.0023
1000 1000 0.00001 0.0010
0.0009
1500 500 0.0107
0.0007
0.0006
m
u
bias
Table 3: Simulation results for Algorithm 2 using Pareto distribution and m + u = 2000
storage units
We now want to compare the performance of Algorithms 1 and 2. Recall that Algorithm
2 is expected to have a higher variance, but with significantly less storage. In Theorems 1
and 2 we got the following closed expressions for the relative variance of the algorithms:
1. Algorithm 1:
1 P0 (1−P0 )+P1
l
(1−P0 )2
2. Algorithm 2:
1 2P0 (1−P0 )+P1
u
(1−P0 )2
+
1
.
m
+
1
.
m
Recall that m + l is the total storage used by Algorithm 1 (l is the sample length), and
m + u is the total storage used by Algorithm 2. The probabilities P0 and P1 are determined
according to the sampling rate P and the frequency distribution of the distinct elements in
the stream (see Theorem 1). Therefore, in a given stream, the only parameters that need to
be determined by the user are m in Algorithm 1, and m and u in Algorithm 2. In order to
find the values of m and u that yield the minimal variance for a given input stream, one only
needs to know the sampling rate and then minimize the relative variance function stated
above.
Table 4 presents the simulation results for n = 10, 000, a uniform distribution of element
frequencies, and for several sampling rates. Table 4(a) presents the variance of Algorithm 1.
In each table row we present the sample length l, the value of m, the total storage used by
the algorithm (m + l), and the simulation variance (averaged over 200 different runs). Recall
that in addition to m, Algorithm 1 uses O(l) storage units for the exact computation of |E1 |.
Table 4(b) presents the minimal variance of Algorithm 2 as a function of B. B indicates
the total number of storage units we are willing to spend. In each table row we present the
optimal partition of B between m and u that minimizes the variance of the estimator, and
the simulation variance for these m and u values. For the case where P = 1 (no sampling),
we provide in both tables the simulation variance of HyperLogLog [19], which we use as
Procedure 1. This
nb algorithm is the best known cardinality estimator and it has a relative
variance of Var n ≈ 1.08/m [19]. In this case we do not provide the values of l, m and u
as there is no meaning to these parameters because sampling is not used.
P
1/100
1/500
1/1000
1
m
100
500
1000
100
500
1000
100
500
1000
100
500
1000
storage
l
505,000
101,000
50,500
-
total
505,100
505,500
506,000
101,100
101,500
102,000
50,600
51,000
51,500
100
500
1000
(a) Algorithm 1
variance
(simulation)
0.0116
0.0018
0.0009
0.0095
0.0021
0.0008
0.0099
0.0019
0.0008
0.0101
0.0021
0.0010
P
1/100
1/500
1/1000
1
B
100
500
1000
100
500
1000
100
500
1000
100
500
1000
storage
m
92
460
921
80
401
803
72
363
724
-
u
8
40
79
20
99
197
28
137
276
-
variance
(simulation)
0.0112
0.0022
0.0009
0.0126
0.0027
0.0011
0.0152
0.0031
0.0013
0.0101
0.0021
0.0010
(b) Algorithm 2
Table 4: Simulation results for Algorithms 1 and 2 using uniform distribution
We can easily see from the tables that the storage-variance trade-off of Algorithm 2 is
significantly better than that of Algorithm 1. For example, the same variance (0.011) is
obtained by both algorithms in the first row of P = 1/100. However, in this row Algorithm
1 uses 505,100 storage units whereas Algorithm 2 uses only 100. For P = 1/500, we see that
the same variance (0.002) is obtained by the two algorithms when Algorithm 1 uses 101,500
storage units while Algorithm 2 uses only 500.
6
Conclusions
In this paper we studied the problem of estimating the number of distinct elements in a
stream when only a small sample of the stream is given. We presented Algorithm 1, which
combines a sampling process with a generic cardinality estimation procedure. The proposed
algorithm consists of two steps: (a) cardinality estimation of the sampled stream using
any known cardinality estimator; (b) estimation of the sampling ratio using Good-Turing
frequency. Then we presented an enhanced algorithm that uses subsampling in order to
reduce the memory cost of Algorithm 1. We proved that both algorithms do not affect the
asymptotic unbiasedness of the original estimator. We also analyzed the sampling effect
on the asymptotic variance of the estimators. Finally, we presented simulation results that
validate our analysis and showed how to find the optimal parameter values that yield the
minimal variance.
References
[1] http://www.sflow.org.
[2] S. Acharya, P. B. Gibbons, and V. Poosala. Congressional samples for approximate
answering of group-by queries. In Proceedings of the 2000 ACM SIGMOD International
Conference on Management of Data, pages 487–498.
[3] S. Acharya, P. B. Gibbons, V. Poosala, and S. Ramaswamy. Join synopses for approximate query answering. In SIGMOD 1999, pages 275–286.
[4] B. Babcock, S. Chaudhuri, and G. Das. Dynamic sample selection for approximate
query processing. In Proceedings of the 2003 ACM SIGMOD International Conference
on Management of Data, pages 539–550.
[5] S. Bhattacharyya, A. Madeira, S. Muthukrishnan, and T. Ye. How to scalably and
accurately skip past streams. In ICDE 2007, pages 654–663.
[6] P. Chassaing and L. Gérin. Efficient estimation of the cardinality of large data sets.
In Proceedings of the 4th Colloquium on Mathematics and Computer Science, pages
419–422, 2006.
[7] S. Chaudhuri, G. Das, and V. R. Narasayya. A robust, optimization-based approach for
approximate answering of aggregate queries. In Proceedings of the 2001 ACM SIGMOD
International Conference on Management of Data, pages 295–306.
[8] S. Chaudhuri, G. Das, and V. R. Narasayya. Optimized stratified sampling for approximate query processing. ACM Trans. Database Syst., 32(2):9, 2007.
[9] P. Clifford and I. A. Cosma. A statistical analysis of probabilistic counting algorithms.
Scandinavian Journal of Statistics, 2011.
[10] E. Cohen and H. Kaplan. Tighter estimation using bottom k sketches.
1(1):213–224, 2008.
PVLDB,
[11] G. Cormode, S. Muthukrishnan, and I. Rozenbaum. Summarizing and mining inverse
distributions on data streams via dynamic inverse sampling. In Proceedings of the 31st
International Conference on Very Large Data Bases, VLDB 2005, pages 25–36.
[12] N. G. Duffield. Sampling for passive internet measurement: A review. In Statistical
Science, volume 19, pages 472–498, 2004.
[13] N. G. Duffield and C. Lund. Predicting resource usage and estimation accuracy in an IP
flow measurement collection infrastructure. In Proceedings of the 3rd ACM SIGCOMM
Internet Measurement Conference, IMC 2003, pages 179–191.
[14] N. G. Duffield, C. Lund, and M. Thorup. Charging from sampled network usage. In
Proceedings of the 1st ACM SIGCOMM Workshop on Internet Measurement 2001, pages
245–256.
[15] N. G. Duffield, C. Lund, and M. Thorup. Estimating flow distributions from sampled
flow statistics. In Proceedings of the ACM SIGCOMM 2003 Conference on Applications,
Technologies, Architectures, and Protocols for Computer Communication, pages 325–
336.
[16] N. G. Duffield, C. Lund, and M. Thorup. Flow sampling under hard resource constraints. In Proceedings of the International Conference on Measurements and Modeling
of Computer Systems, SIGMETRICS 2004, pages 85–96.
[17] C. Estan, G. Varghese, and M. E. Fisk. Bitmap algorithms for counting active flows on
high-speed links. IEEE/ACM Trans. Netw., 14(5):925–937, 2006.
[18] W. W. Esty. A normal limit law for a nonparametric estimator of the coverage of a
random sample. The Annals of Statistics, 11(3):905–912, 1983.
[19] P. Flajolet, É. Fusy, O. Gandouet, and F. Meunier. Hyperloglog: the analysis of a
near-optimal cardinality estimation algorithm. In Analysis of Algorithms (AofA) 2007.
[20] G. Frahling, P. Indyk, and C. Sohler. Sampling in dynamic data streams and applications. Int. J. Comput. Geometry Appl., 18(1/2):3–28, 2008.
[21] É. Fusy and F. Giroire. Estimating the number of active flows in a data stream over a
sliding window. In ANALCO 2007, pages 223–231.
[22] W. A. Gale and G. Sampson. Good-turing frequency estimation without tears. Journal
of Quantitative Linguistics, 2(3):217–237, 1995.
[23] S. Ganguly. Counting distinct items over update streams.
378(3):211–222, 2007.
Theor. Comput. Sci.,
[24] S. Ganguly, M. N. Garofalakis, R. Rastogi, and K. K. Sabnani. Streaming algorithms
for robust, real-time detection of ddos attacks. In ICDCS 2007.
[25] P. B. Gibbons. Distinct sampling for highly-accurate answers to distinct values queries
and event reports. In VLDB 2001, pages 541–550.
[26] P. B. Gibbons and Y. Matias. New sampling-based summary statistics for improving
approximate query answers. In SIGMOD 1998, pages 331–342.
[27] F. Giroire. Directions to use probabilistic algorithms for cardinality for dna analysis.
Journés Ouvertes Biologie Informatique Mathématiques, 2006.
[28] F. Giroire. Order statistics and estimating cardinalities of massive data sets. Discrete
Applied Mathematics, 157:406–427, 2009.
[29] S. Heule, M. Nunkesser, and A. Hall. Hyperloglog in practice: Algorithmic engineering
of a state of the art cardinality estimation algorithm. In Proceedings of the EDBT 2013
Conference.
[30] Z. B.-Y. T. S. Jayram, R. Kumar, D. Sivakumar, and L. Trevisan. Counting distinct
elements in a data stream. In Proceedings of the 6th International Workshop on Randomization and Approximation Techniques, RANDOM 2002, pages 1–10.
[31] A. Kumar and J. J. Xu. Sketch guided sampling - using on-line estimates of flow size
for adaptive data collection. In INFOCOM 2006.
[32] J. Lumbroso. An optimal cardinality estimation algorithm based on order statistics and
its full analysis. In Analysis of Algorithms (AofA) 2010.
[33] A. Metwally, D. Agrawal, and A. E. Abbadi. Why go logarithmic if we can go linear?:
Towards effective distinct counting of search traffic. In Proceedings of the 11th International Conference on Extending Database Technology: Advances in Database Technology,
EDBT 2008, pages 618–629.
[34] T. Mori, T. Takine, J. Pan, R. Kawahara, M. Uchida, and S. Goto. Identifying heavyhitter flows from sampled flow statistics. IEICE Transactions, 90-B(11):3061–3072,
2007.
[35] T. Mori, M. Uchida, R. Kawahara, J. Pan, and S. Goto. Identifying elephant flows
through periodically sampled packets. In Proceedings of the 4th ACM SIGCOMM Conference on Internet Measurement 2004, pages 115–120.
[36] C. Qian, H. Ngan, Y. Liu, and L. M. Ni. Cardinality estimation for large-scale RFID
systems. IEEE Trans. Parallel Distrib. Syst., 22(9):1441–1454, 2011.
[37] B. F. Ribeiro, D. F. Towsley, T. Ye, and J. Bolot. Fisher information of sampled packets:
an application to flow size estimation. In Proceedings of the 6th ACM SIGCOMM
Conference on Internet Measurement 2006, pages 15–26.
[38] J. Shao. Mathematical Statistics. Springer, 2nd edition, 2003.
[39] J. S. Vitter. Random sampling with a reservoir. ACM Trans. Math. Softw., 11(1):37–57,
1985.
| 8cs.DS
|
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
arXiv:1611.00936v4 [math.GR] 21 Dec 2017
Journal of Knot Theory and Its Ramifications
c World Scientific Publishing Company
SIMPLY CONNECTED LATIN QUANDLES
MARCO BONATTO
Department of Algebra, Faculty of Mathematics and Physics, Charles University
Sokolovská 83, 18675 Praha, Czech Republic
[email protected]
PETR VOJTĚCHOVSKÝ
Department of Mathematics, University of Denver
2280 S Vine St, Denver, Colorado 80112, U.S.A.
[email protected]
ABSTRACT
A (left) quandle is connected if its left translations generate a group that acts transitively on the underlying set. In 2014, Eisermann introduced the concept of quandle
coverings, corresponding to constant quandle cocycles of Andruskiewitsch and Graña.
A connected quandle is simply connected if it has no nontrivial coverings, or, equivalently, if all its second constant cohomology sets with coefficients in symmetric groups
are trivial.
In this paper we develop a combinatorial approach to constant cohomology. We
prove that connected quandles that are affine over cyclic groups are simply connected
(extending a result of Graña for quandles of prime size) and that finite doubly transitive
quandles of order different from 4 are simply connected. We also consider constant
cohomology with coefficients in arbitrary groups.
Keywords: Quandle, connected quandle, latin quandle, quandle cohomology, non-abelian
cohomology, constant cocycle, simply connected quandle, covering.
Mathematics Subject Classification 2000: Primary 05C38, 15A15; Secondary 05A15,
15A18.
1. Introduction
Quandles were introduced by Joyce [13] in 1982 as algebraic objects whose elements
can be used to color oriented knots. In more detail, let K be an oriented knot and
X = (X, ⊲, ⊲) a set with two binary operations. Given a diagram of K, an assignment
of elements of X to arcs of K is consistent if the relations
y
y
x
x
x⊲y
x⊲y
1
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
2
FILE
Marco Bonatto and Petr Vojtěchovský
are satisfied at every crossing of the diagram. If, in addition, the assignment remains
consistent when Reidemeister moves are applied to the diagram of K in all possible
ways, we say that the assignment is a coloring of K. We denote by colX (K) the
number of colorings of K by elements of X in which more than one element is used.
Whether colX (K) 6= 0 is a delicate question that depends on both K and
X = (X, ⊲, ⊲). However, when the consistency conditions imposed by Reidemeister
moves are universally quantified, they force ⊲ to be the inverse operation to ⊲ (in
the sense that x ⊲ (x ⊲ y) = y and x ⊲ (x ⊲ y) = y) and they force precisely the
quandle axioms onto ⊲. In particular, if (X, ⊲) is a quandle and there is a consistent
assignment of elements of X to arcs of a diagram of K, it is also a coloring of K.
It is therefore customary to color arcs of oriented knots by elements of quandles.
For an oriented knot K, the knot quandle of K is the quandle freely generated
by the oriented arcs of K subject only to the above crossing relations. It was shown
early on by Joyce [13] and Matveev [17] that knot quandles are complete invariants
of oriented knots up to mirror image and reversal of orientation.
It has been conjectured that for any finite collection K of oriented knots there
exists a finite collection of finite quandles X1 , . . . , Xn such that the n-tuples
(colX1 (K), . . . , colXn (K)) distinguish the knots of K up to mirror image and reversal of orientation. (See [4] for a more precise formulation.) This conjecture has
been verified in [4] for all prime knots with at most 12 crossings, using a certain
collection of 26 quandles. Many oriented knots can be distinguished by a coarser
invariant, namely by merely checking whether colX (K) 6= 0 [8,14].
Whenever an oriented knot K is colored by a quandle X, the elements of X
actually used in the coloring form a connected subquandle of X. Consequently, for
the purposes of quandle colorings, it is sufficient to consider connected quandles.
Although far from settled, the theory of connected quandles is better understood
than the general case [11], and connected quandles have been classified up to size
47 [11,19].
Our work is primarily concerned with simply connected quandles which can be
defined in several equivalent ways.
In a long paper [6], Eisermann developed the theory of quandle coverings and
offered a complete categorical characterization of coverings for a given quandle.
In [1], Andruskiewitsch and Graña introduced an extension theory for quandles.
Their (dynamical) quandle cocycles can be used to color arcs of oriented knots,
giving rise to knot invariants [2]. (These were first defined in [3] as an analog of
the Dijkgraaf-Witten invariants for 3-manifolds [10].) The quandle cocycle condition (see Definition 2.6) is rather difficult to work with, involving a 3-parameter
mapping. Constant quandle cocycles are precisely those quandle cocycles in which
one parameter is superfluous (see Definition 2.14). Even constant quandle cocycles
yield powerful knot invariants [2, Section 5].
The following conditions are equivalent for a connected quandle X (see Proposition 2.16):
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
3
• the fundamental group of X is trivial,
• every covering of X is equivalent to the trivial covering of X over some set
S,
• for every set S, the second constant cohomology set of X with coefficients
in the symmetric group SS is trivial.
If a connected quandle X satisfies any of these equivalent conditions, we say that
X is simply connected.
In this paper we develop a combinatorial approach to constant cohomology with
emphasis on simply connected quandles. From an algebraic point of view, our main
result is as follows:
Theorem 1.1. Let X be a finite connected quandle that is affine over a cyclic
group, or a finite doubly transitive quandle of size different from 4. Then X is
simply connected.
We offer two proofs of Theorem 1.1. The first proof is combinatorial in nature
and mostly self-contained. The second proof (whose main idea was suggested to us
by the referee of an earlier version of this paper) is much shorter and relies on an
explicit description of the fundamental group of affine quandles from an unpublished
note of Clauwens [5].
We also investigate constant cohomology with coefficients in arbitrary groups
and we prove:
Theorem 1.2. Let X be a latin quandle. Then the following conditions are equivalent:
(i) X is simply connected,
(ii) Hc2 (X, G) = 1 for every group G.
We can then easily obtain the following knot-theoretical corollary:
Corollary 1.3. Let X be a simply connected latin quandle. Then every conjugacy
quandle cocycle invariant based on X is trivial.
The paper is organized as follows. Basic results about quandles and their extensions, coverings, cohomology and constant cohomology are recalled in Section
2. Let X be a latin quandle, u ∈ X, and let S be a nonempty set. Every constant
quandle cocycle β : X × X → SS with coefficients in the symmetric group SS is cohomologous to a normalized (constant) quandle cocycle βu satisfying βu (x, u) = 1
for every x ∈ X, as is recalled in Section 3. In Section 4 we introduce three bijections f , g, h of X × X under which every u-normalized cocycle βu is invariant,
that is, βu (k(x, y)) = βu (x, y) for k ∈ {f, g, h}. To prove that a given connected
quandle is simply connected, it then suffices to show that for every (x, y) ∈ X × X
there exists some (x0 , y0 ) ∈ X × X and k ∈ hf, g, hi such that βu (x0 , y0 ) = 1
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
4
FILE
Marco Bonatto and Petr Vojtěchovský
and k(x0 , y0 ) = (x, y). We therefore study the orbits of f , g, h in Section 5, and
again in Section 6 in the restricted case of connected affine quandles. Theorem 1.1
is proved in Section 7. Clauwens’ explicit description of the fundamental group of
affine quandles is recalled in Section 8 and then Theorem 1.1 is proved once more.
Finally, constant cohomology with coefficients in arbitrary groups is introduced in
Section 9, where we also prove Theorem 1.2 and Corollary 1.3.
2. Basic Results and Quandle Extensions
2.1. Quandles
For a groupoid (X, ·) and x ∈ X, let
Lx : X → X,
x 7→ x · y,
Rx : X → X,
x 7→ y · x
be the left translation by x and the right translation by x, respectively.
We will often suppress the binary operation while talking about groupoids and
denote them by X rather than by (X, ·). We denote by Aut(X) the automorphism
group of X. When X is merely a set, then Aut(X) = SX , the symmetric group on
X.
Definition 2.1. A groupoid (X, ⊲) is a quandle if it is a left quasigroup that is left
distributive and idempotent. That is, (X, ⊲) is a quandle if it satisfies the following
axioms:
Lx ∈ SX ,
(2.1)
x ⊲ (y ⊲ z) = (x ⊲ y) ⊲ (x ⊲ z),
(2.2)
x⊲x=x
(2.3)
for every x, y, z ∈ X.
Note that the identity (2.1) is equivalent to X having a left division operation
defined by
x\y = L−1
x (y),
and that the two identities (2.1) and (2.2) jointly state that Lx ∈ Aut(X) for every
x ∈ X.
Every quandle is flexible, indeed, x ⊲ (y ⊲ x) = (x ⊲ y) ⊲ (x ⊲ x) = (x ⊲ y) ⊲ x, and
it is therefore safe to write x ⊲ y ⊲ x.
For any groupoid X and ϕ ∈ Aut(X) we have ϕLx ϕ−1 = Lϕ(x) for every x ∈ X.
In particular, if X is a quandle then
Ly Lx L−1
y = Ly⊲x for every x, y ∈ X.
(2.4)
Example 2.2. Quandles appear naturally as the following examples illustrate.
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
5
(i) The one element groupoid is called the trivial quandle. More generally, any
projection groupoid on a set X (that is, a groupoid satisfying x ⊲ y = y for
every x, y ∈ X) is a quandle, the projection quandle PX over X.
(ii) Let G be a group and H a union of (some) conjugacy classes of G. For x,
y ∈ H, let x ⊲ y = xyx−1 . Then (H, ⊲) is a quandle, the conjugation quandle
on H.
(iii) Let G be a group, α ∈ Aut(G) and H ≤ Fix(α) = {x ∈ G | α(x) = x}. Let
G/H be the set of left cosets {xH | x ∈ G}. Then G/H with multiplication
xH ⊲ yH = xα(x−1 y)H
is a quandle Q(G, H, α) = (G/H, ⊲), called the coset quandle (also known as
homogeneous quandle or Galkin quandle).
(iv) A coset quandle Q(G, H, α) with H = 1 is called principal and will be denoted
by Q(G, α). If, in addition, G is an abelian group, then Q(G, α) is an affine
quandle.
Definition 2.3. For a quandle X, we call the set
LX = {Lx | x ∈ X}
the left section of X, and the group
LMlt(X) = hLX i ≤ Aut(X)
the left multiplication group of X. The group LMlt(X) is often denoted by Inn(X)
and called the inner automorphism group of X.
The left section LX is closed under conjugation by (2.4), and the corresponding
conjugation quandle on LX will be denoted by L(X). Note that the mapping X →
L(X), x 7→ Lx is a homomorphism of quandles.
Definition 2.4. A quandle X is latin if Rx ∈ SX for every x ∈ X.
In a latin quandle we can define right division by x/y = Ry−1 (x). A latin
quandle X is therefore a quasigroup (X, ⊲, \, /) in which the multiplication ⊲ is
left distributive and idempotent. As in any quasigroup, a homomorphism of a
latin quandle (X, ⊲) is automatically a homomorphism of (X, ⊲, \, /). For instance,
x ⊲ (y/z) = Lx (y/z) = Lx (y)/Lx (z) = (x/y) ⊲ (x/z) holds in any latin quandle.
Definition 2.5. A quandle X is connected (or, sometimes, transitive) if LMlt(X)
acts transitively on X, and doubly transitive if LMlt(X) acts doubly transitively on
X.
All latin quandles are connected. Indeed, if X is a latin quandle and x, y ∈ X,
then Lx/y (y) = x. All finite quandles can be built from connected quandles but the
extension theory is not well understood, cf. [1, Proposition 1.17] or [12].
In order to simplify notation, we will from now on denote the quandle multiplication by · or by juxtaposition rather than by ⊲.
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
6
FILE
Marco Bonatto and Petr Vojtěchovský
2.2. Quandle extensions
The notion of quandle extensions was introduced in [1]. Let X be a groupoid, S
a nonempty set, and suppose that (X × S, ·) is a groupoid. Then the canonical
projection π : X × S → X, (x, s) 7→ x is a homomorphism of groupoids if and only
if there is a mapping β : X × X × S × S → S such that
(x, s) · (y, t) = (xy, β(x, y, s, t)).
We will then denote (X × S, ·) by X ×β S.
Suppose now that X is a quandle. It is then easy to see that X ×β S is also a
quandle if and only if β(x, y, s) : S → S, t 7→ β(x, y, s, t) is a bijection for every x,
y ∈ X, s ∈ S, and the quandle cocycle conditions
β(xy, xz, β(x, y, s)(t))β(x, z, s) = β(x, yz, s)β(y, z, t),
(2.5)
β(x, x, s)(s) = s
(2.6)
hold for every x, y, z ∈ X and every s, t ∈ S. In the context of quandles, we will
therefore consider β as a mapping X × X × S → SS .
Definition 2.6 ([1, Definition 2.2]). Let X be a quandle and S a nonempty set.
A mapping β : X × X × S → SS is a quandle cocycle if (2.5) and (2.6) hold. The
set of all quandle cocycles X × X × S → SS will be denoted by Z 2 (X, SS ).
Proposition 2.7 ([1]). The following conditions are equivalent for quandles X,
Y:
(i) Y is a quandle defined on X × S for some set S and the canonical projection
Y → X is a quandle homomorphism.
(ii) Y ∼
= X ×β S for some set S and some quandle cocycle β ∈ Z 2 (X, SS ).
(iii) X ∼
= Y /α for a uniform congruence α of Y , that is, a congruence α of Y such
that all blocks of α have the same cardinality.
Proof. We have already shown above that (i) and (ii) are equivalent. Suppose that
(ii) holds and Y = X ×β S for some β ∈ Z 2 (X, SS ). Then X ∼
= Y /ker(π), where
π : X ×β S → X is the canonical projection. Clearly, ker(π) is a uniform congruence,
each block having cardinality |S|. Conversely, let α be a uniform congruence on Y ,
X = Y /α, and let S be a set of the same cardinality as any of the blocks of α. Let
{h[x] : [x] → S | [x] ∈ X} be a family of bijections indexed by the blocks of α.
Then the mapping β : X × X × S → SS defined by
β([x], [y], s) = h[xy] Lh−1 (s) h−1
[y]
[x]
is a quandle cocycle and the mapping
Y → X ×β S,
is an isomorphism of quandles.
x 7→ ([x], h[x] (x))
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
7
We therefore define:
Definition 2.8. Let X, Y be quandles. Then Y is an extension of X if X ∼
= Y /α
for some uniform congruence α of Y .
For a quandle X, let H(X) denote the class of all homomorphic images of X.
We have:
Proposition 2.9. Let Y be a connected quandle. Then the following conditions are
equivalent for a quandle X:
(i) X ∈ H(Y ),
(ii) Y is an extension of X.
Proof. By the Fundamental Homomorphism Theorem, X ∈ H(Y ) if and only if
X∼
= Y /α for some congruence α of Y . Since Y is connected, it is easy to show that
every congruence of Y is uniform.
The following equivalence relation partitions Z 2 (X, SS ) so that any two cocycles
in the same block give rise to isomorphic quandles. In a suitably defined category
of quandle extensions (see [1] or Proposition 2.11), two cocycles belong to the same
block of this partition if and only if the two quandle extensions are isomorphic.
Definition 2.10. Let X be a quandle and S a nonempty set. We say that β,
β ′ ∈ Z 2 (X, SS ) are cohomologous, and we write β ∼ β ′ , if there exists a mapping
γ : X → SS
such that
β ′ (x, y, s) = γ(xy)β(x, y, γ(x)−1 (s))γ(y)−1
holds for every x, y ∈ X and s ∈ S. The factor set
H 2 (X, SS ) = Z 2 (X, SS )/ ∼
is the second (non-abelian) cohomology set of X with coefficients in SS .
The following result makes clear the relationship between cohomologous cocycles
and isomorphisms of quandle extensions.
Proposition 2.11 ([1, pp. 194–195]). The following conditions are equivalent
for a quandle X, a set S and cocycles β, β ′ ∈ Z 2 (X, SS ):
(i) β ∼ β ′ ,
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
8
FILE
Marco Bonatto and Petr Vojtěchovský
(ii) there exists an isomorphism φ : X ×β S −→ X ×β ′ S such that the following
diagram is commutative
φ
X ×β ′
π
// X .
♦♦77
♦
♦
♦
π♦♦♦♦
♦
♦
♦♦
♦♦♦
S
X ×β S
2.3. Quandle coverings and constant cocycles
We are interested in a special class of quandle extensions, so-called quandle coverings.
Definition 2.12 ([6, Definition 1.4]). A connected quandle Y is a covering of
a quandle X if there is a surjective quandle homomorphism f : Y → X such that
the left translations Lx , Ly of Y coincide whenever f (x) = f (y).
For a quandle Y , let ker(LY ) denote the equivalence relation on Y induced by
equality in the left section LY , that is, (x, y) ∈ ker(LY ) if and only if Lx = Ly . Then
ker(LY ) is in fact a congruence on Y , thanks to (2.4). Moreover, if Y is a connected
quandle then ker(LY ) is a uniform congruence, and hence Y is an extension of
Y /ker(LY ). Therefore, a connected quandle Y is a covering of a quandle X if and
only if X ∼
= Y /α, where α is some uniform congruence of Y that refines ker(LY ).
Here, we say that a congruence α refines a congruence β if (x, y) ∈ β whenever
(x, y) ∈ α.
Here are some nontrivial examples of quandle coverings:
Proposition 2.13. Let X1 = Q(G, H1 , α) and X2 = Q(G, H2 , α) be two coset
quandles such that H1 ≤ H2 . Then X1 is a covering of X2 .
Proof. Define ψ : X1 → X2 by ψ(xH1 ) = xH2 . The mapping ψ is surjective and
every block of ker(ψ) has the same cardinality as H2 /H1 . For x, y ∈ G we have
ψ(xH1 · yH1 ) = ψ(xα(x−1 y)H1 ) = xα(x−1 y)H2 = xH2 · yH2 = ψ(xH1 ) · ψ(yH1 ),
so ψ is a homomorphism.
Suppose that ψ(xH1 ) = ψ(yH1 ), i.e., x = yh for some h ∈ H2 ≤ Fix(α). Then
xα(x−1 ) = yhα(h−1 y −1 ) = yα(y −1 ) and thus
xH1 · zH1 = xα(x−1 z)H1 = yα(y −1 z)H1 = yH1 · zH1
for every z ∈ G. This shows that LxH1 = LyH1 in X1 , so X1 is a covering of X2 .
We proceed to identify those quandle cocycles that correspond to quandle coverings.
Definition 2.14 ([1, Definition 2.2]). Let X be a quandle and S a nonempty
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
9
set. A quandle cocycle β ∈ Z 2 (X, SS ) is a constant quandle cocycle if
β(x, y, r) = β(x, y, s)
for every x, y ∈ X and r, s ∈ S. Since the value of β(x, y, s) is then independent of
s ∈ S, we will think of constant cocycles as mappings β : X × X → SS .
The set of all constant quandle cocycles X × X → SS will be denoted by
2
Zc (X, SS ). The equivalence relation ∼ on Z 2 (X, SS ) induces an equivalence relation
on Zc2 (X, SS ), and we define
Hc2 (X, SS ) = Zc2 (X, SS )/ ∼,
the second constant cohomology set of X with coefficients in SS .
We see immediately from (2.5) and (2.6) that a mapping β : X × X → SS is a
constant quandle cocycle if and only if it satisfies
β(xy, xz)β(x, z) = β(x, yz)β(y, z),
β(x, x) = 1
(CC)
(CQ)
for every x, y, z ∈ X.
Note that (CC) implies
β(xy, xz) = β(x, yz) ⇔ β(x, z) = β(y, z)
(WCC)
for every x, y, z ∈ X. We will call (WCC) the weaker cocycle condition.
Just as quandle cocycles parametrize quandle extensions, the constant cocycles
parametrize quandle coverings.
Proposition 2.15 ([2, Lemma 5.1]). Let X, Y be connected quandles. Then the
following conditions are equivalent:
(i) Y is a covering of X,
(ii) Y ∼
= X ×β S for some set S and β ∈ Zc2 (X, SS ).
Proof. Let Y be an extension of X, say Y = X ×β S for β ∈ Z 2 (X, SS ). Then
(x, r) · (y, t) = (x, s) · (y, t) for every x, y ∈ X, r, s, t ∈ S if and only if β(x, y, r) =
β(x, y, s) for every x, y ∈ X, r, s ∈ S, which is equivalent to β ∈ Zc2 (X, SS ).
Let X be a quandle and S a nonempty set. The mapping defined by
X × X −→ SS ,
(x, y) 7→ 1
is a constant cocycle, called the trivial cocycle and denoted by 1. It is easy to see
that X ×1 S is the direct product of X and the projection quandle over S. The
covering Y = X ×1 S is called a trivial covering of X.
Two coverings f : Y → X, f ′ : Y ′ → X of X are said to be equivalent if there
is a quandle isomorphism φ : Y → Y ′ such that f ′ ◦ φ = f .
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
10
FILE
Marco Bonatto and Petr Vojtěchovský
Let X = (X, ·) be a quandle. The adjoint group Adj(X) of X is the group with
generators {ex | x ∈ X} and presenting relations {ex·y = e−1
x ey ex | x, y ∈ X}. Following [6, Definitions 1.7, 1.10], let ǫ : Adj(X) → Z be the unique homomorphism
such that ǫ(ex ) = 1 for every x ∈ X. Let Adj(X)◦ be the kernel of ǫ. The fundamental group of X based at x ∈ X is defined as π1 (X, x) = {g ∈ Adj(X)◦ | xg = x}. By
[6, Proposition 5.8], π1 (X, x) is conjugate to π1 (X, y) whenever x, y are in the same
orbit of LMlt(X). In particular, if X is a connected quandle then the isomorphism
type of π1 (X, x) is independent of the base point x, and it is safe to write π1 (X)
instead of π1 (X, x).
Proposition 2.16. The following conditions are equivalent for a connected quandle
X:
(i) π1 (X) is trivial,
(ii) every covering Y → X is equivalent to the trivial covering of X over some set
S,
(iii) Hc2 (X, SS ) is trivial for every set S.
Proof. The equivalence of (i) and (ii) is established in [6, Proposition 5.15]. Let
us prove the equivalence of (ii) and (iii).
By Proposition 2.15, any covering of X is of the form π : X ×β S → S for some
nonempty set S. If X ×β S → S, X ×β ′ S ′ → S ′ are two equivalent coverings of
X, then S and S ′ have the same cardinality. It therefore suffices to investigate two
coverings X ×β S → S and X ×β ′ S → S with β, β ′ ∈ Zc2 (X, SS ). By Proposition
2.11, these two coverings are equivalent if and only if β ∼ β ′ .
3. Normalized Constant Cocycles
In this section we start computing the constant cohomology set Hc2 (X, SS ) of a
latin quandle X. The situation is greatly simplified in the latin case because every
cocycle of Hc2 (X, SS ) can be assumed to be normalized:
Definition 3.1 (compare [9, Lemma 5.1]). Let X be a latin quandle, S a
nonempty set and u ∈ X. Then β ∈ Zc2 (X, SS ) is said to be u-normalized if
β(x, u) = 1
for every x ∈ X.
For β ∈ Zc2 (X, SS ) and σ ∈ SS define β σ ∈ Zc2 (X, SS ) by β σ (x, y) =
σβ(x, y)σ −1 .
Proposition 3.2. Let X be a latin quandle, S a nonempty set and u ∈ X. For
β ∈ Zc2 (X, SS ) define βu ∈ Zc2 (X, SS ) by
βu (x, y) = β((xy)/u, u)−1 β(x, y)β(y/u, u).
Then {βuσ | σ ∈ SS } is the set of all u-normalized cocycles cohomologous to β.
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
11
Proof. We have δ ∼ β if and only if δ(x, y) = γ(xy)β(x, y)γ(y)−1 for some mapping
γ : X → SS . The following conditions are then equivalent for δ: δ is u-normalized,
1 = δ(x, u) = γ(xu)β(x, u)γ(u)−1 for every x ∈ X, γ(xu) = γ(u)β(x, u)−1 for every
x ∈ X,
γ(x) = γ(u)β(x/u, u)−1
(3.1)
for every x ∈ X, where we have used the latin property in the last step. Conversely,
given σ = γ(u) ∈ SS , the formula (3.1) defines a map γ : X → SS (it is well defined
since γ(u)β(u/u, u)−1 = γ(u)β(u, u)−1 = γ(u) by (CQ)). Then
δ(x, y)=γ(xy)β(x, y)γ(y)−1 =σβ((xy)/u, u)−1 β(x, y)β(y/u, u)σ −1 =σβu (x, y)σ −1 ,
so δ = βuσ .
Corollary 3.3. Let X be a latin quandle, S a nonempty set, and u ∈ X. Let β,
β ′ ∈ Zc2 (X, SS ), and let δ, δ ′ be u-normalized cocycles such that δ ∼ β and δ ′ ∼ β ′ .
Then β ∼ β ′ if and only if δ ′ = δ σ for some σ ∈ SS . Moreover, the following
conditions are equivalent:
(i) Hc2 (X, SS ) is trivial,
(ii) if β ∈ Zc2 (X, SS ), δ ∼ β and δ is u-normalized then δ = 1,
(iii) βu = 1 for every β ∈ Zc2 (X, SS ).
Proof. The first statement follows immediately from Proposition 3.2. Suppose that
(i) holds, let β ∈ Zc2 (X, SS ), and let δ ∼ β be u-normalized. Since 1 is also unormalized and 1 ∼ β by triviality of Hc2 (X, SS ), we have δ = 1 by the first
statement, establishing (ii). Clearly, (ii) implies (iii). Finally, if (iii) holds then
β ∼ βu = 1 for every β ∈ Zc2 (X, SS ), so Hc2 (X, SS ) is trivial.
Many identities for normalized cocycles can be derived from (CC) and (CQ).
We will later need:
Lemma 3.4. Let X be a latin quandle and let β be a u-normalized cocycle. Then
β (u/ (u/x) , x) = β (u/x, x)
for every x ∈ X. Moreover u/(u/x) · x = u if and only if x = u.
Proof. Setting x = u/y and y = u/z in (CC), we get β(u/(u/z), z) = β(u/z, z)
for every z ∈ X. Moreover,
u/(u/x) · x = u ⇔ u/(u/x) = u/x ⇔ u/x = u ⇔ x = u.
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
12
FILE
Marco Bonatto and Petr Vojtěchovský
4. Three Bijections on X × X that Preserve Normalized Cocycles
Given a mapping α : A → B and a bijection ℓ : A → A, we say that α is ℓ-invariant
if α(x) = α(ℓ(x)) for every x ∈ A.
In this section we introduce three bijections of X × X (where X is a latin quandle) under which all normalized cocycles are invariant. We will use these bijections
throughout the rest of the paper.
Let X be a latin quandle and u ∈ X. Define
f : X × X → X × X,
(x, y) 7→ (x · y/u, xu),
g : X × X → X × X,
(x, y) 7→ (ux, uy),
h : X × X → X × X,
(x, y) 7→ (y/(x\u) · x, y).
(4.1)
The element u on which f , g, h depend will always be understood from the context.
Proposition 4.1. Let X be a latin quandle and u ∈ X. Then f ∈ SX×X and every
u-normalized cocycle is f -invariant.
Proof. It is easy to see that f has an inverse, namely the mapping (x, y) 7→
(y/u, (y/u)\x·u). Let β be a u-normalized cocycle. Then (WCC) implies β(xy, xu) =
β(x, yu) for every x, y ∈ X. With z = yu, we obtain β(x, z) = β(x, yu) =
β(xy, xu) = β(x(z/u), xu) = β(f (x, z)).
Lemma 4.2. Let X be a latin quandle, u ∈ X and β a constant cocycle. Then the
following conditions are equivalent:
(i) β(u, x) = 1 for every x ∈ X,
(ii) β is g-invariant.
Proof. If (i) holds, then
(i)
(CC)
(i)
β(ux, uy) = β(ux, uy)β(u, y) = β(u, xy)β(x, y) = β(x, y)
for every x, y ∈ X. Conversely, if (ii) holds, let x = u/y and verify
(CC)
β(u, y) = β(u · u/y, uy)−1 β(u, u/y · y)β(u/y, y)
= β(u · u/y, uy)−1 β(u, u)β(u/y, y)
(CQ)
(ii)
= β(u · u/y, uy)−1β(u/y, y) = 1
for every y ∈ X.
We remark that (i) implies (ii) in Lemma 4.2 even if X is an arbitrary quandle,
not necessarily latin.
Proposition 4.3. Let X be a latin quandle and u ∈ X. Then g ∈ SX×X and every
u-normalized cocycle is g-invariant.
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
13
Proof. Since g = Lu × Lu , we obviously have g ∈ SX×X for any quandle X.
Suppose now that β is a u-normalized cocycle. In view of Lemma 4.2, it suffices to
prove that β(u, x) = 1 for every x ∈ X. Now,
(CC)
β(u, xu) = β(u, xu)β(x, u) = β(ux, u)β(u, u) = 1,
and we are done because Ru is a bijection.
Lemma 4.4. The following identities hold in a latin quandle:
(xy)/z = x(y/(x\z)),
(4.2)
(x/y)(zy) = ((x/y)z)x.
(4.3)
Proof. For (4.2), substitute x\z for z in x(y/z) = (xy)/(xz). The identity (4.3)
follows immediately from left distributivity and (x/y)y = x.
Proposition 4.5. Let X be a latin quandle and u ∈ X. Then h ∈ SX×X and every
u-normalized cocycle is h-invariant.
Proof. We will show that
k(x, y) = (u/((xy/u)\y), y)
is the inverse of h. (The mapping k was found by automated deduction [18].) It
suffices to show that h, k are inverse to each other in the first coordinate. We will
freely use the quasigroup identities x/(y\x) = y and (x/y)\x = y.
The first coordinate of h(k(x, y)) is equal to
y/[(u/((xy/u)\y))\u] · u/((xy/u)\y)
= y/[(xy/u)\y] · u/((xy/u)\y)
= (xy/u) · u/((xy/u)\y),
which is an expression of the form a · b/(a\c) and therefore, by (4.2), equal to
((xy/u)u)/y = xy/y = x.
The first coordinate of k(h(x, y)) is equal to
u/[(((y/(x\u) · x)y)/u)\y].
(4.4)
Since (y/(x\u)·x)y is of the form (a/b)c·a, it is by (4.3) equal to (y/(x\u))(x·x\u) =
y/(x\u) · u, and substituting this back into (4.4) yields
u/[((y/(x\u) · u)/u)\y] = u/[(y/(x\u))\y] = u/(x\u) = x.
We have proved that k is the inverse of h.
Let β be a u-normalized cocycle. Then we have
(CC)
β(u/y, y) = β(u/y·x, u)β(u/y, y) = β(u/y·x, u/y·y)β(u/y, y) = β(u/y, xy)β(x, y),
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
14
FILE
Marco Bonatto and Petr Vojtěchovský
and also
(CC)
β(u/y, y) = β(x, u)β(u/y, y) = β(x, u/y · y)β(u/y, y) = β(x · u/y, xy)β(x, y).
Therefore β(u/y, xy) = β(x · u/y, xy), and with u/y = z, xy = v we have x = v/y =
v/(z\u) and β(z, v) = β(v/(z\u) · z, v) = β(h(z, v)).
5. Orbits of the Three Bijections on X × X
For ℓ ∈ SA and a ∈ A let Oℓ (a) be the orbit of a under the natural action of hℓi on
A, and let Oℓ = {Oℓ (a) | a ∈ A} be the collection of all ℓ-orbits.
In this section we study the orbits of the bijections f , g, h on X × X defined in
(4.1), and also the orbits of the induced action of hf, hi on Og .
Let X be a quandle. Denote by p the product mapping
p : X × X → X,
(x, y) 7→ xy.
For z ∈ X, the fiber p−1 (z) is equal to
p−1 (z) = {(x, y) ∈ X × X | xy = z} = {(x, x\z) | x ∈ X}
and hence has cardinality |X|. Moreover, since
p(f (x, y)) = p(x · y/u, xu) = (x · y/u)(xu) = x(y/u · u) = xy = p(x, y),
every fiber is a union of f -orbits.
We have Og (u, u) = {(u, u)}, and we collect some additional orbits of g by
setting
Ogf = {Og (x, xu) | u 6= x ∈ X},
Ogu = {Og (x, x\u) | u 6= x ∈ X}.
S
The notation is explained as follows. By Lemma 5.7 below, Ogf = {(x, xu) |
S u
u 6= x ∈ X} and Og = {(x, x\u) | u 6= x ∈ X}. By Proposition 5.2 then,
S
S
{(u, u)} ∪ Ogf are precisely the fixed points of f , while {(u, u)} ∪ Ogu is the fiber
p−1 (u).
We will ultimately prove that certain quandles are simply connected by the following strategy, exploiting the invariance under f , g and h of u-normalized cocycles
(see Propositions 4.1, 4.3 and 4.5). For a u-normalized cocycle β, we first partition
its domain X × X into g-orbits Og on which hf, hi acts by Proposition 5.5. By
Corollary 5.8, f acts on both Ogu and Ogf , while h most definitely does not. The
bijection h is much easier to understand in affine connected quandles, cf. Lemma
6.6. The affine-over-cyclic case of Theorem 1.1 then easily follows. In the doubly
transitive case, we will show that there are at most five orbits of hf, g, hi, namely
{Og (u, u)}, Ogu , Ogf and certain sets Og1 , Og2 introduced later. (We note that Ogf ,
Ogu need not be disjoint, but their intersection is easy to understand, cf. Lemma
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
15
5.7.) A careful analysis of the orbit sizes of h then shows that the four sets Ogu , Ogf ,
Og1 and Og2 must be linked by h (as long as |X| 6= 4), establishing the main result.
Lemma 5.1. Let X be a latin quandle. Then for every x, y ∈ X and every k ∈ Z
we have
f k (x, y) = (fk (x, y), fk−1 (x, y)u),
where
fk (x, y) =
(
k
(Lx Ly/u ) 2 (x), if k is even,
k+1
(Lx Ly/u ) 2 (y/u), if k is odd.
(5.1)
Proof. Fix x, y ∈ X. Let ϕ = Lx Ly/u and define fk , fk′ by f k (x, y) = (fk , fk′ ). Then
′
′
(fk+1 , fk+1
) = f (fk , fk′ ) = (fk · fk /u, fk u), so fk+1
= fk u and fk+1 = fk · fk′ /u =
fk fk−1 for every k. Note that ϕk+1 (y/u) = ϕk (ϕ(y/u)) = ϕk (x(y/u · y/u)) =
ϕk (x · y/u).
For the base step, we will show that (5.1) holds for k ∈ {0, 1}. Indeed, f0 = x =
ϕ0 (x) and f1 = x · y/u = x(y/u · y/u) = ϕ(y/u).
For the ascending induction, suppose that (5.1) holds for k − 1 and k. If k is
even, we have
fk+1 = fk fk−1 = ϕk/2 (x)ϕk/2 (y/u) = ϕk/2 (x · y/u) = ϕ(k+2)/2 (y/u).
If k is odd, we have
fk+1 = fk fk−1 = ϕ(k+1)/2 (y/u)ϕ(k−1)/2 (x) = ϕ(k−1)/2 (x · y/u)ϕ(k−1)/2 (x)
= ϕ(k−1)/2 ((x · y/u)x) = ϕ(k−1)/2 (x(y/u · x)) = ϕ(k+1)/2 (x),
where we have used flexibility.
For the descending induction, suppose that (5.1) holds for k and k + 1. If k is
even then
fk−1 = fk \fk+1 = ϕk/2 (x)\ϕk/2 (x · y/u) = ϕk/2 (x\(x · y/u)) = ϕk/2 (y/u).
If k is odd then
fk−1 = fk \fk+1 = ϕ(k+1)/2 (y/u)\ϕ(k+1)/2 (x)
= ϕ(k+1)/2 ((y/u)\x) = ϕ(k−1)/2 ϕ((y/u)\x) = ϕ(k−1)/2 (x),
finishing the proof.
Proposition 5.2. Let X be a latin quandle and x, y ∈ X. Then, using the notation
of Lemma 5.1, the following conditions are equivalent: f k (x, y) = (x, y), fk (x, y) =
x, fk−1 (x, y) = y/u. In particular,
(i) |Of (x, y)| = 1 if and only if y = xu,
(ii) |Of (x, y)| 6= 2,
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
16
FILE
Marco Bonatto and Petr Vojtěchovský
(iii) |Of (x, y)| ≤ |X|.
Proof. Fix x, y ∈ X and let fk = fk (x, y). Clearly, f k (x, y) = (x, y) holds if and
only if both fk = x and fk−1 = y/u hold. Since xy = p(x, y) = p(f k (x, y)) =
p(fk , fk−1 u) = fk · fk−1 u, we have fk = x if and only if fk−1 u = y.
Part (i) now follows. Suppose that f 2 (x, y) = (x, y). The equality x = f2 says
x = x · (y/u · x), which implies x = x · y/u, x = y/u, y = xu. But then |Of (x, y)| = 1
by (i). Finally, (iii) follows from the above-mentioned fact that every fiber of p has
cardinality |X| and is a union of f -orbits.
Proposition 5.3. Let X be a finite latin quandle and x, y ∈ X. Then |Og (x, y)| =
lcm(|OLu (x)|, |OLu (y)|). Moreover, |Og (x, y)| = 1 if and only if (x, y) = (u, u).
Proof. This follows immediately from g k (x, y) = (Lku (x), Lku (y)).
Lemma 5.4. Let X be a latin quandle and x, y ∈ X. Then the following conditions
are equivalent:
(i) |Oh (x, y)| = 1,
(ii) p(h(x, y)) = p(x, y),
(iii) y = u.
Proof. Each of (i), (ii) is equivalent to y/(x\u) · x = x, which is equivalent to
y/(x\u) = x, that is, to y = u.
The action of hf, hi on X × X in fact induces an action on Og , the orbits of g:
Proposition 5.5. Let X be a latin quandle. Then
k(Og (x, y)) = Og (k(x, y))
for every k ∈ hf, hi.
Proof. It suffices to show that f and h commute with g. Let x, y ∈ X. Since
Lu ∈ Aut(X), we have
f (g(x, y)) = f (ux, uy) = (ux · (uy)/u, uxu) = (u · (x · (y/u)), u · xu) = g(f (x, y))
and
h(g(x, y))=h(ux, uy)=(uy/(ux\u) · ux, uy)=(u · (y/(x\u) · x), u · y)=g(h(x, y)).
Remark 5.6. The mappings f and h never commute on a nontrivial latin quandle.
More precisely, f h(x, y) = hf (x, y) if and only if x = y = u. Indeed, we certainly
have f (u, u) = (u, u) = h(u, u), so f h(u, u) = hf (u, u). Suppose now that
f h(x, y) = f (y/(x\u) · x, y) = ((y/(x\u) · x)(y/u), (y/(x\u) · x)u)
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
17
is equal to
hf (x, y) = h(x · y/u, xu) = (xu/((x · y\u)\u) · x(y\u), xu).
By comparing the second coordinates we see that y = u, and substituting this into
the first coordinates we arrive at xu = xu/(xu\u) · xu, that is, x = u.
Lemma 5.7. Let X be a latin quandle and x ∈ X. Then
(i) Og (x, xu) = {(y, yu) | y ∈ OLu (x)},
(ii) Og (x, x\u) = {(y, y\u) | y ∈ OLu (x)}.
In particular, Og (x, y) ∈ Ogu ∩ Ogf if and only if y = xu and x · xu = u.
Proof. We have g(x, xu) = (ux, uxu), g −1 (x, xu) = (u\x, u\(xu)) =
(u\x, (u\x)u), and (i) follows by induction. Similarly, g(x, x\u) = (ux, u · x\u) =
(ux, (ux)\u) and g −1 (x, x\u) = (u\x, u\(x\u) = (u\x, (u\x)\u) prove (ii). Hence
Og (x, y) ∈ Ogu ∩ Ogf if and only if (x, y) = (z, zu) = (w, w\u) for some z, w ∈ X,
which is equivalent to z = w = x, y = xu = x\u.
Corollary 5.8. Let X be a latin quandle. Then
(i)
(ii)
(iii)
(iv)
h(Ogu ) ∩ Ogu = ∅,
h(Ogf ) ∩ Ogf = ∅,
f (Og (x, xu)) = Og (x, xu) for every x ∈ X, in particular, f (Ogf ) = Ogf ,
f (Ogu ) = Ogu .
Proof. For (i), suppose that h(Og (x, x\u)) ∈ Ogu for some x 6= u. By Lemma 5.7,
h(x, x\u) = (y, y\u) for some y ∈ X. But then x = y, h(x, x\u) = (x, x\u), and
thus x = u by Lemma 5.4, a contradiction.
The proof of (ii) is similar: If h(Og (x, xu)) ∈ Ogf for some x 6= u, then h(x, xu) =
(y, yu) for some y by Lemma 5.7, hence x = y, h(x, xu) = (x, xu), xu = u, x = u,
a contradiction.
For (iii), recall that f (x, xu) = (x, xu) by Proposition 5.2.
Finally, for (iv), note that p(f (x, x\u)) = p(x, x\u) = u, hence f (x, x\u) must
be equal to some (y, y\u). Moreover, if x 6= u then y 6= u because f fixes (u, u).
A permutation is said to be semiregular if all its nontrivial cycles are of the
same finite length. A quandle X is semiregular if there is a positive integer s such
that every nontrivial cycle of any left translation Lx has length s.
Clearly, if X is a connected quandle and Lu is semiregular for some u ∈ X, then
X is semiregular. In particular, a latin quandle is semiregular if and only if Lu is
semiregular for some u ∈ X.
Let us denote a typical orbit of the induced action of f on Og by Of (Og (x, y)),
cf. Proposition 5.5. Then |Of (Og (x, y))| ≤ |Of (x, y)| and the strict inequality can
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
18
FILE
Marco Bonatto and Petr Vojtěchovský
occur in examples.
Lemma 5.9. Let X be a latin semiregular quandle. Then |Of (Og (x, y))| =
|Of (x, y)| and |Oh (Og (x, y))| = |Oh (x, y)| for every x, y ∈ X. Hence f (Og (x, y)) =
Og (x, y) if and only if Og (x, y) ∈ Ogf .
Proof. Suppose that f k (Og (x, y)) = Og (x, y) and k > 0 is smallest possible.
Then f k (x, y) = (Lru (x), Lru (y)) for some r ∈ Z. Therefore, xy = p(x, y) =
p(Lru (x), Lru (y)) = Lru (xy). But then semiregularity implies that |Lu | divides r,
f k (x, y) = (x, y), and |Of (Og (x, y))| ≥ |Of (x, y)|.
Similarly, suppose that hk (Og (x, y)) = Og (x, y) and k > 0 is smallest possible. Then hk (x, y) = (z, y) = (Lru (x), Lru (y)) for some r ∈ Z and some z ∈ X.
Then Lru (y) = y, hence Lru (x) = x by semiregularity, hk (x, y) = (x, y), and
|Oh (Og (x, y))| ≥ |Oh (x, y)|.
6. The Orbits on Connected Affine Quandles
In this section we take advantage of the affine representation to arrive at explicit
expressions for the mappings f and h in terms of the underlying group and the
automorphism α. Moreover, we compute the orbit lengths for f and h.
We will use additive notation for the underlying groups and set u = 0 in affine
quandles. We therefore also write Og0 for Ogu .
The results from previous sections apply to finite affine connected quandles
thanks to the following, well-known result.
Proposition 6.1 ([15, Proposition 1]). Let X = Q(A, α) be a finite affine
quandle. Then the following conditions are equivalent:
(i) X is latin,
(ii) X is connected,
(iii) 1 − α ∈ Aut(A).
Note that Proposition 6.1 implies that in any finite connected affine quandle
X = Q(A, α) we have L0 (y) = (1 − α)(0) + α(y) = α(y), that is, L0 = α.
Proposition 6.2. Let A be an abelian group and let α ∈ Aut(A) be such that
1 − α ∈ Aut(A). Then for every x ∈ A and every positive integer n we have
αn (x) = x ⇔
n−1
X
αk (x) = 0.
k=0
Pn−1 k
Proof. We have 1 − αn = (1 − α) k=0
α . Since 1 − α ∈ Aut(A), we deduce that
Pn−1 k
n
α (x) = x if and only if k=0 α (x) = 0.
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
19
Lemma 6.3. Let X = Q(A, α) be an affine quandle. Then for every x, y ∈ X and
k ≥ 0, the element fk (x, y) from (5.1) is equal to
fk (x, y) = x +
k
X
(−1)j αj (x − y/0).
(6.1)
j=1
Proof. Fix x, y ∈ X and let fk = fk (x, y). The formula (5.1) yields f0 = x and
f1 = x · (y/0) = x − α(x) + α(y/0), in agreement with (6.1). Suppose that (6.1)
holds for k − 1 and k, and recall that fk+1 = fk fk−1 . Then with z = x − y/0 we
have
fk+1 = (x +
k
X
(−1)j αj (z)) · (x +
k
X
j=1
=x+
k
X
(−1)j αj (z))
j=1
j=1
= x − α(x) +
k−1
X
(−1)j αj (z) −
k
X
(−1)j αj+1 (z) + α(x) +
k−1
X
(−1)j αj+1 (z)
j=1
j=1
(−1)j αj (z) − (−1)k αk+1 (z) = x +
k+1
X
(−1)j αj (z).
j=1
j=1
Proposition 6.4. Let X = Q(A, α) be a finite connected affine quandle. Then for
every x, y ∈ X we have
Pn
(i) |Of (x, y)| = min{n ∈ N | j=1 (−1)j αj (x − y/0) = 0},
(ii) (−1)|Of (x,y)| α|Of (x,y)| (x − y/0) = x − y/0,
(iii) if n = |Of (x, y)| then |OL0 (x − y/0)| divides (2n)/gcd(2, n),
(iv) if 2(x − y/0) = 0 then |Of (x, y)| = |OL0 (x − y/0)|.
Proof. (i) By Proposition 5.2, |Of (x, y)| is the smallest positive k such that
fk (x, y) = x, or, equivalently, fk−1 (x, y) = y/0. By Lemma 6.3, fk (x, y) = x if
Pk
and only if j=1 (−1)j αj (x − y/0) = 0.
(ii) Let z = y/0 and n = |Of (x, y)|. By Lemma 6.3 and the above remarks, we
have
n
X
(−1)j αj (x − z),
x = fn (x, y) = x +
j=1
z = fn−1 (x, y) = x +
n−1
X
(−1)j αj (x − z).
j=1
Taking the difference of these two equations yields (−1)n αn (x − z) = x − z.
(iii) Since L0 = α, part (ii) shows that |OL0 (x − z)| divides 2n. If n is even, (ii)
in fact shows that |OL0 (x − z)| divides n.
(iv) Suppose that 2(x − z) = 0. Then 2αj (x − z) = 0 for every j. Part (i)
Pn
Pn
Pn−1
then yields 0 = j=1 (−1)j αj (x − z) = j=1 αj (x − z) = j=0 αj (x − z), and
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
20
FILE
Marco Bonatto and Petr Vojtěchovský
Proposition 6.2 implies αn (x−z) = x−z. As n is minimal with 0 =
we deduce |OL0 (x − z)| = n.
Pn−1
j=0
αj (x−z),
We will now express |Of (x, y)| as a function of |OL0 (x − y/0)| and the order of
a certain element of A. We present only the case when |OL0 (x − y/0)| is even (since
that is all we need later), but the argument in the odd case is similar.
Lemma 6.5. Let X = Q(A, α) be a finite connected affine quandle and let x,
y ∈ X. Suppose that ℓ = |OL0 (x − y/0)| is even. Then
kℓ,
if |Of (x, y)| is even,
|Of (x, y)| =
k ′ ℓ/2, otherwise,
Pℓ−1
Pℓ/2−1
where k = | j=0 (−1)j αj (x − y/0)| and k ′ = | j=0 (−1)j αj (x − y/0)|.
Proof. Let z = x − y/0, ℓ = |OL0 (z)| and n = |Of (x, y)|. Suppose that n is even.
Then, by Proposition 6.4(iii), n = rℓ for some r. By Proposition 6.4(i), we have
0=
rℓ
X
(−1)j αj (z) =
rℓ−1
X
(−1)j αj (z) = r
(−1)j αj (z),
j=0
j=0
j=1
ℓ−1
X
where we have used rℓ even in the second equality and (−1)ℓ αℓ (z) = z (due to
Pℓ−1
ℓ even) in the third equality. Moreover, | j=0 (−1)j αj (z)| = r because rℓ is the
Prℓ
smallest positive integer for which j=1 (−1)j αj (z) = 0.
Suppose now that n is odd. Then, by Proposition 6.4(iii), n = (2s + 1)ℓ/2 for
some s. Since n is odd, we have ℓ/2 odd and Proposition 6.4(iii) yields −z = αn (z) =
α(2s+1)ℓ/2 (z) = αℓ/2 (z) and therefore (−1)ℓ/2 αℓ/2 (z) = z. Using these observations,
Proposition 6.4(i) implies
X
ℓ/2−1
(2s+1)ℓ/2−1
(2s+1)ℓ/2
0=
(−1)j αj (z) =
(−1)j αj (z) = (2s + 1)
j=0
j=1
We again have |
X
Pℓ/2−1
j=0
X
(−1)j αj (z).
j=0
(−1)j αj (z)| = 2s + 1.
We conclude this section by explicitly calculating h and |Oh (x, y)| in a connected
affine quandle.
Lemma 6.6. Let X = Q(A, α) be a connected affine quandle and let β be a 0normalized cocycle. Then
h(x, y) = (y + x, y),
β(ny + x, y) = β(x, y),
β(nx, x) = 1
for every integer n and every x, y ∈ X. In particular, |Oh (x, y)| = |y|.
(6.2)
(6.3)
(6.4)
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
21
Proof. Let x, y ∈ X, set z = x\0 and note that z = x − α−1 (x). Then
h(x, y) = (y/(x\0) · x, y) = (y/z · x, y)
= ((z + (1 − α)−1 (y − z)) · x, y) = ((1 − α)(z) + y − z + α(x), y)
= (−α(z) + y + α(x), y) = (−α(x) + x + y + α(x), y)
= (y + x, y).
The h-invariance of β (cf. Proposition 4.5) then yields (6.3). With y = x, (6.3)
yields β((n + 1)x, x) = β(x, x) = 1, which is (6.4). Finally, |Oh (x, y)| = |y| is an
immediate consequence of (6.2).
7. Two Classes of Simply Connected Quandles
In this section we show that every finite connected affine quandle over a cyclic group
is simply connected (extending a result of Graña for connected quandles of prime
order), and that every finite doubly transitive quandle of order different from 4 is
simply connected.
7.1. Connected affine quandles over cyclic groups
Graña showed:
Proposition 7.1 ([9, Lemma 5.1]). Let q be a prime and X = Q(Zq , α) an
affine quandle. Then X is simply connected.
Proof. Let x, y ∈ X be nonzero elements and let β be a 0-normalized cocycle.
Then y = nx for a suitable n and (6.4) yields β(y, x) = β(nx, x) = 1.
It is known that every connected quandle of prime order q is isomorphic to an
affine quandle of the form Q(Zq , α), cf. [7]. Proposition 7.1 therefore states that
every connected quandle of prime order is simply connected.
Every automorphism of Zm is of the form λn for some n with gcd(m, n) = 1,
where λn (x) = nx. Suppose that gcd(m, n) = 1. As an immediate consequence of
Proposition 6.1, we see that the affine quandle Q(Zm , λn ) is connected if and only
if gcd(m, 1 − n) = 1. Note that the conditions gcd(m, n) = 1 = gcd(m, n − 1) imply
that m is odd.
Let U(Zm ) denote the group of units in the ring of integers modulo m.
Proposition 7.2. Let X = Q(Zm , λn ) be a connected affine quandle and let β
be a 0-normalized cocycle. Then β is u-normalized for every u ∈ U(Zm ), that is,
β(x, u) = 1 for every x ∈ X, u ∈ U(Zm ). In addition, β(u, x) = 1 and β(u·x, u·y) =
β(x, y) for every x, y ∈ X, u ∈ U(Zm ).
Proof. Let u ∈ U(Zm ) and x ∈ X. Then x = nu for a suitable n and we have
β(x, u) = β(nu, u) = 1 by (6.4), showing that β is u-normalized. By Proposition
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
22
FILE
Marco Bonatto and Petr Vojtěchovský
4.3, β(u · x, u · y) = β(x, y) for every x, y ∈ X. Then by Lemma 4.2, β(u, x) = 1 for
every x ∈ X.
Lemma 7.3. Let X = Q(Zm , α) be a connected affine quandle with m odd. Then
for every x ∈ X there are u, v ∈ U(Zm ) such that x = u · v.
Proof. Since u + v = (1 − α)−1 (u) · α−1 (v) and α and 1 − α are automorphisms
of (Zm , +), it suffices to prove that for every x ∈ Zm there are u, v ∈ U(Zm ) such
that x = u + v. This is well-known and can be established as follows:
Let m = pn1 1 · · · pnr r , where p1 , . . . , pr are distinct primes. By the Chinese
remainder theorem, Zm ∼
= U(Zpn1 1 ) × · · · ×
= Zpn1 1 × · · · × Zpnr r and also U(Zm ) ∼
n
r
U(Zpr ). It therefore suffices to prove the claim when m = pn for a prime p.
Consider x = ap + b, where 0 ≤ b < p. If x is invertible (that is, b 6= 0), then so is
2x (since m is odd), and x = 2x + (−x) does the job. If x is not invertible, then
x = ap = (ap + 1) + (−1) finishes the proof.
Theorem 7.4. Let X = Q(Zm , λn ) be a connected affine quandle. Then X is
simply connected.
Proof. Let x, y ∈ X and let β be a 0-normalized cocycle. By Proposition 7.3,
we can write x = u · v for some invertible elements u, v. By Proposition 7.2,
β(x, y) = β(u · v, u · u\y) = β(v, u\y) = 1.
7.2. Doubly transitive quandles
Finite doubly transitive quandles can be characterized as follows:
Theorem 7.5 ([20, Corollary 4]). Let X be a finite quandle. Then LMlt(X) is
doubly transitive if and only if X ∼
= Q(Znq , α) for some prime q, some n > 0 and
n
α ∈ Aut(Zq ) with |α| = |X| − 1.
Lemma 7.6. Let X be a finite idempotent groupoid with a doubly transitive automorphism group. Then X is semiregular and the parameter s (length of any nontrivial orbit of any left translation) is a divisor of |X| − 1.
Proof. For any x ∈ X we have Lx (x) = x by idempotence. Given x 6= y with
Lkx (y) = y and some v 6= w, let ϕ ∈ Aut(X) be such that ϕ(x) = v, ϕ(y) = w. Then
w = ϕ(y) = ϕLkx (y) = Lkϕ(x) (ϕ(y)) = Lkv (w). Hence X is semiregular, and the rest
follows.
Combining Theorem 7.5 and Lemma 7.6, we see that if X ∼
= Q(Znq , α) is a finite
doubly transitive quandle, then α is an (|X| − 1)-cycle (since α cannot be trivial
by Proposition 6.1).
Proposition 7.7. Let X be a finite doubly transitive quandle. Then there is an
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
23
integer F > 1 such that
|Of (x, y)| =
1, if x = y/0,
F, otherwise.
Proof. By Proposition 5.2, |Of (y/0, y)| = 1. Suppose that x, y, v, w ∈ X are
such that x 6= y/0 and v 6= w/0. Let ϕ ∈ LMlt(X) ≤ Aut(X) be such that
ϕ(x) = v and ϕ(y/0) = w/0. Suppose that |Of (x, y)| = k is even, the odd case
being similar. By Lemma 5.1 and Proposition 5.2, we have (Lx Ly/u )k/2 (x) = x.
Then v = ϕ(x) = ϕ(Lx Ly/0 )k/2 (x) = (Lv Lw/0 )k/2 (v), and thus |Of (v, w)| ≤ k.
Corollary 7.8. Let X = Q(Znq , α) be a doubly transitive quandle. Then
(i)
(ii)
(iii)
(iv)
Og0 and Ogf are singletons,
S
S
( Og0 ) ∩ ( Ogf ) = ∅ if |X| > 3,
if q = 2 then F = |X| − 1,
if q > 2 then F = |X| − 1 if it is even and F = (|X| − 1)/2 if it is odd.
Proof. (i) Since α = L0 is an (|X|−1)-cycle, we have |Og (x, y)| = |X|−1 whenever
(x, y) 6= (0, 0) by Proposition 5.3. The claim follows by Lemma 5.7.
S
S
(ii) By Lemma 5.7, (x, y) ∈ ( Og0 ) ∩ ( Ogf ) if and only if x · (x · 0) = 0 and
y = x · 0. We have x · (x · 0) = (1 + α)(1 − α)(x). Since 1 − α ∈ Aut(A), we have
x · (x · 0) = 0 if and only if α(x) = −x. Then α2 (x) = x, so |X| − 1 = |α| ≤ 2.
(iii) According to Proposition 7.7, it suffices to compute the length of an f orbit for some (x, y) with y 6= x · 0. We have 2x = 0 for every x ∈ X, and hence
F = |X| − 1 by Proposition 6.4(iv).
(iv) Since q > 2, |OL0 (x)| = |X| − 1 is even for every x ∈ X. By Proposition 5.2,
F ≤ |X| − 1. If F is even then F = k(|X| − 1) by Lemma 6.5 and thus F = |X| − 1.
If F is odd then the same lemma yields F = k(|X| − 1)/2, the case F = |X| − 1
cannot occur since |X| − 1 is even, and thus F = (|X| − 1)/2.
Lemma 7.9. Let X = Q(Znq , α) be a doubly transitive quandle and β a 0normalized cocycle.
S
(i) If F = |X| − 1 then β(x, y) = 1 for every (x, y) 6∈ (Ogf ∪ Og0 ).
S f
(ii) If β(x, y) = 1 for every (x, y) 6∈ (Og ∪Og0 ) and, in addition, there is 0 6= z ∈ X
such that β(z, z\0) = 1, then β = 1.
Proof. (i) We have β(x, x) = 1 for any x ∈ X. Suppose that (0, 0) 6= (x, y) ∈
S
(X × X) \ (Ogf ∪ Og0 ). Then (x, y) is not a fixed point of f , and neither is (xy, xy),
since xy 6= 0. The fiber p−1 (xy) contains both (x, y) and (xy, xy), and it is a
union of f -orbits. By assumption, one of the orbits has size |F | − 1, forcing the
remaining element of p−1 (xy) to be a fixed point of f . Hence (x, y) ∈ Of (xy, xy).
Then β(xy, xy) = 1 implies β(x, y) = 1.
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
24
FILE
Marco Bonatto and Petr Vojtěchovský
S
(ii) Suppose that β(x, y) = 1 for every (x, y) 6∈ (Ogf ∪ Og0 ), and β(z, z\0) = 1
S 0
for some 0 6= z ∈ X. Then β = 1 on Og since Og0 is a singleton by Corollary
S
S
7.8. Finally, let (x, y) ∈ Ogf . By Corollary 5.8, h(x, y) 6∈ Ogf and thus β(x, y) =
β(h(x, y)) = 1.
Lemma 7.10. Let X = Q(A, α) be a latin affine quandle and 0 6= x ∈ X. Then
S
(i) h(x, x\0) ∈ Ogf if and only if 2x = 0,
(ii) Og (0/(0/x), x) ∈ Ogf if and only if (α2 + α − 1)(x) = 0.
S
Proof. (i) Recall that (x, y) ∈ Ogf if and only if y = x · 0. We have h(x, x\0) =
(x + x\0, x\0) and x\0 = (1 − α−1 )(x). Therefore h(x, x\0) ∈ Ogf if and only if
(x + x\0) · 0 = x\0, which is equivalent to (1 − α)((1 − α−1 )(x) + x) = (1 − α−1 )(x).
This is easily seen to be equivalent to (1 − α)(2x) = 0. Since 1 − α ∈ Aut(A), the
last condition is equivalent to 2x = 0.
(ii) Note that (0/(0/x)) · 0 = x is equivalent to x/0 · 0/x = 0. Also note that
x/0 = (1 − α)−1 (x) and 0/x = −α(1 − α)−1 (x). Then x/0 · 0/x = 0 holds if and
only if x − α2 (1 − α)−1 (x) = 0, which is equivalent to (α2 + α − 1)(x) = 0.
Theorem 7.11. Let X = Q(Znq , α) be a doubly transitive quandle with q ≥ 3. Then
X is simply connected.
Proof. If n = 1 then X is simply connected by Theorem 7.4. Suppose that n > 1
and let β be a 0-normalized cocycle. Recall that X is semiregular with α an (|X|−1)cycle. Hence |Og (x, y)| = |X| − 1 for every (0, 0) 6= (x, y) ∈ X × X. By Lemma
5.9 and Corollary 7.8, |Of (Og (x, y))| = |Of (x, y)| ∈ {1, F } and F ∈ {|X| − 1,
(|X| − 1)/2}. By Lemmas 5.9 and 6.6, |Oh (Og (x, y))| = |Oh (x, y)| = |y| ∈ {1, q} for
any x, y ∈ X. Moreover, by Corollary 5.8, Og0 and Ogf are disjoint singletons.
Suppose that F = |X| − 1. By Lemma 7.9(i), β = 1 on the complement of
S
O = (Og0 ∪ Ogf ). Let 0 6= z ∈ X. By Lemma 7.9(ii), we will be done if we show
S
that β(z, z\0) = 1. By Corollary 5.8, h(z, z\0) ∈
/ Og0 . Since q ≥ 3, Lemma 7.10(i)
S f
yields h(z, z\0) 6∈ Og . Hence h(z, z\0) ∈ O and β(z, z\0) = β(h(z, z\0)) = 1.
For the rest of the proof suppose that F = (|X| − 1)/2. The sets Og (u, u), Ogf
and Og0 account for 1 + 2(|X| − 1) = 2|X| − 1 elements of X × X and for all fixed
points of f , leaving |X|2 − (2|X| − 1) = (|X| − 1)2 points unaccounted for. The
unaccounted points thus form two hf, gi-orbits, each of size F (|X| − 1), say Og1 and
Og2 . We can certainly take Og1 = Of (Og (0, x)) for some 0 6= x. Since β(0, x) = 1 by
Lemma 4.2, we have β = 1 on Og1 . If we can show that β = 1 on Og2 , too, then we
can finish as in the case F = |X| − 1, completing the proof.
We will now show that if q 6= 3 then the induced action of h on Og has only two
orbits, namely {Og (0, 0)} and its complement. This will finish the proof for q 6= 3.
Since h has no fixed points in the set Og0 ∪ Ogf of size 2, it suffices to consider the
following situations:
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
25
(a) Suppose that h acts on Ogf ∪ Og1 and thus also on Og0 ∪ Og2 , both sets of size
F + 1. Let 0 6= x ∈ X. Since Og (x, 0) is fixed by h and belongs to one of these sets,
we see that h acts on a set of size F , hence q divides F = (|X| − 1)/2, hence q
(being odd) divides |X| − 1, a contradiction. We reach a similar contradiction if h
acts on Ogf ∪ Og2 and Og0 ∪ Og1 .
(b) Suppose that h acts on Og0 ∪ Ogf ∪ Og1 and thus also on Og2 , sets of size F + 2
and F , respectively. Since Og2 contains no fixed-points of h, we reach a contradiction
as in (a).
(c) Suppose that h acts on Og0 ∪ Ogf ∪ Og2 and thus also on Og1 , sets of size F + 2
and F , respectively. Once we account for Og (x, 0) ∈ Og1 , we see that h acts on a set
of size F + 2 = (|X| + 3)/2 and on a set of size F − 1 = (|X| − 3)/2, forcing q = 3.
For the rest of the proof we can therefore assume that q = 3. Let 0 6= x ∈ X.
Setting x = z and y = 0 in (CC) yields β(x · 0, x) = β(x, 0 · x). We also have
(x · 0, x), (x, 0 · x) 6∈ O. Suppose for a while that (x · 0, x) and (x, 0 · x) are in the
same f -orbit, that is,
(x · 0, x) = f k (x, 0 · x) = (fk (x, 0 · x), fk−1 (x, 0 · x) · 0)
for some k ≥ 1. Note that (0·x)/0 = (1−α)−1 α(x). Comparing coordinates, Lemma
6.3 yields
(1 − α)(x) = x · 0 = fk (x, 0 · x) = x +
k
X
(−1)j αj (x − (1 − α)−1 α(x)),
j=1
x = fk−1 (x, 0 · x) · 0 = (1 − α)(x +
k−1
X
(−1)j αj (x − (1 − α)−1 α(x))).
j=1
Applying 1 − α to the first identity and using q = 3 then yields
(1 − α)2 (x) = (1 − α)(x) +
k
X
(−1)j αj ((1 − α)(x) − α(x))
j=1
= (1 − α)(x) +
k
X
(−1)j αj (1 + α)(x),
j=1
while the second identity can be rewritten as
x=(1 − α)(x) +
k−1
k−1
X
X
(−1)j αj (1 + α)(x).
(−1)j αj ((1 − α)(x) − α(x))=(1 − α)(x) +
j=1
j=1
Subtracting the two last identities now gives (1 − α)2 (x) − x = (−1)k αk (1 + α)(x).
Since (1 − α)2 = 1 − 2α + α2 = 1 + α + α2 , we can rewrite this as α(1 + α)(x) =
(−1)k αk (1 + α)(x). Noting that 1 + α ∈ Aut(A) (if α(x) = −x then α2 = 1 and
|X| = 3) and, canceling, we finally get x = (−1)k αk−1 (x). If k is even, we deduce
k ≡ 1 (mod |X| − 1), thus also k ≡ 1 (mod F ), but then (x · 0, x) = f k (x, x · 0) =
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
26
FILE
Marco Bonatto and Petr Vojtěchovský
(x, x · 0) implies x · 0 = x, x = 0, a contradiction. If k is odd, we deduce 2(k − 1) ≡ 0
(mod |X| − 1), therefore k ≡ 1 (mod F ), and we reach the same contradiction.
Consequently, the elements (x, x · 0) and (x · 0, x) are not in the same f -orbit,
S
S
hence one of them lies in Og1 while the other in Og2 , and we see that β = 1 on
Og2 .
Theorem 7.12. Let X = Q(Zn2 , α) be a doubly transitive quandle with n 6= 2.
Then X is simply connected.
Proof. By Corollary 7.8(i), F = |X| − 1. Then by Lemma 7.9, β(x, y) = 1 for
every Og (x, y) 6∈ Og0 ∪ Ogf and it suffices to show that β(z, z\0) = 1 for some z 6= 0.
Note that any element (z, z\0) can be written as (0/y, y) by setting y = z\0. By
Proposition 3.4, β(0/y, y) = β(0/(0/y), y) and Og (0/(0/y), y) ∈
/ Og0 . If we show
that Og (0/(0/y), y) 6∈ Ogf for some y 6= 0, then β(0/y, y) = β(0/(0/y), y) = 1 and
we are through.
By Lemma 7.10(ii), it suffices to show that (α2 +α+1)(y) = (α2 +α−1)(y) 6= 0,
which is equivalent to α3 (y) 6= y, and this follows from the fact that α is an (|X|−1)cycle (here we use |X| 6= 4).
We have now proved Theorem 1.1. We will show in Section 8 that a doubly
transitive quandle of order 4 is not simply connected.
8. A Short Proof of Theorem 1.1
In this section we prove Theorem 1.1 once more, this time using results of Clauwens
[5] on the fundamental group of affine quandles.
Clauwens showed how to explicitly calculate the fundamental groups of affine
quandles. Following [5, Definition 1], let G = (G, +) be an abelian group and
Q(G, α) an affine quandle. Let
I(G, α) = hx ⊗ y − y ⊗ α(x) | x, y ∈ Gi,
S(G, α) = (G ⊗ G)/I(G, α),
F (G, α) = Z × G × S(G, α),
where the operation on F (G, α) is given by
(k, x, a)(m, y, b) = (k + m, αm (x) + y, a + b + (αm (x) ⊗ y + I(G, α))).
Then we have:
Theorem 8.1 ([5, Theorem 1 and page 4]). Let G = (G, +, 0) be an abelian
group and Q(G, α) an affine quandle. Then the groups Adj(Q(G, α)) and F (G, α)
are isomorphic, and the groups π1 (Q(G, α), 0) and S(G, α) are isomorphic.
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
27
We are ready to prove Theorem 1.1. Let X = Q(G, α) be a finite connected
affine quandle. By Proposition 2.16 and Theorem 8.1, X is simply connected if and
only if S(G, α) is trivial.
Suppose first that G = Zn is cyclic. Then G⊗G = Zn ⊗Zn ∼
= Zn is generated by
1⊗1. For y ∈ G we have 1⊗(1−α)(y) = 1⊗y −1⊗α(y) = y ⊗1−1⊗α(y) ∈ I(G, α).
Since X is connected, the homomorphism 1 − α is bijective by Proposition 6.1. In
particular, 1 ⊗ 1 ∈ I(G, α) and S(G, α) is trivial.
Now suppose that X = Q(G, α) is doubly transitive, G is not cyclic, and |X| 6= 4.
By the argument given in Subsection 7.2, G = Znp for some prime p and n > 1, and
α is a cycle of length |X| − 1.
Let us write u ≡ v if u, v ∈ G ⊗ G coincide modulo I(G, α). For every x ∈ G
we have
0 ≡ x ⊗ (x + α(x)) − (x + α(x)) ⊗ α(x)
= x ⊗ x + x ⊗ α(x) − x ⊗ α(x) − α(x) ⊗ α(x) = x ⊗ x − α(x) ⊗ α(x).
Therefore x ⊗ x ≡ α(x) ⊗ α(x) for every x. Since α is a cycle of length |X| − 1, we
conclude that there is e ∈ G such that
x⊗x≡e
for every 0 6= x ∈ G.
(8.1)
If x, y ∈ G are such that x 6= 0 6= y and x 6= y, equation (8.1) implies
e ≡ (x − y) ⊗ (x − y) = x ⊗ x − x ⊗ y − y ⊗ x + y ⊗ y ≡ 2e − x ⊗ y − y ⊗ x.
Therefore
x⊗y+y⊗x≡e
for every x 6= 0 6= y with x 6= y.
(8.2)
We proceed to show that e ≡ 0.
Suppose that p 6= 2. Since |X| > 4, there are x, y ∈ G such that x 6= 0 6= y and
x 6= ±y. Then (8.1) implies
e ≡ (x + y) ⊗ (x + y) = x ⊗ x + x ⊗ y + y ⊗ x + y ⊗ y ≡ 2e + x ⊗ y + y ⊗ x
and we deduce x ⊗ y + y ⊗ x ≡ −e. But (8.2) holds for our choice of x, y as well,
and thus e ≡ 0.
Suppose now that p = 2. Since |X| > 4, there are distinct and nonzero x, y,
z ∈ G such that x + y + z 6= 0. Then (8.1) and (8.2) imply
e ≡ (x + y + z) ⊗ (x + y + z) = x ⊗ x + y ⊗ y + z ⊗ z
+ (x ⊗ y + y ⊗ x) + (x ⊗ z + z ⊗ x) + (y ⊗ z + z ⊗ y) ≡ 6e ≡ 0,
and we again conclude that e ≡ 0.
Let us continue the proof with p arbitrary. We have shown that x ⊗ x ≡ 0 for
every x. The calculations leading to (8.2) can now be repeated for any x, y, and we
obtain x ⊗ y ≡ −y ⊗ x for every x, y. Hence
0 ≡ x ⊗ y − y ⊗ α(x) ≡ x ⊗ y + α(x) ⊗ y = (x + α(x)) ⊗ y
(8.3)
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
28
FILE
Marco Bonatto and Petr Vojtěchovský
for every x, y ∈ G. We claim that 1+α is bijective. Indeed, suppose that (1+α)(x) =
0 for some x 6= 0. Then α(x) = −x and α2 (x) = x, a contradiction with α being a
cycle of length |X| − 1 (using |X| > 3). Now (8.3) shows that x ⊗ y ≡ 0 for every
x, y ∈ G, and Theorem 1.1 is proved.
Example 8.2. This example shows that a doubly transitive quandle Q(G, α) of
order 4 is notsimply connected. We will calculate I(G, α). Let {e1 , e2 } with e1 = 10
and e2 = 01 be a basis of G = Z22 , and suppose without loss of generality that
α = 11 10 . Then α(e1 ) = e1 + e2 , α(e2 ) = e1 and α(e1 + e2 ) = e2 . Calculating in
G ⊗ G, we get
e1 ⊗ e1 + e1 ⊗ α(e1 ) = e1 ⊗ e2 ,
e2 ⊗ e1 + e1 ⊗ α(e2 ) = e1 ⊗ e1 + e2 ⊗ e1 ,
e1 ⊗ e2 + e2 ⊗ α(e1 ) = e1 ⊗ e2 + e2 ⊗ e1 + e2 ⊗ e2 ,
e2 ⊗ e2 + e2 ⊗ α(e2 ) = e2 ⊗ e1 + e2 ⊗ e2 .
Hence I(G, α) is the span of e1 ⊗ e1 + e2 ⊗ e1 , e2 ⊗ e1 + e2 ⊗ e2 and e1 ⊗ e2 . Since
G⊗G∼
= Z42 , the quandle Q(G, α) is not simply connected.
9. Constant Cocycles with Coefficients in Arbitrary Groups
Following [1], we have defined constant quandle cocycles as mappings β : X × X →
SS but the definition makes sense for any group G.
Definition 9.1. Let X be a quandle and G a group. Let
Zc2 (X, G) = {β : X × X → G | β satisfies (CC) and (CQ)}.
For β, β ′ ∈ Zc2 (X, G), we write β ∼ β ′ if there exists a mapping γ : X → G such
that
β ′ (x, y) = γ(xy)β(x, y)γ(y)−1
holds for every x, y ∈ X. Then Hc2 (X, G) = Zc2 (X, G)/ ∼ is the second constant
cohomology set of X with coefficients in G.
A careful look at all our results shows that all calculations can be carried out
over an arbitrary group G, not just over symmetric groups.
Proposition 9.2. Let X be a quandle and G a group. Then Zc2 (X, G) embeds into
Zc2 (X, SG ).
Proof. Let λG : G → SG , λG (g)(h) = gh be the left regular representation of G.
Define
j : Zc2 (X, G) → Zc2 (X, SG ),
Then it is easy to see that j is injective.
j(β)(x, y) = λG (β(x, y)).
(9.1)
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
29
Let us show that the embedding j in (9.1) induces a map j : Hc2 (X, G) →
Suppose that β, β ′ ∈ Zc2 (X, G) are cohomologous and let γ : X → G
be such that γ(xy)β(x, y)γ(y)−1 = β ′ (x, y) for every x, y ∈ X. We claim that
j(β), j(β ′ ) ∈ Zc2 (X, SG ) are cohomologous via j(γ), defined by j(γ)(x) = λG (γ(x)).
Indeed, for every a ∈ G and every x, y ∈ X we have
Hc2 (X, SG ).
j(γ)(xy)j(β)(x, y)(j(γ)(y))−1 (a)=γ(xy)β(x, y)γ(y)−1 a=β ′ (x, y)a=j(β ′ )(x, y)(a).
However, the induced map j : Hc2 (X, G) → Hc2 (X, SG ) is not necessarily an
embedding as we shall see in Example 9.5. We start by having another look at
4-element doubly transitive quandles.
Example 9.3. Let X = Q(Z22 , α) be a doubly transitive quandle and G a group.
Suppose that α, e1 , e2 are as in Example 8.2. We claim that
Hc2 (X, G) = {[βa ]∼ | a ∈ G, a2 = 1},
(9.2)
where βa is given by
0
e1
e1 + e2
e2
0
1
1
1
1
e1 e1 + e2
1
1
1
a
a
1
a
a
e2
1
a,
a
1
(9.3)
and, moreover, βa ∼ βb if and only if a and b are conjugate in G.
To see this, first recall that x·y = x+α(−x+y) = x+α(x+y) and check that βa
defined by (9.3) is a 0-normalized cocycle. Conversely, suppose that β ∈ Zc2 (X, G)
is 0-normalized. Then β(x, 0) = 1 for every x ∈ X. Since β is g-invariant by
Proposition 4.3, we also have β(0, x) = 1 for every x ∈ X by Lemma 4.2, and
β(x, y) = β(g(x, y)) = β(0 · x, 0 · y) = β(α(x), α(y)) for every x, y ∈ X. By
Proposition 4.5, β is h-invariant, and by Lemma 6.6 we have h(x, y) = (y + x, y),
so β(x, y) = (y + x, y) for every x, y ∈ X. Applying g, h as indicated, we get
g
g
h
g
g
β(e1 , e2 ) = β(e1 +e2 , e1 ) = β(e2 , e1 +e2 ) = β(e1 , e1 +e2 ) = β(e1 +e2 , e2 ) = β(e2 , e1 ),
so there is a ∈ G such that β = βa . Setting x = z in (CC) yields β(xy, x) =
β(x, yx)β(y, x) and thus
1 = β(0, e1 ) = β(e1 · (e1 + e2 ), e1 ) = β(e1 , (e1 + e2 ) · e1 )β(e1 + e2 , e1 )
= β(e1 , e2 )β(e1 + e2 , e1 ) = a2 .
Finally, by Proposition 3.2, βa ∼ βb if and only if a and b are conjugate in G.
Remark 9.4. In [16], the second cohomology of X = Q(Z22 , α) was calculated
when G is the additive group of a finite field or G ∈ {Z, Q}. Since G is abelian here,
Hc2 (X, G) is also an abelian group under the operation (β+δ)(a, b) = β(a, b)+δ(a, b).
Our calculations in Example 9.3 agree with those of [16, Example 2 and Corollary
1.1]. We have Hc2 (X, G) = 1 if G ∈ {Z, Q} or G = Zkp with p odd, and Hc2 (X, Zk2 ) ∼
=
k
Z2 .
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
30
FILE
Marco Bonatto and Petr Vojtěchovský
Example 9.5. Let X = Q(Z22 , α) be a doubly transitive quandle and let G = Z22 .
By Example 9.3, every a ∈ G yields βa ∈ Zc2 (X, G), and βa ∼ βb if and only if a = b
since G is abelian. Example 9.3 also shows that every σ ∈ SG with σ 2 = 1 yields
βσ in Zc2 (X, SG ), and βσ = βτ if and only if σ, τ have the same cycle structure.
Consider now the embedding j : Zc2 (X, G) → Zc2 (X, SG ) and note that for every
a ∈ G we have j(βa ) = βλG (a) . If a, b ∈ G are distinct nonzero elements of G then
βa 6∼ βb , but j(βa ) ∼ j(βb ) because λG (a), λG (b) have the same cycle structure.
We conclude that j does not induce an embedding of Hc2 (X, G) into Hc2 (X, SG ).
If X is latin, the embedding j commutes with the normalization procedure
described in Proposition 3.2.
Proposition 9.6. Let X be a latin quandle, G a group and β ∈ Hc2 (X, G). Then
j(βu ) = j(β)u for every u ∈ X.
Proof. We have
j(βu )(x, y) = λG (βu (x, y)) = λG (β((xy)/u, u)−1 β(x, y)β(y/u, u)) =
= λG (β((xy)/u, u))−1 λG (β(x, y))λG (β(y/u, u))) =
= j(β)u (x, y)
for every x, y ∈ X.
We can now prove Theorem 1.2. By Proposition 2.16, a connected quandle X
is simply connected if and only if Hc2 (X, SS ) = 1 for every set S. Let X be a latin
quandle. If Hc2 (X, G) = 1 for every group G, then certainly Hc2 (X, SS ) = 1 for every
set S. Conversely, suppose that Hc2 (X, SS ) = 1 for every set S, let G be a group
and let β ∈ Zc2 (X, G). Let u ∈ X. Since Hc2 (X, SG ) = 1 and j(β) ∈ Zc2 (X, SG ),
Corollary 3.3 implies j(β)u = 1. By Proposition 9.6, λG (βu (x, y)) = j(βu )(x, y) =
j(β)u (x, y) = 1 and therefore βu (x, y) = 1 for every x, y ∈ X. By Corollary 3.3,
Hc2 (X, G) = 1.
Problem 9.7. Let X be a connected quandle. Are the following conditions equivalent?
(i) X is simply connected,
(ii) Hc2 (X, G) = 1 for every group G.
9.1. Conjugacy quandle cocycle invariants
We conclude the paper by establishing Corollary 1.3.
In [2, Section 5], a new family of knot invariants was defined by using constant
quandle cocycles. Let X be a quandle, G a group and β ∈ Zc2 (X, G). Let (τ1 , . . . , τk )
be all the crossings of an oriented knot K encountered while traveling around K
starting from some base point and following the orientation of K. For a crossing τ
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
FILE
Simply connected latin quandles
31
and coloring C ∈ colX (K), let B(τ, C) = β(xτ , yτ )ǫτ , where xτ is the color on the
understrand, yτ is the color on the overstrand, and ǫτ is the sign of the crossing.
Qk
Let ϕ(K, C) = i=1 B(τi , C). For g ∈ G, let [g] be the conjugacy class of g in G.
Then
ϕX,G,β (K) = {[ϕ(K, C)] | C ∈ colX (K)}
is a conjugacy quandle cocycle invariant of K. According to [2, Theorem 5.5], this
is indeed an invariant of oriented knots.
Let us prove Corollary 1.3. Let X be a simply connected latin quandle, G a
group and β ∈ Zc2 (X, G). By [2, Proposition 5.6], if β is cohomologous to the
trivial constant cocycle, then ϕX,G,β (K) is trivial. It therefore suffices to show that
Hc2 (X, G) = 1 and this follows from Theorem 1.2.
Acknowledgment
We thank an anonymous referee for several important comments and in particular
for pointing out the highly relevant preprint of Clauwens [5]. The first author would
like to thank Prof. David Stanovský for fruitful discussions during his visit to Ferrara in January 2016 from which this work was started and Dr. Giuliano Bianco
who introduced him to the topic.
Marco Bonatto was partially supported by Simons Foundation Collaboration
Grant 210176 to Petr Vojtěchovský. Petr Vojtěchovský was partially supported by
University of Denver PROF grant.
References
[1] Andruskiewitsch N., Graña M., From racks to pointed Hopf algebras, Advances in
Mathematics 178 (2) (2003), 177–243 .
[2] Carter J. S., Elhamdadi M., Graña M., Saito M., Cocycle knot invariants from quandle modules and generalized quandle homology, Osaka J. Math. 42 (2005), no. 3,
499–541.
[3] Carter J.S., Jelsovsky D., Kamada S., Langford L., Saito M., Quandle cohomology
and state-sum invariants of knotted curves and surfaces, Trans. Amer. Math. Soc.
355 (2003), 3947–3989.
[4] Clark W.E., Elhamdadi M., Saito M., Yeatman T., Quandle colorings of knots and
applications, J. Knot Theory Ramifications 23/6 (2014), 1450035.
[5] Clauwens F.J.-B.J., The adjoing group of an Alexander quandle, preprint,
https://arxiv.org/abs/1011.1587
[6] Eisermann M., Quandle coverings and their Galois correspondence, Fundamenta
Mathematicae 225 (2014), no. 1, 103–168.
[7] Etingof P., Guralnick R., Soloviev A., Indecomposable set-theoretical solutions to the
Quantum Yang–Baxter Equation on a set with prime number of elements, Journal of
Algebra 242 (2001), 709–719.
[8] Fish A., Lisitsa A., Stanovský D., Combinatorial approach to knot recognition, Embracing Global Computing in Emerging Economies, CCIS 514, Springer, 2015, 64–78.
December
22,
2017 4:22 WSPC/INSTRUCTION
simply˙connected˙latin˙quandles˙v19˙˙final˙
32
FILE
Marco Bonatto and Petr Vojtěchovský
[9] Graña M., Indecomposable racks of order p2 , Beiträge zur Algebra und Geometrie
45 (2004), no. 2, 665–676.
[10] Dijkgraaf R., Witten E., Topological gauge theories and group cohomology, Comm.
Math. Phys. 129 (1990), 393–429.
[11] Hulpke A., Stanovský D., Vojtěchovský P., Connected quandles and transitive groups,
J. Pure Appl. Algebra 220/2 (2016), 735–758.
[12] Jedlička P., Pilitowska A., Stanovský D., Zamojska-Dzienio A., The structure of medial quandles, J. Algebra 443 (2015), 300–334.
[13] Joyce D., A Classifying invariant of knots, the knot quandle, Journal of Pure and
Applied Algebra 23 (1982), 37–65.
[14] Kuperberg G., Knottedness is in NP, modulo GRH., Adv. Math. 256 (2014), 493–
506.
[15] Litherland R.A., Nelson S, The Betti numbers of some finite racks, Journal of Pure
and Applied Algebra 178, Issue 2, (2003), 187-202.
[16] Mochizuki T., Some calculations of cohomology groups of finite Alexander quandles,
Journal of Pure and Applied Algebra 179-3, (2003), 287–330.
[17] Matveev S. V., Distributive groupoids in knot theory, Math. USSR - Sbornik 47/1
(1984), 73–83.
[18] W.
W.
McCune,
Prover9,
version
2009-11A.
http://www.cs.unm.edu/~ mccune/prover9/
[19] Vendramin L., On the classification of quandles of low order, J. Knot Theory Ramifications 21 (2012), no. 9, 1250088, 10 pp.
[20] Vendramin L., Doubly transitive groups and quandles, to appear in J. Math. Soc.
Japan.
| 4math.GR
|
arXiv:1706.01092v2 [math.GR] 4 Jul 2017
Low-complexity computations for nilpotent
subgroup problems
Jeremy Macdonald∗, Alexei Miasnikov†, Denis Ovchinnikov‡
July 5, 2017
Abstract
We solve the following algorithmic problems using T C 0 circuits, or in
logspace and quasilinear time, uniformly in the class of nilpotent groups
with bounded nilpotency class and rank: subgroup conjugacy, computing
the normalizer and isolator of a subgroup, coset intersection, and computing the torsion subgroup. Additionally, if any input words are provided in
compressed form as straight-line programs or in Mal’cev coordinates the
algorithms run in quartic time.
Contents
1 Introduction
1
2 Background
2.1 Nilpotent presentations . . . . . . . . . . . .
2.2 Subgroups . . . . . . . . . . . . . . . . . . . .
2.3 Logspace, TC0 , and compressed words . . . .
2.4 Fundamental algorithms for nilpotent groups
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
3
4
4
7
3 Algorithmic problems
3.1 Subgroup conjugacy and
3.2 Coset intersection . . . .
3.3 Torsion subgroup . . . .
3.4 Isolator . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
10
13
19
21
1
normalizers
. . . . . . .
. . . . . . .
. . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Introduction
This is the third paper in a series on complexity of algorithmic problems in
finitely generated nilpotent groups. In the first paper [MMNV15], we showed
∗ Concordia
University, [email protected]
Institute of Technology, [email protected]
‡ Stevens Institute of Technology, [email protected]
† Stevens
1
that the basic algorithmic problems (normal forms, conjugacy of elements, subgroup membership, centralizers, presentation of subgroups, etc.) can be solved
by algorithms running in logarithmic space and quasilinear time. Further, if the
problems are considered in ‘compressed’ form with each input word provided as
a straight-line program, we showed that the problems are solvable in polynomial time. The second paper [MW17] pushed the complexity of these problems
lower, showing that they may be solved by TC0 circuits. Here we expand the
list of algorithmic problems for nilpotent groups which may be solved in these
low complexity conditions to include several fundamental problems concerning
subgroups.
Note that in group theory algorithmic problems for subgroups of groups
are usually much harder then the basic algorithmic problems mentioned above.
Nevertheless, we present here algorithms for deciding the conjugacy of two subgroups of a finitely generated nilpotent group G, finding the normalizer and
the isolator of a given subgroup of G, finding the torsion subgroup T (G) of G,
and finding the intersection of two cosets of subgroups of G, all of which may
be implemented by TC0 circuits, or run in logarithmic space and quasilinear
time on a (multi-tape) Turing machine. Furthermore, the compressed versions
of these problems are solvable in polynomial (specifically, quartic) time. All
of the algorithms work uniformly over finitely generated nilpotent groups (i.e.
the group may be included in the algorithm’s input), however the complexity
bounds depend on the nilpotency class and the rank (number of generators) of
the presentation. When both are bounded, we solve all the problems uniformly
in TC0 or logspace and quasilinear time.
Algorithmic problems in nilpotent groups have been studied for a long time.
On the one hand, it was shown that many of them are decidable and many
sophisticated decision algorithms were designed (see, for example, the pioneering paper [KRR+ 69] by Kargapolov et al. published in 1969 and the books
[Sim94] and [HEO05] for more recent techniques); on the other hand, there are
some which have been known to be undecidable for some time (for instance,
decidability of equations [Rom77]). Recent work by a variety of authors has
introduced a host of decidable/undecidable problems. New undecidable problems, including the knapsack problem, commutator and rectract problems are
described in [Loh15], [KLZ15], [MT16], and [Rom16], while positive decidability results for direct product decompositions and equations in the Heisenberg
group are described in [BMO16] and [DLS15]. Decidability and undecidability
results for equations over random nilpotent groups are also given in [GMO16a]
and [GMO16b].
However, it seems that this paper together with [MMNV15] and [MW17]
present the first thorough attempt to study the complexity of the problems,
beyond the decidable/undecidable dichotomy. In fact, it seems this is currently
the only known large class of non-abelian groups where the major algorithmic
problems are shown to have low space and time complexity. Another large class
of such groups is, perhaps, the class of finitely generated free groups given by
the standard presentations. Even there, if the free groups are given by arbitrary
finite presentations the complexity of the algorithmic problems is still mostly
2
unknown.
We have not yet mentioned one of the fundamental algorithmic problems in
nilpotent groups: the isomorphism problem. It is decidable due to the famous
result of Grunewald and Segal [GS80]. Nevertheless, not much is known about
its complexity.
Problem 1. Is the isomorphism problem in finitely generated nilpotent groups
decidable in polynomial time? Exponential time?
2
Background
This section describes, summarizing from [MMNV15] and [MW17], how we will
represent finitely generated nilpotent groups (§2.1) and their subgroups (§2.2),
and gives black-box descriptions of several algorithms that we will be using as
subroutines (§2.4). We also give a brief introduction to the TC0 circuit model
of computation, logspace computations, and the use of compressed words in
algorithmic problems over groups (§2.3).
2.1
Nilpotent presentations
Let G be a finitely generated nilpotent group of nilpotency class c. Then G has
lower central series
G = Γ1 ✄ Γ2 ✄ . . . ✄ Γc ✄ Γc+1 = 1
with Γi+1 = [G, Γi ] for i > 1. From this series we derive a presentation for G,
as follows.
Each Γi /Γi+1 is a finitely generated abelian group. We select and fix a finite
generating set asi−1 +1 Γi+1 , . . . , asi Γi+1 for Γi /Γi+1 and put
A = {a1 , a2 , . . . , am }.
For each j = 1, . . . , m, if si−1 + 1 ≤ j ≤ si , we denote by ej the order of aj Γi+1
in Γi /Γi+1 , using ej = ∞ when the order is infinite. Denote
T = {i | ei < ∞}.
Provided that each generating set above is chosen to correspond to a primary or
invariant factor decomposition of Γi /Γi+1 , every element g ∈ G may be written
uniquely in Mal’cev normal form as
αm
1 α2
g = aα
1 a2 · · · am
(1)
where αi ∈ Z and if i ∈ T then 0 ≤ αi < ei . The set A is called a Mal’cev basis
of G and the integers (α1 , . . . , αm ) are the Mal’cev coordinates of g.
For each i = 1, . . . , m, denote Gi = hai , . . . , am i. An essential fact, which
follows from the definition of the lower central series, is that for any i < j,
[ai , aj ] ∈ Gℓ
3
for some ℓ > j. From this it follows that relations of the form
α
α
ij(ℓ+1)
ijm
aj ai = ai aj aℓ ijℓ aℓ+1
· · · aα
m
(2)
βijℓ βij(ℓ+1)
ai a−1
aℓ+1
j aℓ
(3)
a−1
j ai
=
βijm
,
· · · am
with ℓ > j, hold in G. In addition, for each i ∈ T there is a relation of the form
µ
i(ℓ+1)
· · · aµmim
aei i = aµℓ iℓ aℓ+1
(4)
where ℓ > j. The set {a1 , . . . , am }, viewed as an abstract set of symbols, together with relators (2)–(4) then form a presentation for G called a nilpotent presentation. In fact, any presentation of this form defines a nilpotent group. Such
a presentation is called consistent if the order of each ai modulo hai+1 , . . . , am i
is precisely ei . Note that ei = 1 is permitted in a nilpotent presentation.
For low-complexity algorithms, an essential property of nilpotent presentations is the following (see [MMNV15] Thm. 2.3 and Lem. 2.5): if w is any
word over A± , then the length of the Mal’cev normal form (1) of the element g
corresponding to w in G is bounded by a polynomial function of the length of
w, with the degree of the polynomial depending on the nilpotency class c and
number of generators r of G. This fact plays a crucial role in solving efficiently
the fundamental algorithmic problems in finitely generated nilpotent groups.
2.2
Subgroups
All of our results concern subgroups of finitely generated nilpotent groups. For
every subgroup H ≤ G (all of which are, necessarily, finitely generated), one
may define a unique generating set (h1 , . . . , hs ) called the full-form sequence for
H. The precise definition was given in [Sim94] (and is reviewed in [MMNV15]),
but we mention here only three facts about (h1 , . . . , hs ) that we will need.
First, let B be the matrix in which row i is the row vector consisting of the
Mal’cev coordinates of hi . Then B is in row echelon form and does not contain
zero rows. We denote by πi the pivot column of row i of B. Since this column
corresponds to generator aπi , the Mal’cev normal form of hi begins with aπi , so
hi ∈ Gπi = haπi , . . . , am i.
Second, the number of generators s is bounded by the length m of the
Mal’cev basis. Third, every element h ∈ H can be uniquely presented in the
form
h = hβ1 1 · · · hβs s ,
where βi ∈ Z and 0 ≤ βi < eπi if πi ∈ T . Hence
H = {hβ1 1 · · · hβs s | βi ∈ Z and 0 ≤ βi < eπi if πi ∈ T }.
2.3
Logspace, TC0 , and compressed words
Let A be a finite language. We are interested in both decision and search
problems, and we may regard each such problem as a function f : A∗ → A∗ .
4
The set A consists of a set of symbols, say x1 , . . . , xn , which denote group
generators, and a few extra symbols used to separate different parts of the
input (commas to separate relators etc.). We will be computing f (x) using
logarithmic space or using TC0 circuits. We recall both of these notions below.
Logspace. A c-logspace transducer, where c > 0 is a constant, is a multi-tape
Turing machine consisting of the following tapes: an ‘input’ tape which is readonly, a constant number of read-write ‘work’ tapes, and a write-only ‘output‘
tape. For any input of length L, which is provided on the input tape, the
amount of space the transducer is allowed to use on each work tape is c log(L).
The output of the machine is the content of the output tape. A function f is said
to be logspace computable, or more casually the associated problem is solvable
in logarithmic space, if there exists a constant c and a c-logspace transducer
that produces f (x) on the output tape for any input x appearing on the input
tape.
Though computation on a c-logspace transducer puts a bound only on space
resources, a polynomial time bound of O(Lc ) is forced by the fact that the
machine may not enter the same configuration twice (otherwise it will loop
infinitely) and the number of configurations is bounded by a polynomial function
of the input length. The degree c may be very high, and for this reason it is
also desirable to show directly that our algorithms run in low-degree polynomial
time, in particular quasilinear time (i.e. O(L logk (L)) for some constant k).
Most of our algorithms invoke other logspace algorithms as subroutines,
and as such we need to compute compositions f ◦ g of logspace computable
functions. A standard argument shows that f ◦ g is again logspace computable,
but in computing (f ◦ g)(x) in this way, each symbol of g(x) is recomputed each
time it is needed in computation of f , which may give a significant increase in
time complexity. However, if the output g(x) is always of size O(log(L)), one
may simply compute g(x) first, store the output on the work tape, and then
proceed to compute f (g(x)). This is the case in all of our algorithms, so in this
case the time complexity of g is simply added to the overall time complexity.
TC0 circuits. A TC0 circuit with n inputs is a boolean circuit of constant
depth using NOT gates and unbounded fan-in AND, OR, and MAJORITY
gates, such that the total number of gates is bounded by a polynomial function
of n. A MAJORITY gate outputs 1 when more than half of its inputs are 1. A
function f (x) is TC0 -computable (more casually, ‘an algorithm is in TC0 ’) if for
each n there is a TC0 circuit Fn with n inputs which produces f (x) on every
input x of length n. Essential for our purposes is the fact that the composition
of two TC0 -computable functions is again TC0 -computable.
Since this definition of being computable only asserts that such a family
{Fn }∞
n=1 of circuits exists, one normally imposes in addition a uniformity condition stating that each Fn is constructible in some sense. We will only be
concerned here with standard notion of DLOGTIME uniformity, which asserts
that there is a random-access Turing machine which decides in logarithmic time
5
whether in circuit Fn the output of gate number i is connected to the input of
gate j, and determines the types of gates i and j. We refer the reader to [Vol99]
for further details on TC0 .
To put our results in context, we remind the reader of the following inclusions
of complexity classes:
TC0 ⊆ LOGSPACE ⊆ P ⊆ NP.
It is not known whether any of these inclusions is strict. Though every TC0 computable function is also logspace-computable and polynomial-time computable, our algorithm descriptions also give direct proofs of membership in
these classes.
Compressed words. We are also interested in algorithms that run efficiently
when the input is given in compressed format. The use of Mal’cev coordinates
provides a natural compression scheme for elements of G: each g ∈ G may be
encoded by a tuple of integers (its Mal’cev coordinates) written in binary. Notice
that if the size m of the Mal’cev basis is bounded, a normal form of length n
may be encoded by O(log n) bits. Since every finitely generated nilpotent group
has a Mal’cev basis, it is natural to consider algorithmic problems in which
input words represented in this compact way. Of course, such ‘compressed
problems’ are, in terms of computational complexity, more difficult than their
uncompressed siblings.
Since we will consider uniform algorithms, in which a finitely generated
nilpotent group G is given by an arbitrary presentation as part of the input,
we also consider two other compression schemes which do not depend on a the
specification of a Mal’cev basis. First, we may simply allow exponents to be
encoded in binary. In this scheme, a word is encoded as a product of tuples
(g, m), representing g m , where g is a group generator or, recursively, a word of
this form, and m is a binary integer. For example, (x4 y 2 )8 x−6 is encoded as
(((x, 0100)(y, 0010)), 1000)(x−1, 0110).
Second, we consider straight-line programs, that is, context-free grammars
that generate exactly one string. Formally, a straight-line program or compressed
word over an alphabet A consists of a set {A1 , A2 , . . . , An } called the nonterminal symbols and for each non-terminal symbol Ai a production rule either of
the form Ai → Aj Ak with j, k < i, or of the form Ai → a where a ∈ A∪{ǫ} with
ǫ denoting the empty word. The non-terminal An is termed the root, and one
‘expands’ the compressed word by starting with the one-character word An and
successively replacing any non-terminal with the right side of its production rule
until only symbols from A remain. The number n of non-terminal symbols is the
size of the program. Compression arises since a program of size n may expand
to a word of length 2n−1 . We refer the reader to the survey article [Loh12] and
the monograph [Loh14] for further information on compressed words, or to the
introduction of [MMNV15] for some brief remarks.
6
2.4
Fundamental algorithms for nilpotent groups
Throughout this paper, we make extensive use of algorithms described in [MMNV15]
and [MW17]. We give below a summary of some of the most heavily-used ones,
and we will use the names listed here, in bold text, to refer to their use.
• Full-form Sequence: Given H ≤ G, compute the full-form generating
sequence for H.
• Membership: Given g ∈ G and H ≤ G, determine if g ∈ H and if so,
αs
1
compute the unique expression g = hα
1 · · · hs where (h1 , . . . , hs ) is the
full-form sequence for H.
• Subgroup Presentation: Given H ≤ G, compute a consistent nilpotent
presentation for H.
• Conjugacy: Given g, h ∈ G, produce x ∈ G such that g x = h or determine that no such x exists.
• Centralizer: Given g ∈ G, compute a generating set for the centralizer
of g in G.
• Kernel: Given K ≤ G and φ : K → G1 , produce a generating set for the
kernel of φ.
• Preimage: Given K ≤ G, φ : K → G1 , and h ∈ G1 guaranteed to be in
φ(K), produce k ∈ K such that φ(k) = h.
We will need some further details regarding the input/output of these algorithms as well as their complexity.
Input. In each algorithm, we fix in advance two integers c and r. The
ambient nilpotent groups G and G1 are part of the input (thus the algorithms
are ‘uniform’) but must be of nilpotency class at most c and be presented using
at most r generators for the complexity bounds given below to be valid. Group
elements are given as words over the generating set(s), subgroups are specified
by finite generating sets, and φ is given by listing the elements φ(k) for each
given generator k of K. The length L of the input is the sum of the lengths of
all relators in G and G1 plus the lengths of all input words.
Output. Each output word is given as a word over the original generating
set except possibly in Full-form sequence and Membership. In these cases,
the algorithm converts to a nilpotent presentation of G, if one is not already
provided, and provides the output words in the new generators (the isomorphism
may also be provided, see Lemma 3 below). In Centralizer and Kernel, if the
original presentation of G is already a nilpotent presentation, one may assume
that the subgroup generating set in the output is the full-form sequence.
In every case, the total length of each output word is bounded by a polynomial function of L and the number of output words is bounded by a constant.
Optionally, the output words may be given by their Mal’cev coordinates.
7
Complexity. Each algorithm may be implemented on a logspace transducer,
and if so runs in time quasilinear in L. The proofs are given in [MMNV15].
Alternatively, each problem may be solved using TC0 circuits, as proved in
[MW17].
Compressed inputs. Each algorithm may also be run ‘with compressed inputs’. In this case, any input word (including group relators) may be provided
by (binary) Mal’cev coordinates, words with binary exponents, or straight-line
programs, as described in §2.3. We will measure the size of the input in terms
of the number n of input words and the maximum size M of any single input
word (in number of bits or number of non-terminal symbols). The space complexity of each algorithm is then O(M ) (it does not depend on n) and the time
complexity is O(nM 3 ). All output is provided in the corresponding compressed
format. Although each input word, in its expanded form, may have length
O(2M ), the polynomial bound for the length of output words implies that each
output word, in expanded form, has length O(2dM ) where d is the degree of the
aformentioned polynomial bound. Since d is constant, the compressed size of
each output remains O(log(2dM )) = O(M ).
Remark 2. We place no restriction on the number n of input words. In all
of the algorithms, any variable-sized set of input words (e.g. list of subgroup
generators, group relators) will be fed as input to the matrix reduction algorithm
described in Thm. 3.4 of [MMNV15] and processed in the ‘piecewise’ manner
described there, one word at a time. After this, sets of words usually only appear
as full-form sequences for subgroups, the number of which is always bounded
by a constant. The value n contributes a linear factor to the time complexity
of this algorithm (in both uncompressed and compressed cases), but does not
contribute to the space complexity.
While neither these algorithms nor the ones we describe in this paper require
that the input groups G and G1 be given by a nilpotent presentation, this form
is used internally by all of the algorithms. Converting to such a presentation is
accomplished as follows.
Lemma 3. Let c and r be fixed integers. There is an algorithm that, given a
finitely presented nilpotent group G = hX | Ri of nilpotency class at most c and
with |X| ≤ r, a finite set Y ⊂ G, and a word h over X ± guaranteed to be in the
subgroup H = hY i, produces
• a consistent nilpotent presentation hY ′ | Si for H, in which binary numbers
are used to encode exponents in the relators S,
• a map φ : Y ′ → (Y ±1 )∗ which extends to an isomorphism hY ′ | Si → H,
and
• a binary integer tuple h′ giving the Mal’cev coordinates of h relative to Y ′ .
The algorithm runs in space logarithmic in the input length L and time quasilinear in L, or in TC0 , and the (expanded) length of each output word is bounded
8
by a polynomial function of L. If compressed inputs are used (in R, Y , or h),
the space requirement is O(M ) and the time is O(nM 3 ), where n is the total
number of input words and M bounds the size of any single input word.
Proof. Algorithm. Begin by applying Prop. 5.1 of [MMNV15] (or Lem. 5
of [MW17] in the TC0 case) to compute a consistent nilpotent presentation
G = hX ′ | R′ i. Here X ⊂ X ′ , the inclusion X ֒→ X ′ induces an isomorphism
hX | Ri ≃ hX ′ | R′ i, and each element of X ′ \ X is a commutator in elements of
X. Use Subgroup Presentation to compute a nilpotent presentation hY ′ | Si
for H. The generating set Y ′ = {h1 , . . . , hs } is precisely the full-form sequence
for H. The relators have the form (2)-(4), and we encode the exponents appearing on the right sides in binary. To obtain φ, note that each element of Y ′
αm
′
1
has the form xα
1 · · · xm , where X = {x1 , . . . , xm }. We replace each xi with
its definition as a commutator of elements of X and encode the exponents αi
using binary numbers. Finally, use Subgroup Membership with input h and
{h1 , . . . , hs }, which returns an expression h = hγ11 · · · hγs s , giving the Mal’cev
coordinates (γ1 , . . . , γs ).
Complexity. Follows immediately from [MMNV15] and [MW17]. Note that
s is a constant depending on c and r.
We will often use this lemma in the case Y = X to convert from an arbitrary
presentation of G to a nilpotent presentation. In this case, we may assume the
algorithm uses Y ′ = X ′ ⊃ X. We convert all input words into their Mal’cev
coordinates (relative to X ′ ) at the same time, and perform further computations
directly on the Mal’cev coordinates.
Using binary numbers in the output is necessary in order to obtain quasilinear time, since writing down a word in its expanded form takes as many steps as
the length of the word itself, which in this case is only bounded by a polynomial
function of L.
3
Algorithmic problems
Before presenting the algorithms, let us make a few remarks regarding their
complexity analysis. The analysis of most of the algorthims is similar, so we
present here a general argument and fill in any additional details in the proof
of each algorithm.
First, note that the nilpotency class c and maximum number of generators r
of the input group(s) are constant. All other constants are expressible in terms
of c and r.
At the beginning of each algorithm, we convert to a nilpotent presentation,
if necessary, using Lemma 3. We denote the resulting Mal’cev basis by
{a1 , . . . , am }.
Note that m is constant. Word lengths are unchanged during this conversion
(see Lemma 3). We are guaranteed by [MMNV15] Thm. 2.3 that a word of
9
length L has a Mal’cev form of length polynomial in L, hence its coordinates
require O(log L) bits to record.
Our algorithms generally consist of a sequence of subroutine calls, using the
algorithms described in §2.4 as well as those described in this section, with some
minor additional processing. The complexity bounds described in §2.4 also apply
to the algorithms we describe in this section, as we will see. In all cases, we
prove that the total number of subroutine calls and the total number of words
that must be stored in memory at any given time is constant. Consequently, the
entire algorithm can, in principle, be expressed as a composition of a constant
number of functions. Each such function is TC0 -computable, hence so is the
composition. Note that to ‘store x in memory’ in TC0 terms means to add a
parallel computation branch computing x.
Though it follows immediately that we have logspace solutions to these problems, we wish to prove that one may in fact run the algorithms on a logspace
transducer in quasilinear time. To do so, we must show that each subroutine
may be run directly ‘in memory’ on the logspace transducer.
This is achieved by invoking each subroutine in its ‘compressed’ form. Initially, all input words are converted into O(log L)-bit Mal’cev coordinate form.
In this process, any variable-sized set of words (subgroup generators or group
relators) is reduced to a constant-sized set (the full-form sequence). This size is
bounded by m, and we often assume it is precisely m for notational convenience.
Each subroutine is then called with a constant number of O(log L)-bit words. It
will therefore run in space O(log L) and time O(log3 L), and produce a constant
number of O(log n)-bit output words.
For compressed inputs, the argument is similar. As we observed earlier,
the polynomial length bound implies that the compressed size of words remains O(M ) throughout the algorithm. Each subroutine therefore has space
complexity O(M ) and time complexity O(M 3 ), so the overall space and time
complexities are O(M ) and O(nM 3 ).
Finally, let us note that if we have a constant number elements g1 , . . . , gt in
Mal’cev form we can, by [MMNV15] Lem. 2.10, compute the Mal’cev form of
the product g1 · · · gt within the space and time bounds specified above, in both
compressed and uncompressed cases. We use this without mention to maintain
elements in coordinate form.
3.1
Subgroup conjugacy and normalizers
In this section we give an algorithm to determine whether or not two subgroups
of a nilpotent group are conjugate and if so to compute a conjugating element.
A natural by-product of this algorithm is the computation of subgroup normalizers.
We begin with a preliminary lemma solving the simultaneous conjugacy
problem for tuples of commuting elements. In fact, commutation is not required,
but we will obtain this stronger result (Theorem 7) as a corollary of the more
complicated coset intersection algorithm.
10
Lemma 4. Fix positive integers c, r, and l. There is an algorithm that, given
a finitely generated nilpotent group G = hX | Ri of nilpotency class at most c
with |X| ≤ r and two tuples of elements (a1 , . . . , al ) and (b1 , . . . , bl ) such that
[ai , aj ] = [bi , bj ] = 1 for all 1 ≤ i, j ≤ l, decides if there exists g ∈ G such that
agi = bi
for all 1 ≤ i ≤ l. The algorithm produces g if one exists, returns a generating
set for the centralizer of {b1 , . . . , bl }, and may be run in space logarithmic in the
length L of the input and time quasilinear in L, or in TC0 . The length of each
output word is bounded by a polynomial function of L. If compressed inputs are
used, the algorithm uses space O(M ) and time O(nM 3 ), where n = |R| and M
bounds the encoded size of each input word.
Proof. Algorithm. If necessary, use Lemma 3 to convert to a nilpotent presentation. Next, we check conjugacy of a1 with b1 using the Conjugacy Algorithm. If they are not conjugate, we may return ‘No’. Otherwise, we obtain
h such that ah1 = b1 and we compute a generating set for CG (b1 ) using the
Centralizer Algorithm.
If l > 1, we proceed recursively. Notice that g exists if and only if there exists
x ∈ G such that (ahi )x = bi for i = 1, . . . , l, since we may put x = h−1 g. Further,
such x must lie in CG (b1 ) since b1 = (ah1 )x = bx1 . Therefore it suffices to call
Lemma 4 recursively with the (commuting) tuples (ah2 , . . . , ahl ) and (b2 , . . . , bl )
and the subgroup CG (b1 ) in place of G. Before making the recursive call, we
use Lemma 3 to convert to a nilpotent presentation for CG (b1 ) and write each
of ahi , bi relative to this presentation.
If we obtain a conjugator x, we may return g = hx, using the map φ provided
by Lemma 3 to write x in the original generators X. In addition, we obtain
a generating set for the centralizer of {b2 , . . . , bl } in CG (b1 ), which is precisely
a generating set for the centralizer of the complete set {b1 , . . . , bl } in G. As
above, we must use φ to write these words in generators X. If the recursive call
returns ‘No’, then the tuples are not conjugate.
Complexity. The depth l of the recursion is constant and we need only store
h and the (constant-sized) generating set for the centralizer at each step of the
recursion, hence the general argument given at the beginning of the section
applies.
We now give the algorithm for determining conjugacy of two subgroups.
Theorem 5. Fix integers c and r. There is an algorithm that, given a finitely
presented nilpotent group G = hX | Ri of nilpotency class at most c with |X| ≤ r
and two subgroups H and K, determines if there exists g ∈ G such that
Hg = K
and if so finds such an element g as well as
a generating set for the normalizer NG (K)
11
of K. The algorithm runs in space logarithmic in the total length L of the
input and time quasilinear in L, or in TC0 , and the length of every output word
is bounded by a polynomial function of L. If compressed inputs are used, the
space complexity is O(M ) and the time complexity O(nM 3 ) where n is the total
number of input words and M bounds the encoded size of each input word.
Proof. Algorithm. Begin by converting, if necessary, to a nilpotent presentation
of G using Lemma 3.
The algorithm recurses on the maximum j such that H ∩ Γj 6= 1 and H ∩
Γj+1 = 1. To find j, simply compute the Full-form Sequence for H and
observe that if the last element of the sequence begin with the letter ak then j
is the unique index such that ak Γj+1 belongs to the generating set of Γj /Γj+1
(see §2.1). Compute similarly the maximum j ′ such that K ∩ Γj ′ 6= 1 and
K ∩ Γj ′ +1 = 1. If j 6= j ′ , then H and K are not conjugate since their conjugacy
would imply conjugacy of H ∩ Γi with K ∩ Γi for all i (since the Γi are normal
subgroups), hence equality of j and j ′ .
Denote Hj = H ∩ Γj and produce the full-form sequence for this group by
taking the elements of the full-form sequence for H that are in Γj . Proceed
similarly for Kj = K ∩ Γj . Next, we check conjugacy of Hj with Kj .
Conjugacy of Hj with Kj . Let : G → G/Γj+1 be the natural homomorphism. By the definition of central series, G acts trivially by conjugation on Γj .
Hence if Hj and Kj are conjugate then Hj = Kj . We first check if Hj = Kj ,
returning ‘No’ if not. To do so, it suffices to compute the Full-form Sequences
for Hj and Kj , and check them for equality.
Let (h1 , . . . , hl ) be the full-form sequence for Hj , computed above. We now
produce a generating set (k1 , . . . , kl ) for Kj such that hi = ki for all i, as follows.
Use the Preimage algorithm, with the subgroup Kj , the homomorphism :
Kj → G/Γj+1 , and the element hi , to produce each ki . Since Kj ∩ Γj+1 = 1,
(k1 , . . . , kl ) generates K.
We claim for any x ∈ G, Hjx = Kj if and only if hxi = ki for i = 1, . . . , l.
Indeed, since the tuples (h1 , . . . , hl ) and (k1 , . . . , kl ) are generating sets their
conjugacy implies Hj and Kj are conjugate. Conversely, if Hjx = Kj then
hxi ∈ Kj for all i. But hxi = hi = ki , and since ¯ is injective on Kj , we have
hxi = ki for all i. Also observe that Hj is abelian, since
[Hj , Hj ] ≤ Hj ∩ Γ2j ≤ Hj ∩ Γj+1 = 1,
and similary for Kj . Hence (h1 , . . . , hl ) and (k1 , . . . , kl ) are both tuples of
commuting elements. So to determine conjugacy of Hj with Kj it suffices to
use the algorithm of Lemma 4 to determine conjugacy of (h1 , . . . , hl ) and
(k1 , . . . , kl ) and if so find a conjugator x and a generating set Y for CG (Kj ).
In fact, CG (Kj ) = NG (Kj ) since if any element y ∈ G normalizes Kj , then for
each i = 1, . . . , l we have kiy ∈ Kj and hence kiy = ki , arguing as above.
Recursion. If j = 0, then H = Hj and K = Kj and we have already solved
the problem. Otherwise, letting
b : NG (Kj ) → NG (Kj )/Kj
12
cx
be the canonical homomorphism, we reduce the problem to conjugation of H
b in NG (Kj )/Kj , as follows.
and K
An element g such that H g = K exists if and only if there exists y ∈ G such
that (H x )y = K. Such an element y must lie in NG (Kj ), since
Kjy = (Hjx )y = (H ∩ Γj )xy = H xy ∩ Γxy
j = K ∩ Γj = K j .
Now K ≤ NG (Kj ), and H x ≤ NG (Kj ) since
Kj = Hjx = (H ∩ Γj )x = H x ∩ Γj ✂ H x .
cx )yb = K
b for some yb ∈ NG (Kj )/Kj , we claim that (H x )y = K.
Finally, if (H
Indeed, if k ∈ K then for some h ∈ H and k ′ ∈ Kj we have k = y −1 hx yk ′ =
−1
−1
y −1 (hx (k ′ )y )y. But y ∈ NG (Kj ) and Kj ≤ H x , so hx (k ′ )y ∈ H x and the
x
inclusion K ⊆ H follows. The reverse inclusion is proved similarly.
cx and K
b in NG (Kj )/Kj , we
In order to solve the conjugation problem of H
first use Lemma 3, with the generating set Y , to find a nilpotent presentation for NG (Kj ) and to convert the generating sets for H x , K, and Kj into
coordinate form in this presentation. Add the generators of Kj to this presentation to obtain a presentation for NG (Kj )/Kj , and call Theorem 5 with this
cx and K.
b
presentation and the subgroups H
It is essential to prove that the value of j decreases in the recursive call.
Letting Nj denote term j of the lower central series of NG (Kj ), we have that
Nj ≤ Γj , hence K ∩ Nj ≤ K ∩ Γj = Kj , and the intersection is trivial modulo
Kj , hence j must decrease.
cx and K
b are not conjugate, in which
The recursive call either proves that H
case H and K are not conjugate, or returns a conjugator yKj and a generating
b in NG (Kj )/Kj . Note that y (and each
set Z · Kj for the normalizer of K
element of Z) is given as a word over the generating set of NG (Kj ) with binary
exponents. We convert back to the generating set X of G using the map φ
provided by Lemma 3. For the conjugator, we return the word g = xy.
For the normalizer, we append to Z a generating set of Kj to obtain a
generating set Z ′ for the normalizer of K in NG (Kj ). But this is precisely the
normalizer of K in G: if K z = K for some z ∈ G then Kjz = K z ∩Γzj = K ∩Γj =
Kj and so NG (K) ≤ NG (Kj ).
Complexity. The depth of the recursion is bounded by the constant c, and
the number of words to store in memory is constant.
It should be noted that while the algorithm does not compute the normalizer
of K in the event that H and K are not conjugate, one may of course obtain it
by running the algorithm with H = K.
3.2
Coset intersection
We describe an algorithm to compute the intersection of cosets in finitely generated nilpotent groups, and apply it to solving the simultaneous conjugacy
13
problem. Recall that in any group, the intersection g1 H ∩ g2 K of two cosets is,
if non-empty, a coset of the intersection H ∩ K.
Theorem 6. Fix integers c and r. There is an algorithm that, given a finitely
presented nilpotent group G = hX | Ri of nilpotency class at most c with |X| ≤ r,
two subgroups H and K of G, and two elements g1 and g2 of G, determines if
the intersection g1 H ∩ g2 K is non-empty and if so, produces a generating set
for H ∩ K and an element g ′ ∈ g1 H ∩ g2 K, hence
g1 H ∩ g2 K = g ′ (H ∩ K).
The algorithm runs in space logarithmic in the length L of the input and time
quasilinear in L, or in TC0 . If compressed inputs are used, the space complexity
is O(M ) and the time complexity O(nM 3 ) where n is the total number of input
words and M bounds the encoded size of each input word.
Proof. Begin by using Lemma 3 to convert to a nilpotent presentation for G,
if necessary. We proceed by induction on the nilpotency class c.
Base case c=1. In this case, G is abelian. First, we will determine if the
intersection is non-empty and if so find g. Writing
g1 H ∩ g2 K = g2 (g2−1 g1 H ∩ K),
it suffices to determine if there exists h ∈ H such that g2−1 g1 h ∈ K. Since G is
abelian, this occurs if and only if g2−1 g1 ∈ hH ∪ Ki. We use the Membership
algorithm, with the union of the Full-form sequences of H and K as a
generating set for hH ∪ Ki, to determine if this is the case, returning ‘No’ if it is
not. Otherwise, we obtain an expression of g2−1 g1 as a linear combination of the
elements of the full-form sequence for hH ∪ Ki. We can convert to an expression
in terms of the full-form sequences for H and K, thus obtaining an expression
g2−1 g1 = hk for some elements h ∈ H and k ∈ K, by following the procedure
described in Cor. 3.9 of [MMNV15] (essentially, recording an expression of
each matrix row in terms of the given generators during the matrix reduction
process). This corollary gives only polynomial time, but Thm. 14 of [MW17]
gives the corresponding result for TC0 (hence logspace), though we need the
fact that g2−1 g1 and the full-form sequences of H and K are stored using only
O(log L) bits. We now set g = g1 h−1 and obtain g1 H ∩ g2 K = g(H ∩ K).
We will now find a generating set for H ∩ K. Let {u1 , . . . , un } be the generating set for H and consider the homomorphism φ : Zn → G defined by
αn
1
φ(α1 , . . . , αn ) = uα
1 · · · un
αn
1
and the composition φ′ : Zn → G → G/K. An element uα
1 · · · un of H is
also an element of K if and only if (α1 , . . . , αn ) is in the kernel of φ′ , hence
H ∩ K = φ(ker φ′ ). To compute the kernel, add the generators of K to the
relators of G to obtain a presentation of G/K, and pass this group together with
φ′ and the standard presentation of Zn to the Kernel algorithm. Applying
14
φ to each resulting subgroup generator, we obtain a generating set (in fact, the
full-form sequence) for H ∩ K.
Inductive case. Denote by
: G → G/Γc the canonical homomorphism.
Invoke Theorem 6 recursively in G/Γc with inputs H, K, g1 , and g2 . Note
that it suffices to erase all generators of Γc to compute (more formally, one
may use Lemma 3).
If the recursive call determines that g1 H ∩g2 K is empty, then so is g1 H ∩g2 K.
Otherwise, we obtain an element γ ∈ g1 H ∩ g2 K and a generating set w1 , . . . , wl
of H ∩ K, hence
g1 H ∩ g2 K = γhw1 , . . . , wl i.
Denote by Λ (but do not compute) the preimage of H ∩ K under . We will
rewrite the intersection g1 H ∩ g2 K in the form
g1 H ∩ g2 K = (g ′ (Λ ∩ H)) ∩ (g ′ c0 (Λ ∩ K))
(5)
for certain g ′ ∈ G, c0 ∈ Γc defined below. Compute a Preimage x1 of g1−1 γ in
H and a Preimage x2 of g2−1 γ in K. Let
g ′ = g1 x1 .
Since g2 x2 = γ = g1 x1 , it follows that (g1 x1 )−1 g2 x2 = c0 ∈ Γc .
To see that (5) holds, let u be an element of the left side. Then u = g1 h
and for some h ∈ H and u = γλ for some λ ∈ Λ. Then for some c′ ∈ Γc ,
u = g1 x1 λc′ hence λc′ ∈ H since x1 ∈ H. Clearly λc′ ∈ Λ, hence u ∈ g ′ (Λ ∩ H).
Similarly u ∈ g ′ c0 (Λ∩K). Conversely, any element of the right side has the form
g ′ h = g1 x1 h for some h ∈ H hence is in g1 H, and has the form g ′ c0 k = g2 x2 k
for some k ∈ K hence is in g2 K.
We will now find the full-form sequences for Λ ∩ H and Λ ∩ K. Apply
the Preimage algorithm to compute for each wi preimages u′i ∈ H and vi′ ∈
K. Compute a generating set y1′ , . . . , ys′ ′ for H ∩ Γc by finding the Full-form
sequence for H and selecting only those elements that belong to Γc . Similarly,
compute a generating set z1′ , . . . , zt′′ for K ∩ Γc . We now have
= hu′1 , . . . , u′l , y1′ , . . . , ys′ ′ i.
= hv1′ , . . . , vl′ , z1′ , . . . , zt′′ i,
Λ∩H
Λ∩K
Using the generating sets above, find the Full-form sequence (v1 , . . . , vn , z1 , . . . , zs )
for Λ ∩ K, where z1 denotes the first generator of the sequence that lies in Γc .
Likewise find the Full-form sequence (u1 , . . . , un′ , y1 , . . . , yt ) for Λ ∩ H, with
y1 being the first generator in Γc . Since Λ ∩ H and Λ ∩ K have the same image
H ∩ K under , it follows that n′ = n and for all i = 1, . . . , n that ui = vi ci for
some ci ∈ Γc . We now have the full-form sequences
Λ∩H
Λ∩K
=
=
hv1 c1 , . . . , vn cn , y1 , . . . , yt i.
hv1 , . . . , vn , z1 , . . . , zs i,
15
(6)
(7)
The next step produces a generating set of H ∩ K and the element g. The
correctness of this step is argued below. Denote C1 = hc1 , . . . , cn , y1 , . . . , yt i and
consider the intersection
C1 ∩ (Λ ∩ K ∩ Γc )
in the abelian group Γc . Define a homomorphism ψ : Zn+t → Γc by
βt
αn β1
1
ψ(α1 , . . . , αn , β1 , . . . , βt ) = cα
1 · · · cn y 1 · · · y t .
Using the composition ψ ′ : Zn+t → Γc → Γc /(Λ ∩ K ∩ Γc ), we may then use the
Kernel algorithm, as in the base case, to produce a finitely generated subgroup
P = hp1 , . . . , pb i ≤ Zn+t such that
C1 ∩ Λ ∩ K ∩ Γc = ψ(P ).
The sequence (p1 , . . . , pb ) is the full-form sequence for P , so the corresponding
matrix formed is in row-echelon form. We denote pi = (pi1 , . . . , pi(n+t) ) for
i = 1, . . . , b. In addition, we use the Membership algorithm, as described in
the base case, to find
c′0 ∈ C1 ∩ c0 (Λ ∩ K ∩ Γc )
if such an element exists and to write c′0 in the form
α′
α′
β′
β′
c′0 = c1 1 · · · cnn y1 1 · · · yt t .
If c′0 does not exist, we return ‘No’. Otherwise, we define
′
′
β′
β′
h = (v1 c1 )α1 · · · (vn cn )αn y1 1 · · · yt t
and return g = g ′ h. For the generating set of H ∩ K, define the function (it is
not, in general, a homomorphism) θ : Zn+t → Λ ∩ H by
θ(α1 , . . . , αn , β1 , . . . , βt ) = (v1 c1 )α1 · · · (vn cn )αn y1β1 · · · ytβt
and return the Full-form sequence for the subgroup generated by the set
Π = {θ(pi ) | 1 ≤ i ≤ b}.
It remains to prove the correctness of the last step. First, we prove that Π
generates H ∩ K = (Λ ∩ H) ∩ (Λ ∩ K). Take any pi = (α1 , . . . , αn , β1 , . . . , βt ).
Then, using the fact that C1 ≤ Γc is in the center of G,
θ(pi ) = (v1 c1 )α1 · · · (vn cn )αn y1β1 · · · ytβt
=
v1α1
1
· · · vnαn cα
1
n β1
· · · cα
n y1
· · · ytβt .
(8)
(9)
βt
αn β1
1
Line (8) gives θ(pi ) ∈ Λ ∩ H and, since cα
1 · · · cn y1 · · · yt = ψ(pi ) ∈ ψ(P ) ⊂
Λ ∩ K, line (9) gives θ(pi ) ∈ Λ ∩ K. Hence hΠi ≤ H ∩ K.
16
For the opposite inclusion, let (Λ ∩ H)i = hvi ci , . . . , vn cn , y1 , . . . , yt i for 1 ≤
i ≤ n and (Λ ∩ H)n+1 = hy1 , . . . , yt i. We will prove, by induction on i in the
reverse order n + 1, . . . , 1, that
(Λ ∩ H)i ∩ (Λ ∩ K) ≤ hΠi
for all i = n + 1, . . . , 1 (in particular for i = 1). For the base case i =
n + 1, let q ∈ (Λ ∩ H)n+1 ∩ (Λ ∩ K). Then q = c01 · · · c0n y1β1 · · · ytβt for some
(0, . . . , 0, β1 , . . . , βt ) = p. Since q ∈ C1 ∩ Λ ∩ K, we have p ∈ P . Since the
matrix corresponding to P is in row echelon form, we may write p as a linear
combination
b
X
p=
γj p j
j=k
where pj = (0, . . . , 0, pj(n+1) , . . . , pj(n+t) ) for all j ≥ k. Then
p
p
p
p
q = (y1 k(n+1) · · · yt k(n+t) )γk · · · (y1 b(n+1) · · · yt b(n+t) )γb
= θ(pk )γk · · · θ(pb )γb
hence q ∈ hΠi.
For the inductive case, assume that (Λ ∩ H)i+1 ∩ (Λ ∩ K) ≤ hΠi for some
i + 1 ≤ n + 1 and let q ∈ (Λ ∩ H)i ∩ (Λ ∩ K). Then
q = (v1 c1 )0 · · · (vi−1 ci−1 )0 (vi ci )αi · · · (vn cn )αn y1β1 · · · ytβt
for some (0, . . . , 0, αi , . . . , αn , β1 , . . . , βt ) = p′ . Since q ∈ Λ ∩ K and viαi · · · vnαn ∈
βt
αn β1
i
Λ ∩ K, it follows, rewriting q as in (9), that cα
i · · · cn y1 · · · yt ∈ Λ ∩ K and
′
hence p ∈ P . Hence
b
X
p=
γj′ pj
j=k′
for some γj′ and k ′ ≥ 1 such that pj = (0, . . . , 0, pji , . . . , pj(n+t) ) for all j ≥ k ′ .
Now consider the element
′
′
q ′ = qθ(pk′ )−γk′ · · · θ(pb )−γb .
′
′
In the word θ(pk′ )−γk′ · · · θ(pb )−γb , the generators v1 c1 , . . . , vi−1 ci−1 do not appear. Further, the total exponent sum of the generator vi ci is −αi , while in q
it is αi . Since for any i < l ≤ n and 1 ≤ l′ ≤ t the commutators [vi ci , vl cl ] and
[vi ci , yl′ ] are elements of (Λ ∩ H)i+1 we may collect all occurrences of vi ci and
eliminate its occurrence in q ′ . Hence q ′ ∈ (Λ ∩ H)i+1 ∩ (Λ ∩ K). By induction,
q ′ ∈ hΠi hence q ∈ hΠi as well.
Regarding the intersection being non-empty, observe that by (5),
g1 H ∩ g2 K 6= ∅ ⇐⇒ (Λ ∩ H) ∩ c0 (Λ ∩ K) 6= ∅.
17
Now if g = (v1 c1 )α1 · · · (vn cn )αn y1β1 · · · ytβt ∈ Λ∩H is also an element of c0 (Λ∩K)
βt
αn β1
1
then, again rewriting as in (9), cα
1 · · · cn y1 · · · yt is an element of C1 which
also lies in c0 (Λ ∩ K ∩ Γc ) hence c′0 exists. Conversely, if c′0 exists, then we have
α′
β′
β′
γ′
γ′
1
t
1
s
n′
c′0 = c1 1 · · · cα
n y 1 · · · y t = c0 z 1 · · · z s
hence
′
′
β′
β′
α′
α′
γ′
γ′
(v1 c1 )α1 · · · (vn cn )αn y1 1 · · · yt t = c0 v1 1 · · · vn n z1 1 · · · zs s
is an element of (Λ ∩ H) ∩ c0 (Λ ∩ K), hence this intersection is non-empty. This
proves the correctness of the decision problem.
Finally, we must show that g ∈ g1 H ∩ g2 K. Since h ∈ Λ ∩ H, we have
g = g ′ h ∈ g ′ (Λ ∩ H). Since c′0 ∈ c0 (Λ ∩ K), we have c′0 = c0 k for some k ∈ Λ ∩ K
hence
α′
α′
g = g ′ h = g ′ v1 1 · · · vn n c′0 = g ′ c0 (kv1α1 · · · vnαn )
is an element of g ′ c0 (Λ ∩ K), as required.
Complexity. The depth of the recursion is c, which is constant, so the total
number of subroutine calls is constant. The total number of group elements to
record is also constant.
As an application of the intersection algorithm, we may generalize Lemma
4 to solve the simultaneous conjugation problem for tuples in nilpotent groups.
Theorem 7. Fix integers c, r, and l. There is an algorithm that, given a
nilpotent group G = hX | Ri of nilpotency class at most c with |X| ≤ r and two
tuples (a1 , . . . , al ) and (b1 , . . . , bl ) of elements of G, computes:
• an element g ∈ G such that
g −1 ai g = bi
for i = 1, . . . , l
• a generating set for the centralizer CG (b1 , . . . , bl ), or determines that no
such element g exists.
The algorithm runs in space logarithmic in the size L of the input and time
quasilinear in L, or in TC0 , and the length of each output word is bounded by
a polynomial function of L. If compressed inputs are used, the space complexity
is O(M ) and the time complexity O(nM 3 ) where n = |R| and M bounds the
encoded size of each input word.
Proof. Algorithm. Begin by applying Lemma 3 to convert to a nilpotent presentation if necessary. Next, for each i = 1, . . . , l, use the Conjugacy algorithm
to find gi ∈ G such that agi i = bi . If any pair is not conjugate, then g does not
exists and we may return ‘No’. We also use the Centralizer algorithm to find,
for each i, a generating set for CG (bi ).
18
Now for any g ∈ G and any i, the equation
g gi−1 g
agi = ai i
g−1 g
= bi i
shows that agi = bi if and only if gi−1 g ∈ CG (bi ), i.e. g ∈ gi CG (bi ). Hence the
Tl
set of all possible conjugators is precisely the coset intersection i=1 gi CG (bi )
which we may compute by iterating Theorem 6. As a by-product, we obtain
a generating set for
l
\
CG (bi ) = CG (b1 , . . . , bl ).
i=1
Complexity. Since l is fixed, the number of subroutine calls and elements of G
to store is constant.
3.3
Torsion subgroup
In every nilpotent group G the set T consisting of all elements of finite order
forms a subgroup called the torsion subgroup. We give an algorithm to compute,
from a presentation of G, a generating set and presentation for T as well as its
order. We follow an algorithm outlined in [KRR+ 69].
Theorem 8. Fix positive integers c and r. There is an algorithm that, given
a finitely presented nilpotent group G = hX | Ri of nilpotency class at most c
with |X| ≤ r, produces
• a generating set for the torsion subgroup T of G,
• a presentation for T , and
• the order of T .
The algorithm runs in space logarithmic in the size L of the given presentation
and time quasilinear in L, or in TC0 . The length of each output word is bounded
by a polynomial function of L and the number of such words is bounded by a
constant. If compressed inputs are used, the space complexity is O(nM ) and the
time complexity is O(nM 3 ), where n = |R| and M bounds the length of each
relator in R.
Proof. Define inductively a sequence T1 , T2 , . . . of finite normal subgroups of G
as follows. Let T1 = T (Z(G)), which is clearly finite and normal. For i > 1
define the homomorphism φi : G → G/Ti−1 and set
Ti = φi−1 (T (Z (G/Ti−1 ))) .
Since Z(G/Ti−1 ) is abelian and finitely generated, T (Z(G/Ti−1 )) is finite and
hence finiteness of Ti follows from that of Ti−1 . Normality of Ti follows from
normality of Ti−1 in G and of T (Z(G/Ti−1 )) in G/Ti−1 . Since G is Noetherian,
the sequence must stabilize at some Tk . But then T (Z(G/Tk )) is trivial, hence
19
G/Tk is torsion-free (its torsion subgroup must otherwise intersect its center),
hence Tk = T .
Algorithm. We compute the sequence described above. Begin by applying
Lemma 3 to compute a nilpotent presentation G = hA | Si. Since Z(G) is
simply the centralizer of any generating set, we may find the full-form sequence
(h1 , . . . , hm ) for Z(G) using Theorem 7 with the set A. Since Z(G) is abelian,
its torsion subgroup is generated by the set X1 consisting of elements hi such
that i ∈ T . Note that T is determined by examining the relators of the form
(4) is S.
Now assume, by induction, that we have a generating set Xi for Ti . Use
Theorem 7 with the nilpotent group G/Ti = hA | S ∪ Xi i and the set A to
find, as described in the base case, the full-form sequence (τ1 Ti , . . . , τm Ti ) for
T (Z(G/Ti )). Then Xi ∪ {τ1 , . . . , τm } generates Ti+1 , and we compute the the
Full-form sequence Xi+1 of Ti . If Xi+1 = Xi , then Ti+1 = Ti = T . Otherwise,
we proceed with the next step of the induction.
Once we obtain the full-form sequence (t1 , . . . , tn ) for T , it suffices to run
Subgroup Presentation to give a presentation for T . Denote by π1 , . . . , πn
the pivot columns of the matrix associated with (t1 , . . . , tn ) and by αij the (i, j)entry of this matrix. Then every element of T may be expressed uniquely in
the form tβ1 1 · · · tβnn where 0 ≤ βi < eπi /αiπi , and every such expression gives a
different element. Hence the order of T is
|T | =
n
Y
eπi /αiπi .
i=1
Complexity. First, we will prove that the depth of the recursion is bounded
by c. Let Zi = {h ∈ G | [h, g] ∈ Zi−1 for all g ∈ G} be the ith term of the upper
central series of G, with Z1 = Z(G). We claim that
T i ⊇ Zi ∩ T
for all i = 1, . . . , c, hence Tc ⊇ G ∩ T so Tc = T and the depth of the recursion
is bounded by c. We proceed by induction. For i = 1 we have T1 = T (Z(G)) =
Z1 ∩T . Now let g ∈ Zi ∩T and consider φi (g) = g. Let h ∈ G/Ti−1 and consider
[g, h] = [g, h]. Since g ∈ T , we have [g, h] = g −1 g h ∈ T and since g ∈ Zi , we
have [g, h] ∈ Zi−1 . By the inductive assumption, [g, h] ∈ Ti−1 hence [g, h] = 1
hence g ∈ Z(G/Ti−1 ). Clearly g ∈ T (Z(G/Ti−1 )) hence g ∈ Ti , proving the
claim.
Since the depth of the recursion is constant, the total number of subroutine
calls is constant, as is the number of elements kept in memory, since each Xi is
a full-form sequence (hence of bounded length).
In computing the order of T , recall that the numbers ei where i ∈ T , appear as exponents in the nilpotent presentation computed by Lemma 3. Consequently, each is bounded by a polynomial function of L. Since the length
n of the full-form sequence for T is bounded by a constant, the order of T is
polynomially bounded.
20
Corollary 9. If G = hX | Ri is a nilpotent group of nilpotency class c with
|X| ≤ r, then the order of the torsion subgroup of G is bounded by a polynomial
function of the sum of the lengths of the relators R.
3.4
Isolator
Recall that the isolator of H in G is defined by
IsG (H) = {g ∈ G | g n ∈ H for some n 6= 0}
and, in nilpotent groups, forms a subgroup.
Theorem 10. Fix integers c and r. There is an algorithm that, given a finitely
presented nilpotent group G = hX | Ri of nilpotency class at most c with |X| ≤ r,
and a subgroup H ≤ G, computes
a generating set for the isolator IsG (H).
The algorithm runs in space logarithmic in the length L of the input and time
quasilinear in L, or in TC0 , and the length of each generator is bounded by a
polynomial function of L. If compressed inputs are used, the space complexity is
O(M ) and the time complexity O(nM 3 ) where n is the number of input words
and M bounds the encoded size of each input word.
Proof. Algorithm. First, apply Lemma 3 to convert to a nilpotent presentation
G = ha1 , . . . , am | Si.
Let N 0 = H and for i > 0 define N i = NG (N i−1 ), the normalizer of N i−1
in G. It is proved in [KM79] Thm. 16.2.2 that N c = G. Using Theorem 5, we
compute in turn the full-form generating sequences for each of the subgroups
H, N 1 , . . . , N c and using the Subgroup Presentation algorthim we compute
a nilpotent presentation
N i = hXi | Ri i
for each. We now proceed, by induction, to compute for each i = 0, . . . , c a
generating set Yi for IsN i (H). For i = 0, we have IsN 0 (H) = H and we use the
computed full-form sequence X0 for H. Now assume that we have a generating
set Yi−1 for IsN i−1 (H).
The subgroup N i−1 is normal in N i , and we will find the torsion subgroup of
N i /N i−1 . Using Lemma 3, write each element of Xi−1 in its Xi -coordinates.
Appending these elements to Ri we obtain a presentation of N i /N i−1 , which
we pass to Theorem 8 to obtain a generating set {τ1 N i−1 , . . . , τm N i−1 } for
the torsion subgroup. Then IsN i (N i−1 ) is generated by Xi−1 ∪ {τ1 , . . . , τr }.
Converting these elements back to generators of G, we then compute the Fullform sequence Zi for IsN i (N i−1 ) and, using Lemma 3 the corresponding
nilpotent presentation
IsN i (N i−1 ) = hZi | Si i.
We claim that IsN i−1 (H) E IsN i (N i−1 ). Indeed, the property that g n ∈ H
for some n > 0 is unchanged under conjugation, and since N i normalizes N i−1
21
all conjugates remain in N i−1 . Using Lemma 3, write each element of Yi−1 in
terms of the generators Zi and append these words to Si to obtain a presentation
of IsN i (N i−1 )/IsN i−1 (H). Now apply Theorem 8 to compute a generating set
{σ1 IsN i−1 (H), . . . , σr IsN i−1 (H)} for the torsion subgroup.
Set Yi = {σ1 , . . . , σr } ∪ Yi−1 , using the two prior calls to Lemma 3 to write
each σi in generators of G. We claim that Yi generates IsN i (H). Clearly Yi−1 ⊂
IsN i−1 (H) ⊂ IsN i (H). For each σi there exists pi > 0 such that σipi ∈ IsN i−1 (H),
therefore there exists qi > 0 such that σipi qi ∈ H. Hence Yi ⊂ IsN i (H). Now if
g ∈ IsN i (H) then g n ∈ H for some n > 0 and g ∈ IsHi (Hi−1 ), hence g·IsHi−1 (H)
lies in the torsion subgroup of IsHi (Hi−1 )/IsHi−1 (H), and so g is an element of
the subgroup generated by Yi .
Finally, return the Full-form sequence for Yn .
Complexity. Since the number of elements in a full-form sequence is bounded
by m, the total number of elements in the sequences for N i , i = 0, . . . , c, is
constant, as is the number of relators in each Ri and the number of elements
in the generating set for the torsion subgroups. The depth of the recursion is
bounded by c. Hence the total number of elements to store and the number of
subroutine calls is constant.
References
[BMO16]
Gilbert Baumslag, Charles F. Miller, and Gretchen Ostheimer. Decomposability of finitely generated torsion-free nilpotent groups.
Internat. J. Algebra Comput., 26(8):1529–1546, 2016.
[DLS15]
Moon Duchin, Hao Liang, and Michael Shapiro. Equations in nilpotent groups. Proc. Amer. Math. Soc., 143(11):4723–4731, 2015.
[GMO16a] Albert Garreta, Alexei Miasnikov, and Denis Ovchinnikov. Properties of random nilpotent groups. 2016. Preprint. arXiv:1612.01242
[math.GR].
[GMO16b] Albert Garreta, Alexei Miasnikov, and Denis Ovchinnikov. Random nilpotent groups, polycyclic presentations, and diophantine
problems. 2016. arXiv:1612.02651 [math.GR].
[GS80]
Fritz Grunewald and Daniel Segal. Some general algorithms. II.
Nilpotent groups. Ann. of Math. (2), 112(3):585–617, 1980.
[HEO05]
Derek F. Holt, Bettina Eick, and Eamonn A. O’Brien. Handbook
of computational group theory. Discrete Mathematics and its Applications (Boca Raton). Chapman & Hall/CRC, Boca Raton, FL,
2005.
[KLZ15]
Daniel König, Markus Lohrey, and Georg Zetzsche. Knapsack and
subset sum problems in nilpotent, polycyclic, and co-context-free
groups. 2015. arXiv:1507.05145 [math.GR].
22
[KM79]
M. I. Kargapolov and Ju. I. Merzljakov. Fundamentals of the theory
of groups, volume 62 of Graduate Texts in Mathematics. SpringerVerlag, New York-Berlin, 1979. Translated from the second Russian
edition by Robert G. Burns.
[KRR+ 69] M. I. Kargapolov, V. N. Remeslennikov, N. S. Romanovskiı̆, V. A.
Roman′ kov, and V. A. Čurkin. Algorithmic questions for σ-powered
groups. Algebra i Logika, 8:643–659, 1969.
[Loh12]
Markus Lohrey. Algorithms on SLP-compressed strings: a survey.
Groups Complex. Cryptol., 4(2):241–299, 2012.
[Loh14]
Markus Lohrey. The compressed word problem for groups. SpringerBriefs in Mathematics. Springer, New York, 2014.
[Loh15]
Markus Lohrey. Rational subsets of unitriangular groups. Internat.
J. Algebra Comput., 25(1-2):113–121, 2015.
[MMNV15] Jeremy Macdonald, Alexei Miasnikov, Andrey Nikolaev, and Svetla
Vassileva. Logspace and compressed-word computations in nilpotent groups. 2015. Preprint. arXiv:1503.03888 [math.GR].
[MT16]
Alexei Mishchenko and Alexander Treier. Knapsack problem for
nilpotent groups. 2016. Preprint. arXiv:1606.08584 [math.GR].
[MW17]
Alexei Miasnikov and Armin Weiß. TC0 circuits for algorithmic
problems in nilpotent groups. 2017. To appear in MFCS 2017
proceedings. Preprint at arXiv:1702.06616 [math.GR].
[Rom77]
V. A. Roman’kov. Unsolvability of the problem of endomorphic
reducibility in free nilpotent groups and in free rings. Algebra i
Logika, 16(4):457–471, 494, 1977.
[Rom16]
Vitaly A. Roman’kov. Diophantine questions in the class of finitely
generated nilpotent groups. J. Group Theory, 19(3):497–514, 2016.
[Sim94]
Charles C. Sims. Computation with finitely presented groups, volume 48 of Encyclopedia of Mathematics and its Applications. Cambridge University Press, Cambridge, 1994.
[Vol99]
Heribert Vollmer. Introduction to circuit complexity. Texts in
Theoretical Computer Science. An EATCS Series. Springer-Verlag,
Berlin, 1999. A uniform approach.
23
| 4math.GR
|
arXiv:1606.03140v2 [quant-ph] 1 Jul 2016
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
FOR GROUPS ARISING FROM NON-LOCAL GAMES
WILLIAM SLOFSTRA
Abstract. Tsirelson’s problem asks whether the commuting operator model
for two-party quantum correlations is equivalent to the tensor-product model.
We give a negative answer to this question by showing that there are nonlocal games which have perfect commuting-operator strategies, but do not
have perfect tensor-product strategies. The weak Tsirelson problem, which
is known to be equivalent to Connes embedding problem, remains open.
The examples we construct are instances of (binary) linear system games.
For such games, previous results state that the existence of perfect strategies is controlled by the solution group of the linear system. Our main
result is that every finitely-presented group embeds in some solution group.
As an additional consequence, we show that the problem of determining
whether a linear system game has a perfect commuting-operator strategy
is undecidable.
Contents
1. Introduction
2. Linear system games, hypergraphs, and solution groups
3. The embedding theorem and consequences
4. Presentations by involutions
5. The wagon wheel embedding
6. Pictures for groups generated by involutions
6.1. Pictures as planar graphs
6.2. Groups and labellings of pictures
7. Pictures over solution groups and hypergraphs
8. A category of hypergraphs
9. Cycles and outer faces
10. Pictures over suns
11. Stellar cycles and constellations
11.1. Structure of constellations
11.2. Stellar cycles
11.3. Covers versus copies
11.4. Cycles covered by other cycles
11.5. Proof of Theorem 11.4
12. Proof of the embedding theorem
References
1
2
5
7
9
13
15
15
19
21
24
31
34
40
42
42
44
46
47
49
56
2
WILLIAM SLOFSTRA
1. Introduction
In a two-player non-local game, the players, commonly called Alice and
Bob, are physically separated and unable to communicate. They each receive
a question chosen at random from a finite question set, and reply with a
response from a finite answer set. If the joint answers meet a predetermined
winning condition dependent on the joint questions, then Alice and Bob win;
otherwise they lose. The rules of the game, including the winning condition
and distribution on questions, are completely known to Alice and Bob, and
they can arrange in advance a strategy which will maximize their success
probability. However, since they cannot communicate during the game, they
may not be able to play perfectly, i.e. win with probability one.
Classically, Alice and Bob’s strategy for a non-local game is described by
a local hidden variable model. Bell’s famous theorem states that Alice and
Bob can achieve better results than is possible with a local hidden variable
model if they share an entangled quantum state [Bel64]. Since Bell’s discovery, non-local games have been heavily studied1 in physics, mathematics, and
computer science; see [CHSH69, FC72, Cir80, AGR82, CHTW04, BPA+ 08,
NPA08, JPPG+ 10, KV11, KKM+ 11, HBD+ 15] for a small sample of results.
Despite this, a number of foundational questions remain open, chief among
which is Tsirelson’s problem: A quantum strategy for a non-local game can
be described as a set of measurement operators on Hilbert spaces HA and
HB , along with a quantum state in the joint space H = HA ⊗ HB . We refer
to this as the tensor-product model. While the tensor-product model is often the default, there is another choice: a quantum strategy can be described
as a set of measurements and quantum state on a shared Hilbert space H,
with the property that Alice’s measurement operators commute with Bob’s
measurement operators. This commuting-operator model is used, for instance,
in algebraic quantum field theory [HK64]. The observable consequences of a
strategy in either model are captured by the correlation matrix of the strategy.
If H is finite-dimensional, every correlation matrix arising from a commutingoperator strategy can be realized using a tensor-product strategy. Tsirelson’s
problem asks whether this is true for a general Hilbert space.
This problem has an interesting history. Tsirelson originally stated the
problem in a survey on Bell inequalities [Tsi93], and claimed without proof
that the two models gave rise to the same set of correlation matrices. He later
retracted this claim, and posted the question to a list of open problems in
quantum information theory [Tsi06]. Subsequent authors [NPA08, JNP+ 11,
Fri12, NCPGV12, PT15, DP16, PSS+ 16] studying Tsirelson’s problem have
1
Usually under the name “Bell tests” or “Bell inequalities”. The term “non-local games”
is more recent.
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
3
defined the set of tensor-product strategies in several different ways. Paulsen
and Todorov [PT15] (see also Dykema and Paulsen [DP16]) observe that these
variations lead to a hierarchy of sets of correlation matrices
Cq ⊆ Cqs ⊆ Cqa ⊆ Cqc ,
where Cq is the set of correlations arising from tensor-product strategies on
finite-dimensional Hilbert spaces, Cqs is the set of correlations arising from
tensor-product strategies (on possibly infinite-dimensional Hilbert spaces) with
a vector state, Cqa = Cq is the set of correlations which are limits of correlations in Cq , and Cqc is the set of correlations arising from commuting-operator
strategies. If we restrict to non-local games with question sets of size n and
2 2
answer sets of size m, then all these sets are convex subsets of Rm n , and
none of the inclusions were previously known to be strict. Thus for each
t ∈ {q, qs, qa} there is a Tsirelson problem asking whether Cqc = Ct . Ozawa,
building on the work of Junge, Navascués, Palazuelos, Pérez-Garcı́a, Sholz,
and Werner [JNP+ 11] and work of Fritz [Fri12], has shown that Cqc = Cqa if
and only if Connes’ embedding conjecture is true [Oza13]. At the other end
of the hierarchy, if Cq was equal to Cqc then every correlation matrix, whether
commuting operator or tensor-product, would arise from a finite-dimensional
Hilbert space. The “middle” version, which asks whether Cqc = Cqs , seems
closest to Tsirelson’s original problem statement.
The first main result of this paper is that there is a non-local game which can
be played perfectly with a commuting-operator strategy, but which cannot be
played perfectly using a tensor-product strategy with a vector state. Thus we
resolve the middle version of Tsirelson’s problem by showing that Cqc 6= Cqs .
This game is interesting from the perspective of quantum information and
computation, where a non-local game is often regarded as a computational
scenario in which better results can be achieved with entanglement as a resource. From this point of view, it is natural to ask how much entanglement is
needed to play a game optimally, and in particular, whether every game can be
played optimally on a finite-dimensional Hilbert space. The game we construct
shows that this is not possible, at least if we allow commuting-operator strategies. Previously-known examples of this type have involved either quantum
questions [LTW13, RV15] or infinite answer sets [MV14].
One reason we would have desired that every game have an optimal strategy on a finite-dimensional Hilbert space is that it would make it possible to
determine the optimal winning probability of a non-local game over entangled
strategies. At present the only known methods for this task, aside from bruteforce search over strategies, are variants of the Navascués-Pironio-Acı́n (NPA)
hierarchy [NPA08, DLTW08]. Given a non-local game, the NPA hierarchy
provides a sequence of upper bounds which converge to the optimal winning
4
WILLIAM SLOFSTRA
probability in the commuting-operator model. However, the hierarchy does
not provide, outside of special cases, a stopping criterion, i.e. a way to tell if
the value will fall below a given threshold. Our second main result is that it
is undecidable to determine if a non-local game can be played perfectly with
a commuting-operator strategy. In particular, this implies that there is no
stopping criterion for the NPA hierarchy which applies to all games.
The games we consider are binary linear system games, so named because
they arise from linear systems over Z2 . Such games have been studied previously in [CM14, Ark12, Ji13]. Cleve and Mittal implicitly associate a certain
group to every linear system over Z2 , such that perfect tensor-product strategies for the game correspond to certain finite-dimensional representations of
the group [CM14]. We call this group, which is analogous to the solution
space of a linear system, the solution group. In [CLS16] it is shown that
perfect commuting-operator strategies for a binary linear system game correspond to certain possibly-infinite-dimensional representations of the solution
group. Solution groups form an interesting class of groups. They are finitely
presented, but their presentations must satisfy a property which in [CLS16] is
called local compatibility: if
x1 · · · xn = 1
is a relation, where xi , 1 ≤ i ≤ n, are not necessarily distinct generators of
the group, then the presentation must also contain the relations xi xj = xj xi
for all 1 ≤ i, j ≤ n. This condition is natural from the perspective of quantum mechanics, where two observables commute if and only if the observables
correspond to quantities which can be measured (or known) simultaneously.
Group relations of this exact type can be found in contextuality theorems of
Mermin and Peres [Mer90, Per90, Mer93]. Local compatibility is a priori a
strong constraint on group presentations. Our primary result, on which our
other two results are based, is that any finitely-presented group can be embedded in a solution group. This embedding theorem allows us to extend results
from combinatorial group theory, such as the existence of a non-residuallyfinite group, to solution groups. Using non-residually-finite groups to recognize
infinite-dimensional state spaces was previously proposed in [Fri13].
A number of open questions remain, such as whether there are other separations between the correlation sets Ct (m, n), and (assuming that Cqa 6= Cqc )
whether it is decidable to determine the optimal value of a non-local game
over Cqa . The argument used in this paper cannot distinguish between Cqs
and Cq , and we do not expect our methods to help with other separations
between correlation sets.
The rest of this paper is structured as follows. In the next section, we give
some background on binary linear system games, and introduce the solution
group of a linear system. In Section 3, we state the embedding theorem for
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
5
solution groups, and prove our two main results as corollaries. The rest of
the paper is concerned with the proof of the embedding theorem. The main
technical tool used is pictures of groups. We give an overview of pictures
tailored to solution groups in Sections 6 and 7; expert readers will want to skip
or briefly review these sections on first reading. In dealing with pictures, it is
more convenient to use hypergraphs instead of linear systems, and we introduce
hypergraphs into our definitions immediately in Section 2. In Section 8 we
develop a notion of morphisms between hypergraphs; the concept is similar to
graph minors, but differs from standard notions of hypergraph minors as in
[RS10]. The more technical aspects of the proof are contained in Sections 9-11;
we suggest that the reader skim these sections on first reading and proceed to
Section 12. For a second reading, Corollary 10.4 might provide a good initial
target. One thing we do not provide is a definition of the correlation sets
mentioned above, or a definition of binary linear system games; instead we
defer to the excellent references listed above.
1.1. Acknowledgements. I thank Richard Cleve, Jason Crann, Zhengfeng
Ji, Li Liu, Andrew Marks, Vern Paulsen, and Vincent Russo for many helpful
discussions. I thank Richard Cleve and Li Liu for detailed comments on the
manuscript, and Tobias Fritz for helpful suggestions. Any remaining errors
are my own.
2. Linear system games, hypergraphs, and solution groups
Binary linear system games are based on linear systems Ax = b over Z2 .
It is convenient to think of linear systems in terms of hypergraphs. By a
hypergraph, we mean a triple H = (V, E, A), where V = V (H) and E = E(H)
are the sets of vertices and edges respectively, and A ∈ ZV≥0×E is the incidence
matrix between V and E, so Ave ≥ 0 is the degree of incidence between edge
e and vertex v. We sayP
that v and e are incident if Ave > 0.PIf v ∈ V , then
the degree of v is |v| = e Ave . Similarly if e ∈ E then |e| = v Ave . We say
that H is simple if Ave ≤ 1 for all v ∈ V and e ∈ E, k-regular if |v| = k for all
v ∈ V , and a graph if |e| = 2 for all e ∈ E.
Note that this definition of hypergraphs allows both isolated vertices and
isolated edges, i.e. vertices (resp. edges) which are incident to no edges (resp.
vertices). A Z2 -vertex labelling of H is a function b : V → Z2 . With these
conventions, there is a correspondence between linear systems Ax = b and
simple hypergraphs H with a vertex labelling b. From this point of view,
the edges of a hypergraph correspond to the variables of a linear system, and
the vertices correspond to constraints. Similarly, pairs (H, b) where H is not
necessarily simple correspond to linear systems Ax = b over Z2 with a choice
of non-negative integer representatives for the coefficients Ave .
6
WILLIAM SLOFSTRA
To any linear system Ax = b, we can associate a linear system non-local
game G, and a group Γ [CM14, CLS16]. The group Γ is the focus of attention
of this paper, and is defined as follows.
Definition 2.1. Let H = (V, E, I) be a (not necessarily simple) hypergraph
and let b be a function V → Z2 . The solution group Γ = Γ(H; b) associated to
H and b is the group generated by {xe , e ∈ V } ∪ {J}, subject to relations:
(1) x2e = 1 for all e ∈ E and J 2 = 1 (i.e. Γ is generated by involutions)
(2) [xe , J] = 1 for all e ∈ E (i.e. J is central),
(3) [xe , xe′ ] = 1 if there is some vertex v incident to both e and e′ , and
(4)
Y
ve
xA
= J bv for all v ∈ V.
e
e
The null solution group is the group Γ(H) := Γ(H, 0).
We call the last two types of relations commuting relations and linear relations respectively. The definition of the linear relations assumes that E is
ordered, but the choice of order is irrelevant because of the commuting relations. Note that if v and e are incident and Ave is even, then the linear
relations
Y A ′
Y A ′
xe′ve = J bv
xe′ve = J bv and
e′
e′ 6=e
are equivalent. However, the fact that Ave > 0 might still lead to commuting
relations that wouldn’t hold otherwise.
The one-dimensional representations π of Γ(H; b) in which π(J) 6= 1 correspond to the solutions of the linear system Ax = b. Higher-dimensional
representations of Γ(H; b) with this property can be thought of as quantum
solutions of Ax = b. This is justified by the following theorem, which relates
solution groups to non-local games.
Theorem 2.2 ([CM14, CLS16]). Let Ax = b be a linear system over Z2 , where
A is a non-negative integral matrix, let G be the associated linear system nonlocal game, and let Γ be the corresponding solution group. Then:
• G has a perfect quantum commuting-operator strategy if and only if
J 6= 1 in Γ.
• G has a perfect quantum tensor-product strategy if and only if G has
a perfect finite-dimensional quantum strategy, and this happens if and
only if Γ has a finite-dimensional representation π with π(J) 6= 1.
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
7
The first part of this theorem is due to [CLS16], while the second part is
due to [CM14].
3. The embedding theorem and consequences
In light of Theorem 2.2, we would like to understand the structure (or lack
thereof) of solution groups for linear system games. Recall that a presentation
hS : Ri of a group G is a set S and subset R of the free group F (S) generated
by S, such that G = F (S)/(R), where (R) is the normal subgroup generated
by R. A group is finitely presented if it has a presentation hS : Ri where both
S and R are finite. Our primary result is that understanding solution groups
is as hard as understanding finitely presented groups.
Theorem 3.1. Let G be a finitely presented group, let J ′ ∈ G be a central
element with (J ′ )2 = 1, and let w1 , . . . , wn , n ≥ 0 be a sequence of elements in
G such that wi2 = 1 for all 1 ≤ i ≤ n. Then there is a hypergraph H, a vertex
labelling function b : V (H) → Z2 , a sequence of edges e1 , . . . , en in H, and a
homomorphism φ : G → Γ(H, b) such that φ is an embedding, φ(J ′ ) = J, and
φ(wi ) = xei for all 1 ≤ i ≤ n.
The hypergraph H, vertex labelling b and homomorphism φ can be explicitly
constructed from a presentation of G. This is described in Sections 4 and 5.
Theorem 3.1 is proved at the beginning of Section 5 via reduction to another
embedding theorem. The proof of this latter embedding theorem (and hence
the proof of Theorem 3.1) is completed in Section 12.
In the remainder of this section, we prove two consequences of Theorem 3.1.
The first is an answer to Tsirelson’s problem.
Corollary 3.2. There is a linear system non-local game which has a perfect
quantum commuting-operator strategy, but does not have a perfect quantum
tensor-product strategy.
Proof. Suppose that G is a finitely-presented group with a central element J ′
of order two, such that π(J ′ ) = 1 for every finite-dimensional representation
π of G. By Theorem 3.1, there is an embedding of G in a solution group Γ
which identifies J ′ with J. In particular, this implies that J 6= 1 in Γ, so
the associated linear system non-local game G must have a perfect quantum
commuting-operator strategy by the first part of Theorem 2.2. If π is a finitedimensional representation of Γ, then π(J) = π|G (J ′ ) = 1. By the second part
of Theorem 2.2, the associated linear system non-local game does not have a
perfect quantum tensor-product strategy.
To finish the proof, we construct a group G with the above property. Consider Higman’s group
H0 = ha, b, c, d : aba−1 = b2 , bcb−1 = c2 , cdc−1 = d2 , dad−1 = a2 i.
8
WILLIAM SLOFSTRA
It is well-known that H0 has no non-trivial linear representations, and that the
generators a, b, c, d of H0 have infinite order [Hig51, Ber94]. Let H = H0 × Z2 ,
and let J ∈ H denote the generator of the Z2 -factor. Let G be the HNN
extension of H by the automorphism of ha, Ji ∼
= Z × Z2 sending J 7→ J and
a 7→ aJ. By the properties of the HNN extension, H is a subgroup of G, and in
particular J is non-trivial in G. Furthermore, we can construct a presentation
for G from a presentation of H by adding a generator x and relations [x, J] = 1
and [x, a] = J. The former relation implies that J is central in G. Finally,
if π is a finite-dimensional representation of G, then π|H0 is trivial, and in
particular, π(a) = 1. But this implies that
π(J) = π([x, a]) = 1,
as required.
We note that any non-residually-finite group can be used in place of Higman’s group in the above proof.
It would be interesting to know the smallest linear system for which the
corresponding game can be played perfectly only with commuting-operator
strategies. No effort is made to reduce the size of the linear system in the
proof of Theorem 3.1, and the main construction from Section 5 depends on
the number of variables and the total length of the relations in the presentation
of G. If we naively follow the proof through for the group in Corollary 3.2, we
get a linear system with roughly 600 variables and 450 linear relations. Making
some obvious improvements in Section 4 can get this down to 400 variables
and 300 relations. It seems likely that this is far from the smallest possible
example.
The second consequence concerns the difficulty of determining whether a
non-local game has a perfect commuting-operator strategy.
Corollary 3.3. It is undecidable to determine if a binary linear system game
has a perfect commuting-operator strategy.
Proof. By Theorem 2.2, determining if a binary linear system game has a
perfect strategy is equivalent to determining if J 6= 1 in a solution group.
Because Theorem 3.1 is constructive, this is in turn equivalent to the following
decision problem: given a group presentation G = hS : Ri and a word J in the
generators S such that J ∈ Z(G) and J 2 = 1, decide if J = 1 in G.
We claim that the word problem for groups can be reduced to this latter
problem. Indeed, given a finitely presented group K = hS : Ri and a word
w ∈ F (S) in the generators of K, it is possible to recursively construct a
finitely presented group Lw with the property that K is a subgroup of Lw if
w 6= 1, and Lw is trivial if w = 1 (see [LS77, pg. 190], where this construction
is attributed to Rabin). The presentation of Lw can be constructed by adding
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
9
generators and relations to the presentation of K. Let Hw be the result of
applying this construction to the group K × Z, and let z be the generator
of Hw corresponding to the Z-factor in K × Z. If w = 1, then Hw is trival,
and z = 1. If w 6= 1, then K × Z is a subgroup of Hw , and the order of z is
infinite. Finally, construct a group Gw by adding two generators x and J to
the presentation of Hw , along with relations J 2 = [x, J] = 1, [s, J] = 1 for all
generators s of Hw , and [x, z] = J. As in Corollary 3.2, if w 6= 1 then Gw is the
HNN extension of the group Hw × Z2 by the automorphism of the subgroup
hz, Ji ∼
= Z × Z2 sending J 7→ J and z 7→ zJ. Thus, if w 6= 1 then J 6= 1 in
Gw . If w = 1, then z = 1 in Hw , and consequently J = [x, z] = 1 in Gw . This
completes the reduction.
Although not used in either of the above corollaries, Theorem 3.1 also allows
us to embed a finitely-presented group G in a solution group Γ so that a given
set of involutions of G become generators of Γ. This can be used to prove
that other tasks involving solution groups are undecidable. For instance, it
is undecidable to determine if a generator xe of a null solution group is nontrivial.
4. Presentations by involutions
We are interested primarily in groups which (a) have a distinguished central
element of order ≤ two, and (b) are generated by involutions. For clarity in
subsequent sections, we encode these conditions in two formal definitions.
Definition 4.1. A group over Z2 is a group G with a distinguished central
element J = JG such that J 2 = 1.
A morphism G1 → G2 over Z2 is a group homomorphism sending JG1 7→
JG2 . Similarly, an embedding over Z2 is an injective morphism over Z2 .
Note that J is allowed to be the identity in this definition. This is so that
we can construct groups over Z2 by starting with some finite presentation,
picking an element J ′ ∈ F (S), and adding relations (J ′ )2 = 1 and [J ′ , s] = 1
for all s ∈ S. By allowing J = 1, we can do this even if J ′ becomes trivial.
Elements of F (S) are represented by words over {s, s−1 : s ∈ S}. Every
element r ∈ F (S) can be represented uniquely as sa11 · · · sann , where ai ∈ {±1},
and ai = ai+1 whenever si = si+1 . A word meeting these conditions is said
to be reduced. The number n is the length of r. The element r is said to be
cyclically reduced if, in addition, sn = s1 implies that an = a1 .
Definition 4.2. Given a set S, let F2 (S) = hS : s2 = 1, s ∈ Si. A presentation by involutions over Z2 for a group G is a set of generators S and a set
of relations R ⊂ F2 (S) × Z2 such that G = F2 (S) × Z2 /(R), where (R) is
10
WILLIAM SLOFSTRA
the normal subgroup generated by R. We denote presentations of this form by
InvhS : Ri, and write G = InvhS : Ri when the meaning is clear.
We use J (written in multiplicative notation) to denote the generator of the
Z2 -factor in F2 (S) × Z2 . If G = InvhS : Ri, we can regard G as a group over
Z2 by letting J = JG be the image of J ∈ F2 (S) × Z2 in G.
Elements of F2 (S) are represented by words over S. Every element r ∈
F2 (S) × Z2 can be represented uniquely as J a s1 · · · sn , where a ∈ Z2 and
s1 , . . . , sn is a sequence in S with si 6= si+1 . Again, a word of this form is
said to be reduced, and n is called the length of r. If, in addition, sn 6= s1
then we say that r is cyclically reduced. We say that a set of relations R is
cyclically reduced if every element of R is cyclically reduced.
If R is a set of relations, the symmetrization of R is the set of relations
sym
R
containing all relations of the form
J a si si+1 · · · sn s1 · · · si−1 and J a si si−1 · · · s1 sn · · · si+1 , 1 ≤ i ≤ n,
for every relation J a s1 · · · sn in R.
Every group presented by involutions over Z2 has a presentation InvhS : Ri
where R is cyclically reduced. The presentations InvhS : Ri and InvhS :
Rsym i are equivalent (i.e. they define isomorphic groups), and if R is cyclically
reduced then Rsym is cyclically reduced.
By definition, solution groups are examples of groups presented by involutions over Z2 . Theorem 3.1 states that every finitely presented group over Z2
embeds (over Z2 ) in a solution group. The first step in proving Theorem 3.1
is showing that every finitely presented group embeds in a group presented by
involutions.
Proposition 4.3. Suppose (G, J) is a group over Z2 with finite presentation
hS : Ri, and J ′ ∈ F (S) is a representative of JG . Let T be the set of indeterminates {zs1 , zs2 : s ∈ S}, and choose integers ks ≥ 1 for all s ∈ S. Finally,
let φ : F (S) → F2 (T ) × Z2 be the morphism sending s 7→ (zs1 zs2 )ks . Then the
induced morphism
φ : G → K := InvhT : R′ i, where R′ := {φ(r) : r ∈ R} ∪ {JK φ(J ′ )}
is an embedding over Z2 .
Furthermore, if R ∪ {J ′ } is cyclically reduced then R′ is cyclically reduced.
Proof. Let s1 , . . . , sn be a list of the elements of S, and let mi be the order of
si in G. For convenience, we write zij in place of zsi ,j . Let
2
2
Di := hzi1 , zi2 : zi1
= zi2
= (zi1 zi2 )ksi mi = 1i,
the dihedral group of order ksi mi (if mi is infinite, then the last relation is
omitted, so that Di is the infinite dihedral group). We define an increasing
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
11
sequence of groups G = K0 ⊆ K1 ⊆ · · · ⊆ Kn such that JG is in the centre
Z(Ki ) for all i, as follows: Suppose that Ki−1 is defined. Observe that the
subgroup hsi , JG i ⊆ G is abelian, and hence is isomorphic to either Zmi or
m /2
Zmi × Z2 (where Z∞ := Z), depending on whether or not JG = si i . In the
former case, let Ki be the amalgamated product of Ki−1 and Di , where we
identify si ∈ Ki−1 with (zi1 zi2 )ksi in Di . Since G is a subgroup of Ki−1 , this
amalgamated product is well-defined, and since JG ∈ Z(Ki−1 ) is identified
with (zi1 zi2 )ksi ms /2 ∈ Z(Di ), we also get that JG belongs to the centre of Ki .
When hsi , JG i = Zmi × Z2 , we let Ki be the amalgamated product of Ki−1
and Di × Z2 , where we identify si ∈ Ki−1 with (zi1 zi2 )ksi and JG ∈ Ki−1 with
the generator J of Z2 in Di × Z2 . Once again, the amalgamated product is
well-defined and JG ∈ Z(Ki ).
Now it is not hard to see that Km has presentation
hS ∪ {z11 , z12 , . . . , zm1 , zm2 , J} : R ∪ {zij2 = [zij , J] = 1 : 1 ≤ i ≤ m, j = 1, 2}
∪ {sj = (zi1 zi2 )ksj } ∪ {J = J ′ }i
for all 1 ≤ m ≤ n. For Kn , this presentation is equivalent to the presentation
InvhT : R′ i of the group K, and the isomorphism Kn ∼
= K identifies the
inclusion G ⊆ Kn with the morphism φ : G → K.
Finally, it is easy to see that if r ∈ R is cyclically reduced, then φ(r) is
cyclically reduced. If J ′ is cyclically reduced, then Jφ(J ′ ) is also cyclically
reduced.
Definition 4.4. A relation r = J a s1 · · · sn ∈ F2 (S) × Z2 is odd (resp. even)
if a is odd (resp. even). Equivalently, a relation is odd (resp. even) if it is of
the form r ′ = J (resp. r ′ = 1) for some r ′ ∈ F2 (S).
The even part of the relation r = J a s1 · · · sn is r + = s1 · · · sn . If InvhS : Ri
is a presentation by involutions over Z2 , then the corresponding even presentation over Z2 is InvhS : R+ i, where R+ = {r + : r ∈ R}.
Similarly, if G is any group over Z2 , then the even quotient is G+ :=
G/(JG ) × Z2 . The group G+ is regarded as a group over Z2 with JG+ equal to
the generator of the Z2 factor.
It is easy to see that if G = InvhS : Ri, then G+ = InvhS : R+ i. For
instance, the null solution group Γ(H) = Γ(H, 0) of a hypergraph H is the
even quotient Γ(H, b)+ of the solution group Γ(H, b) for any b.
Definition 4.5. Let J a s1 · · · sn be a reduced word for an element r ∈ F2 (S) ×
Z2 . The multiplicity of s ∈ S in r is
mult(s; r) := |{1 ≤ i ≤ n : si = s}|.
We say that s 6= t ∈ S are adjacent in r if either {s, t} = {si , si+1 } for some
i = 1, . . . , n − 1 or {s, t} = {s1 , sn }.
12
WILLIAM SLOFSTRA
The reason that we introduce numbers ks in Proposition 4.3 is that, for the
proof of Theorem 3.1, we would like to work with relations r where mult(s; r)
is even for all s. In fact, we will be able to handle slightly more general
presentations, which we now define.
Definition 4.6. We say that a presentation InvhS : Ri by involutions over Z2
is collegial if
(a) the presentation is finite and cyclically reduced,
(b) R ∩ {1, J} = R ∩ S = ∅, and
(c) if mult(s; r0 ) is odd for some r0 ∈ R, and t is adjacent to s in some
r1 ∈ R, then mult(t; r ′ ) is even for all r ′ ∈ R.
Remark 4.7. Note that if InvhS : Ri is collegial, then every relation r ∈ R
must have length at least four, i.e. r = J a s1 · · · sn where n ≥ 4. This is
because relations of length zero and one are explicitly excluded by condition
(b), relations of the form s2 , sts, s2 t, or ts2 are not cyclically reduced, and
relations st and str, where s, t, r are distinct, do not satisfy condition (c).
Corollary 4.8. Let G be a finitely presented group over Z2 , with a sequence
of elements w1 , . . . , wn ∈ G such that wi2 = 1 for all 1 ≤ i ≤ n. Then there is
collegial presentation InvhS : Ri, and an embedding φ : G → K := InvhS : Ri
over Z2 such that φ(wi ) ∈ S ⊂ K for all 1 ≤ i ≤ n.
Proof. We can find a cyclically reduced presentation hS0 : R0 i for G in which
JG is a generator, 1 6∈ R0 , and each wi has a representative wi′ ∈ F (S0) \ {1}
(it is always possible to find such a presentation, since if necessary we can
add an extra generator z, along with the relation z = 1, and use this as
a representative of the identity). In particular, this gives us a presentation
where JG is represented by a cyclically reduced non-identity element of F (S0 ),
namely itself.
Applying Proposition 4.3 to this presentation with ks = 2 (or any other even
number) for all s ∈ S0 gives us an embedding φ of G in a finite presentation
InvhT : R′ i, where R′ is cyclically reduced. Since all ks ’s are even, mult(t; φ(r))
is even for every r ∈ F (S0) and t ∈ T , and every relation in R′ has length ≥ 4.
We conclude that InvhT : R′ i is collegial.
Now let
S = T ∪ {w1 , . . . , w n },
where w1 , . . . , w n are new indeterminates, and set
R = R′ ∪ {wi φ(wi′ ) : 1 ≤ i ≤ n},
where φ : F (S0 ) → F2 (T ) × Z2 as in Proposition 4.3. Since wi does not appear
in φ(wi′ ), the relation r = w i φ(wi′ ) is cyclically reduced. Furthermore, none
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
13
of the wi ’s are adjacent, and mult(s; r) is even for all s ∈ T and r ∈ R, so
InvhS : Ri is collegial. But InvhS : Ri is plainly equivalent to InvhT : R′ i, so
the corollary follows.
5. The wagon wheel embedding
Using Corollary 4.8, the proof of Theorem 3.1 reduces to the following:
Theorem 5.1. Let G be a group with a collegial presentation I = InvhS, Ri.
Then there is a hypergraph W := W(I) and vertex labelling b := b(I) such
that S ⊂ E(W), and the resulting map
F (S) × Z2 → Γ(W, b) : s 7→ xs
descends to an embedding G ֒→ Γ(W, b) over Z2 .
Proof of Theorem 3.1 using Theorem 5.1. Let G be a group over Z2 with elements w1 , . . . , wn such that wi2 = 1 for i = 1, . . . , n. By Corollary 4.8, there
is a collegial presentation I := InvhS : Ri and an embedding φ1 : G → K :=
InvhS : Ri over Z2 with φ1 (wi ) ∈ S for all i = 1, . . . , n.
By Theorem 5.1, there is an embedding φ2 : K → Γ(W(I), b(I)) over Z2
with φ2 (s) = xs for all s ∈ S. The composition φ2 ◦ φ1 satisfies the conditions
of Theorem 3.1.
Although we are still very far from being able to prove Theorem 5.1, in this
section we shall describe the hypergraph W(I), which we call the wagon wheel
hypergraph of I. The proof of Theorem 5.1 will be given in Section 12.
The wagon wheel hypergraph can be defined for any (not necessarily collegial) presentation InvhS, Ri. Let R = {r1 , . . . , rm }, let ni be the length of ri ,
and write ri = J ai si1 · · · sini , where sij ∈ S. The wagon wheel hypergraph is a
simple hypergraph W with vertex set
V := {(i, j, k) : 1 ≤ i ≤ m, j ∈ Zni , 1 ≤ k ≤ 3},
and edge set
E := S ⊔ {aij , bij , cij , dij : 1 ≤ i ≤ m, j ∈ Zni }.
P
As a result, if M := ki=1 ni , then W has 3M vertices and 4M + |S| edges.
W has the following incidence relations for every 1 ≤ i ≤ m and 1 ≤ j ≤ ni :
• s ∈ S is incident with (i, j, 1) if and only if sij = s,
• aij is incident with (i, j − 1, 2) and (i, j, 1),
• bij is incident with (i, j, 1) and (i, j, 2),
• cij is incident with (i, j, 2) and (i, j, 3), and
• dij is incident with (i, j − 1, 3) and (i, j, 3).
14
WILLIAM SLOFSTRA
s1
a1
b0
sn i
0, 1
a0
1, 1
b1
0, 2
a2
1, 2
c1
c0
2, 1
b2
d1
−1, 2
d0
c−1
−1, 3
0, 3
1, 3
s2
c2
d2
2, 2
a3
2, 3
3, 3
s3
3, 1
d3
b3
c3
3, 2
Figure 1. The portion of the wagon wheel hypergraph containing vertices Vi and all incident edges. To save space, (i, j, k)
is written as j, k, and sij , aij , . . . are written as sj , aj , . . ..
Note that the only edges incident with vertices
Vi := {(i, j, k) : j ∈ Zni , 1 ≤ k ≤ 3}
are the edges in
Ei := {aij , bij , cij , dij : j ∈ Zni }
and the edges si1 , . . . , sini . Furthermore, all the edges in Ei are incident with
exactly two vertices, both belonging to Vi . The portion of the hypergraph W
incident with Vi is shown in Figure 1. An example of a wagon wheel hypergraph
for a small (non-collegial) presentation is shown in Figure 2.
We also need to define the vertex labelling in Theorem 5.1.
Definition 5.2. A I-labelling of W is a vertex labelling b : V → Z2 such that
|b−1 (1) ∩ Vi | = ai mod 2 for all 1 ≤ i ≤ m.
For Theorem 5.1, we can choose any I-labelling. This is because all Ilabellings are equivalent in the following sense:
Lemma 5.3. Let b and b′ be two I-labellings of W. Then there is an isomorphism Γ(W, b) → Γ(W, b′ ) which sends xs 7→ xs for all s ∈ S.
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
y
z
x
v
15
u
Figure 2. An example of the wagon wheel hypergraph W(I)
when I = Invhx, y, z, u, v : xyxz = xuvu = 1i.
Proof. Suppose H is a hypergraph with incidence matrix A(H) and vertex
(1)
labelling b(0) . Given e ∈ E(H), let b(1) be the vertex labelling with bv =
(0)
bv + A(H)ve (i.e. we toggle the sign of all vertices incident with e according
to multiplicity). Then there is an isomorphism
(
xf
f 6= e
Γ(H, b(0) ) → Γ(H, b(1) ) : xf 7→
.
Jxe f = e
For W, since |b−1 (1) ∩ Vi | and |(b′ )−1 (1) ∩ Vi | have the same parity, it is easy to
see that b|Vi can be transformed to b′ |Vi by toggling signs of vertices incident
to edges e ∈ Ei as necessary. The lemma follows.
6. Pictures for groups generated by involutions
In this section we give an overview of the main technical tool used in the
proof of Theorem 5.1: pictures of groups. These pictures, which are dual to
the somewhat better known van Kampen diagrams, are a standard tool in
combinatorial group theory. The purpose of pictures is to encode derivations
of group identities from a set of starting relations; see [Sho07] for additional
background. Here we introduce a variant adapted to groups generated by
involutions.
6.1. Pictures as planar graphs. By a curve, we shall mean the image of a
piecewise-smooth function γ from a closed interval [a, b] (where a < b) to either
the plane or the sphere. A curve γ is simple if γ(s) 6= γ(t) for all a ≤ s < t ≤ b,
except possibly when s = a and t = b. The points γ(a) and γ(b) are called the
endpoints of the curve. A curve has either one or two endpoints; if γ(a) = γ(b)
16
WILLIAM SLOFSTRA
then the curve is said to be closed. A connected region (in the plane or on the
sphere) is simple if its boundary is a simple closed curve.
Definition 6.1. A picture is a collection (V, E, D), where
(a) D is a closed simple region,
(b) V is a finite collection of points, called vertices, in D,
(c) E is a finite collection of simple curves, called edges, in D, and
(d) for all edges e ∈ E and points p of e,
(i) if e is not closed and p is an endpoint of e, then either p ∈ V , or p
belongs to the boundary of D and is not the endpoint of any other
edge;
(ii) if e is closed and p is an endpoint of e, then p does not belong to
the boundary of D;
(iii) if p is not an endpoint of e, then p 6∈ V , and p does not belong to
any other edge or the boundary of D.
If an edge e contains a vertex v, then we say that e and v are incident. If e
contains a point of the boundary of D, then we say that e is incident with the
boundary. A picture is closed if no edges are incident with the boundary of
D. The size of a G-picture P is the number of vertices in P.
According to this definition, a picture is a type of planar embedding of a
graph, albeit a graph where we can have multiple edges between vertices, loops
at a vertex, and even closed loops which are not incident to any vertex. From
this point of view, the boundary of D can be regarded as a special type of
vertex; if we think of the picture as drawn on a sphere, then this vertex would
naturally be drawn at infinity. An illustration of these two equivalent points
of view is shown in Figure 3. However, it is more convenient not to include the
boundary of D in the vertex set of a picture, and we stick with the convention
of treating the boundary separately. In particular, the picture in Figure 3 has
size 7.
There is one important exception where we want to forget the boundary of
D, and that is when the picture is closed. If this happens, we often want to
think of the picture as embedded in the sphere, without the point at infinity
being marked. An example of a closed picture on the sphere, as seen from
two different positions, is given in Figure 4. To handle this case, we allow
D to be the whole sphere, in which case the boundary is empty. Also note
that we consider two pictures equal if they differ up to isotopy, either in the
plane or on the sphere as appropriate. Such isotopies are allowed to move the
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
17
Figure 3. A picture embedded in a disk (left) and on the plane
with the exterior of the disk shrunk down to a special vertex at
infinity (right).
1
2
1
3
3
4
2
4
Figure 4. A closed picture embedded in the sphere seen (up to
isotopy) with two different choices for the location of the point
at infinity. Faces are distinguished by different colours.
boundary of the simple region, as well as the location of endpoints of edges on
the boundary, as long as endpoints are not identified.
If P is a picture in D0 and D is a closed simple subregion of D0 , then the
portion of P contained in D0 can be interpreted as a picture inside D. We do,
however, have to make sure that boundary edges of the picture inside D do not
have a common endpoint. This leads to two natural notions of the restriction
of P to D.
Definition 6.2. Let P be a picture in D0 , and let D be a closed simple region
in D0 with interior D o . Given ǫ > 0, let D ǫ denote the ǫ-relaxation of D, and
let D −ǫ denote the ǫ-contraction.
Recall that a curve intersects the boundary of D transversally if, in every
small disk around the intersection point, there are points of the curve which
18
WILLIAM SLOFSTRA
=⇒
Figure 5. Pictures can be restricted to a region homotopic to
a disk and bounded by a simple closed curve which is transverse
to the picture.
lie both on the interior and the exterior of D. We say that D is transverse to
P if every edge which intersects the boundary of D does so transversally, and
the boundary of D does not contain any vertices of P.
The restriction of P to a transverse region D is the picture res(P, D) with
vertex set V (P) ∩ D, and whose edges are the closures of the connected components of e ∩ D o , for e ∈ E(P). In other words, edges are cut off at the
boundary, and edges which intersect the boundary at multiple points may be
cut into multiple edges.
For a general region D, the contraction D −ǫ will be transverse to P for small
enough ǫ > 0. The restrictions res(P, D −ǫ ) are thus well-defined, and can be
identified via isotopy with pictures in D. These pictures belong to a single
isotopy class res(P, D), which we identify as the restriction of P to D.
Similarly, the germ of D in P is the isotopy class germ(P, D) of res(P, D ǫ )
for small ǫ > 0.
Unless otherwise noted, we assume that subregions are closed. If D is transverse to P, then res(P, D) and germ(P, D) agree. An example of this type
of restriction is shown in Figure 5. In general, vertices in the boundary of D
will not appear in res(P, D), but are preserved, along with all their outgoing
edges, in germ(P, D). One way to get a simple region is to take a region D
enclosed by a simple cycle in P. An example of this type, in which the germ
is different than the restriction, is given in Figure 6.
Definition 6.3. Let P be a picture in D0 . A simple cycle in P is a collection
of edges whose union is a simple closed curve.
A closed loop is an edge of P which is not incident to any vertex or to the
boundary, and thus forms a simple cycle by itself.
A face of P is an open connected region D of D0 which does not contain any
points of P, and such that the boundary of D is a union of points of P and
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
⇒
19
⇒
Figure 6. The germ of the region enclosed by a simple cycle
is computed by first taking an ǫ-relaxation of the region. In this
case the cycle is facial, so the restriction would be empty.
points in the boundary of D0 . An outer face is a face whose boundary contains
points of the boundary of D0 .
A simple cycle is facial if it is the boundary of a face.
Every simple cycle in the disk bounds a unique simple region (the interior
of the cycle), while a simple cycle on the sphere bounds two simple regions.
A face does not have to be simple, but a facial cycle always bounds a simple
face by definition.
6.2. Groups and labellings of pictures.
Definition 6.4. Let G = InvhS : Ri. A G-picture is a picture P with every
vertex v labelled by a relation r(v) ∈ R and every edge e labelled by a generator
s(e) ∈ S, such that if e1 , . . . , en is the sequence of edges incident to v, read
in counter-clockwise order with multiplicity from some starting point, then
s(e1 )s(e2 ) · · · s(en ) ∈ {r(v)+ }sym .
The boundary of P is the cyclic word bd(P) = s(e1 ) · · · s(en ) over S, where
e1 , . . . , en is the list of edges incident with the boundary, read in counterclockwise order around the boundary of the disc, with multiplicity. If P is
closed then we say that bd(P) = 1, the empty word.
Two pictures P1 and P2 are equivalent if bd(P1 ) = bd(P2 ).
The sign of a picture P is sign(P) = |{v ∈ V (P) : r(v) is odd}| mod 2.
If D is a simple region and P is a G-picture, then res(P, D) and germ(P, D)
both inherit the structure of a G-picture from P via restricting the labelling
functions.
Typically pictures use directed edges to represent inverses of generators,
but this is not necessary for groups generated by involutions. As previously
mentioned, the point of pictures is that they capture relations in the group,
in the following sense:
20
WILLIAM SLOFSTRA
Proposition 6.5 (van Kampen lemma). Let G = InvhS : Ri, let r be a word
over S, and let a ∈ Z2 . Then r = J a in G if and only if there is a G-picture
P with bd(P) = r and sign(P) = a.
The original version of the van Kampen lemma goes back to [VK33]. The
proof of this version is not substantially different than the proof of the original
version, and we omit it.
Example 6.6. A G-picture encodes a specific derivation of a group relation
in the given presentation. Consider the Coxeter group S4 × Z2 over Z2 . This
group has presentation
G = Invhs1 , s2 , s3 : s1 s3 = s3 s1 , s1 s2 s1 = s2 s1 s2 , s2 s3 s2 = s3 s2 s3 i,
and in particular is presented by involutions. An example of a G-picture showing that s1 s2 s3 s2 s1 = s3 s2 s1 s2 s3 is given in Figure 7.
s2
s3
s2
s1
s1
s3
s3
s2
s1
s1
s3
s3
s2
s1
s2
Figure 7. A G-picture for G = S4 × Z2 showing that
s1 s2 s3 s2 s1 = s3 s2 s1 s2 s3 . Vertex labels are omitted, as they can
be deduced from the edge labels.
Definition 6.7. Let D be a simple region of a G-picture P0 , and let P ′ =
res(P, D) or germ(P, D). If P ′ is equivalent to a picture P ′′ , then we can cut
out P ′ and glue in P ′′ in its place to get a new picture P1 . We refer to the
process P0 ⇒ P1 as surgery.
If P ′ has size zero,2 then we call P0 ⇒ P1 a null surgery. We say that P0
and P1 are equivalent via null surgeries if there is a sequence of null surgeries
transforming P0 to P1 .
An example of a surgery is shown in Figure 8.
2Size
zero pictures do not have vertices, but they can still have edges.
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
s3
s2
21
s3
s1
s1
s1
s2
⇒
s3
s1
s3
s1
s1
s2
s2
s2
s2
s1
=⇒
=⇒
s1
s3
s3
s2
s3
s1
s1
s1
s1
s3
s1
s2
s2
s1
s1
s3 s1
s3
s1
Figure 8. Surgery for an S4 × Z2 -picture with boundary relation s1 s2 s3 s3 s1 = s2 s1 s2 . Generators and relations are the same
as in Example 6.6.
7. Pictures over solution groups and hypergraphs
Let H be a hypergraph with vertex labelling function b : V (H) → Z2 .
The solution group Γ(H, b) is finitely presented by involutions over Z2 ; to
talk about Γ-pictures, we just need to pick a presentation of Γ(H, b). One
candidate is the presentation from Definition 2.1. Q
This presentation contains
two types of relations: linear relations of the form xe = J a , and commuting
relations of the form xe xe′ = xe′ xe . However, it will be more convenient to use
a presentation without commuting relations:
Definition 7.1. As a group presented by involutions over Z2 , we let Γ(H, b) =
InvhS, Ri, where S = {xe : e ∈ E(H)} and
R = {J bv xe1 · · · xen : all v ∈ V and all orderings e1 , . . . , en
of the edges incident to v listed with multiplicity}
Clearly any linear relation in R can be recovered from one linear relation and
the commuting relations, while a commuting relation can be recovered from
two linear relations as shown in Figure 9. Consequently, the relations in R do
22
WILLIAM SLOFSTRA
x1
x1
x2
⇒
x2
x1
x3
x2
x2
x4
x1
Figure 9. In a solution group with linear relation x1 x2 x3 x4 =
J a , the commuting relation x1 x2 = x2 x1 can be represented pictorially in two different ways, depending on whether we use the
presentation from Definition 2.1 (on the left), or the presentation
from Definition 7.1 (on the right).
indeed give a presentation for Γ(H, b). We will always use the presentation in
Definition 7.1 when working with Γ-pictures.
Definition 7.2. Let H be a hypergraph. An H-picture is a picture P with
a pair of labelling functions hV : V (P) → V (H) and hE : E(P) → E(H),
such that for all v ∈ V (P) and e′ ∈ E(H), if we list the edges e1 , . . . , en of P
incident to v with multiplicity then Ah(v)e′ = |{1 ≤ i ≤ n : h(ei ) = e′ }|.
The boundary of an H-picture P is the cyclic word bd(P) = h(e1 ) · · · h(en )
over E(H), where as before e1 , . . . , en is the sequence of edges incident with
the boundary, read counter-clockwise with multiplicity. The character of P is
V (H)
the vector ch(P) ∈ Z2
with ch(P)v = |h−1
V (v)| mod 2.
It is easy to see that there is a one-to-one correspondence between Γ(H, b)pictures and H-pictures. If P is an H-picture, then the sign of the corresponding Γ-picture is the standard dot product ch(P) · b. The leads to the following
restatement of the van Kampen lemma for H-pictures.
Proposition 7.3 (van Kampen lemma). Let Γ(H, b) be a solution group. Then
xe1 · · · xen = J a in Γ(H, b) if and only if there is an H-picture P with bd(P) =
e1 · · · en and ch(P) · b = a.
Example 7.4. Consider the solution group for the linear system
(1)
x+y+z =1
(2)
x+y+z =0
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
23
The underlying hypergraph H of this system is shown below.
x
1
y
2
z
This drawing of the hypergraph is also a closed H-picture P with character
ch(P) = (1, 1). Since b = (1, 0) and ch(P) · b = 1, van Kampen’s lemma tells
us that J = 1.
Remark 7.5. Given a hypergraph H with incidence matrix A, let X ⊂ ZV2 be
the set of vertex labellings b such that J 6= 1 in Γ(H, b), and let Y ⊂ ZV2 be the
set of characters ch(P) of H-pictures P. It is not hard to see that X and Y
are subspaces of ZV2 . Proposition 7.3 states that Y is the orthogonal subspace
to X with respect to the standard bilinear product. By Theorem 2.2, X can be
regarded as a quantum-information-theoretic analogue of the columnspace of
A. From this point of view, Y is then an analogue of the left nullspace of A.
If edge e is incident to vertex v in an H-picture P, then h(e) will be incident
to h(v) in H. Consequently, the labelling function h in Definition 7.2 can be
seen as a type of weak hypergraph homomorphism. If P is closed, and H and
P are simple loopless graphs, then h will be an actual graph homomorphism if
and only if h(v) 6= h(v ′ ) for all adjacent vertices v and v ′ in P. Furthermore,
if this happens then h must be a planar graph cover. This motivates the
following definition:
Definition 7.6. Let H be a hypergraph. A closed H-picture P is a cover of
H if every edge of P is incident with two distinct vertices v and v ′ such that
h(v) 6= h(v ′ ).
Size, equivalence, restriction to a region, surgery, and null surgery are all
defined for H-pictures via the correspondence with Γ-pictures. For instance,
the size of an H-picture is simply the number of vertices in the picture.
Definition 7.7. Let H be a hypergraph, and let b : V (H) → Z2 be a vertexlabelling function. Two H-pictures P0 and P1 are b-equivalent (resp. characterequivalent) if bd(P0 ) = bd(P1 ) and ch(P0 ) · b = ch(P1 ) · b (resp. ch(P0 ) =
ch(P1 )).
An H-picture P is b-minimal (resp. character-minimal) if P has minimum
size among all b-equivalent (resp. character-equivalent) pictures.
Two pictures are character-equivalent if and only if they are b-equivalent
for all vertex-labelling functions b. Thus a b-minimal picture is also characterminimal.
24
WILLIAM SLOFSTRA
If P0 ⇒ P1 is a surgery in which a region P ′ is replaced by a b-equivalent
(resp. character equivalent) region P ′′ , then P0 and P1 will be b-equivalent
(resp. character-equivalent). As a result, if P is b-minimal (resp. characterminimal) then res(P, D) and germ(P, D) will be b-minimal (resp. characterminimal) for all simple regions D.
We can also remove closed loops without changing the character-equivalence
class:
Lemma 7.8. Suppose P is a picture, and let P ′ be the same picture but with
all closed loops deleted. Then P and P ′ are character-equivalent and have the
same size.
8. A category of hypergraphs
Definition 8.1. Let H = (V, E, A) be a hypergraph. A subhypergraph of H
is a hypergraph H′ = (V ′ , E ′ , A′ ) with V ′ ⊂ V , E ′ ⊂ E, and A′ve = Ave for all
v ∈ V ′ and e ∈ E ′ .
In other words, a subhypergraph is simply a subset of V (H) ∪ E(H). Although this definition is substantially less restrictive than other notions of
subhypergraphs in the literature, it is natural in the context of hypergraphs
with isolated edges.
Definition 8.2. If H′ is a subhypergraph of H, then the neighbourhood N (H′ )
of H′ is the subhypergraph with V (N (H′)) = V (H′ ), and
E(N (H′ )) = E(H′ )∪{e ∈ E(H) : e is incident in H to some vertex v ∈ V (H′ )}.
We say that H′ is open if H′ = N (H′).
The proof of the following proposition is elementary, and we omit it.
Proposition 8.3. Let H be a hypergraph. The collection of open subhypergraphs of H forms a topology on V (H) ∪ E(H). A subhypergraph H′ is closed
in this topology if and only if, for all v ∈ V (H), if v is incident to e ∈ E(H′ )
then v ∈ V (H′ ).
Definition 8.4. Let H1 and H2 be hypergraphs. A generalized morphism
φ : H1 → H2 consists of a pair of morphisms
φV : V (H1 ) → V (H2 ) ∪ {ε} and φE : E(H1 ) → E(H2 ) ∪ {ε},
such that, for all v ∈ V (H1 ),
(1) if φV (v) 6= ε, then
X
′
e∈φ−1
E (e )
for all e′ ∈ E(H2 ), and
A(H1 )ve = A(H2 )φ(v)e′ ,
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
(2) if φV (v) = ε, then
X
25
A(H1 )ve
e∈E(H1 )\φ−1
E (ε)
is even, and φE (e1 ) = φE (e2 ) for all edges e1 , e2 ∈ E(H1 ) \ φ−1
E (ε)
incident to v.
When there is no confusion, we write φ for both φV and φE . The composition
φ2 ◦ φ1 of two morphisms φ1 : H1 → H2 and φ2 : H2 → H3 is defined by setting
φ2 (ε) = ε.
Proposition 8.5.
(a) If v and e are incident in H1 , and φ : H1 → H2 is a generalized
morphism with φ(v) ∈ V (H2 ), φ(e) ∈ E(H2 ), then φ(v) and φ(e) are
incident in H2 .
(b) If φ1 : H1 → H2 and φ2 : H2 → H3 are generalized morphisms, then
φ2 ◦ φ1 is a generalized morphism.
Proof. For part (a), let e′ = φ(e) ∈ E(H2 ). If v and e are incident in H1 , and
φ(v) 6= ε, then A(H2 )φ(v)e′ ≥ A(H1 )ve > 0. So e′ and φ(v) are incident.
For part (b), let φ = φ2 ◦ φ1 . If φ(v) 6= ε, then φ1 (v) 6= ε, so
X
X
X
A(H1 )ve
A(H1 )ve =
e∈φ−1 (e′ )
−1 ′′
′
e′′ ∈φ−1
2 (e ) e∈φ1 (e )
=
X
A(H2 )φ1 (v)e′′ = A(H3 )φ(v)e′
′
e′′ ∈φ−1
2 (e )
for all e′ ∈ E(H3 ).
Next, suppose φ(v) = ε and φ1 (v) 6= ε. If e1 , e2 ∈ E(H1 ) \ φ−1 (ε) are both
incident to v, then φ(e1 ) and φ(e2 ) belong to E(H2 ) \ φ−1
2 (ε) and are incident
to φ1 (v) by part (a). Thus φ(e1 ) = φ(e2 ). Similarly,
X
X
X
A(H1 )ve =
A(H1 )ve
e∈E(H1 )\φ−1 (ε)
−1 ′
e′ ∈E(H2 )\φ−1
2 (ε) e∈φ1 (e )
=
X
A(H2 )φ1 (v)e′
e′ ∈E(H2 )\φ−1
2 (ε)
is even, since φ2 (φ1 (v)) = ε.
Finally, suppose φ(v) = φ1 (v) = ε. If E(H1 ) \ φ−1 (ε) does not contain
any edges incident with v, then part (2) of Definition 8.4 is trivially satisfied.
Suppose on the other hand that E(H1 ) \ φ−1 (ε) contains edges incident with
v. All such edges belong to E(H1 ) \ φ−1
1 (ε), and hence are sent by φ1 to
26
WILLIAM SLOFSTRA
i
a
1
2
c
e
b
3′
d
g
4
f
5
a′
1′
3
=⇒
e′
g′
c′
h
6
j
5′
7
j′
6′
k′
7′
k
8
Figure 10. A generalized morphism φ between two hypergraphs. Hyperedges are drawn as shaded regions. The morphism
sends an edge or vertex x to x′ , with the following exceptions:
φ(i) = φ(2) = φ(4) = ε, φ(b) = a′ , φ(d) = φ(f ) = φ(h) = c′ , and
φ(8) = 6′ .
some common edge e′ ∈ E(H2 ), where φ2 (e′ ) 6= ε. Conversely, any edge
′
of E(H1 ) \ φ−1
1 (ε) incident with v is sent by φ1 to e , and hence belongs to
E(H1 ) \ φ−1 (ε). We conclude that all edges of E(H1 ) \ φ−1 (ε) incident with v
are sent to φ2 (e′ ), and that
X
X
A(H1 )ve
A(H1 )ve =
e∈E(H1 )\φ−1 (ε)
e∈E(H1 )\φ−1
1 (ε)
is even. Consequently φ is a generalized morphism.
Example 8.6. If e is a hyperedge in a hypergraph H, we can delete e to
get a new hypergraph H \ e. There is a generalized morphism H → H \ e
which sends e 7→ ε. Similarly, we construct other generalized morphisms by
identifying edges, deleting isolated vertices, and collapsing vertices incident to
an even number of edges (deleting a vertex and identifying all incident edges).
An example of a series of these operations is shown in Figure 10.
Another way we can construct generalized morphisms is through subhypergraphs.
Proposition 8.7. Let H be a hypergraph.
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
27
(a) If H′ is closed subhypergraph, then the function
(
x x ∈ V (H′ ) ∪ E(H′ )
r : V (H) ∪ E(H) → V (H′ ) ∪ E(H′ ) ∪ {ε} : x 7→
ε otherwise
is a generalized morphism H → H′ .
(b) If H′ is an open subhypergraph, then the inclusion map
ι : V (H′ ) ∪ E(H′ ) → V (H) ∪ E(H) : x 7→ x
is a generalized morphism H′ → H.
Proof. If H′ is any subhypergraph, then part (1) of Definition 8.4 holds for r
and part (2) holds vacuously for ι. If H′ is closed and r(v) = ǫ, then r(e) = ǫ
for all edges e incident to v, so part (2) of Definition 8.4 holds for r. If H′ is
open, v ∈ V (H′ ), and e′ ∈ E(H) is incident to v, then e′ belongs to H′ , and
hence part (1) holds for ι.
Definition 8.8. A subhypergraph H′ of H is a retract if there is a generalized
morphism r : H → H′ such that r|H′ is the identity.
Part (a) of Proposition 8.7 shows that every closed subhypergraph is a retract.
Morphisms can also be constructed by gluing together morphisms over open
subhypergraphs.
Proposition
8.9. Let {Hi }i∈I be a family of open subhypergraphs of H such
S
that Hi = H, and let {φi }i∈I be a family of generalized morphisms φi :
Hi → H′ such that φi |Hi ∩Hj = φj |Hi ∩Hj . Then there is a unique generalized
morphism φ : H → H′ such that φ|Hi = φi .
Proof. Clearly φ is uniquely defined as a function. Given v ∈ V (H), find Hi
with v ∈ V (Hi ), so φ(v) = φi (v). Since Hi is open, if Ave (H) > 0 then e ∈ Hi
and A(Hi )ve = A(H)ve . Consequently, if φ(v) 6= ε then
X
X
X
A(Hi )ve = A(H′ )φ(v)e′
A(H)ve =
A(Hi )ve =
e∈φ−1 (e′ )
e∈φ−1 (e′ )
e∈E(Hi )
′
e∈φ−1
i (e )
since φi is a morphism. Similarly, if φ(v) = ε, then
X
X
A(Hi )ve =
A(H)ve =
e∈E(H)\φ−1 (ε)
e∈E(Hi )\φ−1 (ε)
X
A(Hi )ve
e∈E(Hi )\φ−1
i (ε)
is even, and φ(e1 ) = φi (e1 ) = φi (e2 ) = φ(e2 ) for all edges e1 , e2 ∈ E(H)\φ−1 (ε)
incident to v.
The functor from hypergraphs to null solution groups is natural with respect
to generalized morphisms.
28
WILLIAM SLOFSTRA
Proposition 8.10. Let φ : H1 → H2 be a generalized morphism. Then there
is a morphism φ : Γ(H1 ) → Γ(H2 ) defined by
(
1
φE (e) = ε
φ(xe ) =
.
xφ(e) otherwise
Proof. The morphism, if it exists, will be uniquely determined by the values
φ(xe ), e ∈ E(H1 ). To show that the morphism is well-defined, we need to
show that
n
Y
φ(xei ) = 1
(8.1)
i=1
for every vertex v of H1 and ordering e1 , . . . , en of the edges incident to v,
listed with multiplicity. Suppose that φ(v) = ε. If φ(ei ) = ε for all 1 ≤ i ≤ n,
then φ(xei ) = 1 for all i, so Equation (8.1) holds. If φ(ei ) 6= ε for some i, then
n
Y
Y
φ(xej ) =
j=1
xφ(ej ) = xK
φ(ei ) = 1,
φ(ej )6=ε
P
where the last equality holds because K = e∈E(H1 )\φ−1 (ε) Ave is even. Hence
E
Equation (8.1) holds in this case as well.
If φ(v) 6= ε, then for all 1 ≤ i ≤ n, either φ(ei ) = ε or φ(ei ) is incident to
φ(v). We conclude that φ(xe1 ), . . . , φ(xen ) commute in Γ(H2 ). Consequently
n
Y
i=1
φ(xei ) =
Y
Y
φ(xe )Ave =
e′ ∈E(H2 ) e∈φ−1 (e′ )
E
Y
A
′
xe′φ(v)e = 1.
e′ ∈E(H2 )
We conclude that the morphism φ is well-defined.
As a consequence, open retracts are special:
Corollary 8.11. If H′ is an open subhypergraph of H and a retract of H then
Γ(H′ ) is a (semidirect factor) subgroup of Γ(H).
Proof. Let r be the retraction morphism H → H′ , and let ι be the inclusion
H′ → H. Then r ◦ ι is the identity on H′ , so the composition r ◦ ι : Γ(H′ ) →
Γ(H) → Γ(H′ ) is the identity morphism, and consequently ι : Γ(H′ ) → Γ(H)
must be injective. It also follows immediately that Γ(H) = N ⋊ Γ(H′ ), where
N is the kernel of r : Γ(H) → Γ(H′ ).
Example 8.12. Consider the graph G of a cube, shown on the left in Figure 11, with vertices numbered for reference. The open neighbourhood of
{1, 2, 3, 4} is the subhypergraph H with vertex set {1, 2, 3, 4} and edge set
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
5
5
6
8
1
29
7
1
2
4
3
6
8
7
2
4
9
3
Figure 11. In the hypergraph of the cube (left), the open
neighbourhood of the bottom face is a retract. In the hypergraph on the right, this is no longer the case.
{12, 23, 34, 14, 15, 26, 37, 48}. The morphism G → H which is the identity
on H, and sends
5 7→ 1, 6 7→ 2, 7 7→ 3, 8 7→ 4, and
56 7→ 12, 67 7→ 23, 78 7→ 34, 58 7→ 14,
is a retract of G onto H.
On the other hand, if we subdivide the edge 12 with a vertex as shown on
the right of Figure 11, then the open neighbourhood H of {1, 2, 3, 4, 9} is no
longer a retract of G, since any retract must send 6 7→ 2 and 7 7→ 3, but 23
is not an edge of H. However, the open neighbourhood of {5, 6, 7, 8} is still a
retract, since we can define the morphism as above but with edges 29 and 39
sent to 67, and vertex 9 sent to ε.
Finally, we can also apply generalized morphisms to pictures.
Proposition 8.13. Let φ : H1 → H2 be a generalized morphism, and let P be
an H1 -picture with bd(P) = e1 · · · en . Construct a new picture P ′ as follows:
(1) If e is an edge of P such that φ(h(e)) = ε, then delete e from P.
(2) For all remaining edges e of P, change the label from h(e) ∈ E(H1 ) to
φ(h(e)) ∈ E(H2 ).
(3) If v is a vertex of P such that φ(h(v)) = ε, then at this point there is
an even number of edges incident with v, and all have the same label.
Delete v, and connect up the remaining incident edges so that no pair
of edges cross.
(4) For all remaining vertices v of P, change the label from h(v) ∈ E(H1 )
to φ(h(v)) ∈ E(H2 ).
30
WILLIAM SLOFSTRA
g
b
i
3
2
d
c
a
7
f
4
′
a
5
1
j
5
f
=⇒
j
k′
c
c′
′
5′
1′
6
8
7′
3′
h
e
e
g′
k
j′
6′
e′
k
6′
e′
k′
j′
5′
c′
Figure 12. The generalized morphism from Figure 10 applied
to a picture. A choice must be made in regards to the edges
formerly incident to vertex 4. We have chosen to connect 1′ − 3′
and 5′ − 7′ instead of 3′ − 7′ and 1′ − 5′ .
Then P ′ is an H2 -picture of size less than or equal to the size of P, with
bd(P ′ ) = φ(e1 ) · · · φ(en ), where ε is regarded as the empty word.
Proof. To show that this picture is an H2 -picture, we need to show that there
are A(H2 )v′ e′ edges labelled by e′ incident to any vertex labelled by v ′ in P ′ .
But this follows immediately from the construction and part (1) of Definition
8.4. Since the construction does not add any vertices, the size of P ′ must be
at most the size of P.
An example is given in Figure 12. Since step (3) in Proposition 8.13 requires
choosing a matching on edges, the graph P ′ we end up with is not unique.
One exception is when φ is one of the morphisms r or ι from Proposition
8.7; in this case, there is always a unique choice, and hence P ′ is uniquely
determined. However, in general the different choices can defer by a sequence
of null-surgeries.
Definition 8.14. If φ : H1 → H2 is a generalized morphism and P is an
H1 -picture, we use φ(P) to denote either the null-surgery equivalence class of
pictures constructed in Proposition 8.13, or some arbitrarily chosen representative of this class.
If r : H → H′ is the retraction morphism onto a closed subhypergraph, and
P is a H-picture, then we also denote r(P) by P[H′ ].
If ι is the inclusion of an open subhypergraph H′ in H, and P is an H′ picture, then P and ι(P) are essentially identical. In this case, Proposition
8.13 states the obvious fact that every picture over H′ can be regarded as a
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
31
picture over H. Note that if such a picture is character (or b)-minimal as an
H-picture, then it is also minimal as an H′ -picture; however, the converse is
not true.
9. Cycles and outer faces
In this section, we lay the foundation for the proof of Theorem 5.1 by looking
at the interaction between pictures and cycles in hypergraphs.
Definition 9.1. A cycle in a hypergraph H is a closed subhypergraph C which is
a simple connected 2-regular graph. A cycle C is cubic if every vertex v ∈ V (C)
has degree three in H.
A C-cycle in an H-picture is a simple cycle C such that every edge of C is
labelled by an edge of C.
Lemma 9.2. Let C be a cycle in a hypergraph H, and suppose P is an Hpicture such that the edges of C do not appear in bd(P). Then every connected
component of P[C] is a C-cycle.
Proof. P[C] is a closed C-picture in which every vertex has degree two. As a
result, every connected component will be a simple closed curve.
In several upcoming proofs, we will use the following measure of the complexity of P.
Definition 9.3. If C is a cycle in H, and P is a H-picture, then we let
#Cycle(P, C) denote the number of C-cycles in P. If Φ is a collection of
cycles, we let
X
#Cycle(P, Φ) =
#Cycle(P, C).
C∈Φ
If we include incident vertices, then a C-cycle C is a closed C-picture, and
thus the definition of cover from Definition 7.6 applies. If C is both facial and
a cover of C, then we say that C is a facial cover. Note that closed loops are
not covers.
Proposition 9.4. If C is a cubic cycle in H, and P is a character-minimal
H-picture with no closed loops, then every facial C-cycle in P is a facial cover.
Proof. Suppose that C is a facial C-cycle. Since C is not a loop at a vertex,
and P has no closed loops, every edge of C must be incident to two distinct
vertices of P. Suppose that C has an edge e connecting two vertices u0 and
u1 with h(u0 ) = v = h(u1 ). By hypothesis, v has degree 3 in H, and thus
is incident with e′ = h(e), another edge f ′ of C, and an edge g ′ which is not
in C. Hence each vertex ui is incident with an edge fi belonging to C with
h(fi ) = f ′ , and an edge gi not in C with h(gi ) = g ′ . Note that the edges f0
32
WILLIAM SLOFSTRA
and f1 could be equal, as could g0 and g1 . Since C is facial, g0 and g1 must
lie on the same side of C. Thus there is a surgery which removes u0 and u1 ,
connecting g0 with g1 and f0 with f1 , as shown below:
g′
f′
v
g′
e′
v
⇒
f′
g′
f′
But this means that P is not character-minimal.
A connected closed cover C of a cycle C is determined up to isotopy by an
orientation of C, and the ply, that is, the size of h−1 (v) for any v ∈ V (C),
where h : C → C is the labelling function.
Definition 9.5. A C-cycle C is a copy of C if the labelling function h : C → C
is a graph isomorphism, or equivalently if |h−1 (v)| = 1 for all v ∈ V (C).
We are also interested in how different cycles interact.
Lemma 9.6. Let C be a cycle in H, and let D be a simple region in a picture
P. If C is a facial C-cycle in P, then either C is contained in D, or the edges
of germ(C, D) form the boundaries of simple outer faces of germ(P, D).
Proof. Suppose C is not contained in D. If germ(C, D) is empty, then the
lemma is vacuously true. Otherwise, choose ǫ > 0 such that germ(P, D) =
res(P, D ǫ ), and in particular such that C intersects the boundary of D ǫ transversally. Then C ∩ D ǫ is divided into a number of segments which start and end
on the boundary of D ǫ . Thus every face of germ(C, D) will be a simple region
whose boundary contains some (possibly disconnected) portion of the boundary of D ǫ . Each edge of germ(C, D) will be incident to exactly two such faces,
and one of these two faces will be contained in F ∩ D ǫ , where F is a face in P
bounded by C. Since F is a face, any face of germ(C, D) contained in F ∩ D ǫ
will be a face of germ(P, D).
Lemma 9.7. Let C be a simple cycle in a picture P, let D be a simple region
bounded by C, and let F be a simple outer face of germ(P, D). Then the edges
of germ(P, D) in the boundary of F form a single simple path P with at least
two edges.
Furthermore, if we write the edges of P in order as e1 , . . . , en , then e1 and
en are incident with the boundary of D, and e2 , . . . , en−1 belong to C.
Proof. Every edge of bd(germ(P, D)) is incident to a vertex in C. If F is
simple, then bd(germ(P, D)) must contain at least two edges. If these edges
are incident with two or more vertices of C, then it follows from the definition
of germ(P, D) that every outer face is a simple region with boundary of the
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
33
Figure 13. A picture in the disk, with four outer faces, two of
which (shaded) form outer quadrilaterals.
required form. If all the edges of bd(germ(P, D)) are incident with a single
vertex in C, then all but one of the outer faces is simple, and the simple outer
faces are bounded by a path formed by two edges of bd(germ(P, D)).
Definition 9.8. Let P be a picture in D0 , where D0 is not the whole sphere.
An outer quadrilateral of P is the closure of a simple outer face F of P, whose
boundary contains exactly three edges of P, together forming a path of length
three.
An outer quadrilateral must have a fourth side consisting of points in the
boundary of D0 , hence the name. A picture with highlighted outer quadrilaterals is shown in Figure 13.
Proposition 9.9. Let C be a cycle in H, and let C ′ be a cubic cycle such that
|E(C) ∩ E(C ′ )| ≤ 1. If D is a region bounded by a C ′ -cycle C ′ , and C is a
facial C-cycle, then either C is contained in D, or the edges of germ(C, D)
form outer quadrilaterals of germ(P, D).
Proof. By Lemma 9.6, we can assume that F is a simple outer face of germ(P, D)
formed by the edges of germ(C, D). By Lemma 9.7, the edges of germ(C, D)
in the boundary of F form a single path e0 , e1 , . . . , en+1 , where e0 and en+1 are
incident with the boundary and e1 , . . . , en belong to C ′ , as shown below:
e0
en+1
e1
e2
en
Since C ′ is cubic, it follows as well that n ≥ 1. Since |E(C) ∩ E(C ′ )| ≤ 1 and
the edges e1 , . . . , en belong to both C and C ′ , we must have h(e1 ) = . . . =
h(en ) = e′ , the unique element of E(C) ∩ E(C ′ ). Since C (and C ′ ) are simple
graphs, the edge e′ is incident with exactly two vertices v and v ′ , and both
Ave′ = Av′ e′ = 1. Thus n must be one, and we conclude that the closure of F
is an outer quadrilateral.
34
WILLIAM SLOFSTRA
f1
f2
2
e1
1
e2
e6
f6
3
6
f3
e3
e5
4
e4
5
f5
f4
Figure 14. The sun of size 6.
10. Pictures over suns
In this section, we look at pictures over a specific family of hypergraphs:
Definition 10.1. The sun of size n, where n ≥ 3, is the hypergraph with vertex
set {1, . . . , n}, edge set {ei , fi : 1 ≤ i ≤ n}, and incidence relation
(
(
1 i=j
1 i∼
= j or i ∼
= j + 1 mod n
Aifj =
and Aiej =
.
0 i 6= j
0 otherwise
Note that a sun has a unique cycle. The sun of size n = 6 is shown in Figure
14.
Proposition 10.2. Let H be a sun, and let P be a character-minimal Hpicture such that bd(P) does not contain any edges from the cycle C of H. Then
P is character-equivalent to a character-minimal picture P ′ with no closed
loops, in which every C-cycle is a facial cover. Furthermore, P ′ can be chosen
so that every outer quadrilateral of P is an outer quadrilateral of P ′ .
The last statement in this proposition implies in particular that if an edge
e is part of an outer quadrilateral in P, then e is also an edge of P ′ , with the
same label. The proof of Proposition 10.2 uses the following lemma.
Lemma 10.3. If P is a closed picture over a sun H, then ch(P) = 0.
Proof. Let n be the size of H, and let U be the closure of the complement of the
cycle C of H, or in other words, the subhypergraph with vertex set {1, . . . , n}
and edge set {f1 , . . . , fn }. Then every vertex in P[U] has degree one, so if P
is closed then P[U] is a matching. Furthermore, if e is an edge of P[U] with
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
35
h(e) = fi , and endpoints at vertices v and v ′ , then h(v) = h(v ′ ) = i. We
conclude that P has an even number of vertices labelled by i, so ch(P)i = 0
for every i.
Proof of Proposition 10.2. Without loss of generality, we can assume that every picture lies in a bounded region, rather than a sphere. As a consequence,
every simple closed curve will bound a unique simple region. Suppose that C
is a C-cycle in some H-picture P, and let D be the simple region bounded by
C. For the purpose of this proof, we let NE(C, P) be the number of edges e
in res(P, D) with h(e) ∈ U, where U is the closure of the complement of C, as
in the proof of Lemma 10.3. We then set
NE(P) :=
X
NE(C, P).
C a C-cycle
Now suppose we start with some character-minimal picture P = P0 with no
edges from C in bd(P0 ). Our strategy will be to reduce NE(P0 ) via a sequence
of surgeries. Let C be a C-cycle in P0 , and let D be the simple region bounded
by C. Suppose e is an edge contained in D and incident with C, such that
h(e) = fi for some i. Clearly e cannot be incident with the boundary of P0 ,
and since e is not a closed loop, we conclude that e is incident with two distinct
vertices v0 and v1 with h(v0 ) = h(v1 ) = i. Let e, a0 , a1 be the sequence of
edges incident with v0 , as they appear in counter-clockwise order, and let e,
b0 , b1 be the edges incident with v1 as they appear in clockwise order. Then
a0 , e, and b0 all lie in the boundary of a common face, as do a1 , e, and b1 (all
these edges may lie in a common non-simple face, so we have to be careful
with wording here). If h(a0 ) = h(b0 ) and h(a1 ) = h(b1 ), then we could delete
e, v0 , and v1 , and connect ai with bi (shown in the diagram below, where we
assume without loss of generality that h(a0 ) = ei )
ei−1
i
ei
⇒
fi
ei−1
i
ei−1
ei
ei
to get a character-equivalent picture of smaller size, contradicting minimality.
We conclude that we must have {h(e), h(aj ), h(bj )} = {ei−1 , ei , fi } for j = 0, 1.
Thus we can perform the surgery where we make vj incident to e, aj , and bj ,
36
WILLIAM SLOFSTRA
for j = 0, 1, (shown below, again with the assumption that h(a0 ) = ei )
ei−1
i
ei
ei−1
⇒
fi
i
ei
i
ei
i
fi
ei
ei−1
ei−1
to get a new picture P1 . Since the surgery did not change the number or
labels of vertices, P1 will be character-minimal and character-equivalent to
P0 . Since bd(P0 ) does not contain any edges from C, the boundary of an outer
quadrilateral of P0 must contain edges e1 , e2 , e3 , where h(e2 ) ∈ C, and e1 and
e3 are the two unique edges of h−1 (U) which are incident with the endpoints
of e2 . Since e is not incident to the boundary, the edges a0 , a1 , b0 , and b1 and
the vertices v0 and v1 do not belong to an outer quadrilateral of P0 , and hence
every outer quadrilateral of P0 will be an outer quadrilateral of P1 .
Recall that e lies in D. Without loss of generality, we can assume that v0
belongs to C. By Lemma 9.2, v1 must also belong to a C-cycle. This creates
two possible outcomes for NE(P1 ). First, suppose that both v0 and v1 lie
on C. Then the surgery will “pinch” the region bounded by C, creating two
C-cycles C0 and C1 connected by e, as shown below.
a1
C
a0
C1
⇒
e
b1
a1
e
b1
b0
C0
a0
b0
Any edge e′ 6= e in D will end up in the region bounded by C0 or the region
bounded by C1 , so NE(C, P0 ) = NE(C1 , P1 ) + NE(C2 , P1 ) + 1.
The other possibility is that v1 lies on a different C-cycle C ′ , where (since
C-cycles cannot cross) C ′ lies in the interior of the region bounded by C. In
this case, the surgery will connect C and C ′ to form a new cycle C ′′ , as shown
below.
a1
C
a0
a1
e
a0
e
b1
b1
b0
b0
⇒
C′
C ′′
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
37
The edge e, along with all the edges inside the cycle C ′ , will end up on the
outside of C ′′ . The only edges remaining in the region bounded by C ′′ are
edges that belonged to D, so NE(C ′′ , P1 ) = NE(C, P0 ) − NE(C ′ , P0 ) − 1. In
both cases, no other cycles are changed by the surgery, so we conclude that
NE(P0 ) > NE(P1 ).
Iterating this procedure, we get a sequence P0 , P1 , P2 , . . . of character-minimal
pictures, all character-equivalent, such that the outer quadrilaterals of Pi are
outer quadrilaterals of Pi+1 , and NE(Pi ) > NE(Pi+1 ). Since NE(Pi ) cannot
decrease indefinitely, this process must terminate at a picture Pn with the
property that if e ∈ h−1 (U) is incident with a C-cycle C, then e is not contained in the region bounded by C. Equivalently, we can say that res(Pn , D)
is closed for every simple region D bounded by a C-cycle.
Let P ′ be the picture Pn with all closed loops deleted. By Lemma 7.8,
′
P is character-minimal and character-equivalent to Pn (and hence P0 ). In
addition, it is easy to see that all outer quadrilaterals of Pn will be outer
quadrilaterals of P ′ , and that res(P ′ , D) will be closed for every simple region
D bounded by a C-cycle in P ′ . Suppose D is a simple region bounded by a Ccycle such that res(P ′ , D) is non-empty. Since P ′ does not contain any closed
loops, res(P ′ , D) must contain a vertex. But since res(P ′ , D) is closed, we must
have ch(res(P ′ , D)) = 0 by Lemma 10.3, and hence we can delete res(P ′ , D)
from P ′ to get a character-equivalent picture of smaller size, contradicting
the minimality of P ′ . We conclude that res(P ′ , D) must be empty for every
simple region bounded by a C-cycle, and hence every C-cycle in P ′ is facial.
By Proposition 9.4, every C-cycle in P ′ is a facial cover, as required.
The following collorary is not needed for the proof of Theorem 5.1, but it
does serve as a good example of how we will apply Proposition 10.2 in the
following section.
Corollary 10.4. Let H be theP
sun of size n, let b : V (H) → Z2 be a vertex labelling function, and set b0 := ni=1 bi . Then the subgroup of Γ = Γ(H, b) generated by S = {xf1 , . . . , xfn } is isomorphic to K = Invhf1 , . . . , fn : f1 · · · fn =
J b0 i.
Proof. Clearly there is an H-picture P with bd(P) = f1 · · · fn and ch(P) =
(1, . . . , 1), so xf1 · · · xfn = J b0 in Γ by Proposition 7.3.
Conversely, suppose that xfi1 · · · xfik = J c holds in Γ. By Proposition 7.3
again, there is an H-picture P with bd(P) = fi1 · · · fik and ch(P) · b = c.
Let P be the set of character-minimal pictures which are character-equivalent
to P, and in which every C-cycle is a facial cover. By Proposition 10.2, P
is non-empty, and since all elements of P have the same number of vertices,
#Cycle(P ′ , C) is bounded for P ′ ∈ P. Let P ′ be an element of P which
maximizes #Cycle(P ′ , C). If C is a C-cycle in P ′ which is not a copy of C,
38
WILLIAM SLOFSTRA
then it is possible to cut C into two C-cycles as shown below, where the interior
of C is a face of P.
C
1
2
1
2
2
1
⇒
2
1
Since this surgery does not change the character or modify any other C-cycle
in P, we get an element of P with more C-cycles than P ′ , a contradiction. We
conclude that every C-cycle in P ′ is a facial copy.
Since C is cubic, each C-cycle in P ′ must bound a unique face, even if P is
a picture in a sphere. Let P ′′ be the picture constructed by contracting each
one of these faces to a vertex, as shown below (for n = 6).
f2
e2
f3
f1
2
e1
1
3
e3
f4
e4
5
f1
e6
6
4
f2
f6
⇒
f3
f6
e5
f5
f4
f5
Since every edge in P ′ labelled by an ei must belong to a C-cycle by Lemma
9.2, every edge remaining in P ′′ is labelled by an fi for some 1 ≤ i ≤ n. If
we label each vertex by the single relation J b0 f1 · · · fn , then P ′′ is a K-picture
with bd(P ′′ ) = fi1 · · · fik and sign(P ′′ ) = c. By Proposition 6.5, the relation
fi1 · · · fik = J c holds in K.
The proof techniques of Proposition 10.2 and Corollary 10.4 are illustrated
by the following example.
Example 10.5. Let H be the sun of size 4, and let K = hf1 , . . . , f4 : f1 f2 f3 f4 =
1i. It is not hard to see that the relation w := f1 f2 f3 f4 (f1 f2 f3 )2 = 1 holds in
K. Given a K-picture with boundary word w, we can replace every vertex
with a C-cycle to get an equivalent H-picture, for instance as shown in Figure 15c. However, there are H-pictures with boundary word w, as shown in
Figure 15a, which do not come from a K-picture in this way. Nonetheless, if
P is a minimal H-picture whose boundary contains only fi ’s, then the proofs
of Proposition 10.2 and Corollary 10.4 provide a method to transform P to
an equivalent picture which does come from a K-picture. The transformation
process is shown in Figure 15.
To finish the section, we prove one more technical lemma:
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
f3
f2
e3
e2
f4
f1
f3
e4
e1
e3
e2
f4
e1
e4
e2
f1
f2
f3
f3
f2
f1
e3
f4
e4
e4
e3
f3
e2
e1
e4
e1
e2
e3
e3
f4
e4
f2
f3
f1
f3
f2
f1
f3
e4
e2
f4
e3
e2
f2
e1
e2
e1
e1
e2
e4
f1
f2
e1
f2
f3
(a) A single C-cycle with an
edge labelled by fi in its interior.
f1
f1
e3
39
e3
f4
e4
f1
f2
e1
(b) We remove the edge in
the interior of the C-cycle using the surgery highlighted in
blue. We now have a picture
with two facial covers of C.
(c) By cutting and reattaching the edges highlighted in
red, we can turn the two facial
covers into three facial copies
of C.
Figure 15. Using surgery, we turn a picture over the sun of
size 4 into an equivalent picture where all C-cycles are facial
copies.
Lemma 10.6. Let P be a character-minimal picture with no closed loops over
the sun H of size n. Then there is no cycle in P with edge labels contained in
{ei , fi , fi+1 } for some fixed 1 ≤ i ≤ n (where fn+1 := f1 ).
Proof. Suppose C is a cycle of this form, and let Ui be the closure of the
subhypergraph {fi , fi+1 }. Since P has no closed loops, C must have at least
one vertex. Since H is simple, C must have at least two vertices, and there
cannot be two consecutive edges of C with the same label. Thus C[Ui ] is a
(non-empty) matching between the vertices of C, where paired vertices must
have the same label (either i or i+1, depending on whether the edge connecting
them is labelled by an fi or fi+1 ). Since every vertex of C occurs in C[Ui ], C
must consist of a sequence of vertices v1 , . . . , v2n , v2n+1 = v1 , such that v2i−1
and v2i are connected by an edge bi with h(bi ) ∈ {fi , fi+1 }, and v2i and v2i+1
are connected by an edge labelled by ei , i = 1, . . . , n. Let ai be the edge
40
WILLIAM SLOFSTRA
incident to vi which is not in C. Since every vertex vi is incident to an edge
of C labelled by ei , we conclude that
(
ei−1 h(vi ) = i
h(ai ) =
,
ei+1 h(vi ) = i + 1
where e0 := en and en+1 := e1 . But since h(v2i−1 ) = h(v2i ), this means that
h(a2i−1 ) = h(a2i ) for all i = 1, . . . , n. Thus if we delete all the vertices and
edges of C, we can connect a2i−1 and a2i along the path previously taken by
bi to get a new H-picture P ′ which is character-equivalent to P. But the
size of P ′ would then be strictly smaller than the size of P, contradicting
minimality.
11. Stellar cycles and constellations
In this section, we extend the results for suns from the previous section
to more complicated hypergraphs through the notion of a constellation. The
result is the “constellation theorem”, which will be the key to the proof of
Theorem 5.1.
Definition 11.1. Let H be a hypergraph with vertex labelling function b :
V (H) → Z. A cycle C in H is b-stellar if
(a) the neighbourhood N (C) is isomorphic to a sun,
(b) N (C) is a retract of H, and
(c) bv = 0 for all v ∈ V (C).
Definition 11.2. Let H be a hypergraph with vertex labelling b : V (H) →
Z2 . A b-constellation is a collection Φ of subhypergraphs of H satisfying the
following properties:
(a) If C ∈ Φ, then C is a cycle, the neighbourhood N (C) is isomorphic to
a sun, and C is either:
(i) b-stellar, or
(ii) a sequence of edges e1 e2 · · · en (in order), n ≥ 3, such that ek
belongs to a b-stellar cycle C ′ ∈ Φ for all 3 ≤ k ≤ n.
(b) For every element C ∈ Φ, either:
(i) there is an edge e in C which does not belong to any cycle in Φ\{C},
or
(ii) there is another cycle C ′ ∈ Φ such that E(C) ∩ E(C ′ ) 6= ∅, and C ′
contains an edge e which does not belong to any cycle in Φ \ {C ′ }.
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
41
5
6
8
1
7
2
9
4
3
Figure 16. A hypergraph based on Figure 11, but where the
open neighbourhood of each face is a sun.
(c) If C0 , C1 ∈ Φ, where C0 6= C1 , then |E(C0 ) ∩ E(C1 )| ≤ 1, and if neither
C0 or C1 is b-stellar, then E(C0 ) ∩ E(C1 ) = ∅.
If Φ is a b-constellation, then a cycle C in a picture P is a Φ-cycle if C is a
C-cycle for some C ∈ Φ.
Roughly speaking, property (a) in Definition 11.2 says that every cycle C in
Φ is either stellar or (mostly) covered by other stellar cycles, while properties
(b) and (c) state that the cycles in Φ do not overlap too much.
Example 11.3. Consider the hypergraph H shown in Figure 16, and let b be
the vertex labelling function with b9 = 1 and bv = 0 for v 6= 9. Let C1 , C2 ,
and C3 be the cycles with vertex sets {1, 2, 5, 6}, {1, 4, 5, 8}, and {3, 4, 7, 8}
respectively, and let C4 be the cycle with vertex set {1, 2, 3, 4, 9}. As in Example
8.12, the neighbourhoods N (Ci ),i = 1, 2, 3, are retracts of H, and hence C1 ,C2 ,
and C3 are b-stellar. The cycle C4 is not a retract, nor is bv = 0 for all
v ∈ V (C4 ), so C4 is not b-stellar. But N (C4 ) is a sun, and the edges 12, 14,
and 34 belong to C1 , C2 , and C3 respectively. Thus Φ = {C1 , . . . , C4 } is a
b-constellation.
The cycle with vertex set {5, 6, 7, 8} is also b-stellar, and could be added to
this constellation, but the cycle with vertex set {2, 3, 6, 7, 9} cannot be added
since it is not b-stellar and shares edges with C4 .
We can now state the main theorem of this section:
Theorem 11.4 (Constellation theorem). Let H be a hypergraph with vertex
labelling b, and let Φ be a b-constellation. Let P be an H-picture such that:
(p.1) bd(P) does not contain any edges from any cycle C ∈ Φ, and
(p.2) either b = 0 or P is closed.
42
WILLIAM SLOFSTRA
Then P is b-equivalent to a picture P ′ such that all Φ-cycles in P ′ are facial
copies.
The rest of this section is concerned with the proof of Theorem 11.4. To
aid the reader, the proof is split into a number of lemmas, which are in turn
grouped into subsections. In all the lemmas, H will be a hypergraph and b will
be a vertex labelling. We refer to the hypotheses (p.1) and (p.2) of Theorem
11.4 as necessary.
11.1. Structure of constellations. We start by proving two lemmas about
constellations.
Lemma 11.5. If Φ is a b-constellation in H, and e is an edge of H, then there
are at most two cycles in Φ containing e.
Proof. Suppose C is an element of Φ containing e. Since N (C) is a sun, C is
cubic. Thus if v is an endpoint of e, then there is a unique edge f incident to
v and not contained in C. If C ′ ∈ Φ \ {C} contains e, then C ′ must also contain
f , since |E(C) ∩ E(C ′ )| ≤ 1. Hence if C ′ and C ′′ ∈ Φ \ {C} both contain e, then
{e, f } ⊂ E(C ′ ) ∩ E(C ′′ ), and consequently C ′ = C ′′ .
Lemma 11.6. Let Φ be a b-constellation in H, and let Φ′ ⊆ Φ. If all elements
of Φ′ are b-stellar or Φ′ contains all b-stellar elements of Φ, then Φ′ is a bconstellation.
Proof. Every subset of Φ satisfies part (c) of Definition 11.2. If all elements
of Φ′ are b-stellar, or Φ′ contains all b-stellar cycles in Φ, then Φ′ also satisfies
part (a). If e is an edge of C ∈ Φ′ such that e does not belong to any element
of Φ \ {C}, then clearly e does not belong to any element of Φ′ \ {C}. Suppose
every edge of C ∈ Φ′ belongs to some element of Φ \ {C}. Then by definition
there is a cycle C ′ ∈ Φ such that E(C) ∩ E(C ′ ) 6= ∅ and C ′ contains an edge
e which does not belong to any cycle in Φ \ {C ′ }. If C ′ belongs to Φ′ , then
we are done. If C ′ does not belong to Φ′ , then let e′ be the unique edge of
E(C) ∩ E(C ′ ). By Lemma 11.5, the only cycles of Φ which contain e′ are C and
C ′ , and hence no element of Φ′ \ {C} contains e′ .
11.2. Stellar cycles. Now we turn to the core of the argument: showing that
C-cycles can be turned into facial covers if C is stellar. The proof relies on
hypothesis (p.2) in the following way:
Lemma 11.7. Suppose C is a b-stellar cycle in H, and P is a b-minimal Hpicture satisfying hypothesis (p.2). Then every C-cycle in P bounds a simple
region D such that ch(germ(P, D)) · b = 0.
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
43
Proof. If b = 0, then the lemma is vacuously true. If P is closed and ch(P)·b =
0, then P is b-equivalent to the empty picture. But since P is b-minimal, in
this case P must have size zero, and again the lemma is vacuously true.
Suppose that P is closed and ch(P) · b = 1. Since P is closed, we can
think of P as a picture in the sphere, in which case every C-cycle bounds two
simple regions D1 and D2 . Every vertex of P appears in one of germ(P, D1 )
or germ(P, D2 ), with only the vertices of C appearing in both. Hence
ch(P) = ch(germ(P, D1 )) + ch(germ(P, D2 )) − ch(C).
Since C is b-stellar, bv = 0 for all v ∈ V (C), so ch(C) · b = 0. Consequently,
ch(germ(P, D1 )) · b + ch(germ(P, D2 )) · b = ch(P) · b = 1
and we conclude that one of ch(germ(P, Di )) · b, i = 1, 2 must be 0.
Lemma 11.8. Let Φ be a b-constellation in which every cycle C ∈ Φ is bstellar. Let P be an H picture satisfying (p.1) and (p.2). Then P is bequivalent to a b-minimal picture P ′ with no closed loops, such that every
Φ-cycle in P ′ is facial.
Proof. The proof is by induction on the size of Φ. If Φ is empty, then the
lemma is true by Definition 7.7 and Lemma 7.8. Suppose the lemma is true
for all b-constellations of size m, where m ≥ 0, and let Φ be a b-constellation of
size m + 1 in which every cycle is b-stellar. If we pick an element C ∈ Φ, then
Φ′ := Φ \ {C} will be a b-constellation of size m by Lemma 11.6, so that every
H-picture satisfying (p.1) and (p.2) (with respect to Φ) will be b-equivalent
to some b-minimal picture P, also satisfying (p.1) and (p.2), such that every
Φ′ -cycle in P is facial. Thus, to show that the lemma holds for Φ, we can
assume that we start with a picture P of this form.
Let NF (P) denote the number of non-facial C-cycles in P, where C ∈ Φ
is the cycle chosen above. Similarly to the proof of Proposition 10.2, our
strategy will be to perform a sequence of surgeries, starting from P0 := P,
each of which decreases NF (P). Suppose that P has a non-facial C-cycle
C. By Lemma 11.7, there is a simple region D bounded by C such that
ch(germ(P, D)) · b = 0. Let f : H → N (C) be a retract onto N (C), so that
b := f (germ(P, D)) is an N (C)-picture. Recall that, since N (C) is open,
P
b can be regarded as an H-picture. By the definition
an N (C)-picture like P
of germ, the labels of edges and vertices in the boundary of the outer faces
of germ(P, D) must belong to N (C). Since f is a retract, we conclude that
the closures of the outer faces of germ(P, D) will be identical to the closures
b and in particular, bd(P)
b = bd(germ(P, D)). The
of the outer faces of P,
construction in Proposition 8.13 relabels or deletes vertices not in N (C), so
b might not be character-equivalent to germ(P, D). But since C is b-stellar,
P
44
WILLIAM SLOFSTRA
b · b = 0. We conclude that
bv = 0 for all v ∈ V (C), and consequently, ch(P)
b is b-equivalent to germ(P, D). Since the size of P
b is less than or equal to
P
the size of germ(P, D), and germ(P, D) is b-minimal, we also conclude that
b is b-minimal as an H-picture. It follows that Pb is b-minimal (and hence
P
character-minimal) as an N (C)-picture. From, again, the definition of germ,
b does not contain any edges from
we know that bd(germ(P, D)) = bd(P))
C. Finally, N (C) is a sun, so we can apply Proposition 10.2 to Pb to get an
b′ such that P
b′ is character-equivalent to and of the same size
N (C)-picture P
b every C-cycle in P
b′ is facial, every outer quadrilateral of P
b is an outer
as P,
′
′
′
b , and P
b has no closed loops. In particular, P
b is b-minimal
quadrilateral of P
and b-equivalent to germ(P, D), and every outer quadrilateral of germ(P, D)
b′ .
is an outer quadrilateral of P
b′ . Clearly P1 is bLet P1 be the result of replacing germ(P, D) with P
′
′
minimal and b-equivalent to P. Suppose C is a C -cycle in P1 , where C ′ ∈ Φ′ .
Now bd(P) = bd(P1 ) does not contain any edges from the cycles of Φ, so by
Lemma 9.2, every edge of P which is labelled by an edge of C ′ belongs to a
unique C ′ -cycle, and the same holds for P1 . If C ′ contains a boundary edge e of
b′ , then there is a C ′ -cycle C ′′ in P which also contains e. By hypothesis, C ′′
P
is facial, and hence by Proposition 9.9, the edges of germ(C ′′ , D) form outer
quadrilaterals in germ(P, D). But since the outer quadrilaterals of germ(P, D)
b′ , the edges of C ′′ are unchanged in P1 . Since
are also outer quadrilaterals of P
e belongs to a unique C ′ -cycle in P1 , we must have C ′ = C ′′ , and hence C ′
is facial in P1 . On the other hand, if C ′ does not contain a boundary edge
b′ , then C ′ either does not intersect D, or lies entirely inside of D. In
of P
the former case, C ′ will also be a facial C ′ -cycle in P, and will remain facial
in P1 . In the latter case, C ′ would have to consist only of edges labelled by
E ′ = E(C ′ ) ∩ E(N (C)). Since N (C) is a sun and |E(C ′ ) ∩ E(C)| ≤ 1, the
intersection E ′ is either empty, or is equal to (in the notation of Definition
b′ is character-minimal and has no
10.1) {ei , fi , fi+1 } for some i. But since P
closed loops, Lemma 10.6 implies that Pb′ does not contain any cycles of this
form. We conclude that every C ′ -cycle in P1 remains facial. Finally, every
C-cycle in P1 belongs either to Pb′ or is inherited unchanged from P, with the
consequence that NF (P1 ) < NF (P).
Iterating this procedure, we get a sequence P0 = P, P1 , . . . , Pk of b-minimal
pictures, all b-equivalent, such that all Φ′ -cycles in Pi are facial for 1 ≤ i ≤ k,
and such that all Φ-cycles in Pk are facial. Deleting all closed loops from Pk
finishes the proof.
11.3. Covers versus copies. For the next lemma, we show that if all Φ-cycles
are facial covers, then we can turn Φ-cycles into facial copies.
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
45
f
f
C
a
a
b
e
⇒
b
a
e
b
e
e
a
b
f
f
Figure 17. Surgery in the proof of Lemma 11.9. Edges labelled
by e and f are cut and reconnected to turn three Φ-cycles into
four Φ-cycles. The edges of C are dashed, while the edges of C1
and C2 are dotted. Interiors of cycles are faces in P ′ and P ′′ .
Lemma 11.9. Let Φ be a b-constellation. Suppose that P is an H-picture
satisfying (p.1), and such that all Φ-cycles in P are facial covers. Then P is
character-equivalent to a picture P ′ in which all Φ-cycles are facial copies.
Proof. The proof is similar to the proof of Corollary 10.4. Let P be the set
of pictures which are character-equivalent to P, have the same size as P, and
in which all Φ-cycles are facial covers. Since all elements of P have the same
number of vertices, #Cycle(P ′ , Φ) is bounded across P ′ ∈ P. Let P ′ be an
element of P which maximizes #Cycle(P ′ , Φ).
Suppose that C is a C-cycle in P ′ which is not a copy of C, where C ∈ Φ
has an edge e which does not belong to any other cycle in Φ. Since C is a
cover, there are two distinct edges e1 and e2 in C with h(e1 ) = h(e2 ) = e. As
in the proof of Corollary 10.4, we can cut C at e1 and e2 to get a new picture
P ′′ , character-equivalent to P ′ , in which C has been replaced by two facial
covers. By Lemma 9.2 and the hypothesis on e, the edges e1 and e2 are not
contained in any other Φ-cycle. Thus all other Φ-cycles in P ′ are unchanged
in P ′′ , and P ′′ will be an element of P with #Cycle(P ′ , Φ) < #Cycle(P ′′ , Φ),
a contradiction. We conclude that if C ∈ Φ has an edge which does not belong
to any other cycle in Φ, then all C-cycles in P ′ are facial copies.
Now suppose that C is a C-cycle in P ′ which is not a copy of C, where every
edge of C ∈ Φ belongs to some other cycle of Φ. By Definition 11.2, part (b),
there is another cycle C ′ ∈ Φ such that E(C) ∩ E(C ′ ) = {e}, and C ′ contains
an edge e′ which does not belong to any cycle in Φ \ {C ′ }. As above, there are
46
WILLIAM SLOFSTRA
distinct edges e1 and e2 in C such that h(e1 ) = h(e2 ) = e. By Lemma 9.2, each
edge ei belongs to a unique C ′ -cycle Ci , i = 1, 2. Since C1 and C2 are copies
of C ′ , we must have C1 6= C2 . Let fi be the unique edge in Ci with h(fi ) = e′ .
Thus we can construct a new picture P ′′ by cutting and reconnecting ei and
fi , i = 1, 2, as shown in Figure 17.
By Lemmas 9.2 and 11.5, C and Ci are the only Φ-cycles containing ei , while
Ci is the only Φ-cycle containing fi . Thus all Φ-cycles in P ′′ , aside from those
shown in Figure 17, are Φ-cycles in P ′ , and hence will be facial covers. Because
C is a cover and C1 and C2 are copies, the new cycles created in Figure 17
will be facial covers. We conclude again that P ′′ will be an element of P with
#Cycle(P ′ , Φ) < #Cycle(P ′′ , Φ), a contradiction. Thus every Φ-cycle in P ′ is
a facial copy.
11.4. Cycles covered by other cycles. Finally, we prove two lemmas that
will allow us to handle non-stellar cycles.
Lemma 11.10. Let Φ′ ⊆ Φ be a pair of b-constellations. Suppose that P is
an H-picture satisfying (p.1) with respect to Φ, such that all Φ′ -cycles in P
are facial covers. Let C0 be a connected closed subhypergraph of C ∈ Φ \ Φ′ ,3
such that every edge of C0 is contained in an element of Φ′ . If C is a C-cycle
in P, then:
(a) C[C0 ] is a cover of C0 , and
(b) if C0 is a connected component of C[C0 ], then all edges not contained
in C and incident with a vertex of C0 lie on the same side of C.
Proof. Let e be an edge of C with h(e) ∈ E(C0 ). By hypothesis and Lemma
9.2, e belongs to a C ′ -cycle C ′ , where C ′ ∈ Φ′ . Since C ′ is a facial cover, e has
two distinct endpoints a and b with h(a) 6= h(b). We conclude that C[C0 ] is a
cover of C0 .
Now let x and y be the edges of P incident to a and b respectively, but not
contained in C. Since |E(C) ∩ E(C ′ )| ≤ 1 and C is simple, we conclude that x
and y belong to C ′ . Let D be a simple region bounded by C. By Proposition
9.9, since C ′ is facial the edges x,y, and e either belong to D, or form an outer
quadrilateral in germ(P, D). It follows that x and y lie on the same side of C.
Since C0 is connected, all edges not contained in C and incident to C0 lie on
the same side of C.
Lemma 11.11. Let Φ′ ⊆ Φ be a pair of b-constellations. Suppose that P is
a b-minimal picture with no closed loops satisfying (p.1) and (p.2), such that
every Φ′ -cycle is a facial cover. If every edge of C ∈ Φ is contained in some
element of Φ′ , then every C-cycle in P is a facial cover.
3In
other words, C0 is either equal to C, or a path in C.
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
47
Proof. Let C be a C-cycle. Applying Lemma 11.10 with C0 = C, we get
immediately that C is a cover, and that the edges incident to C all lie on
the same side of C.
Suppose b = 0, and let D be a simple region bounded by C. If all the edges
incident to C lie in D, then germ(P, D) is closed, and ch(germ(P, D)) · b = 0.
Since P is b-minimal, germ(P, D) must be b-minimal, and this implies that
germ(P, D) must have size zero, in contradiction of the fact that germ(P, D)
contains C. Thus all the edges incident to C lie outside the interior of C,
and we conclude that res(P, D) is closed. But once again, this implies that
res(P, D) must have size zero, and since P has no closed loops, this implies
that res(P, D) is empty. We conclude that C is facial.
Now suppose that P is closed. Similarly to the proof of Lemma 11.7, the
fact that P is b-minimal with size greater than zero implies that ch(P) · b = 1.
Now C bounds two simple regions D1 and D2 in the sphere, where we assume
that all edges incident to C are contained in D1 , so res(P, D2 ) is closed. If
ch(res(P, D2 )) · b = 1, then res(P, D2 ) is b-equivalent to P, contradicting the
b-minimality of P. Thus ch(res(P, D2 )) · b = 0, and as above, res(P, D2 ) must
be empty. We conclude again that C is facial.
11.5. Proof of Theorem 11.4. Given a b-constellation Φ, let Φ0 be the set
of b-stellar cycles in Φ, and let Φ1 be the set of cycles C ∈ Φ such that every
edge of C belongs to a cycle in Φ0 . Then Φ0 ⊆ Φ1 , and by Lemma 11.6, Φ0
and Φ1 are b-constellations.
Given a picture P0 satisfying (p.1) and (p.2), Lemma 11.8 states that we
can find a b-equivalent picture P1 which is b-minimal and has no closed loops,
such that every Φ0 -cycle in P1 is facial. By Proposition 9.4, every Φ0 -cycle
in P1 is a facial cover. By Lemma 11.11, every Φ1 -cycle in P1 is also a facial
cover.
This leaves the cycles in Φ \ Φ1 . By definition, any element of Φ \ Φ1 is
non-stellar. If a non-stellar cycle C ∈ Φ shares an edge with another cycle
C ′ ∈ Φ, then C ′ must be b-stellar. Hence Φ \ Φ1 consists of the non-stellar
cycles C ∈ Φ which have an edge e not contained in any element of Φ \ {C}.
For the purpose of this proof, we say that such an edge e ∈ C is independent.
By part (a) of Definition 11.2, every element C ∈ Φ \ Φ1 has either one or two
independent edges. In the latter case, the two edges will be incident with a
common vertex of C.
Suppose that C is a C-cycle in P1 , where C ∈ Φ \ Φ′ . Let C0 be the path
containing the non-independent edges of C, regarded as a closed subhypergraph. By Lemma 11.10, C[C0 ] is a cover of C0 , and since C0 is a path rather
than a cycle, all connected components of C[C0 ] are copies of C0 . If C[C0 ] is
non-empty, we can write C as a sequence C1 C2 · · · C2k−1C2k of paths Ci , where
48
WILLIAM SLOFSTRA
C (2)
C
e
⇒
C1
e
a1
a2
e
C (1)
e
a1
a2
(a) one independent edge.
C (2)
C
e1
e1
a1
g
⇒
C1
a2
e2
e2
b
C
b
(1)
e1
a1
e2
a2
(b) two independent edges.
Figure 18. Surgery to cut apart the cycle C in the proof of
Theorem 11.4.
k ≥ 1, the path C2i−1 is a connected component of C[C0 ] for all i = 1, . . . , k,
and the edges of C2i are labelled by independent edges of C for all i = 1, . . . , k.
Let v1 (resp. v2 ) be the first (resp. last) vertex of C1 , and let f1 (resp. f2 )
be the edge of C2k (resp. C2 ) which is incident to v1 (resp. v2 ). The labels
ai = h(vi ) are the endpoints of the path C0 in C. If C has one independent edge
e, then e will join a1 and a2 , and h(f1 ) = h(f2 ) = e. By Lemma 11.10, every
edge incident to C1 lies on the same side of C, so we can cut and reconnect f1
and f2 as shown in Figure 18a. If C has two independent edges, then we can do
something similar. In this case, there will be an independent edge ei = h(fi )
incident to ai , i = 1, 2, and both edges will be incident with a third vertex b.
Let g be the edge incident to b not in C. Since C is cubic, any cycle containing
g must also contain either e1 or e2 . Since these edges are independent, g is
not contained in any element of Φ. We can then cut f1 and f2 and reconnect
them by adding two vertices labelled by b, and a edge labelled by g, as shown
in Figure 18b.
These surgeries are not as well behaved as those considered previously:
When C has two independent edges, we end up increasing the size, so the
result will no longer be b-minimal. When C has a single independent edge, it
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
49
is possible that f1 = f2 , in which case we create a closed loop. However, in
both cases we split C into two cycles C (1) and C (2) , where C (1) is a facial copy
of C, and C (2) [C0 ] has fewer connected components than C[C0 ]. Furthermore,
we do not change any other C-cycle. And since we only change or add edges
whose labels are not contained in any element of Φ\{C}, we conclude that C is
the only Φ-cycle changed by this surgery. As a result we may repeat this type
of surgery to get a picture P2 (not necessarily b-minimal, and possibly containing closed loops) which is character-equivalent to P1 , and in which every
Φ-cycle is either a facial cover, or labelled only by independent edges.
Let P be the collection of pictures which are character-equivalent to P2 , and
in which every Φ-cycle is either a facial cover or labelled only by independent
edges. Let P3 be an element of P of minimum size, and let P4 be the picture
P3 with all closed loops deleted. Clearly P4 is also an element of P of minimum
size. Suppose P4 has a C-cycle C which is not a facial cover for some C ∈ Φ.
By definition, C is labelled by independent edges of C. Since P4 has no closed
loops and C is simple, C has at least two edges. Consequently C must have
two independent edges, say e1 and e2 . As above, let b be the vertex incident
to both e1 and e2 , and let g be the edge incident to b and not in C. Since
every edge of C is labelled by e1 or e2 , every vertex of C must be labelled by
b. We can now argue similarly to Lemma 10.6: C must consist of a sequence
of edges f1 , . . . , f2k , where h(f2i−1 ) = e1 and h(f2i ) = e2 for all 1 ≤ i ≤ k. Let
v1 , . . . , v2k be the vertices of C in order, so fi has endpoints vi and vi+1 , where
v2k+1 := v1 , and let gi be the edge of P4 incident to vi with h(gi ) = g. Let P5
be the picture P4 with all the edges and vertices of C deleted, and g2i−1 and
g2i joined into a single edge along the path taken by f2i−1 . Since the edges e1 ,
e2 , and g do not belong to any element of Φ \ {C}, this process does not create
or change any other Φ-cycle. Thus P5 ∈ P, in contradiction of the minimality
of P4 . We conclude that every Φ-cycle in P4 is a facial cover, and the theorem
follows from Lemma 11.9.
12. Proof of the embedding theorem
In this section we finish the proof of Theorem 5.1 (and thus complete the
proof of Theorem 3.1). We continue with the notation from Section 5, so in
particular I := InvhS : Ri is a presentation by involutions over Z2 , G is the
group with presentation I, W := W(I) is the corresponding wagon wheel
hypergraph, and R = {r1 , . . . , rm } is an ordered set of relations. Although
we do not yet assume that I is collegial, for convenience we assume that the
length ni of the relation ri is at least 4 for all 1 ≤ i ≤ m (by Remark 4.7,
this assumption holds if I is collegial). In addition, we make the following
definitions:
50
WILLIAM SLOFSTRA
Bi
⇒
Bi
⇒
Bi
Figure 19. To retract N (Wi ) onto N (Bi ), we remove intermediary vertices to get a simplified wagon wheel shape, and then
fold this wagon wheel onto the central cycle. In the example
shown above, ni = 4.
• Let Wi be the closed subhypergraph of W containing vertices Vi and
edges Ei . (The open neighbourhood N (Wi ) is shown in Figure 1.)
• Let Ai , 1 ≤ i ≤ m be the cycle containing edges ai1 , bi1 , ai2 , bi2 , . . . , aini ,
bini .
• Let Bi , 1 ≤ i ≤ m, be the cycle containing edges di1 , . . . , dini .
• Let Cij , 1 ≤ i ≤ m, j ∈ Zni , be the cycle containing edges aij , bij , cij ,
dij , and ci,j−1.
• Let Φ = {Cij : 1 ≤ i ≤ m, j ∈ Zni } ∪ {Bi : 1 ≤ i ≤ m}.
Before we can prove Theorem 5.1, we need some preliminary lemmas.
Lemma 12.1. If b is an I-labelling of W, then there is a well-defined morphism G → Γ(W, b) over Z2 sending s 7→ xs for all s ∈ S.
Proof. There is a well-defined morphism F (S)×Z2 → Γ(W, b) over Z2 sending
s 7→ xs . As can be seen from Figure
P 1, there is a N (Wi )-picture P with
bd(P) = si1 · · · sini and ch(P) · b = v∈Vi bv = ai . By Proposition 7.3, the
relation ri holds in Γ(W, b) for all 1 ≤ i ≤ n.
Lemma 12.2. N (Bi ) is a retract of W for all 1 ≤ i ≤ n.
Proof. Define r : W → N (Bi ) by
i′ 6= i
ε
r((i′ , j, k)) = ε
i = i′ and k = 1 , and
(i, j, 3) i = i′ and k = 2, 3
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
ε
ε
r(e) =
cij
dij
51
e∈S
e ∈ Ei′ with i 6= i′
.
e = cij
e ∈ {aij , bij , dij }
It is clear that r is the identity on N (Bi ). The only vertices mapped to ε which
are incident with edges in E(W) \ r −1 (ε) are the vertices (i, j, 1) for j ∈ Zni .
For these vertices, the incident edges aij and bij are identified as required by
condition (2) of Definition 8.4.
Since r is the identity on N (Bi ), condition (1) of Definition 8.4 holds for
the vertices (i, j, 3), 1 ≤ j ≤ ni . The vertices (i, j, 2) are incident with three
edges of E(W)\r −1 (ε), namely ai,j+1 , cij , and bi,j , and these edges are mapped
to the three edges di,j+1, cij , and di,j incident to (i, j, 3). We conclude that
condition (1) of Definition 8.4 also holds for the vertices (i, j, 2), and hence r
is a generalized morphism.
The map r can be visualized as deleting everything outside of Wi to get a
simplified wagon wheel shape, and then folding this wagon wheel onto N (Bi ).
This is depicted in Figure 19.
Lemma 12.3. Let s = sij for some 1 ≤ i ≤ m and j ∈ Zni . If R is cyclically
reduced, and mult(s; ri′ ) is even for all 1 ≤ i′ ≤ m, then N (Cij ) is a retract of
W.
Proof. We start by showing that N (Cij ) is a retract of N (Wi ). Since Wi
depends only on the cyclic order of si1 · · · sini , we can assume without loss of
generality that j = 1. Suppose that mult(s; ri ) = 2k, and let 1 = j1 < j2 <
· · · < j2k ≤ ni be a list of the indices 1 ≤ l ≤ ni such that sil = s. Since R
is cyclically reduced, ji+1 > ji + 1 for all i = 1, . . . , 2k − 1, and j2k < ni . For
convenience, let
Jr = {1, j2 , . . . , j2k }, Jl = {0, j2 − 1, . . . , j2k − 1},
Jrodd
= {1, j3 , j5 . . . , j2k−1 }, Jlodd = {0, j3 − 1, j5 − 1 . . . , j2k−1 − 1},
Jreven = {j2 , j4 , . . . , j2k },
and Jleven = {j2 − 1, j4 − 1, . . . , j2k − 1}.
These sets represent the indices of vertices and edges on the right and left of
the cycles Cijp . To talk about edges which do not belong to these cycles, we
also define
J r = {j1 + 1, j1 + 2, . . . , j2 − 1, j3 + 1, . . . , j4 − 1, . . . , j2k−1 + 1, . . . , j2k − 1} and
J l = {j2 + 1, j2 + 2, . . . , j3 − 1, j4 + 1, . . . , j5 − 1, . . . , j2k + 1, . . . , ni }.
52
WILLIAM SLOFSTRA
Define qi : N (Wi ) → N (Ci1 ) by
ε
(i, 1, 1)
qi ((i, j, k)) = ε
(i, 0, k)
(i, 1, k)
k
k
k
k
k
= 1 and j 6∈ Jr
= 1 and j ∈ Jr
,
= 2, 3 and j 6∈ Jr ∪ Jl
= 2, 3 and j ∈ Jlodd ∪ Jreven
= 2, 3 and j ∈ Jrodd ∪ Jleven .
ai1 j ∈ Jrodd
b
j ∈ Jreven
i1
qi (aij ) =
,
bi0 j ∈ J l
ai2 j ∈ J r
odd
even
ci0 j ∈ Jl ∪ Jr
qi (cij ) = ci1 j ∈ Jrodd ∪ Jleven ,
ε
j 6∈ Jr ∪ Jl
(
s s′ = s
qi (s′ ) =
.
ε s′ ∈ S \ {s}
ai1
b
i1
qi (bij ) =
bi0
ai2
di1
qi (dij ) = di0
d
i2
j
j
j
j
∈ Jreven
∈ Jrodd
,
∈ Jl
∈ Jr
j ∈ Jr
j ∈ Jl ,
j ∈ Jr
The map qi can be visualized as deleting vertices and edges cij (the spokes of
the wagon wheel) not in the cycles Cijp , and then folding up the cycles Cijp
onto Ci1 , alternating the directions of the folds after each Cijp like a napkin.
The smallest example, when k = 1, is depicted in Figure 20. As in the proof
of Lemma 12.2, it follows that qi is a retract.
Now we look at Wi′ for i′ 6= i. If mult(s; ri′ ) = 0, then define
qi′ : N (Wi′ ) → N (Cij ) : x 7→ ε.
If mult(s; ri′ ) > 0, then find j ′ such that si′ j ′ = s, and let f : N (Wi′ ) →
N (Ci′ j ′ ) be the retract defined above onto N (Ci′ j ′ ). Now N (Ci′ j ′ ) and N (Cij )
are both suns, so there is an isomorphism g : N (Ci′ j ′ ) → N (Cij ) with g(s) = s,
and we let qi′ = g ◦ f .
The morphisms qi′ , 1 ≤ i′ ≤ m all send s′ ∈ S to either ε if s′ 6= s, or to s
if s′ = s. If i′ 6= i′′ , then the open subhypergraphs N (Wi′ ) and N (Wi′′ ) have
no vertices in common. All common edges of N (Wi′ ) and N (Wi′′ ) belong
to S, so qi′ and qi′′ agree on the intersection. Every vertex of W belongs
to some Wi . There may be elements s′ of S which do not appear in any
relation ri′ , and hence do not belong to any N (Wi ); for these edges, we can
add additional morphisms which send s′ 7→ ε. By Proposition 8.9, there is a
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
s
53
s
Ci1
Ci1
s
Ci1
⇒
s
⇒
s
Figure 20. To retract N (Wi ) onto N (Ci1 ), we remove intermediary vertices and edges, and then fold up the remaining cycles
like a napkin. In the example above, ni = 4 and mult(s; ri ) = 2.
In general, if mult(s; ri ) = 2k then we make k folds.
morphism q : W → N (Cij ) which agrees with qi′ on N (Wi′ ), and in particular
is a retract.
Lemma 12.4. Suppose I is collegial. Then there is an I-labelling b of W
such that Φ is a b-constellation.
Proof. Let b be any I-labelling such that
• |b−1 (1) ∩ V (Wi )| ≤ 1 for all 1 ≤ i ≤ m,
• b((i, j, 2)) = b((i, j, 3)) = 0 for all 1 ≤ i ≤ m and j ∈ Zni , and
• if b((i, j, 1)) = 1, then either mult(sij , ri′ ) is odd for some 1 ≤ i′ ≤ m,
or mult(sij ′ , ri′ ) is even for all j ′ ∈ Zni and 1 ≤ i′ ≤ m.
We will show that Φ is a b-constellation. First, we observe that Bi is b-stellar
for all 1 ≤ i ≤ n. Indeed, N (Bi ) is a sun, and by Lemma 12.2, N (Bi ) is a
retract of W. Finally, b|Bi = 0, so Bi is b-stellar. Similarly, Lemma 12.3 implies
that Cij will be b-stellar for 1 ≤ i ≤ m and j ∈ Zni as long as b((i, j, 1)) = 0
and mult(sij ; ri′ ) is even for all 1 ≤ i′ ≤ m.
Suppose Cij is not b-stellar. If mult(sij ; ri′ ) is odd for some 1 ≤ i′ ≤ m,
then by Definition 4.6, mult(si,j+1 ; ri′′ ) and mult(si,j−1 ; ri′′ ) are even for all
1 ≤ i′′ ≤ m. By the definition of b, we must have b((i, j + 1, 1)) = b((i, j −
1, 1)) = 0. Thus Ci,j+1 and Ci,j−1 are b-stellar, so Cij (which consists of edges
aij , bij , cij , dij , ci,j−1) shares edges cij , dij , and ci,j−1 with b-stellar cycles. If
mult(sij ; ri′ ) is even for all 1 ≤ i′ ≤ m, then we must have b((i, j, 1)) = 1. By
the definition of b, this means that mult(sij ′ , ri′ ) is even for all 1 ≤ i′ ≤ m and
j ′ ∈ Zni , and b((i, j + 1, 1)) = b((i, j − 1, 1)) = 0. It follows that Ci,j+1 and
54
WILLIAM SLOFSTRA
Ci,j−1 are b-stellar, and once again Cij will share edges cij , dij , and ci,j−1 with
b-stellar cycles. Thus Φ satisfies condition (a) of Definition 11.2.
The cycle Cij is the only cycle in Φ containing edges aij and bij . Every edge
dij of Bi is also contained in Cij , so Φ satisfies condition (b) of Definition 11.2.
Finally, it is easy to see that |E(Cij ) ∩ E(Di′ )| ≤ 1 for all 1 ≤ i, i′ ≤ m and
1 ≤ i′ ≤ m, and (since ni ≥ 4 for all 1 ≤ i ≤ m) that |E(Cij ) ∩ E(Ci′ j ′ )| ≤ 1
for all distinct (i, j) and (i′ , j ′ ). We showed above that if Cij is not b-stellar,
then Ci,j+1 and Ci,j−1 are b-stellar, so there is no pair of non-b-stellar cycles in
Φ with a common edge. Thus Φ satisfies condition (c) of Definition 11.2.
Lemma 12.5. Let P be a W-picture in which all Φ-cycles are facial copies,
and such that all edges in bd(P) belong to S. Let G be the group with presentation I = InvhS; Ri, and let b be an I-labelling. Then there is a G-picture
P ′ with bd(P ′ ) = bd(P) and sign(P ′ ) = ch(P) · b.
Proof. Let W \ S denote the closed subhypergraph containing all vertices of
W and all edges except those in S. Equivalently, W \ S is the subhypergraph
with connected components Wi , 1 ≤ i ≤ m. Suppose that P0 is a connected
component of P[W \ S]. Since bd(P) does not contain any edges from outside
S, P0 must be closed. Since P0 is connected, there must be some 1 ≤ i ≤ m
such that P0 is a Wi -picture. By Lemma 9.2, P0 contains a Φ-cycle. By
hypothesis, every Φ-cycle in P0 is a facial copy, so every Cij -cycle will contain
an edge labelled by dij . Thus P0 contains an edge labelled by dij for some
j ∈ Zni , and by Lemma 9.2 again, P0 contains a Bi -cycle B. Since Bi ∈ Φ,
B is also a facial copy, and hence consists of edges dˆj , j ∈ Zni , such that
h(dˆj ) = dij . Let ĉj , j ∈ Zni , be the third edge incident to the common
endpoint of dˆj and dˆj+1. Every edge dˆj is contained in a unique Cij -cycle Cj ,
which will consist of edges ĉj , dˆj , ĉj−1, and two additional edges âj and b̂j with
h(âj ) = aij and h(b̂j ) = bij . Let
P0′ = B ∪
[
Cj .
j∈Zni
The common endpoint of âj and b̂j has degree two in P[W \ S], while the
endpoints of ĉj have degree three and are incident with ĉj , dˆj , dˆj+1 and ĉj , âj+1 ,
b̂j respectively. Hence P0′ contains every edge of P[W \ S] incident to a vertex
of P0′ . Thus P0′ is a maximal connected subgraph of P[W \ S], so P0′ = P0 . In
particular, we conclude that |h−1 (v) ∩ V (P0 )| = 1 for all v ∈ V (Wi ). Since b
is an I-labelling, it follows that ch(P0 ) · b = ai .
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
55
Now let D(B) be the closure of the face bounded by B, let D(Cj ) be the
closure of the face bounded by Cj , j ∈ Zni , and let
D(P0 ) := D(B) ∪
[
D(Cj ).
j∈Zni
Clearly P0 is contained in D(P0 ), and conversely every edge or vertex of P
in D(P0 ) belongs to P0 . Since âj and b̂j belong to the boundary of D(Cj )
and are not contained in B or any Cj ′ , j ′ 6= j, we conclude that âj and b̂j
belong to the boundary of D(P0 ). Every other edge of Cj belongs either to
B, to Cj+1, or to Cj−1, so the boundary of D(P0 ) does not contain any other
edges of Cj . Similarly, the boundary does not contain any edges of B, or
any of the vertices (i, j, 3), j ∈ Zni . We conclude that D(P0 ) is bounded by
the Ai -cycle A = â1 b̂1 · · · âni b̂ni . In particular, D(P0 ) is a simple region, and
bd(germ(P, D(P0 ))) = si1 · · · sini or sini · · · si1 depending on the orientation of
A.
Since D(P0 ) contains only edges and vertices of P0 , if P1 is another connected component of P[W \ S], then D(P0 ) and D(P1 ) are completely disjoint.
Thus, as in the proof of Corollary 10.4, we can collapse each region D(P0 ) to
a single vertex labelled by ri = J ai si1 · · · sini to form a G-picture P ′ . Since
the edges of S will be unchanged, bd(P) = bd(P ′ ). Since every vertex of
P belongs to a unique connected component of P[W \ S], we conclude that
sign(P ′ ) = ch(P) · b.
Proof of Theorem 5.1. Let G be the group with presentation InvhS : Ri, and
suppose that I = InvhS : Ri is collegial. Let I + = InvhS : R+ i, and let G+ be
the even quotient of G, as in Definition 4.4. By Lemma 12.4, we can choose
an I-labelling b of W such that Φ is a b-constellation. Every b-stellar cycle
is also 0-stellar, so Φ is also a 0-constellation by Definition 11.2. By Lemma
12.1, there are morphisms φ : G → Γ(W, b) and φ+ : G+ → Γ(W, 0), both
sending s 7→ xs (note that 0 is an I + -labelling).
To show that φ is injective, we start with φ+ . If φ+ (w) = 1 for some w ∈
F2 (S), then by Proposition 7.3 there is a W-picture P with bd(P) = w. By
Theorem 11.4 (b = 0 case), we can choose P so that all Φ-cycles in P are facial
copies. By Lemma 12.5, there is a G+ -picture P ′ such that bd(P ′ ) = bd(P),
so w = 1 in G+ . Since φ+ (J a w) = 1 for a ∈ Z2 and w ∈ F2 (S) if and only if
a = 0 and φ+ (w) = 1, it follows that φ+ is injective.
56
WILLIAM SLOFSTRA
Now there is a commutative diagram
G
φ
Γ(W, b)
q1
G+
q2
φ+
,
Γ(W, 0)
where q1 and q2 are the quotient maps G → G+ = (G/(JG )) × Z2 and
Γ(W, b) → Γ(W, 0) = (Γ(W, b)/(JΓ )) × Z2 by JG and JΓ := JΓ(W,b) respectively. Since JG is central of order ≤ two, we conclude that ker q1 = {1G , JG }.
Since φ+ is injective, if φ(w) = 1 for w ∈ G, then q1 (w) = 1, and consequently
w ∈ {1, JG }. Thus it remains only to show that φ(JG ) = 1 if and only if JG = 1
in G. By definition, φ(JG ) = JΓ , and if JΓ = 1, then by Proposition 7.3 there
is a closed W-picture P with ch(P) · b = 1. Since P is closed, Theorem 11.4
again implies that we can choose P so that all Φ-cycles in P are facial copies.
By Lemma 12.5, there is a closed G-picture P ′ such that sign(P ′ ) = 1, and we
conclude that JG = 1 in G. Thus φ is injective.
References
[AGR82]
Alain Aspect, Philippe Grangier, and Gérard Roger, Experimental Realization
of Einstein-Podolsky-Rosen-Bohm gedankenexperiment : A New Violation of
Bell’s Inequalities, Physical Review Letters 49 (1982), no. 2, 91–94.
[Ark12]
Alex Arkhipov, Extending and Characterizing Quantum Magic Games,
arXiv:1209.3819 [quant-ph].
[Bel64]
J. S. Bell, On the Einstein Podolsky Rosen paradox, Physics 1 (1964), no. 3,
195–200.
[Ber94]
A. J. Berrick, Groups with no nontrivial linear representations, Bull. Austral.
Math. Soc. 50 (1994), 1–11.
[BPA+ 08] Nicolas Brunner, Stefano Pironio, Antonio Acı́n, Nicolas Gisin, André Allan
Méthot, and Valerio Scarani, Testing the Dimension of Hilbert Spaces, Physical
Review Letters 100 (2008), no. 21, 210503–4.
[CHSH69] John F. Clauser, Michael A. Horne, Abner Shimony, and Richard A. Holt,
Proposed Experiment to Test Local Hidden-Variable Theories, Physical Review
Letters 23 (1969), no. 15, 880–884.
[CHTW04] R. Cleve, P. Høyer, B. Toner, and J. Watrous, Consequences and limits of nonlocal strategies, 19th IEEE Annual Conference on Computational Complexity,
2004. Proceedings, 2004, pp. 236–249.
[Cir80]
B. S. Cirel’son, Quantum generalizations of Bell’s inequality, Letters in Mathematical Physics 4 (1980), no. 2, 93–100.
[CLS16]
Richard Cleve, Li Liu, and William Slofstra, Perfect commuting-operator
strategies for linear system games, arXiv:1606.02278.
[CM14]
Richard Cleve and Rajat Mittal, Characterization of Binary Constraint System
Games, Automata, Languages, and Programming, Lecture Notes in Computer
TSIRELSON’S PROBLEM AND AN EMBEDDING THEOREM
57
Science, no. 8572, Springer Berlin Heidelberg, 2014, arXiv:1209.2729, pp. 320–
331.
[DLTW08] Andrew C. Doherty, Yeong-Cherng Liang, Ben Toner, and Stephanie Wehner,
The Quantum Moment Problem and Bounds on Entangled Multi-prover
Games, 2012 IEEE 27th Conference on Computational Complexity (Los Alamitos, CA, USA), vol. 0, IEEE Computer Society, 2008, pp. 199–210.
[DP16]
Kenneth J. Dykema and Vern Paulsen, Synchronous correlation matrices and
Connes’ embedding conjecture, Journal of Mathematical Physics 57 (2016),
no. 1, 015214.
[FC72]
Stuart J. Freedman and John F. Clauser, Experimental Test of Local HiddenVariable Theories, Physical Review Letters 28 (1972), no. 14, 938–941.
[Fri12]
Tobias Fritz, Tsirelson’s problem and Kirchberg’s conjecture, Reviews in Mathematical Physics 24 (2012), no. 05, 1250012.
[Fri13]
, On infinite-dimensional state spaces, Journal of Mathematical Physics
54 (2013), no. 5, 052107.
[HBD+ 15] B. Hensen, H. Bernien, A. E. Dréau, A. Reiserer, N. Kalb, M. S. Blok,
J. Ruitenberg, R. F. L. Vermeulen, R. N. Schouten, C. Abellán, W. Amaya,
V. Pruneri, M. W. Mitchell, M. Markham, D. J. Twitchen, D. Elkouss,
S. Wehner, T. H. Taminiau, and R. Hanson, Loophole-free Bell inequality violation using electron spins separated by 1.3 kilometres, Nature 526 (2015),
no. 7575, 682–686.
[Hig51]
Graham Higman, A Finitely Generated Infinite Simple Group, Journal of the
London Mathematical Society s1-26 (1951), no. 1, 61–64.
[HK64]
Rudolf Haag and Daniel Kastler, An Algebraic Approach to Quantum Field
Theory, Journal of Mathematical Physics 5 (1964), no. 7, 848–861.
[Ji13]
Zhengfeng Ji, Binary Constraint System Games and Locally Commutative Reductions, arXiv:1310.3794 [quant-ph].
[JNP+ 11]
M. Junge, M. Navascués, C. Palazuelos, D. Pérez-Garcı́a, V. B. Scholz, and
R. F. Werner, Connes’ embedding problem and Tsirelson’s problem, Journal of
Mathematical Physics 52 (2011), no. 1, 012102.
[JPPG+ 10] M. Junge, C. Palazuelos, D. Pérez-Garcı́a, I. Villanueva, and M. M. Wolf,
Unbounded Violations of Bipartite Bell Inequalities via Operator Space Theory,
Communications in Mathematical Physics 300 (2010), no. 3, 715–739.
[KKM+ 11] J. Kempe, H. Kobayashi, K. Matsumoto, B. Toner, and T. Vidick, Entangled
Games Are Hard to Approximate, SIAM Journal on Computing 40 (2011),
no. 3, 848–877.
[KV11]
Julia Kempe and Thomas Vidick, Parallel Repetition of Entangled Games, Proceedings of the Forty-third Annual ACM Symposium on Theory of Computing
(New York, NY, USA), STOC ’11, ACM, 2011, pp. 353–362.
[LS77]
R.C. Lyndon and P.E. Schupp, Combinatorial group theory, Classics in Mathematics, Springer, 1977.
[LTW13]
Debbie Leung, Ben Toner, and John Watrous, Coherent state exchange in
multi-prover quantum interactive proof systems, Chicago Journal of Theoretical
Computer Science 19 (2013), no. 1, 1–18.
[Mer90]
N. David Mermin, Simple unified form for the major no-hidden-variables theorems, Physical Review Letters 65 (1990), no. 27, 3373–3376.
, Hidden variables and the two theorems of John Bell, Reviews of Mod[Mer93]
ern Physics 65 (1993), no. 3, 803–815.
58
WILLIAM SLOFSTRA
[MV14]
Laura Manc̆inska and Thomas Vidick, Unbounded Entanglement Can Be
Needed to Achieve the Optimal Success Probability, Automata, Languages, and
Programming, Lecture Notes in Computer Science, no. 8572, Springer Berlin
Heidelberg, 2014, pp. 835–846.
[NCPGV12] M. Navascués, T. Cooney, D. Pérez-Garcı́a, and N. Villanueva, A Physical
Approach to Tsirelson’s Problem, Foundations of Physics 42 (2012), no. 8,
985–995.
[NPA08]
Miguel Navascués, Stefano Pironio, and Antonio Acı́n, A convergent hierarchy
of semidefinite programs characterizing the set of quantum correlations, New
Journal of Physics 10 (2008), no. 7, 073013.
[Oza13]
Narutaka Ozawa, About the Connes embedding conjecture, Japanese Journal of
Mathematics 8 (2013), no. 1, 147–183.
[Per90]
Asher Peres, Incompatible results of quantum measurements, Physics Letters
A 151 (1990), no. 3, 107–108.
[PSS+ 16]
Vern I. Paulsen, Simone Severini, Daniel Stahlke, Ivan G. Todorov, and Andreas Winter, Estimating quantum chromatic numbers, Journal of Functional
Analysis 270 (2016), no. 6, 2188–2222.
[PT15]
Vern I. Paulsen and Ivan G. Todorov, Quantum Chromatic Numbers Via Operator Systems, The Quarterly Journal of Mathematics 66 (2015), no. 2, 677–692.
[RS10]
Neil Robertson and Paul Seymour, Graph minors XXIII. Nash-Williams’ immersion conjecture, Journal of Combinatorial Theory, Series B 100 (2010),
no. 2, 181–205.
[RV15]
Oded Regev and Thomas Vidick, Quantum XOR Games, ACM Trans. Comput.
Theory 7 (2015), no. 4, 15:1–15:43.
[Sho07]
Hamish Short, Diagrams and groups, The Geometry of the Word Problem for
Finitely Generated Groups, Birkhäuser Basel, Basel, 2007.
[Tsi93]
B.S. Tsirelson, Some results and problems on quantum Bell-type inequalities,
Hadronic Journal Supplement 8 (1993), 329–345.
[Tsi06]
B.S. Tsirelson, Bell inequalities and operator algebras, problem statement for website of open problems at TU Braunschweig (2006), available at
http://web.archive.org/web/20090414083019/http://www.imaph.tu-bs.de/qi/problems/33.htm
[VK33]
E. R. Van Kampen, On Some Lemmas in the Theory of Groups, American
Journal of Mathematics 55 (1933), no. 1, 268–273.
Institute for Quantum Computing, University of Waterloo, Canada
E-mail address: [email protected]
| 4math.GR
|
arXiv:1703.02690v1 [stat.ML] 8 Mar 2017
Leveraging Sparsity for Efficient
Submodular Data Summarization
Erik M. Lindgren, Shanshan Wu, Alexandros G. Dimakis
The University of Texas at Austin
Department of Electrical and Computer Engineering
[email protected], [email protected], [email protected]
Abstract
The facility location problem is widely used for summarizing large datasets and
has additional applications in sensor placement, image retrieval, and clustering. One
difficulty of this problem is that submodular optimization algorithms require the calculation of pairwise benefits for all items in the dataset. This is infeasible for large
problems, so recent work proposed to only calculate nearest neighbor benefits. One
limitation is that several strong assumptions were invoked to obtain provable approximation guarantees. In this paper we establish that these extra assumptions are not
necessary—solving the sparsified problem will be almost optimal under the standard
assumptions of the problem. We then analyze a different method of sparsification that
is a better model for methods such as Locality Sensitive Hashing to accelerate the
nearest neighbor computations and extend the use of the problem to a broader family
of similarities. We validate our approach by demonstrating that it rapidly generates
interpretable summaries.
1
Introduction
In this paper we study the facility location problem: we are given sets V of size n, I of size
m and a benefit matrix of nonnegative numbers C ∈ RI×V , where Civ describes the benefit
that element i receives from element v. Our goal is to select a small set A of k columns in
this matrix. Once we have chosen A, element i will get a benefit equal to the best choice out
of the available columns, maxv∈A Civ . The total reward is the sum of the row rewards, so the
optimal choice of columns is the solution of:
X
arg max
max Civ .
(1)
{A⊆V :|A|≤k} i∈I v∈A
A natural application of this problem is in finding a small set of representative images in
a big dataset, where Civ represents the similarity between images i and v. The problem is to
select k images that provide a good coverage of the full dataset, since each one has a close
representative in the chosen set.
1
Throughout this paper we follow the nomenclature common to the submodular optimization for machine learning literature. This problem is also known as the maximization version
of the k-medians problem. A number of recent works have used this problem for selecting
subsets of documents or images from a larger corpus [27, 39], to identify locations to monitor
in order to quickly identify important events in sensor or blog networks [24, 26], as well as
clustering applications [23, 34].
We can naturally interpret Problem 1 as a maximization of a set function F (A) which
takes as an input the selected set of columns and returns the total reward of that set.
Formally, let F (∅) = 0 and for all other sets A ⊆ V define
X
F (A) =
max Civ .
(2)
i∈I
v∈A
The set function F is submodular, since for all j ∈ V and sets A ⊆ B ⊆ V \ {j}, we have
F (A ∪ {j}) − F (A) ≥ F (B ∪ {j}) − F (B), that is, the gain of an element is diminishes
as we add elements. Since the entries of C are nonnegative, F is monotone, since for all
A ⊆ B ⊆ V , we have F (A) ≤ F (B). We also have F normalized, since F (∅) = 0.
The facility location problem is NP-Hard, so we consider approximation algorithms. Like
all monotone and normalized submodular functions, the greedy algorithm guarantees a (1 −
1/e)-factor approximation to the optimal solution [35]. The greedy algorithm starts with the
empty set, then for k iterations adds the element with the largest reward. This approximation
is the best possible—the maximum coverage problem is an instance of the facility location
problem, which was shown to be NP-Hard to optimize within a factor of 1 − 1/e + ε for all
ε > 0 [13].
The problem is that the greedy algorithm has super-quadratic running time Θ(nmk) and
in many datasets n and m can be in the millions. For this reason, several recent papers
have focused on accelerating the greedy algorithm. In [26], the authors point out that if the
benefit matrix is sparse, this can dramatically speed up the computation time. Unfortunately,
in many problems of interest, data similarities or rewards are not sparse. Wei et al. [40]
proposed to first sparsify the benefit matrix and then run the greedy algorithm on this new
sparse matrix. In particular, [40] considers t-nearest neighbor sparsification, i.e., keeping for
each row the t largest entries and zeroing out the rest. Using this technique they demonstrated
an impressive 80-fold speedup over the greedy algorithm with little loss in solution quality.
One limitation of their theoretical analysis was the limited setting under which provable
approximation guarantees were established.
Our Contributions: Inspired by the work of Wei et al. [40] we improve the theoretical
analysis of the approximation error induced by sparsification. Specifically, the previous analysis assumes that the input came from a probability distribution where the preferences of
each element of i ∈ I are independently chosen uniformly at random. For this distribution,
when k = Ω(n), they establish that the sparsity can be taken to be O(log n) and running the
greedy algorithm on the sparsified problem will guarantee a constant factor approximation
with high probability. We improve the analysis in the following ways:
• We prove guarantees for all values of k and our guarantees do not require any assumptions on the input besides nonnegativity of the benefit matrix.
2
• In the case where k = Ω(n), we show that it is possible to take the sparsity of each
row as low as O(1) while guaranteeing a constant factor approximation.
• Unlike previous work, our analysis does not require the use of any particular algorithm
and can be integrated to many algorithms for solving facility location problems.
• We establish a lower bound which shows that our approximation guarantees are tight
up to log factors, for all desired approximation factors.
In addition to the above results we propose a novel algorithm that uses a threshold based
sparsification where we keep matrix elements that are above a set value threshold. This type
of sparsification is easier to efficiently implement using nearest neighbor methods. For this
method of sparsification, we obtain worst case guarantees and a lower bound that matches
up to constant factors. We also obtain a data dependent guarantee which helps explain why
our algorithm empirically performs better than the worst case.
Further, we propose the use of Locality Sensitive Hashing (LSH) and random walk methods to accelerate approximate nearest neighbor computations. Specifically, we use two types
of similarity metrics: inner products and personalized PageRank (PPR). We propose the use
of fast approximations for these metrics and empirically show that they dramatically improve
running times. LSH functions are well-known but, to the best of our knowledge, this is the
first time they have been used to accelerate facility location problems. Furthermore, we utilize personalized PageRank as the similarity between vertices on a graph. Random walks can
quickly approximate this similarity and we demonstrate that it yields highly interpretable
results for real datasets.
2
Related Work
The use of a sparsified proxy function was shown by Wei et al. to also be useful for finding
a subset for training nearest neighbor classifiers [41]. Further, they also show a connection
of nearest neighbor classifiers to the facility location function. The facility location function
was also used by Mirzasoleiman et al. as part of a summarization objective function in [32],
where they present a summarization algorithm that is able to handle a variety of constraints.
The stochastic greedy algorithm was shown to get a 1 − 1/e − ε approximation with
runtime O(nm log 1ε ), which has no dependance on k [33]. It works by choosing a sample set
from V of size nk log 1ε each iteration and adding to the current set the element of the sample
set with the largest gain.
Also, there are several related algorithms for the streaming setting [5] and distributed
setting [6, 25, 31, 34]. Since the objective function is defined over the entire dataset, optimizing the facility location function becomes more complicated in these memory limited
settings. Often the function is estimated by considering a randomly chosen subset from the
set I.
2.1
Benefits Functions and Nearest Neighbor Methods
For many problems, the elements V and I are vectors in some feature space where the
benefit matrix is defined by some similarity function sim. For example, in Rd we may use
3
2
the RBF kernel sim(x, y) = e−γkx−yk2 , dot product sim(x, y) = xT y, or cosine similarity
xT y
.
sim(x, y) = kxkkyk
There has been decades of research on nearest neighbor search in geometric spaces. If
the vectors are low dimensional, then classical techniques such as kd-trees [7] work well and
are exact. However it has been observed that as the dimensions grow that the runtime of all
known exact methods does little better than a linear scan over the dataset.
As a compromise, researchers have started to work on approximate nearest neighbor
methods, one of the most successful approaches being locality sensitive hashing [15, 20].
LSH uses a hash function that hashes together items that are close. Locality sensitive hash
functions exist for a variety of metrics and similarities such as Euclidean [11], cosine similarity
[3, 9], and dot product [36, 38]. Nearest neighbor methods other than LSH that have been
shown to work for machine learning problems include [8, 10]. Additionally, see [14] for efficient
and exact GPU methods.
An alternative to vector functions is to use similarities and benefits defined from graph
structures. For instance, we can use the personalized PageRank of vertices in a graph to define
the benefit matrix [37]. The personalized PageRank is similar to the classic PageRank, except
the random jumps, rather than going to anywhere in the graph, go back to the users “home”
vertex. This can be used as a value of “reputation” or “influence” between vertices in a graph
[17].
There are a variety of algorithms for finding the vertices with a large PageRank personalized to some vertex. One popular one is the random walk method. If πi is the personalized
PageRank vector to some vertex i, then πi (v) is the same as the probability that a random
walk of geometric length starting from i ends on a vertex v (where the parameter of the
geometric distribution is defined by the probability of jumping back to i) [4]. Using this
approach, we can quickly estimate all elements in the benefit matrix greater than some value
τ.
3
Guarantees for t-Nearest Neighbor Sparsification
We associate a bipartite support graph G = (V, I, E) by having an edge between v ∈ V and
i ∈ I whenever Cij > 0. If the support graph is sparse, we can use the graph to calculate
the gain of an element much more efficiently, since we only need to consider the neighbors
of the element versus the entire set I. If the average degree of a vertex i ∈ I is t, (and we
use a cache for the current best value of an element i) then we can execute greedy in time
O(mtk). See Algorithm 1 in the Appendix for pseudocode. If the sparsity t is much smaller
than the size of V , the runtime is greatly improved.
However, the instance we wish to optimize may not be sparse. One idea is to sparsify
the original matrix by only keeping the values in the benefit matrix C that are t-nearest
neighbors, which was considered in [40]. That is, for every element i in I, we only keep the
top t elements of Ci1 , Ci2 , . . . , Cin and set the rest equal to zero. This leads to a matrix with
mt nonzero elements. We then want the solution from optimizing the sparse problem to be
close to the value of the optimal solution in the original objective function F .
n
m
Our main theorem is that we can set the sparsity parameter t to be O( αk
log αk
)—which is
a significant improvement for large enough k—while still having the solution to the sparsified
4
problem be at most a factor of
1
1+α
from the value of the optimal solution.
Theorem 1. Let Ot be the optimal solution to an instance of the facility location problem with
n
n
log αk
),
a benefit matrix that was sparsified with t-nearest neighbor. For any t ≥ t∗ (α) = O( αk
1
we have F (Ot) ≥ 1+α OPT.
Proof Sketch. For the value of t chosen, there exists a set Γ of size αk such that every element
of I has a neighbor in the t-nearest neighbor graph; this is proven using the probabilistic
method. By appending Γ to the optimal solution and using the monotonicity of F , we can
move to the sparsified function, since no element of I would prefer an element that was
zeroed out in the sparsified matrix as one of their top t most beneficial elements is present
in the set Γ. The optimal solution appended with Γ is a set of size (1 + α)k. We then bound
the amount that the optimal value of a submodular function can increase by when adding
αk elements. See the appendix for the complete proof.
Note that Theorem 1 is agnostic to the algorithm used to optimize the sparsified function,
ρ
and so if we use a ρ-approximation algorithm, then we are at most a factor of 1+α
from the
optimal solution. Later this section we will utilize this to design a subquadratic algorithm for
optimizing facility location problems as long as we can quickly compute t-nearest neighbors
and k is large enough.
If m = O(n) and k = Ω(n), we can achieve a constant factor approximation even when
taking the sparsity parameter as low as t = O(1), which means that the benefit matrix C has
only O(n) nonzero entries. Also note that the only assumption we need is that the benefits
between elements are nonnegative. When k = Ω(n), previous work was only able to take
t = O(log n) and required the benefit matrix to come from a probability distribution [40].
Our guarantee has two regimes depending on the value of α. If we want the optimal
solution to the sparsified function to be a 1 − ε factor from the optimal solution to the
m
n
log εk
) suffices. Conversely, if we want to take
original function, we have that t∗ (ε) = O( εk
n
m
the sparsity t to be much smaller than k log k , then this is equivalent to taking α very large
and we have some guarantee of optimality.
In the proof of Theorem 1, the only time we utilize the value of t is to show that there
exists a small set Γ that covers the entire set I in the t-nearest neighbor graph. Real datasets
n
m
often contain a covering set of size αk for t much smaller than O( αk
log αk
). This observation
yields the following corollary.
Corollary 2. If after sparsifying a problem instance there exists a covering set of size αk in
the t-nearest neighbor graph, then the optimal solution Ot of the sparsified problem satisfies
1
F (Ot ) ≥ 1+α
OPT.
In the datasets we consider in our experiments of roughly 7000 items, we have covering
sets with only 25 elements for t = 75, and a covering set of size 10 for t = 150. The size of
covering set was upper bounded by using the greedy set cover algorithm. In Figure 2 in the
appendix, we see how the size of the covering set changes with the choice of the number of
neighbors chosen t.
It would be desirable to take the sparsity parameter t lower than the value dictated by
∗
t (α). As demonstrated by the following lower bound, is not possible to take the sparsity
1
approximation in the worst case.
significantly lower than α1 nk and still have a 1+α
5
Proposition 3. Suppose we take
t = max
1
1
,
2α 1 + α
n−1
.
k
There exists a family of inputs such that we have F (Ot ) ≤
1
OPT.
1+α−2/k
The example we create to show this has the property that in the t-nearest neighbor graph,
the set Γ needs αk elements to cover every element of I. We plant a much smaller covering set
that is very close in value to Γ but is hidden after sparsification. We then embed a modular
function within the facility location objective. With knowledge of the small covering set,
an optimal solver can take advantage of this modular function, while the sparsified solution
would prefer to first choose the set Γ before considering the modular function. See the
appendix for full details.
Sparsification integrates well with the stochastic greedy algorithm [33]. By taking t ≥
t∗ (ε/2) and running stochastic greedy with sample sets of size nk ln 2ε , we get a 1 − 1/e − ε
m
approximation in expectation that runs in expected time O( nm
log 1ε log εk
). If we can quickly
εk
1/3
sparsify the problem and k is large enough, for example n , this is subquadratic. The
following proposition shows a high probability guarantee on the runtime of this algorithm
and is proven in the appendix.
Proposition 4. When m = O(n), the stochastic greedy algorithm [33] with set sizes of
size nk log 2ε , combined with sparsification with sparsity parameter t, will terminate in time
n
m
log εk
), this algoO(n log 1ε max{t, log n}) with high probability. When t ≥ t∗ (ε/2) = O( εk
rithm has a 1 − 1/e − ε approximation in expectation.
4
Guarantees for Threshold-Based Sparsification
Rather than t-nearest neighbor sparsification, we now consider using τ -threshold sparsification, where we zero-out all entries that have value below a threshold τ . Recall the definition
of a locality sensitive hash.
Definition. H is a (τ, Kτ, p, q)-locality sensitive hash family if for x, y satisfying sim(x, y) ≥
τ we have Ph∈H (h(x) = h(y)) ≥ p and if x, y satisfy sim(x, y) ≤ Kτ we have Ph∈H (h(x) =
h(y)) ≤ q.
We see that τ -threshold sparsification is a better model than t-nearest neighbors for LSH,
as for K = 1 it is a noisy τ -sparsification and for non-adversarial datasets it is a reasonable
approximation of a τ -sparsification method. Note that due to the approximation constant
K, we do not have an a priori guarantee on the runtime of arbitrary datasets. However we
would expect in practice that we would only see a few elements with threshold above the
value τ . See [2] for a discussion on this.
One issue is that we do not know how to choose the threshold τ . We can sample elements
of the benefit matrix C to estimate how sparse the threshold graph will be for a given
6
threshold τ . Assuming the values of C are in general position1 , by using the DvoretzkyKiefer-Wolfowitz-Massart Inequality [12, 28] we can bound the number of samples needed
to choose a threshold that achieves a desired sparsification level.
We establish the following data-dependent bound on the difference in the optimal solutions of the τ -threshold sparsified function and the original function. We denote the set of
vertices adjacent to S in the τ -threshold graph with N(S).
Theorem 5. Let Oτ be the optimal solution to an instance of the facility location problem
with a benefit matrix that was sparsified using a τ -threshold. Assume there exists a set S of
size k such that in the τ -threshold graph we have the neighborhood of S satisfying |N(S)| ≥
µn. Then we have
−1
1
F (Oτ ) ≥ 1 +
OPT.
µ
For the datasets we consider in our experiments, we see that we can keep just a 0.01−0.001
fraction of the elements of C while still having a small set S with a neighborhood N(S) that
satisfied |N(S)| ≥ 0.3n. In Figure 3 in the appendix, we plot the relationship between the
number of edges in the τ -threshold graph and the number of coverable element by a a set of
small size, as estimated by the greedy algorithm for max-cover.
Additionally, we have worst case dependency on the number of edges in the τ -threshold
graph and the approximation factor. The guarantees follow from applying Theorem 5 with
the following Lemma.
1 2 2
c 1
edges has a set S of size k such that the
Lemma 6. For k ≥ 1−2c
2 δ , any graph with 2 δ n
neighborhood N(S) satisfies
|N(S)| ≥ cδn.
To get a matching lower bound, consider the case where the graph has two disjoint cliques,
one of size δn and one of size (1 − δ)n. Details are in the appendix.
5
Experiments
5.1
Summarizing Movies and Music from Ratings Data
We consider the problem of summarizing a large collection of movies. We first need to create
a feature vector for each movie. Movies can be categorized by the people who like them,
and so we create our feature vectors from the MovieLens ratings data [16]. The MovieLens
database has 20 million ratings for 27,000 movies from 138,000 users. To do this, we perform
low-rank matrix completion and factorization on the ratings matrix [21, 22] to get a matrix
X = UV T , where X is the completed ratings matrix, U is a matrix of feature vectors for
each user and V is a matrix of feature vectors for each movie. For movies i and j with
vectors vi and vj , we set the benefit function Cij = viT vj . We do not use the normalized dot
product (cosine similarity) because we want our summary movies to be movies that were
1
By this we mean that the values of C are all unique, or at least only a few elements take any particular
value. We need this to hold since otherwise a threshold based sparsification may exclusively return an empty
graph or the complete graph.
7
highly rated, and not normalizing makes highly rated movies increase the objective function
more.
We complete the ratings matrix using the MLlib library in Apache Spark [29] after
removing all but the top seven thousand most rated movies to remove noise from the data.
We use locality sensitive hashing to perform sparsification; in particular we use the LSH in
the FALCONN library for cosine similarity [3] and the reduction from a cosine simiarlity
hash to a dot product hash [36]. As a baseline we consider sparsification using a scan over
the entire dataset, the stochastic greedy algorithm with lazy evaluations [33], and the greedy
algorithm with lazy evaluations [30]. The number of elements chosen was set to 40 and for
the LSH method and stochastic greedy we average over five trials.
We then do a scan over the sparsity parameter t for the sparsification methods and a scan
over the number of samples drawn each iteration for the stochastic greedy algorithm. The
sparsified methods use the (non-stochastic) lazy greedy algorithm as the base optimization
algorithm, which we found worked best for this particular problem2 . In Figure 1(a) we see
that the LSH method very quickly approaches the greedy solution—it is almost identical
in value just after a few seconds even though the value of t is much less than t∗ (ε). The
stochastic greedy method requires much more time to get the same function value. Lazy
greedy is not plotted, since it took over 500 seconds to finish.
(a) Fraction of Greedy Set Value vs. Runtime
0.99
0.98
0.97
0.96
0.95
0.94
0.93
0.92
0.91
0.90
(b) Fraction of Greedy Set Contained vs. Runtime
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0
25 50 75 100 125 150
Runtime (s)
Exact top-t
LSH top-t
Stochastic Greedy
0
25
50 75 100 125 150
Runtime (s)
Figure 1: Results for the MovieLens dataset [16]. Figure (a) shows the function value as the runtime
increases, normalized by the value the greedy algorithm obtained. As can be seen our algorithm is within
99.9% of greedy in less than 5 seconds. For this experiment, the greedy algorithm had a runtime of 512
seconds, so this is a 100x speed up for a small penalty in performance. We also compare to the stochastic
greedy algorithm [33], which needs 125 seconds to get equivalent performance, which is 25x slower.
Figure (b) shows the fraction of the set that was returned by each method that was common with the set
returned by greedy. We see that the approximate nearest neighbor method has 90% of its elements common
with the greedy set while being 50x faster than greedy, and using exact nearest neighbors can perfectly
match the greedy set while being 4x faster than greedy.
A performance metric that can be better than the objective value is the fraction of
elements returned that are common with the greedy algorithm. We treat this as a proxy
for the interpretability of the results. We believe this metric is reasonable since we found
the subset returned by the greedy algorithm to be quite interpretable. We plot this metric
2
When experimenting on very larger datasets, we found that runtime constraints can make it necessary
to use stochastic greedy as the base optimization algorithm
8
Table 1: A subset of the summarization outputted by our algorithm on the MovieLens dataset, plus the
elements who are represented by each representative with the largest dot product. Each group has a natural
interpretation: 90’s slapstick comedies, 80’s horror, cult classics, etc. Note that this was obtained with only
a similarity matrix obtained from ratings.
Happy Gilmore
Nightmare on Elm Street
Star Wars IV
Shawshank Redemption
Tommy Boy
Billy Madison
Dumb & Dumber
Ace Ventura Pet Detective
Road Trip
American Pie 2
Black Sheep
Friday the 13th
Halloween II
Nightmare on Elm Street 3
Child’s Play
Return of the Living Dead II
Friday the 13th 2
Puppet Master
Star Wars V
Raiders of the Lost Ark
Star Wars VI
Indiana Jones, Last Crusade
Terminator 2
The Terminator
Star Trek II
Schindler’s List
The Usual Suspects
Life Is Beautiful
Saving Private Ryan
American History X
The Dark Knight
Good Will Hunting
Pulp Fiction
The Notebook
Pride and Prejudice
The Godfather
Reservoir Dogs
American Beauty
A Clockwork Orange
Trainspotting
Memento
Old Boy
No Country for Old Men
P.S. I Love You
The Holiday
Remember Me
A Walk to Remembe
The Proposal
The Vow
Life as We Know It
Anne of Green Gables
Persuasion
Emma
Mostly Martha
Desk Set
The Young Victoria
Mansfield Park
The Godfather II
One Flew Over the Cuckoo’s Nest
Goodfellas
Apocalypse Now
Chinatown
12 Angry Men
Taxi Driver
against runtime in Figure 1(b). We see that the LSH method quickly gets to 90% of the
elements in the greedy set while stochastic greedy takes much longer to get to just 70%
of the elements. The exact sparsification method is able to completely match the greedy
solution at this point. One interesting feature is that the LSH method does not go much
higher than 90%. This may be due to the increased inaccuracy when looking at elements
with smaller dot products. We plot this metric against the number of exact and approximate
nearest neighbors t in Figure 4 in the appendix.
We include a subset of the summarization and for each representative a few elements
who are represented by this representative with the largest dot product in Table 1 to show
the interpretability of our results.
5.2
Finding Influential Actors and Actresses
For our second experiment, we consider how to find a diverse subset of actors and actresses in
a collaboration network. We have an edge between an actor or actress if they collaborated in
a movie together, weighted by the number of collaborations. Data was obtained from [19] and
an actor or actress was only included if he or she was one of the top six in the cast billing. As
a measure of influence, we use personalized PageRank [37]. To quickly calculate the people
with the largest influence relative to someone, we used the random walk method[4].
We first consider a small instance where we can see how well the sparsified approach
works. We build a graph based on the cast in the top thousand most rated movies. This
graph has roughly 6000 vertices and 60,000 edges. We then calculate the entire PPR matrix
using the power method. Note that this is infeasible on larger graphs in terms of time and
memory. Even on this moderate sized graph it took six hours and takes 2GB of space.
We then compare the value of the greedy algorithm using the entire PPR matrix with the
sparsified algorithm using the matrix approximated by Monte Carlo sampling using the two
9
metrics mentioned in the previous section. We omit exact nearest neighbor and stochastic
greedy because it is not clear how it would work without having to compute the entire PPR
matrix. Instead we compare to an approach where we choose a sample from I and calculate
the PPR only on these elements using the power method. As mentioned in Section 2, several
algorithms utilize random sampling from I. We take k to be 50 for this instance. In Figure 5
in the appendix we see that sparsification performs drastically better in both function value
and percent of the greedy set contained for a given runtime.
We now scale up to a larger graph by taking the actors and actresses billed in the top
six for the twenty thousand most rated movies. This graph has 57,000 vertices and 400,000
edges. We would not be able to compute the entire PPR matrix for this graph in a reasonable
amount of time or space. However we can run the sparsified algorithm in three hours using
only 2 GB of memory, which could be improved further by parallelizing the Monte Carlo
approximation.
We run the greedy algorithm separately on the actors and actresses. For each we take the
top twenty-five and compare to the actors and actresses with the largest (non-personalized)
PageRank. In Figure 2 of the appendix, we see that the PageRank output fails to capture
the diversity in nationality of the dataset, while the facility location optimization returns
actors and actresses from many of the worlds film industries.
Acknowledgements
This material is based upon work supported by the National Science Foundation Graduate
Research Fellowship under Grant No. DGE-1110007 as well as NSF Grants CCF 1344179,
1344364, 1407278, 1422549 and ARO YIP W911NF-14-1-0258.
References
[1] N. Alon and J. H. Spencer. The Probabilistic Method. Wiley, 3rd edition, 2008.
[2] A.
Andoni.
Exact
algorithms
from
approximation
algorithms?
(part
2).
Windows
on
Theory,
2012.
https://windowsontheory.org/2012/04/17/exact-algorithms-from-approximation-algorithm
(version: 2016-09-06).
[3] A. Andoni, P. Indyk, T. Laarhoven, I. Razenshteyn, and L. Schmidt. Practical and
optimal LSH for angular distance. In NIPS, 2015.
[4] K. Avrachenkov, N. Litvak, D. Nemirovsky, and N. Osipova. Monte Carlo methods in
PageRank computation: When one iteration is sufficient. SIAM Journal on Numerical
Analysis, 2007.
[5] A. Badanidiyuru, B. Mirzasoleiman, A. Karbasi, and A. Krause. Streaming submodular
maximization: Massive data summarization on the fly. In KDD, 2014.
[6] R. Barbosa, A. Ene, H. L. Nguyen, and J. Ward. The power of randomization: Distributed submodular maximization on massive datasets. In ICML, 2015.
10
[7] J. L. Bentley. Multidimensional binary search trees used for associative searching. Communications of the ACM, 1975.
[8] A. Beygelzimer, S. Kakade, and J. Langford. Cover trees for nearest neighbor. In ICML,
2006.
[9] M. S. Charikar. Similarity estimation techniques from rounding algorithms. In STOC,
2002.
[10] J. Chen, H.-R Fang, and Y. Saad. Fast approximate k–NN graph construction for high
dimensional data via recursive Lanczos bisection. JMLR, 2009.
[11] M. Datar, N. Immorlica, P. Indyk, and V. S. Mirrokni. Locality-sensitive hashing scheme
based on p-stable distributions. In Symposium on Computational Geometry, 2004.
[12] A. Dvoretzky, J. Kiefer, and J. Wolfowitz. Asymptotic minimax character of the sample distribution function and of the classical multinomial estimator. The Annals of
Mathematical Statistics, 1956.
[13] U. Feige. A threshold of ln n for approximating set cover. Journal of the ACM, 1998.
[14] V. Garcia, E. Debreuve, F. Nielsen, and M. Barlaud. K-nearest neighbor search: Fast
GPU-based implementations and application to high-dimensional feature matching. In
ICIP, 2010.
[15] A. Gionis, P. Indyk, and R. Motwani. Similarity search in high dimensions via hashing.
In VLDB, 1999.
[16] GroupLens. MovieLens 20M dataset, 2015. http://grouplens.org/datasets/movielens/20m/.
[17] P. Gupta, A. Goel, J. Lin, A. Sharma, D. Wang, and R. Zadeh. WTF: The who to
follow service at Twitter. In WWW, 2013.
[18] W. Hoeffding. Probability inequalities for sums of bounded random variables. Journal
of the American Statistical Association, 1963.
[19] IMDb. Alternative interfaces, 2016. http://www.imdb.com/interfaces.
[20] P. Indyk and R. Motwani. Approximate nearest neighbors: towards removing the curse
of dimensionality. In STOC, 1998.
[21] P. Jain, P. Netrapalli, and S. Sanghavi. Low-rank matrix completion using alternating
minimization. In STOC, 2013.
[22] Y. Koren, R. Bell, and C. Volinsky. Matrix factorization techniques for recommender
systems. Computer, 2009.
[23] A. Krause and R. G. Gomes. Budgeted nonparametric learning from data streams. In
ICML, 2010.
11
[24] A. Krause, J. Leskovec, C. Guestrin, J. VanBriesen, and C. Faloutsos. Efficient sensor
placement optimization for securing large water distribution networks. Journal of Water
Resources Planning and Management, 2008.
[25] R. Kumar, B. Moseley, S. Vassilvitskii, and A. Vattani. Fast greedy algorithms in
MapReduce and streaming. ACM Transactions on Parallel Computing, 2015.
[26] J. Leskovec, A. Krause, C. Guestrin, C. Faloutsos, J. VanBriesen, and N. Glance. Costeffective outbreak detection in networks. In KDD, 2007.
[27] H. Lin and J. A. Bilmes. Learning mixtures of submodular shells with application to
document summarization. In UAI, 2012.
[28] P. Massart. The tight constant in the Dvoretzky-Kiefer-Wolfowitz inequality. The
Annals of Probability, 1990.
[29] X. Meng, J. Bradley, B. Yavuz, E. Sparks, S. Venkataraman, D. Liu, J. Freeman, D. B.
Tsai, M. Amde, S. Owen, D. Xin, R. Xin, M. J. Franklin, R. Zadeh, M. Zaharia, and
A. Talwalkar. MLlib: Machine learning in Apache Spark. JMLR, 2016.
[30] M. Minoux. Accelerated greedy algorithms for maximizing submodular set functions.
In Optimization Techniques. Springer, 1978.
[31] V. Mirrokni and M. Zadimoghaddam. Randomized composable core-sets for distributed
submodular maximization. STOC, 2015.
[32] B. Mirzasoleiman, A. Badanidiyuru, and A. Karbasi. Fast constrained submodular
maximization: Personalized data summarization. In ICML, 2016.
[33] B. Mirzasoleiman, A. Badanidiyuru, A. Karbasi, J. Vondrak, and A. Krause. Lazier
than lazy greedy. In AAAI, 2015.
[34] B. Mirzasoleiman, A. Karbasi, R. Sarkar, and A. Krause. Distributed submodular
maximization: Identifying representative elements in massive data. In NIPS, 2013.
[35] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher. An analysis of approximations for
maximizing submodular set functions-I. Mathematical Programming, 1978.
[36] B. Neyshabur and N. Srebro. On symmetric and asymmetric LSHs for inner product
search. In ICML, 2015.
[37] L. Page, S. Brin, R. Motwani, and T. Winograd. The PageRank citation ranking:
Bringing order to the web. Stanford Digital Libraries, 1999.
[38] A. Shrivastava and P. Li. Asymmetric LSH (ALSH) for sublinear time maximum inner
product search (MIPS). In NIPS, 2014.
[39] S. Tschiatschek, R. K. Iyer, H. Wei, and J. A. Bilmes. Learning mixtures of submodular
functions for image collection summarization. In NIPS, 2014.
12
[40] K. Wei, R. Iyer, and J. Bilmes. Fast multi-stage submodular maximization. In ICML,
2014.
[41] K. Wei, R. Iyer, and J. Bilmes. Submodularity in data subset selection and active
learning. In ICML, 2015.
6
Appendix: Additional Figures
Algorithm 1 Greedy algorithm with sparsity graph
Input: benefit matrix C, sparsity graph G = (V, I, E)
define N(v): return the neighbors of v in G
for all i ∈ I:
# cache of the current benefit given to i
βi ← 0
A←∅
for k iterations:
for all v ∈ V :
# calculate the gain of element v
gv ← 0
for all i ∈ N(v):
# add the gain of element v from i
gv ← gv + max(Civ − βi , 0)
∗
v ← arg maxV gv
A ← A ∪ {v ∗ }
for all i ∈ N(v ∗ )
# update the cache of the current benefit for i
βi ← max(βi , Civ∗ )
return A
13
(b) IMDb
(a) MovieLens
60
Size of covering set Γ
Size of covering set Γ
70
50
40
30
20
10
0
0
25
50
75 100
Sparsity t
125
150
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
0
25
50
75 100
Sparsity t
125
150
Figure 2: (a) MovieLens Dataset [16] and (b) IMDb Dataset [19] as explained in the Experiments Section. We see that for sparsity t significantly smaller than the n/(αk) lower bound
we can still find a small covering set in the t-nearest neighbor graph.
(b) IMDb
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
Fraction of elements coverable
Fraction of elements coverable
(a) MovieLens
0.001
0.01
0.1
Fraction of elements kept
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.0001
0.001
0.01
Fraction of elements kept
Figure 3: (a) MovieLens Dataset [16] and (b) IMDb Dataset [19], as explained in the Experiments Section. We see that even with several orders of magnitude fewer edges than the
complete graph we still can find a small set that covers a large fraction of the dataset. For
MovieLens this set was of size 40 and for IMDb this set was of size 50. The number of
coverable was estimated by the greedy algorithm for the max-coverage problem.
14
Fraction of Greedy Set Contained vs. Sparsity
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
Exact top-t
0.1
LSH top-t
0.0
0
50
100
150
200
Sparsity (t)
Figure 4: The fraction of the greedy solution that was contained as the value of the sparsity
t was increased for exact nearest neighbor and approximate LSH-based nearest neighbor
on the MovieLens dataset. We see that the exact method captures slightly more of greedy
solution for a given value of t and the LSH value does not converge to 1. However LSH still
captures a reasonable amount of the greedy set and is significantly faster at finding nearest
neighbors.
(b) Fraction of Greedy Set Contained vs. Runtime
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0
500 1000 1500 2000 2500
Runtime (s)
(a) Fraction of Greedy Set Value vs. Runtime
1.0
0.9
0.8
0.7
0.6
0.5
Monte Carlo PPR
Sample PPR
0.4
0.3
0
500 1000 1500 2000 2500
Runtime (s)
Figure 5: Results for the IMDb dataset [19]. Figure (a) shows the function value as the
runtime increases, normalized by the value the greedy algorithm obtained. As can be seen
our algorithm is within 99% of greedy in less than 10 minutes. For this experiment, the
greedy algorithm had a runtime of six hours, so this is a 36x acceleration for a small penalty
in performance. We also compare to using a small sample of the set I as an estimate of the
function, which does not perform nearly as well as our algorithm even for much longer time.
Figure (b) shows the fraction of the set that was returned by each method that was common
with the set returned by greedy. We see that the approximate nearest neighbor method has
90% of its elements common with the greedy set while being 18x faster than greedy.
15
Table 2: The top twenty-five actors and actresses generated by sparsified facility location
optimization defined by the personalized PageRank of a 57,000 vertex movie personnel collaboration graph from [19] and the twenty-five actors and actresses with the largest (nonpersonalized) PageRank. We see that the classical PageRank approach fails to capture the
diversity of nationality in the dataset, while the facility location results have actors and
actresses from many of the worlds film industries.
Actors
Actresses
Facility Location
PageRank
Facility Location
PageRank
Robert De Niro
Jackie Chan
Gérard Depardieu
Kemal Sunal
Shah Rukh Khan
Michael Caine
John Wayne
Samuel L. Jackson
Bud Spencer
Peter Cushing
Toshirô Mifune
Steven Seagal
Moritz Bleibtreu
Jean-Claude Van Damme
Mads Mikkelsen
Michael Ironside
Amitabh Bachchan
Ricardo Darı́n
Charles Chaplin
Sean Bean
Louis de Funès
Tadanobu Asano
Bogdan Diklic
Nassar
Lance Henriksen
Jackie Chan
Gérard Depardieu
Robert De Niro
Michael Caine
Samuel L. Jackson
Christopher Lee
Donald Sutherland
Peter Cushing
Nicolas Cage
John Wayne
John Cusack
Christopher Walken
Bruce Willis
Kemal Sunal
Harvey Keitel
Amitabh Bachchan
Shah Rukh Khan
Sean Bean
Steven Seagal
Jean-Claude Van Damme
Morgan Freeman
Christian Slater
Val Kilmer
Liam Neeson
Gene Hackman
Julianne Moore
Susan Sarandon
Bette Davis
Isabelle Huppert
Kareena Kapoor
Juliette Binoche
Meryl Streep
Adile Naşit
Catherine Deneuve
Li Gong
Helena Bonham Carter
Penélope Cruz
Naomi Watts
Masako Nozawa
Drew Barrymore
Charlotte Rampling
Golshifteh Farahani
Hanna Schygulla
Toni Collette
Kati Outinen
Edna Purviance
Monica Bellucci
Kristin Scott Thomas
Catherine Keener
Kyôko Kagawa
Julianne Moore
Susan Sarandon
Juliette Binoche
Isabelle Huppert
Catherine Deneuve
Kristin Scott Thomas
Meryl Streep
Bette Davis
Nicole Kidman
Charlotte Rampling
Helena Bonham Carter
Kathy Bates
Naomi Watts
Cate Blanchett
Drew Barrymore
Helen Mirren
Michelle Pfeiffer
Penélope Cruz
Sigourney Weaver
Toni Collette
Catherine Keener
Heather Graham
Sandra Bullock
Kirsten Dunst
Miranda Richardson
16
7
7.1
Appendix: Full Proofs
Proof of Theorem 1
We will use the following two lemmas in the proof of Theorem 1, which are proven later
in this section. The first lemma bounds the size of the smallest set of left vertices covering
every right vertex in a t-regular bipartite graph.
Lemma 7. For any bipartite graph G = (V, I, E) such that |V | = n, |I| = m, every vertex
i ∈ I has degree at least t, and n ≤ mt, there exists a set of vertices Γ ⊆ V such that every
vertex in I has a neighbor in Γ and
n
mt
|Γ| ≤
1 + ln
.
(3)
t
n
The second lemma bounds the rate that the optimal solution grows as a function of k.
Lemma 8. Let f be any normalized submodular function and let O2 and O1 be optimal
solutions for their respective sizes, with |O2 | ≥ |O1 |. We have
f (O2) ≤
|O2 |
f (O1 ).
|O1 |
We now prove Theorem 1.
Proof. We will take t∗ (α) to be the smallest value of t such that |Γ| ≥ αk in Equation (3).
n
n
It can be verified that t∗ (α) ≤ ⌈4 αk
max{1, ln αk
}⌉.
Let Γ ⊆ V be a set such that all elements of I has a t-nearest neighbor in Γ. By Lemma
7, one is guaranteed to exists of size at most αk for t ≥ t∗ (α). Let O be the optimal set of
(1+α)k
size k and let F (t) be the objective function of the sparsified function. Let Otk and Ot
be the optimal solutions to F (t) of size k and (1 + α)k. We have
F (O) ≤ F (O ∪ Γ)
= F (t) (O ∪ Γ)
(1+α)k
≤ F (t) (Ot
).
(4)
The first inequality is due to the monotonicity of F . The second is because every element
of I would prefer to choose one of their t nearest neighbors and because of Γ they can. The
(1+α)k
third inequality is because |O ∪ Γ| ≤ (1 + α)k and Ot
is the optimal solution for this
size.
(1+α)k
Now by Lemma 8, we can bound the approximation for shrinking from Ot
to Otk .
Applying Lemma 8 and continuing from Equation (4) implies
F (O) ≤ (1 + α)F (t) (Otk ).
Observe that F (t) (A) ≤ F (A) for any set A to obtain the final bound.
17
7.2
Proof of Proposition 3
Define Πn (t) to be the n × (n + 1) matrix where for i = 1, . . . , n we have column i equal to 1
for positions i to i + t − 1, potentially cycling the position back to the beginning if necessary,
and then 0 otherwise. For column n + 1 make all values 1 − 1/2n. For example,
1 0 0 0 1 1 11/12
1 1 0 0 0 1 11/12
1 1 1 0 0 0 11/12
Π6 (3) =
0 1 1 1 0 0 11/12 .
0 0 1 1 1 0 11/12
0 0 0 1 1 1 11/12
We will show the lower bound in two parts, when α < 1 and when α ≥ 1.
Proof of case α ≥ 1. Let F be the facility location function defined on the benefit matrix
C = Πn (δ nk ). For t = δ nk , the sparsified matrix C (t) has all of its elements except the n + 1st
row. With k elements, the optimal solution to F (t) is to choose the k elements that let
us cover δn of the elements of I, giving a value of δn. However if we chose the n + 1th
δ
element, we would have gotten a value of n − 1/2, giving an approximation of 1−1/(2n)
.
Setting δ = 1/(1 + α) and using α ≤ n/k implies
F (Ot ) ≤
when we take t =
1 |V |−1
1+α k
1
OPT
1 + α − 1/k
(note that for this problem |V | = n + 1).
Proof of case α < 1. Let F be the facility location function defined on the benefit matrix
Πn ( α1 nk )
0
C=
1n
1
0
− 2n
In×n
αk
For t = α1 nk , the optimal solution to F (t) is to use αk elements to cover all the elements of
Πn , then use the remaining (1 − α)k elements in the identity section of the matrix. This has
a value of less than α1 n. For F , the optimal solution is to choose the n + 1st element of Πn ,
then use the remaining k − 1 elements in the identity section of the identity section of the
1
− n1 ), and therefore an approximation of
matrix. This has a value of more than n(1 + α1 − kα
1
. Note that in this case |V | = 2n + 1 and so we have
1+α−1/k−1/n
F (Ot) ≤
when we take t =
1
OPT
1 + α − 1/k − 1/n
1 |V |−1/2
.
2α
k
18
7.3
Proof of Proposition 4
Proof. The stochastic greedy algorithm works by choosing a set of elements Sj each iteration
of size nk log 1ε . We will assume m = n and ε = 1/e to simplify notation. We want to show
that
k X
X
dv = O(nt)
j=1 v∈Sj
with high probability, where dv is the degree of element v in the sparsity graph. We will
show this using Bernstein’s Inequality: given n i.i.d. random variables X1 , . . . , Xn such that
E(Xℓ ) = 0, Var(Xℓ ) = σ 2 , and |Xℓ | ≤ c with probability 1, we have
!
n
X
nλ2
P
Xℓ ≥ λn ≤ exp − 2 2
.
2σ + 3 cλ
ℓ=1
We will take Xℓ to be the degree of the ℓth element of V chosen uniformly at random,
shifted by the mean of t. Although in the stochastic greedy algorithm the elements are not
chosen i.i.d. but instead iterations in k iterations of sampling without replacement, treating
them as i.i.d. random variables for purposes of Bernstein’s Inequality is justified by Theorem
4 of [18].
We have |Xℓ | ≤ n, and Var(Xℓ ) ≤ tn, where the variance bound is because variance for
a given mean t on support [0, m] is maximized by putting mass nt on n and 1 − nt on 0.
If t ≥ ln n, then take λ = 38 t. If t < ln n, take λ = 83 ln n. This yields
r
k X
X
1
8 m
max{nt, ln n} ≤ .
P
dv ≥ nt +
3 n
n
j=1 v∈S
j
7.4
Proof of Lemma 7
We now prove Lemma 7, which is a modification of Theorem 1.2.2 of [1].
Proof. Choose a set X by picking each element of V with probability p, where p is to be
decided later. For every element of I without a neighbor in X, add one arbitrarily. Call
this set Y . We have E(|X ∪ Y |) ≤ np + m(1 − p)t ≤ np + me−pt . Optimizing for p yields
t
p = 1t ln mt
. This is a valid probability when mt
≥ 1, which we assumed, and when m
≤ et
n
n
n
(we do not need to worry about the latter case because if it does not hold then it implies an
inequality weaker than the trivial one |Γ| ≤ n).
7.5
Proof of Lemma 8
Before we prove Lemma 8, we need the following Lemma.
Lemma 9. Let f be any normalized submodular function and let O be an optimal solution
for its respective size. Let A be any set. We have
|A|
f (O ∪ A) ≤ 1 +
f (O).
|O|
19
We now prove Lemma 8.
Proof. Let
A∗ =
arg max
f (A)
{A⊆O2 :|A|≤|O1|}
and let A′ = O2 \ A∗ . Since A∗ is optimal for the function when restricted to a ground set
O2 , by Lemma 9 and the optimality of O1 for sets of size |O1 |, we have
f (O2) = f (A∗ ∪ A′ )
|A′ |
≤ 1 + ∗ f (A∗ )
|A |
|O2 |
=
f (A∗ )
|O1 |
|O2 |
≤
f (O1 ).
|O1 |
We now prove Lemma 9.
Proof. Define f (v | A) = f ({v} ∪ A) − f (A). Let O = {o1 , . . . ok }, where the ordering is
arbitrary except that
f (ok | O \ {ok }) = arg min f (oi | O \ {oi }).
i=1,...,k
Let A = {a1 , . . . , aℓ }, where the ordering is arbitrary except that
f (a1 | O) = arg max f (ai | O).
i=1,...,ℓ
We will first show that
f (a1 | O) ≤ f (ok | O \ {ok }).
By submodularity, we have
f (a1 | O) ≤ f (a1 | O \ {ok }).
If it was true that
f (a1 | O \ {ok }) > f (ok | O \ {ok }),
then we would have
f ((O \ {ok }) ∪ {a1 }) = f (a1 | O \ {ok }) +
k−1
X
f (oi | {o1 , . . . , oi−1 })
i=1
≥
k
X
f (oi | {o1 , . . . , oi−1 })
i=1
= f (O),
20
(5)
contradicting the optimality of O, thus showing that Inequality 5 holds.
Now since for all i ∈ {1, 2, . . . , k}
f (a1 | O) ≤ f (ok | O \ {ok })
≤ f (oi | O \ {oi })
≤ f (oi | {o1 , . . . , oi−1 }),
it is worse than the average of f (oi | {o1 , . . . , oi−1 }), which is
and showing that
1
f (a1 | O) ≤ f (O).
k
Finally, we have
f (O ∪ A) = f (O) +
≤ f (O) +
ℓ
X
i=1
ℓ
X
1
k
Pk
i=1
f (oi | {o1 , . . . , oi−1 }),
(6)
f (ai | O ∪ {a1 , . . . , ai−1 })
f (ai | O)
i=1
≤ f (O) + ℓf (a1 | O)
ℓ
f (O),
≤ 1+
k
which is what we wanted to show.
7.6
Proof of Theorem 5
Proof. Let O be the optimal solution to the original problem. Let Fτ and F τ be the functions
defined restricting to the matrix elements with benefit at least τ and all remaining elements,
respectively. If there exists a set S of size k such that µn elements have a neighbor in S,
then we have
F (O) ≤ Fτ (O) + F τ (O)
≤ Fτ (O) + nτ
1
≤ Fτ (O) + Fτ (S)
µ
1
Fτ (Oτ )
≤ 1+
µ
where the last inequality follows from Oτ being optimal for Fτ .
7.7
Proof of Lemma 6
Proof. Consider the following algorithm:
B←∅
S←∅
21
while |B| ≤ cδn
v ∗ ← arg max |N(v)|
add v ∗ to S
add N(v ∗ ) to B
remove N(v ∗ ) ∪ {v ∗ } from G
c
We will show that after T = (1−2c
2 )δ iterations this algorithm will terminate. When it does,
S will satisfy |N(S)| ≥ cδn since every element of B has a neighbor in S.
If there exists a vertex of degree cδn, then we will terminate after the first iteration.
Otherwise all vertices have degree less than cδn. Assuming all vertices have degree less than
cδn, until we terminate the number of edges incident to B is at most |B|cδn ≤ c2 δ 2 n2 . At
each iteration, the number of edges in the graph is at least ( 21 −c2 )δ 2 n2 , thus in each iteration
we can find a v ∗ with degree at least (1 − 2c2 )δ 2 n. Therefore, after T iterations, we will have
terminated with the size of S is at most T and |N(S)| ≥ cδn.
We see that this is tight up to constant factors by the following proposition.
Proposition 10. There exists an example where for ∆ = δ 2 n, the optimal solution to the
sparsified function is a factor of O(δ) from the optimal solution to the original function.
Proof. Consider the following benefit matrix.
1
)I
1δn×δn + (1 + k−1
C=
0
(1 −
0
1
)1
(1−δn)×(1−δn)
(1−δ)n
The sparsified optimal would only choose elements in the top left clique and would get a
value of roughly δn, while the true optimal solution would cover both cliques and get a value
of roughly n.
22
| 8cs.DS
|
1
One-Bit Sphere Decoding for Uplink Massive MIMO
Systems with One-Bit ADCs
arXiv:1709.04163v2 [cs.IT] 19 Dec 2017
Yo-Seb Jeon, Namyoon Lee, Song-Nam Hong, and Robert W. Heath, Jr.
Abstract
This paper presents a low-complexity near-maximum-likelihood-detection (near-MLD) algorithm
called one-bit-sphere-decoding for an uplink massive multiple-input multiple-output (MIMO) system
with one-bit analog-to-digital converters (ADCs). The idea of the proposed algorithm is to estimate the
transmitted symbol vector sent by uplink users (a codeword vector) by searching over a sphere, which
contains a collection of codeword vectors close to the received signal vector at the base station in terms
of a weighted Hamming distance. To reduce the computational complexity for the construction of the
sphere, the proposed algorithm divides the received signal vector into multiple sub-vectors each with
reduced dimension. Then, it generates multiple spheres in parallel, where each sphere is centered at
the sub-vector and contains a list of sub-codeword vectors. The detection performance of the proposed
algorithm is also analyzed by characterizing the probability that the proposed algorithm performs worse
than the MLD. The analysis shows how the dimension of each sphere and the size of the sub-codeword
list are related to the performance-complexity tradeoff achieved by the proposed algorithm. Simulation
results demonstrate that the proposed algorithm achieves near-MLD performance, while reducing the
computational complexity compared to the existing MLD method.
I. I NTRODUCTION
Wireless systems with ultra low-precision analog-to-digital converter (ADC) are a power and
cost efficient solution for future cellular networks that support wide bandwidths and a large
Y.-S. Jeon and N. Lee are with the Department of Electrical Engineering, POSTECH, Pohang, Gyeongbuk, 37673 Korea
(e-mail: [email protected], [email protected]).
S.-N. Hong is with the Department of Electrical and Computer Engineering, Ajou University, Suwon, Gyeonggi, 16499 Korea
(e-mail: [email protected]).
R. W. Heath, Jr. is with the Wireless Networking and Communications Group, The University of Texas at Austin, Austin, TX
78712 USA (e-mail: [email protected]).
2
number of antennas at the base station (BS) [1]–[8]. For the multiple-input multiple-output
(MIMO) system with B-bit ADCs, where B > 1, finding an optimal data detection method is
challenging. The challenge arises from multiple quantization levels at the ADCs, which can be
differently chosen to minimize the detection error probability according to an input constellation
and the variance of noise. Numerous sub-optimal data detection and channel estimation methods
have been proposed assuming fixed quantization levels [9]–[15].
The use of one-bit ADCs in MIMO systems is interesting from both practical and theoretical
perspectives [6], [16]–[19]. One major implementation advantage is the simplification of the
circuit complexity by removing automatic gain control [20]. In addition, the characterization of
the channel capacity becomes tractable due to a fixed quantization level (e.g., zero-threshold
comparator). For example, some capacity bounds of the MIMO system with one-bit ADCs were
characterized when employing channel state information at the transmitter for a noise-free case
[6]. Beside a capacity characterization, it is also possible to analytically derive the maximum
likelihood detection (MLD) for the MIMO systems with one-bit ADCs [17], which yields the
minimum error probability of detecting transmit symbols.
The MLD problem for MIMO systems with one-bit ADCs differs from that for the MIMO
system with infinite-precision ADCs. The MLD for the conventional MIMO systems under
Gaussian noise reduces to the minimum Euclidean distance detection problem over a finite
constellation set [21]–[24]. In contrast, the MLD for the MIMO system with one-bit ADCs
finds an integer vector that maximizes the product of Q-functions [17] instead of solving the
least-squares problems. Nevertheless, the computational complexity of both MLD problems is
NP-hard due to the integer constraint on the feasible set.
Some low-complexity detection methods have been developed for MIMO systems with one-bit
ADCs [16]–[18]. For instance, a heuristic zero-forcing detection (ZFD) method using one-bit
quantized measurements was introduced in [16]. A drawback of ZFD is that the number of
receive antennas should be much larger than all possible numbers of transmit symbol vectors
to reliably detect the transmitted data symbols. In other words, for a given number of receive
antennas, the constellation size and/or the number of uplink users sending the data should be
small to achieve a target level of the data detection performance. A near-MLD method using a
convex relaxation technique was proposed in [10]. This approach was also extended to devise
3
ML data detection and channel estimation methods for the MIMO systems employing one-bit
ADCs [17]. The common idea is to convert the non-convex optimization problem that finds
the optimal integer vector to a convex-optimization problem, and then find the solution using
gradient-decent algorithms.
Sphere decoding is a low-complexity detection method for MIMO systems with infiniteprecision ADCs [21]–[26]. The basic idea of sphere decoding is to search over only integer
input vectors that lie in a certain sphere of radius d around an initial estimate of input vector x̂.
It diminishes the computational complexity by reducing the search space, while achieving nearMLD performance. The sphere decoding algorithms in [21]–[24], do not extend to MIMO system
with one-bit ADCs. This is because they find a set of integer vectors within the sphere with
radius of d in terms of the Euclidian distance between a received vector y and the product of the
channel matrix H and the initial estimate of transmit vector x̂ [21]–[23]. When employing one-bit
ADCs, the initial estimate of x̂ with one-bit measurements is inaccurate using ZFD. Furthermore,
constructing the sphere using the Euclidian distance is not optimal when the received signal at
the BS is quantized, as proven in [13] for one-bit ADCs.
In this paper, a low-complexity detection algorithm inspired by sphere decoding is presented
for an uplink massive MIMO system with one-bit ADCs. The major contributions of this paper
are summarized as follows.
•
We develop a near-optimal detection method for the uplink of multi-user MIMO system
with one-bit ADCs. The proposed method is a variant of minimum weighted-Hammingdistance detection (MWD) that was originally introduced in [13]. Unlike [13] in which the
weights are defined in an integral form, the proposed MWD exploits closed-form weights
when computing weighted-Hamming-distances by approximating the Q-function.
•
We propose a low-complexity near-MLD algorithm called one-bit-sphere-decoding (OSD).
The key idea of the OSD is to perform the proposed MWD over a sphere, which is a reduced
set of all possible symbol (codeword) vectors that are close to the received signal at the BS
in terms of the weighted Hamming distance. To diminish the computational complexity for
the construction of codeword list in the sphere, we divide the received signal vector into
multiple sub-vectors each with a reduced dimension. Then, we generate multiple spheres in
parallel, where each sphere is centered at the sub-vector and contains a list of sub-codeword
4
vectors. We compare the detection complexities between the proposed OSD and the existing
MLD, and show the gains in the deduction of the complexity for OSD over MLD.
•
We quantify the detection performance loss of the proposed OSD compared to the optimal
performance. To this end, we characterize an upper bound of the probability that the
proposed OSD performs worse than the MWD. In the characterization, we first show that
this probability is upper bounded by the sphere-list-error-probability (SEP), which is the
probability that the index of the transmitted codeword does not belong to the constructed
list in the sphere. We then derive an analytical expression for the upper bound of the SEP
in terms of the relevant system parameters: 1) the number of uplink users, 2) the number
of receive antennas at the BS, 3) the size of the codeword list in the sphere, and 4) the
dimension of the sub-vector. Our result reveals how the multi-user detection error behaves
with these relevant parameters.
•
Using simulations, we compare the detection performance of the OSD with those of the
MLD and the MWD for both uncoded and coded MIMO systems with one-bit ADCs.
Simulation results show that for the uncoded system, the OSD has near-MLD detection
performance, while achieving a significant reduction in the detection complexity compared
to the MLD. For the coded MIMO system, the OSD is implemented with a soft-output
decoder by applying the technique in [28] and is shown to achieve a significant frameerror-rate (FER) reduction compared to a hard-output decoder.
Notation: Upper-case and lower-case boldface letters denote matrices and column vectors,
respectively. E[·] is the statistical expectation, P(·) is the probability, (·)⊤ is the transpose, | · | is
the absolute value, Re(·) is the real part, Im(·) is the imaginary part, and ⌊·⌋ is the floor function.
1n is an n-dimensional vector whose elements are all ones. I(A) is an indicator function that
equals one if an event A is true and zero otherwise.
II. S YSTEM M ODEL
In this section, we present a model for an uplink massive MIMO system with one-bit ADCs
and provide definitions that will be used in the sequel.
5
…
User
#ʹ
Base station
#ͳ
#ʹ
Re
Im
One-Bit
ADC
Low-bit
ADC
Re
Im
One-Bit
ADC
Low-bit
ADC
…
User
#ͳ
#ܰ
Re
Im
Baseband
processing
One-Bit
ADC
Low-bit
ADC
User
#ܷ
Fig. 1. Illustration of a U-user uplink massive MIMO system that operates with one-bit ADCs.
A. System Model
We consider an uplink massive MIMO system in which U uplink users, each equipped with
a single transmit antenna, send data symbols to a BS equipped with N receive antennas, as
illustrated in Fig. 1. We denote a (data) symbol vector sent by the uplink users at time slot t as
x̄[t] = [ x̄1 [t], x̄2 [t], · · · , x̄U [t]]⊤ ∈ CU , where each data symbol x̄u [t] is drawn from a constellation
set X̄ with size M, i.e., | X̄| = M, and satisfies E[| x̄u [t]| 2 ] = 1. In other words, the symbol vector
x̄[t] collects the transmitted signals from all users at time slot t. We define X as a constellation
√
set for real or imaginary part such that X̄ = {xR + j xI | xR, xI ∈ X} and |X| = M. Then X 2U
represents a symbol vector set that contains all possible combinations of transmit symbols sent
by the U uplink users.
We assume a frequency-flat MIMO channel. Let h̄u ∈ CN×1 be the channel vector from the
uth uplink user to the BS. Then, the channel impulse response is given by a channel matrix
H̄ = [h̄1, . . ., h̄U ] ∈ CN×U . We also make two assumptions: 1) perfect synchronization at the
BS and 2) perfect power control across all uplink users. Under these assumptions, the received
signal vector at time slot t before ADC quantization is
r̄[t] = H̄x̄[t] + z̄[t] ∈ CN ,
(1)
where z̄[t] = [z̄1 [t], z̄2 [t], · · · , z̄ N [t]]⊤ has elements independently drawn from a complex Gaus-
6
sian distribution with zero mean and variance σ 2 . We assume a block fading model in which
the channel is time-invariant during coherence time interval. We denote Td as the duration (the
number of time slots) for data detection. The complex received signal in (1) can be equivalently
rewritten in a real-form as
Re(r̄[t]) Re(H̄) −Im(H̄) Re(x̄[t]) Re(z̄[t])
,
=
+
Im(r̄[t]) Im(H̄) Re(H̄) Im(x̄[t]) Im(z̄[t])
| {z } |
{z
} | {z } | {z }
r[t]
H
x[t]
(2)
z[t]
where Re(A) and Im(A) denote the real and complex parts of a complex matrix A, respectively.
We consider the use of one-bit ADCs at each receive antenna, which implies that the real
and the imaginary components of the received signal are separately quantized to binary levels.
In this paper, sign(·) is the quantization function, which essentially maps a positive value to 1
and a negative value to -1. Then the received signal after the ADCs at time slot t is defined as
y[t] = [y1 [t], y2 [t] · · · , y2N [t]]⊤ ∈ {+1, −1}2N with
yi [t] = sign(ri [t]) = sign hi⊤ x[t] + zi [t] ,
(3)
for all i ∈ I = {1, 2, . . ., 2N}, where hi⊤ is the i-th row of the channel matrix H, and zi [t] is the
i-th element of z[t].
B. Definitions
We provide some definitions that will be used in the sequel.
Definition 1 (Codewords and codebook [13], [14]): We define ck = sign(Hx k ) ∈ {−1, 1}2N
as the k-th (binary) codeword vector corresponding to the k-th symbol vector x k ∈ X 2U . For
k ∈ K = 1, 2, . . ., K = M U , each codeword vector ck = sign(Hx k ) can be interpreted as a
noise-free received signal when a symbol vector x k ∈ X 2U is transmitted via the channel matrix
H. We also define a codebook by a collection of codeword vectors, i.e., C = {c1, c2, . . . , cK }.
Note that similar notions for the codeword vectors and the codebook are considered in [13],
[14].
Definition 2 (Weighted Hamming distance between codewords): Let ck = [ck,1, ck,2, · · · , ck,2N ]⊤
and c j = [c j,1, c j,2, · · · , c j,2N ]⊤ be binary codeword vectors in a codebook C. In addition, let
w = [w1, w2, · · · , w2N ]⊤ ∈ R2N and w̃ = [w̃1, w̃2, · · · , w̃2N ]⊤ ∈ R2N be weight vectors that consist
7
of positive elements. The weight vector w is assigned when measuring the distance between the
elements of ck and c j that have different signs. Whereas, the weight vector w̃ is assigned when
measuring the distance between the elements of ck and c j that have the same sign. Then, the
weighted Hamming distance between ck and c j with respect to w and w̃ is defined as
dw (ck , c j ; w, w̃) =
2N
Õ
i=1
2N
Õ
wi kck,i −c j,i k0 + w̃i (1 − kck,i −c j,i k0 ),
(4)
i=1
where kak0 is the zero norm that denotes the number of nonzero elements in a vector a. Note
that when w̃i > 0 for any i, the weighted Hamming distance between two same codewords can
be non-zero.
III. MLD
FOR
MIMO S YSTEM
WITH
O NE -B IT ADC S
In this section, we first review MLD for uplink massive MIMO systems with one-bit ADCs.
We then show that the MLD is equivalent to MWD by leveraging the weighted Hamming distance
defined in Section II-B. We finally develop a new MWD method which tightly approximates
the MWD that is equivalent to the MLD. The developed MWD will be used as a baseline for a
low-complexity near-MLD algorithm in Section IV.
A. Maximum-Likelihood Detection (MLD)
We present the MLD for uplink multi-user MIMO systems with one-bit ADCs that was
originally introduced in [17]. Let p (y[t]|x k ) be the likelihood function with the received signal
y[t] when the k-th symbol vector, x k ∈ X 2U , was sent at time slot t. Then, p (y[t]|x k ) is given
by [17]
where Q(x) =
∫∞
x
p (y[t]|x k ) =
2N
Ö
i=1
p (yi [t]|x k ) =
2N
Ö
i=1
1−Q
r
2
yi [t]hi⊤ x k
2
σ
!!
,
(5)
2
t
√1 e− 2
2π
dt is a standard Q-function. Using (5), MLD for a MIMO system with
one-bit ADCs is represented by the following optimization problem:
x̂MLD [t] = argmax p (y[t]|x k )
xk ∈X 2U
= argmax
xk ∈X 2U
2N
Ö
i=1
1−Q
(6)
r
2
yi [t]hi⊤ x k
2
σ
!!
.
(7)
8
B. Minimum Weighted-Hamming-Distance Detection (MWD) as an Exact MLD
We show that the MLD is equivalent to MWD. To this end, we demonstrate that the loglikelihood function can be rewritten in the form of the weighted Hamming distance in (4). Let
Nke [t] = {i : yi [t] , ck,i } be the index set of the received signal elements that have different
signs with the elements of the k-th codeword vector under the premise that x k ∈ X 2U was sent
at time slot t. Using this index set, we rewrite the likelihood function in (5) as
!
!!
r
r
Ö
Ö
2
2
Q −
p (y[t]|x k ) =
1−Q
y [t]hi⊤ x k
y [t]hi⊤ x k ,
2 i
2 i
σ
σ
i∈N e [t]
i<N e [t]
k
(8)
k
where the equality is obtained by applying the property of the Q-function: Q(x) = 1 − Q(−x).
Then we take the logarithm of (8) which yields
!
!!
r
r
Õ
Õ
2
2
⊤
⊤
yi [t]hi x k +
ln 1 − Q
yi [t]hi x k .
ln Q −
ln (p (y[t]|x k )) =
σ2
σ2
i<N e [t]
i∈N e [t]
(9)
k
k
To simplify, we define two weights
w′k,i
w̃′k,i
and
r
w′k,i , − ln Q −
and
w̃′k,i , − ln 1 − Q
as
!
2
yi [t]hi⊤ x k > 0,
σ2
r
2
yi [t]hi⊤ x k
σ2
!!
> 0.
(10)
(11)
Using these weights, the log-likelihood function in (9) can be expressed as
Õ
Õ
′
wk,i −
w̃′k,i
ln (p (y[t]|x k )) = −
i∈Nke [t]
=−
2N
Õ
w′k,i
i=1
i<Nke [t]
yi [t] − ck,i
where the last equality holds because
0
−
2N
Õ
i=1
w̃′k,i (1 − yi [t] + ck,i 0 ),
(12)
1, i ∈ Nke [t],
0, i < Nke [t].
From Definition 2 in Section II-B, the log-likelihood function is expressed in the form of the
yi [t] − ck,i
0
=
weighted Hamming distance:
ln (p (y[t]|x k )) = −dw y[t], ck ; w′k, w̃′k .
(13)
9
By applying the result in (13) to the definition of MLD, we can show that the MWD is equivalent
to the MLD:
x̂MLD [t] = argmax ln (p (y[t]|x k )) = argmin dw y[t], ck ; w′k, w̃′k .
(14)
xk ∈X 2U
xk ∈X 2U
C. MWD as a Near MLD
Based on the MLD representation in (14), we develop a new MWD method that provides a
near-MLD solution. The key idea of the developed MWD is to use closed-form weight vectors
that tightly approximate the weight vectors in (14). In this approximation, we adopt a Q-function
approximation in [27] which demonstrates that Q̂(x) = 12 e−0.374x
2 −0.777x
tightly approximates the
Q-function for non-negative x with the absolute error less than 10−3 , i.e.,
|Q(x) − Q̂(x)| ≤ 10−3, for x ≥ 0.
By applying Q(x) ≈ 21 e−0.374x
2 −0.777x
(15)
for x ≥ 0 to both (10) and (11), we obtain two closed-form
weights wk,i and w′k,i that approximate the original weights w′k,i and w̃′k,i , respectively, i.e.,
√
b
2 ⊤
2a
|h x k | + ln 2 > 0,
(16)
w′k,i ≈ wk,i , 2 |hi⊤ x k | 2 +
σ i
σ
w̃′k,i ≈ w̃k,i , − ln (1 − e−wk,i ) > 0,
(17)
where a = 0.374 and b = 0.777. By defining w k = [wk,1, · · · , wk,2N ]⊤ and w̃ k = [w̃k,1, · · · , w̃k,2N ]⊤ ,
we also obtain an approximation of the weighted Hamming distance in (14) as
dw y[t], ck ; w′k, w̃′k ≈ dw (y[t], ck ; w k, w̃ k ) ,
(18)
for k ∈ K. By leveraging the above approximation, we develop the MWD method that has
closed-form weights, unlike the MWD in (14). The detection rule for the developed MWD
method is given by
x̂MWD [t] = argmin dw (y[t], ck ; w k, w̃ k ) ,
(19)
xk ∈X 2U
where two weight vectors in (19) can be computed at the BS from (16) and (17) when CSIR is
available. The developed MWD in (19) is expected to provide near-MLD solution because the
developed MWD tightly approximates the MWD in (14) which has been shown to be equivalent
to the MLD in the previous subsection. The main advantage of the developed MWD is that it
10
does not require the Q-function calculation which necessarily relies on a mapping table for the
computation in a practical system.
Remark 1 (Comparison to MLD for a conventional MIMO system): For a conventional
MIMO system where the input-output relation is linear under Gaussian noise, the optimal MLD
is equivalent to a minimum Euclidean distance detection. For a MIMO system with one-bit ADCs
where the input-output relation is non-linear, minimizing the weighted Hamming distance obtains
near-MLD performance. Specifically, in the MIMO system with one-bit ADCs, the weighted
Hamming distance is measured between noisy and quantized received signal, y[t], and noisefree but quantized received signal, ck = sign (Hx k ) for k ∈ K. When measuring the distance
between the i-th elements of y[t] and ck , different weights are assigned by taking into account
both 1) the sign alignment between yi [t] and ck,i and 2) the reliability information provided by
a received SNR
2
|h⊤
i xk |
.
2
σ
More precisely, if yi [t] , ck,i , the weight wk,i in (16) is assigned which
is an increasing function of the received SNR. Whereas, if yi [t] = ck,i , the weight w̃k,i in (17)
in assigned which is a decreasing function of the received SNR.
Remark 2 (High SNR regime): When the received SNR is sufficiently large, w̃k,i approaches
zero. This fact implies that when computing the weighted distance at high SNR, the receiver can
ignore the elements of y[t] and ck that have different signs. In this case, the weighted Hamming
Í2N
distance is computed as i=1
wk,i yi [t]−ck,i 0 . This motivates us to further simplify the detection
rule for MWD as follows:
x̂MWD [t] ≈ argmin dw (y[t], ck ; w k, 0) = argmin
k∈K
k∈K
2N
Õ
i=1
wk,i yi [t] − ck,i
0
.
(20)
As can be seen in (20), the computational complexity of the MWD in (19) can be reduced
because the weights are computed only for the elements corresponding to yi [t] , sign(hi⊤ x k ).
IV. O NE -B IT S PHERE D ECODING
In this section, based on the MWD in Section III-C, we propose a low-complexity near-MLD
algorithm for uplink massive MIMO systems with one-bit ADCs, referred to as one-bit sphere
decoding (OSD). We also compare the computational complexity of the proposed OSD with
those of the MLD and the MWD, to show a significant reduction in the complexity achieved by
the proposed OSD.
11
A. Proposed Algorithm
The key idea of the OSD is to construct a list of codeword vectors in the sphere for each
possible received signal and then to perform the MWD only over the codeword list in the sphere.
The major differences of the OSD to conventional sphere decoding algorithms in [21], [23]–[26]
are two folds:
•
The list of codeword vectors in the sphere is constructed using preprocessing based on
CSIR. This preprocessing is only possible when the BS receives the signal vector in a finite
set due to the use of one-bit ADCs. This differs from the conventional sphere decoding
algorithms in which the codeword list is constructed during data detection processing.
•
The OSD measures the weighted Hamming distance when constructing the codeword list,
the conventional sphere decoding algorithms measure the Euclidean distance when finding
the codewords in the sphere.
The OSD consists of two parts: list construction in the sphere and detection over the sphere.
Detailed procedures of each part are given below.
List construction in the sphere: The receiver constructs and saves a list of codeword vectors
in the sphere for each received signal, using preprocessing based on CSIR. This list contains the
indices of the codeword vectors that are close to the received signal in terms of the weighted
Hamming distance. If the codeword list is constructed for all possible received signals, i.e., {y ∈
{−1, +1}2N }, the receiver requires to construct total 22N codeword lists. Then the complexity of
the list construction could not be affordable in a practical system when N is large. To resolve this
problem, our strategy is to divide the received signal into G ≥ 1 sub-vectors, each with dimension
of Ns =
2N
G ,
and then to construct 2 Ns possible sub-lists for each sub-vector y ∈ {−1, +1} Ns in
parallel.
(g)
Let y p ∈ {−1, +1} Ns be the p-th possible vector of the g-th sub-vector, where p ∈ {1, 2, . . ., 2 Ns }
and g = {1, 2, . . ., G}. Also let Ig be the index set for the elements of the g-th sub-vector, namely,
Ig = {(g − 1)Ns + 1, (g − 1)Ns + 2, . . ., gNs }, for g ∈ {1, 2, . . ., G}.
(21)
Using this index set, the g-th sub-vector of ck is defined as
⊤
(g)
ck = ck,Ig (1), ck,Ig (2), · · · , ck,Ig (Ns ) ,
(22)
12
(g)
while the weight vectors associated with ck are defined as
⊤
⊤
(g)
(g)
w k = wk,Ig (1), wk,Ig (2), · · · , wk,Ig (Ns ) , and w̃ k = w̃k,Ig (1), w̃k,Ig (2), · · · , w̃k,Ig (Ns ) , (23)
respectively. Note that the above weight vectors can be computed at the receiver from (16) and
(g)
g (ℓ,p)
(17) when CSIR is available. Let πg (ℓ, p) ∈ K be an index function indicating that cπ
(g)
is the
ℓ-th closest sub-codeword vector to y p , i.e.,
(g) (g)
(g)
(g)
(g) (g)
(g)
(g)
dw y p , cπ (ℓ,p) ; wπ (ℓ,p), w̃π (ℓ,p) ≤ dw y p , cπ (t,p) ; wπ (t,p), w̃π (t,p) ,
g
g
g
g
g
g
(g)
for t ∈ {ℓ + 1, ℓ + 2, . . ., K}. Then the sub-list associated with y p is determined as the indices
(g)
of the sub-codeword vectors that are the L closest to y p , that is
(g)
Sg y p , L = πg (1, p), πg (2, p), . . ., πg (L, p) ,
(24)
where p ∈ 1, 2, . . . , 2 Ns and g ∈ {1, 2, . . ., G}. The overall list-construction procedure of the
OSD is illustrated in Fig. 2. Note that this procedure is performed only once during a channel
coherence block.
Detection over the sphere: During data detection processing, the receiver estimates the
codeword vector (the transmitted symbol vector) by searching over the list in the sphere generated
during the list-construction process. Specifically, the receiver performs the MWD in (19) over
the list, to find the codeword vector that has the minimum weighted Hamming distance to the
received signal.
When the received signal, y[t], is observed at time slot t, the receiver divides y[t] into G
sub-vectors, namely y(1) [t], y(2) [t], . . ., y(G) [t] . Then for each sub-vector y(g) [t], the receiver
obtains the sub-list of the L nearest sub-codeword vectors, i.e., Sg y(g) [t], L , that is generated
during the list construction process. Using the obtained G sub-lists, the receiver generates a total
list of the codeword vectors as the union of these G sub-lists, i.e.,
G
Ø
(g)
S(y[t]) =
Sg y [t], L .
(25)
g=1
Note that the cardinality of S(y[t]) is bounded between L and GL, i.e., L ≤ |S(y[t])| ≤ GL
where G and L are chosen to be GL ≪ K. Using S(y[t]), the receiver finds the index of the
transmitted symbol vector by applying the detection rule for the MWD in (19) over S(y[t]), i.e.,
⋆
[t] = argmin dw (ck , y[t]; w k, w̃ k ) .
kOSD
k∈S(y[t])
(26)
13
Channel
Estimation
Detection over the Sphere
Estimated Channel Matrix ۶
Memory
Estimated
Channel Matrix ۶
Memory
List Construction in the Sphere
Sub-Codeword Vector Generation
Codeword
Vector
Generation
Sub-List Construction
in the Sphere
Detection
Over
the Sphere
Memory
Received signal
at time slot ݊
Sub-List Construction
in the Sphere
Sub-vector
Generation
Sub-List Construction
in the Sphere
Fig. 2. The proposed OSD consists of two parts: 1) list construction in the sphere and 2) detection over the sphere, when N = 6,
Ns = 4, and G = 3.
Once the best index is found, the receiver obtains the estimate of the transmitted symbol vector
⋆ [t] . The overall detection procedure of the OSD is depicted in Fig. 2.
x̂OSD [t] = x kOSD
We present a simple example to illustrate the operation of the OSD.
Example 1: Suppose a case in which U = 2, N = 2, Ns = 2, L = 1, and BPSK modulation
per user is assumed. We also consider a channel matrix that is given by
0.8 0.2 h⊤
1
⊤
0.1 0.9 h2
=
H =
⊤ .
−0.7 0.3 h3
0.4 −0.6 h⊤
4
(27)
The receiver first constructs the list in the sphere. In this example, all possible transmit symbol
vectors sent by the two uplink users are
−1
−1
1
1
(28)
x1 = , x2 = , x3 = , x4 = .
−1
1
−1
1
Since ck = sign(Hx k ) and Ns = 2, the receiver generates four sub-codeword vectors as follows:
1 (1) 1 (1) −1 (1) −1
(1)
(29)
c1 = , c2 = , c3 = , c4 = ,
−1
1
−1
1
14
and
−1 (2) −1 (2) 1 (2) 1
= , c2 = , c3 = , c4 = .
1
−1
1
−1
The receiver also generates four possible sub-vectors for the received signal:
1 (g) 1 (g) −1 (g) −1
(g)
y1 = , y2 = , y3 = , y4 = ,
−1
1
−1
1
(2)
c1
(30)
(31)
(g)
where y p is the p-th possible vector in a finite set {−1, +1} Ns and g = {1, 2}. Because we
(g)
consider L = 1 case, the sub-list for y p contains only one index of the sub-codeword vector
(g)
(g)
cπ(1,p) that has the minimum weighted Hamming distance from y p . Suppose that eight sub-lists
are
(1)
(1)
(1)
(1)
S1 y1 , 1 = {1} , S1 y2 , 1 = {2} , S1 y3 , 1 = {3} , S1 y4 , 1 = {4} ,
(2)
(2)
(2)
(2)
S2 y1 , 1 = {4} , S2 y2 , 1 = {3} , S2 y3 , 1 = {2} , S2 y4 , 1 = {1} .
The receiver constructs the codeword list in the sphere once for a channel coherence block.
Now, the receiver performs the detection over the codeword list. Suppose that at time slot
t, the receiver observes a signal vector given by y[t] = [1, −1, −1, 1]⊤ . Then two sub-vectors
corresponding to y[t] are y(1) [t] = [1, −1]⊤ and y(2) [t] = [−1, 1]⊤ . Using these two sub-vectors,
the receiver determines the codeword list in the sphere for y[t] as the union of S1 y(1) [t], 1
(2)
and S2 y [t], 1 , i.e.,
S(y[t]) =
G
Ø
g=1
(g)
Sg y [t], 1 = {2, 3} .
(32)
Then the receiver finds the index of the transmitted symbol vector by applying the detection rule
for the MWD in (19) over the codeword list in S(y[t]):
⋆
kOSD
[t] = argmin dw (ck , y[t]; w k, w̃ k ) .
(33)
k∈{2,3}
⋆ [t] .
Finally, the receiver obtains the estimate of the transmitted symbol vector as x̂OSD [t] = x kOSD
In this example, under the premise that each y[t] ∈ {−1, +1}2N is generated with the equal
probability, the average number of codeword search for the proposed OSD is
Õ
1
|S(y[t])| = 1.75.
22N
2N
y[t]∈{−1,+1}
(34)
15
Therefore, the OSD achieves a 56% reduction of the computational complexity compared to the
MWD which computes four different weighted Hamming distances for every y[t] ∈ {−1, +1}2N .
Remark 3 (The interplay between Ns and L): The dimension of sub-vector Ns and the
list size L determine the tradeoff between the detection performance and the computational
complexity of OSD. If we set Ns to be large, the size of the sub-list, L, can be reduced because
(g)
(g)
(g) (g)
(g)
(g)
the weighted Hamming distance between y p and cπ(ℓ,p), i.e., dw y p ,cπ(ℓ,p);wπ(ℓ,p), w̃π(ℓ,p) for
(g)
large Ns provides enough information to reliably find the best codeword in the set S(y p , L)
with a small number of L. Whereas, if we set Ns to be small, the weighted Hamming distance
(g)
(g)
between y p and cπ(ℓ,p) does not provide reliable information to correctly find the best codeword
(g)
in the set S(y p , L). Therefore, in this case, we need to choose a large size of L to improve
the detection performance. Note that one can also modify the algorithm by choosing a different
dimension of Ns per sub-vector to further optimize the tradeoff between the detection performance
and the computational complexity of the OSD.
Remark 4 (Extension to multi-precision ADCs): The proposed OSD can be extended for the
case with multi-precision ADCs. Suppose that a B-bit scalar quantizer is independently applied
to the real and imaginary parts of the received signal, while Y = {q1, q2, . . ., q2B } is the set of
all possible outputs of the quantizer, and SQ : R → Y is the quantization function of the scalar
quantizer. In [12], it is shown that the log-likelihood function of this system is given by
2N
Õ
l(yi [t]) − hi⊤ x k
u(yi [t]) − hi⊤ x k
ln (p (y[t]|x k )) =
ln Q
−Q
,
σ/2
σ/2
i=1
(35)
where u(y) and l(y) are the upper and the lower bin boundaries associated with the quantized
output y ∈ Y. Define ck = SQ(Hx k ) as the k-th codeword vector associating with the k-th
symbol vector. Then, similar to (9), the log-likelihood function in (35) is expressed as
Õ
l(ck,i ) − hi⊤ x k
u(ck,i ) − hi⊤ x k
ln Q
ln (p (y[t]|x k )) =
−Q
σ/2
σ/2
e
i<Nk [t]
+
Õ
i∈Nke [t]
≤−
Õ
i<Nke [t]
l(yi [t]) − hi⊤ x k
u(yi [t]) − hi⊤ x k
ln Q
−Q
σ/2
σ/2
w̃′k,i −
Õ
w′k,i
i∈Nke [t]
= −dw y[t], ck ; w′k, w̃′k ,
(36)
16
where Nke [t] = {i : yi [t] , ck,i }, w′k = [wk,1, wk,2, · · · , wk,2N ]⊤ with
l(y) − hi⊤ x k
u(y) − hi⊤ x k
′
wk,i = − max ln Q
−Q
,
σ/2
σ/2
y∈Y, y,ck,i
and w̃′k = [w̃k,1, w̃k,2, · · · , w̃k,2N ]⊤ with
u(ck,i ) − hi⊤ x k
l(ck,i ) − hi⊤ x k
′
−Q
.
w̃k,i = − ln Q
σ/2
σ/2
(37)
(38)
Motivated by the inequality in (36), the MLD of the MIMO systems with multi-precision ADCs
can be approximated as
x̂MLD [t] = argmax ln (p (y[t]|x k )) ≈ argmin dw y[t], ck ; w′k, w̃′k .
(39)
xk ∈X 2U
xk ∈X 2U
Except for the definition of the weight vectors, the detection rule in (39) is exactly the same
with the detection rule in (14). Therefore, the proposed OSD can also be extended to the
MIMO systems with multi-precision ADCs, simply by using the weights in (37) and (38) when
computing the weighted Hamming distance. Note that although the detection rule in (39) is an
approximate MLD, it achieves the exact MLD as the number of precision bits at the ADCs
decreases.
Remark 5 (Extension to frequency-selective channels): Although the proposed OSD is
developed under the assumption of frequency-flat channels, it can also be applied to frequencyselective channels with some modifications. Suppose that the number of channel-impulse-response
(CIR) taps of the channel is given by L ≥ 1. For this channel, consider multiple block transmissions; each consists of B successive data symbols followed by L − 1 zeros at the end. Then the
received signal vector at the t-th time slot of the b-th block transmission is expressed as
!
L−1
Õ
yb [t] = sign
H[ℓ]xb[t − ℓ] + zb [t] ,
(40)
ℓ=0
where H[ℓ] ∈ R2N×2U is the real-domain channel matrix that consists of the ℓ-th CIR taps, and
xb[t] ∈ R2U and zb[t] ∈ R2N are the transmitted symbol and the noise at the t-th time slot of the
17
TABLE I
T HE NUMBER OF REAL MULTIPLICATIONS REQUIRED FOR VARIOUS DETECTION METHODS WHEN N ≫ 1.
Detection method
Preprocessing
Data detection processing
MLD in (7)
-
(4U + 6)NKTd
MWD in (19)
-
(4U + 14)NKTd
OSD in (26)
2 Ns (4U + 14)NK
2N L (4U
Ns
+ 14)NTd
b-th block transmission, respectively. By concatenating the received signals during Tb + L − 1
time slots, the total received signal vector of the b-th block transmission is given by
y[1]
H[0]
©
H[1]
y[2]
.
..
= sign ..
.
0
y[B + L − 2]
y[B + L − 1]
« 0
0
H[0]
..
.
···
H[L − 1] H[L − 2]
···
0
H[L − 1]
···
..
.
..
.
0
0
..
.
x[1] z[1] ª
®
x[2] z[2] ®®
®
..
..
®
+
® .
.
.
®
x[B − 1] z[B − 1] ®®
®
x[B] z[B]
¬
(41)
The received signal in (41) is equivalent to the received signal of an uplink MIMO system that
has UB uplink users and N(B + L − 1) receive antennas at the BS. Therefore, the proposed OSD
is directly applicable to frequency-selective channels by assuming that there exists UL uplink
users and N(B + L − 1) receive antennas. Note that the above extension requires a significant
detection complexity when both U and B are large, even for the proposed OSD; thereby, as
future work, it would be interesting to develop a lower complexity method for the use in a
practical system with frequency-selective channels.
B. Computational Complexity Comparison
We compare the computational complexity of three detection methods: MLD, MWD, and
OSD. To this end, we compute the number of real multiplications required for each method,
which is summarized in Table I. Specifically, for the OSD, we consider the worst case in which
18
the size of the codeword list in the sphere is maximized, i.e., GL =
2N
Ns L,
as can be seen
from (25). Table I shows that the number of real multiplications required for the proposed OSD
L
2 Ns
2N L
2 Ns
4U+14
+
of
that
for
the
MLD
and
+
is 2N
×
of that for the MWD, even in the
Ns K
Td
4U+6
Ns K
Td
worst case. These results imply that if two design parameters of the OSD, L and Ns , are properly
set, the OSD has a less detection complexity than both the MLD and the MWD do. Therefore,
by setting L ≪ K and Ns ≪ log2 Td , the proposed OSD achieves a significant reduction in the
detection complexity compared to both methods. Note that a similar result also holds for the
comparison of the numbers of real additions.
V. D ETECTION P ERFORMANCE
OF
O NE -B IT S PHERE D ECODING
In this section, we analyze the detection performance of the proposed OSD by characterizing
an upper bound of the probability that the proposed OSD performs worse than the MWD.
We first demonstrate that this probability is upper bounded by the sphere-list-error-probability
(SEP), which is the probability that the index of the transmitted codeword does not belong to
the constructed list in the sphere.
Let Ploss be the probability that the detection error occurs using the proposed OSD while the
detection is correct using the MWD. Then Ploss is expressed as
Ploss =
K
Õ
k=1
=
K
Õ
k=1
+
Pr (x̂MWD [t] = x k, x̂OSD [t] , x k , x[t] = x k )
Pr (x̂MWD [t] = x k, x̂OSD [t] , x k , k ∈ S(y[t]), x[t] = x k )
K
Õ
k=1
Pr (x̂MWD [t] = x k , x̂OSD [t] , x k, k < S(y[t]), x[t] = x k ) .
(42)
By the detection rule for the OSD in (26), if the transmitted codeword index does not belong to the
codeword list inside of the sphere, the OSD fails to detect the correct symbol vector; thereby, the
error event {k < S(y[t]), x[t] = x k } is a subset of the event {x̂OSD [t] , x k , x[t] = x k }. In addition,
if the MWD finds the transmitted symbol index, the OSD also finds the transmitted symbol index,
provided that this index is in the codeword list in the sphere; thereby, the intersection of two
19
event sets x̂MWD [t] = x k } and {x̂OSD [t] , x k , k ∈ S(y[t])} is an empty set. From this fact, we
rewrite Ploss in (42) as
Ploss =
K
Õ
k=1
≤
Pr (x̂MWD [t] = x k, k < S(y[t]), x[t] = x k )
K
Õ
k=1
Pr (k < S(y[t]), x[t] = x k ) = PSEP,
(43)
where PSEP is the SEP.
Now, we characterize the upper bound of the SEP using the following theorem.
Theorem 1. For a fixed channel matrix H ∈ R2N×2U , SEP of the proposed OSD is
PSEP =
K
Õ
k=1
Pr (k < S (y[t]), x[t] = x k )
K Ns
Õ
1 ÕÖ
⊤ (g)
⊤ (g)
exp −e w k − (1 Ns − e) w̃ k ,
/
K k=1 g=1 (g)
2N
(44)
e∈Ek (L)
where
(g)
E k (L)
n
= e:
(g)
dmin,k (e,
L) ≤
(g)
1⊤Ns w̃ k , e
∈ {0, 1}
Ns
o
(45)
,
o
n
(g)
(g)
(g) (g)
(g)
(g)
(g)
(g)
, dk, j = dw c j , ck ; w j , w̃ j , and
dmin,k (e, L) is the L-th smallest element of dk, j + e⊤ ∆k, j
j,k
(g)
(g)
(g)
(g)
(g)
(g)
N
s
∆k, j ∈ R is a vector whose i-th element is ∆k, j,i = w j,i − w̃ j,i ck,i c j,i − wk,i − w̃k,i .
Proof: See Appendix A
From Theorem 1, we can show that the SEP of the proposed OSD decreases with both the
dimension of a sub-vector, Ns , and the size of a sub-list in the sphere, L. To see this, it should
(g)
first be noticed that the upper bound in (44) decreases with dmin,k (e, L) because the size of a set
(g)
(g)
E k (L) is reduced by increasing dmin,k (e, L). This parameter can be shown to be an increasing
(g)
(g)
function of both Ns and L; first, dmin,k (e, L) is defined as the L-th smallest value, so dmin,k (e, L)
increases with L; next, increasing the dimension of each sub-codeword vector increases the term
(g)
(g)
(g)
dk, j in the definition of dmin,k (e, L), so dmin,k (e, L) also increases with Ns .
We also present a numerical example to show the tightness of the approximate upper bound
derived in Theorem 1.
20
100
10-1
10-2
10-3
10-4
10-5
-3
0
3
6
9
12
Fig. 3. Comparison of the SEP obtained by simulations with the approximate upper bound calculated from (44) when U = 2,
N = 8, 4-QAM is adopted, and CSIR is perfect.
Example 2: In Fig. 3, we compare the SEP obtained by simulations with the approximate
upper bound of the SEP calculated by (44) when U = 2, N = 8, and 4-QAM is used. Simulation
results are averaged over 5000 random realizations of channel coefficients that are independently
drawn from a complex Gaussian distribution with zero mean and unit variance. Channel stat
information at the receiver (CSIR) is assumed to be perfect. Fig. 3 shows that the approximate
upper bound is very tight to the simulated SEP regardless of the values of Ns and L; thereby, this
result validates our analysis in Theorem 1. Another important observation in Fig. 3 is that the
SEP of the proposed OSD decreases with both Ns and L, as we have expected from Theorem 1.
Specifically, it is shown that the SEP obtained when (N s; L) = (8; 4) is significantly lower than
the SEP obtained when (N s; L) = (4; 2), while the computational complexity of the former case
is only 12.5% higher than that of the latter case when Td = 4096 (see Table I). This observation
implies that the determination of Ns and L has a considerable impact on the tradeoff between
the SEP and the computational complexity when using the proposed OSD.
VI. S IMULATION R ESULTS
In this section, using simulations, we evaluate the detection performance of the proposed OSD
for uplink massive MIMO systems with one-bit ADCs. All simulation results are averaged over
5000 random realizations of channel coefficients that are independently drawn from a complex
Gaussian distribution with zero mean and unit variance.
21
10 0
Uncoded SER
U=3, N=28, 16-QAM
10
-1
10
-2
10
-3
U=6, N=32, 4-QAM
U=12, N=100, BPSK
Proposed, OSD (Varying Ns , L = 32)
Proposed, WMD
Conventional, MLD
10 -4
-20
-15
-10
-5
0
5
10
15
20
SNR ρ [dB]
Fig. 4. The SER vs. SNR of the proposed OSD, the proposed MWD, and the conventional MLD for various U, N, and
constellation sets with perfect CSIR.
A. Uncoded Performance
We evaluate the detection performance of the proposed OSD for an uncoded system. For a
comparison, we also present the performances of the conventional MLD in (7) and the proposed
MWD in (19).
Fig. 4 compares the symbol-error-rate (SER) of the OSD with those of the MLD and the
MWD when CSIR is perfect. For the OSD, the dimension of the sub-codeword vector is set
as Ns = 7, 8, 10 for N = 28, 32, 100, respectively. Fig. 4 shows that the OSD has a negligible
SER loss compared to the MLD regardless of the number of receive antennas, the number of
users, the constellation set, and the SNR. Meanwhile, the OSD reduces the detection complexity
(i.e., the number of real multiplications) of the MLD by 89%, 88%, 68% when N = 28, 32, 100,
respectively (for Td = 8192 case, see Table I). It is also noticeable that this complexity reduction
further increases as Td increases. These results show that the proposed OSD provides a good
performance-complexity tradeoff for the uplink massive MIMO systems with one-bit ADCs.
Although the MWD shows almost the same SER performance to the MLD, it does not provide
any reduction in the detection complexity as seen in Table I.
Fig. 5 compares the SER of the OSD with those of the MLD and the MWD when pilot-based
channel estimation is applied with various lengths of pilot signals (i.e., Tt = 20 or 100). For the
channel estimation, we apply a ML-based estimation method developed in [17]. Fig. 5 shows
22
10 0
Proposed, OSD (Ns = 8, L = 32)
Proposed, WMD
Conventional, MLD
10 -1
T =20
Uncoded SER
t
10
-2
10
-3
T =100
t
10 -4
-10
-5
0
5
10
SNR ρ [dB]
Fig. 5. The SER vs. SNR of the proposed OSD, the proposed MWD, and the conventional MLD when pilot-based channel
estimation is applied with various lengths of pilot signals for the case of U = 6, N = 32, and 4-QAM modulation.
1
0.9
Relative SER Performance
0.8
L=128
0.7
128
64
48
32
24
L=16
0.6
0.5
0.4
L=16
0.3
Ns
Ns
Ns
Ns
0.2
0.1
= 8,
= 4,
= 8,
= 4,
SNR
SNR
SNR
SNR
=
=
=
=
-5 dB
-5 dB
5 dB
5 dB
0
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Relative Detection Complexity
Fig. 6. The performance-complexity tradeoff achieved by the proposed OSD for various Ns and L with U = 6, N = 32, 4-QAM
modulation, and perfect CSIR.
that the OSD achieves a near-optimal SER with a reduced detection complexity, regardless of the
length of pilot signals. This result implies that the improvement of the performance-complexity
tradeoff achieved by the OSD is also robust to the channel estimation error.
Fig. 6 plots the performance-complexity tradeoff achieved by the OSD for various dimensions
of a sub-codeword vector, Ns , and also for various sizes of a sub-list in the sphere, L, when
CSIR is perfect. The relative SER performance in the y-axis is computed as the ratio of the SER
achieved by the MLD to that achieved by the OSD, while a relative complexity in the x-axis is
computed as the ratio of the number of real multiplications required by the OSD to that required
23
by the MLD. Fig. 6 shows that as L increases, both the relative SER performance and the relative
complexity increase; this result shows the tradeoff relation between the performance and the
complexity when using the OSD. In addition, the detection performance of the OSD with Ns = 8
is much higher than that with Ns = 4, which implies that the performance-complexity tradeoff
is very sensitive to the choice of Ns . Another interesting observation is that the complexity
required to achieve the optimal SER performance reduces as the SNR increases. Based on this
observation, the effectiveness of the OSD can be improved with the operating SNR of the system.
B. Coded Performance
We also evaluate the detection performance of the proposed OSD for a coded system. As an
underlying channel code, we adopt a 1/2-rate LDPC code of the blocklength NB = 672 from the
IEEE 802.11ad standardization [29]. For the conventional MLD, we only employ a hard-input
bit-flipping decoder [30] as in [17], since they produce hard-decision outputs. For both the MWD
and the OSD, besides the hard-decision outputs, we also derive soft outputs using the technique
in [28], which enables to use a soft-input belief-propagation decoder [31]. For completeness, we
briefly explain how to compute soft outputs from the hard-decision measurement y[t] for the
proposed OSD. Without loss of generality, we only focus on the u-th channel decoder to decode
the user u’s message. Recall that S(y[t]) contains the all codewords in the search-space. We first
partition the S(y[t]) into the four subsets which are defined as
S u (y[t]|i) = {k ∈ S(y[t]) : x̄k,u = X̄(i)},
(46)
for i ∈ {1, 2, 3, 4}, where x̄k,u denotes the u-th element of the x̄ k and X̄(i) denotes the i-th element
of the 4-QAM constellation set X̄. Using this definition, the log-likelihood ratios (LLRs), i.e.,
the inputs of the belief-propagation decoder, are computed as
u
L2n−1
(y[t]) =
min
k∈S u (y[t]|2)∪S u (y[t]|3)
−
u
L2n
(y[t]) =
min
k∈S u (y[t]|0)∪S u (y[t]|1)
min
k∈S u (y[t]|1)∪S u (y[t]|3)
−
dw (y[t], ck ; w k, w̃ k )
min
dw (y[t], ck ; w k, w̃ k )
dw (y[t], ck ; w k, w̃ k )
k∈S u (y[t]|0)∪S u (y[t]|2)
dw (y[t], ck ; w k, w̃ k ),
24
10 0
Proposed, MWD, Soft
Proposed, OSD (Ns = 8, L = 32), Soft
Proposed, MWD, Hard
Proposed, OSD (Ns = 8, L = 32), Hard
Conventional, MLD, Hard
FER
10 -1
10 -2
10 -3
-16
-12
-8
-4
0
4
SNR ρ [dB]
Fig. 7. The FER vs. SNR of the proposed OSD, the proposed MWD, and the conventional MLD when either a soft detector
or a hard detector is applied with U = 6, N = 32, 4-QAM modulation, and perfect CSIR.
u
u (y[t]) : n = 1, . . ., N } are
for n ∈ {1, . . . , NB }. The resulting 2NB LLRs {L2n−1
(y[t]), L2n
B
embedded into the soft belief-propagation decoder as the soft inputs.
Fig. 7 compares the frame error rate (FER) of the OSD with those of the MLD and the MWD
when CSIR is perfect. Fig. 7 shows that for both the MWD and the OSD, almost 10-dB FER
reduction is achieved by using the soft-input belief-propagation decoder instead of using the
hard-input bit-flipping decoder. This result implies that the availability of the soft outputs has a
significant impact on the detection performance when using the one-bit ADCs. In this context,
the MWD and the OSD are suitable for the coded MIMO system with one-bit ADCs since soft
outputs are available for both methods. Comparing two methods, the FER gap between them is
less than 1 dB while the OSD reduces the detection complexity of the MWD by 72.6% when
Td = NB (see Table I). Therefore, the OSD provides a better performance-complexity tradeoff
than the MWD also for the coded system.
VII. C ONCLUSION
In this paper we have proposed a new sphere decoding method for an uplink massive MIMO
system with one-bit ADCs. One salient observation we found is that the weighted Hamming
distance should be exploited to construct a list of codewords for sphere decoding due to the
discrete nature of received signals. We have also characterized the performance-complexity
25
tradeoff achieved by the proposed OSD, in terms of its design parameters. Using simulations,
we have shown that the proposed OSD effectively reduces the detection complexity of the MLD,
while achieving near-MLD detection performance for both coded and uncoded systems.
An important direction for future research is to extend the proposed algorithm to frequencyselective channels in order to improve the practicality of the proposed algorithm. Another
interesting extension is to develop a sphere decoding algorithm for an uplink massive MIMO
system that uses low-resolution ADCs beyond one-bit precision. It would also be interesting to
develop a low-complexity list construction method that may further reduce the complexity of
the proposed algorithm.
A PPENDIX A
P ROOF
OF
T HEOREM 1
In this proof, we omit the index t of time slot for ease of exposition. Suppose that the channel
matrix of the system is given by H. Then the weight vectors w k and w̃ k of the proposed OSD
are deterministic vectors from (16) and (17), respectively. In this case, the probability that the
true symbol index does not belong to the codeword list in the sphere is expressed as
K
Õ
k=1
Pr (k < S(y), x = x k ) =
K
1Õ
Pr k < S(y) x = x k ,
K k=1
provided that all possible symbol vectors are transmitted with an equal probability
(47)
1
K.
By the
definition of the codeword list in the sphere given in (25), the pair-wise probability in (47) is
G
G
Ø
©
ª Ö
(g)
(g)
Pr (k < S(y)|x = x k ) = Pr k <
Sg y , L x = x k ® =
Pr k < Sg y , L x = x k ,
g=1
g=1
«
¬
(48)
where the equality in (48) is obtained from the statistical independence of the noise vector in (3).
An event {k < Sg y(g), L } in (48) implies that the maximum weighted Hamming distance of
the sub-codeword vector in Sg y(g), L to y(g) is less than that of the k-th sub-codeword vector.
Using this fact, we rewrite the pair-wise probability in (48) as
Pr k < Sg y(g), L x = x k
)
(
(g)
(g)
(g)
(g)
(g)
(g)
≤ Pr
max dw c j , y(g) ; w j , w̃ j ≤ dw ck , y(g) ; w k , w̃ k x = x k ,
j∈Sg y(g),L
(49)
26
Pr k < Sg y(g), L x = xk
)
(
Õ
(g) (g)
(g)
(g)
(g) (g)
(g)
(g)
Pr
max dw c j , y ; w j , w̃ j ≤ dw ck , y ; wk , w̃k x = xk , Ek (e) Pr (Ek (e))
≤
e∈{0,1} Ns
Õ
=
I
e∈{0,1} Ns
(
j ∈Sg y(g), L
max
(g)
(g)
j ∈Sg ck −2e◦ck , L
Ns
Õ
(g)
(g)
(g) (g)
(g)
(g)
w
−
w̃
+
,
c
;
w
,
w̃
c
d
w
j,i
j,i ck,i c j,i ei
j
j
j
k
i=1
≤
Õ
=
I
e∈{0,1} Ns
j ∈Sg
max
(g)
(g)
ck −2e◦ck , L
(g) (g)
(g)
(g)
dw c j , ck ; w j , w̃ j
+
Ns
Õ
(g)
wk,i ei
i=1
(g)
e⊤ ∆k, j
≤
+
(g)
w̃k,i (1
(g)
1⊤Ns w̃k
)
− ei ) x = xk Pr (Ek (e))
x = xk Pr (Ek (e)) .
(42)
where the inequality in (49) is due to the equality condition.
We simplify (49) by introducing the notion of an error vector e = [e1, e2, · · · , e2N ]⊤ ∈ {0, 1} Ns
where ei = 1 represents that the sign of the received signal is flipped due to the noise at the
i-th position. Using the error vector, we denote E k (e) as an event that the received signal for the
g-th sub-vector is given by
(g)
yi
(g)
−ck,i ,
ei = 1,
(g)
ck,i
,
=
ei = 0,
for i ∈ {1, 2, . . ., Ns }.
(50)
(g)
Then we can rewrite (49) as (42) given at the top of this page, where ∆k, j ∈ RNs is a vector
whose i-th element is
(g)
∆k, j,i
=
(g)
(g)
w j,i
−
(g)
w̃ j,i
ck,i c j,i −
(g)
wk,i
−
(g)
w̃k,i
(43)
.
Let Ē k (L) be a set of all vectors in {0, 1} Ns that satisfy the inequality condition of the indicator
function in (42), i.e.,
(g)
where
form:
(g)
dk, j
=
(
Ē k (L) = e :
max
(g)
(g)
j∈Sg ck −2e◦ck ,L
(g)
(g)
(g) (g)
dw c j , ck ; w j , w̃ j
)
Ns
⊤
,
dk, j + e ∆k, j , e ∈ {0, 1}
(g)
(g)
(44)
for all j, k, g. Using this set, (42) is expressed in a simplified
Pr k < Sg y(g), L x = x k ≤
Õ
(g)
e∈Ēk (L)
Pr (E k (e)) .
(45)
27
(g)
We derive the upper bound of (45) by constructing an extended set E k (L) which contains the set
(g)
(g)
(g)
(g)
Ē k (L) in (45) as a subset, i.e., Ē k (L) ⊂ E k (L). For this, let dmin,k (e, L) be the L-th smallest
o
n
(g)
(g)
element of a set dk, j + e⊤ ∆ k, j , j ∈ K \ {k} . Then because
(g)
dmin,k (e, L) ≤
(g)
(g)
⊤
dk, j + e ∆k, j ,
(g)
max
(g)
j∈Sg ck −2e◦ck ,L
(g)
(g)
we can construct the extended set E k (L) of Ē k (L) as
(g)
(g)
Ns
⊤ (g)
E k (L) = e : dmin,k (e, L) ≤ 1 Ns w̃ k , e ∈ {0, 1} .
Using this extended set, we rewrite (45) as
(g)
Pr k < Sg y , L x = x k ≤
Õ
(g)
e∈Ek (L)
Pr (E k (e)) .
(46)
(47)
(48)
Now, the remaining term in (48) is the probability of the event E k (e). By defining a set
I(e) = {i : ei = 1}, this probability can be represented as
Pr (E k (e)) =
=
Ö
Pr(ck,i , yi |x = x k )
i∈I(e)
Ö
Q
i∈I(e)
r
Ö
Pr(ck,i = yi |x = x k )
(
!)
r
Ö
2 ⊤
2 ⊤
1−Q
|h x k |
|h x | .
2 i k
σ2 i
σ
i<I(e)
!
i<I(e)
(49)
From Lemma 1, we can approximate the right-hand-side of (49) using two weights wk,i and w̃k,i
as follows:
Õ
© Õ
ª
(g)
(g)
Pr (E k (e)) ≈ exp −
wk,i −
w̃k,i ® = exp −e⊤ w k − (1 Ns − e)⊤ w̃ k .
(50)
i<I(e)
« i∈I(e)
¬
It is noticeable that the approximation used in (50) is tight because the Q-function approximation
in Lemma 1 has a bounded error less than 10−3 . Plugging (50) into (48) and then applying the
result to (48) and (47) yields the approximate upper bound in (44); this completes the proof.
R EFERENCES
[1] B. Murmann, “ADC performance survey 1997-2016,” [Online]. Available: http://web.stanford.edu/∼murmann/adcsurvey.html
[2] R. H. Walden, “Analog-to-digital converter survey and analysis,” IEEE J. Sel. Areas Commun., vol. 17, no. 4, pp. 539–550,
Apr. 1999.
[3] J. A. Nossek and M. T. Ivrlac, “Capacity and coding for quantized MIMO systems,” in Proc. IEEE Int. Wireless Commun.
Mobile Computing Conf. (IWCMC), Washington, DC, July 2006.
28
[4] A. Mezghani and J. Nossek, “On ultra-wideband MIMO systems with 1-bit quantized outputs: Performance analysis and
input optimization,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Nice, France, June 2007.
[5] J. Singh, O. Dabeer, and U. Madhow, “On the limits of communication with low-precision analog-to-digital conversion at
the receiver,” IEEE Trans. Commun., vol. 57, no. 12, pp. 3629–3639, Dec. 2009.
[6] J. Mo and R. W. Heath, Jr., “Capacity analysis of one-bit quantized MIMO systems with transmitter channel state
information,” IEEE Trans. Signal Process., vol. 63, no. 20, pp. 5498–5512, Oct. 2015.
[7] N. Liang and W. Zhang, “Mixed-ADC massive MIMO,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 983–997, Apr.
2016.
[8] S. Jacobsson, G. Durisi, M. Coldrey, U. Gustavsson, and C. Studer, “Throughput analysis of massive MIMO uplink with
low-resolution ADCs,” IEEE Trans. Wireless Commun., vol. 16, no. 6, pp. 4038–4051, Apr. 2017.
[9] O. Dabeer and U. Madhow, “Channel estimation with low-precision analog-to-digital conversion,” in Proc. IEEE Int. Conf.
Commun. (ICC), Cape Town, South Africa, May 2010.
[10] S. Wang, Y. Li, and J. Wang, “Convex optimization based multiuser detection for uplink large-scale MIMO under lowresolution quantization,” in Proc. IEEE Int. Conf. Commun., June 2014.
[11] S. Wang, Y. Li, and J. Wang, “Multiuser detection in massive spatial modulation MIMO with low-resolution ADCs,” IEEE
Trans. Wireless Commun., vol. 14, no. 4, pp. 2156–2168, Apr. 2015.
[12] C. Studer and G. Durisi, “Quantized massive MU-MIMO-OFDM uplink,” IEEE Trans. Commun., vol. 64, no. 6, pp.
2387–2399, June 2016.
[13] S.-N. Hong, S. Kim, and N. Lee, “Uplink multiuser massive MIMO systems with low-resolution ADCs: A coding-theoretic
approach,” arXiv:1704.03287 [cs.IT], Apr. 2017. [Online]. Available: http://arxiv.org/abs/1704.03287
[14] Y.-S. Jeon, S.-H. Hong, and N. Lee, “Supervised-learning-aided communication framework for massive MIMO systems
with low-resolution ADCs,” arXiv:1610.07693 [cs.IT], Apr. 2017. [Online]. Available: http://arxiv.org/abs/1610.07693
[15] C.-K. Wen, C.-J. Wang, S. Jin, K.-K. Wong, and P. Ting, “Bayes-optimal joint channel-and-data estimation for massive
MIMO with low-precision ADCs,” IEEE Trans. Signal Process., vol. 64, no. 10, pp. 2541–2556, May 2016.
[16] C. Risi, D. Persson, and E. G. Larsson, “Massive MIMO with 1-bit ADC,” arXiv:1404.7736 [cs.IT], Apr. 2014. [Online].
Available: http://arxiv.org/abs/1404.7736
[17] J. Choi, J. Mo, and R. W. Heath, Jr., “Near maximum-likelihood detector and channel estimator for uplink multiuser
massive MIMO systems with one-bit ADCs,” IEEE Trans. Commun., vol. 64, no. 5, pp. 2005–2018, May 2016.
[18] C. Mollén, J. Choi, E. G. Larsson, and R. W. Heath, Jr., “Uplink performance of wideband massive MIMO with one-bit
ADCs,” IEEE Trans. Wireless Commun., vol. 16, no. 1, pp. 87–100, Jan. 2017.
[19] Y. Li, C. Tao, G. Seco-Granados, A. Mezghani, A. L. Swindlehurst, and L. Liu, “Channel estimation and performance
analysis of one-bit massive MIMO systems,” IEEE Trans. Signal Process., vol. 65, no. 15, pp. 4075–4089, Aug. 2017.
[20] J. Singh, S. Ponnuru, and U. Madhow ,“Multi-gigabit communication: The ADC bottleneck,” in Proc. IEEE Int. Conf.
Ultra-Wideband, Sep. 2009, pp. 2227.
[21] M. O. Damen, H. E. Gamal, and G. Caire, “On maximum-likelihood detection and the search for the closest lattice point,”
IEEE Trans. Inf. Theory, vol. 49, no. 10, pp. 2389–2402, Oct. 2003.
[22] T. Kailath, H. Vikalo, and B. Hassibi, “MIMO receiver algorithms,” in Space-Time Wireless Systems, H. Bolcskei, D.
Gesbert, C. Papadias, andA. van der Veen, Eds. Cambridge, U.K.: Cambridge Univ. Press, 2006
29
[23] J. W. Choi, B. Shim, A. C. Singer, and N. I. Cho, “Low-complexity decoding via reduced dimension maximum-likelihood
search,” IEEE Trans. Signal Process., vol. 58, no. 3, pp. 1780–1793, Mar. 2010.
[24] Z. Guo and P. Nilsson, “Algorithm and implementation of the K-best sphere decoding for MIMO detection,” IEEE J. Sel.
Areas Commun., vol. 24, no. 3, pp. 491–503, Mar. 2006.
[25] U. Fincke and M. Pohst, “Improved methods for calculating vectors of short length in a lattice, including a complexity
analysis,” Mathematics of Computation, vol. 44, no. 170, pp. 463–471, Apr. 1985.
[26] B. Hassibi and H. Vikalo, “On the sphere-decoding algorithm I. Expected complexity,” IEEE Trans. Signal Process., vol.
53, no. 8, pp. 2806–2818, Aug. 2005.
[27] A. Mastin and P. Jaillet, “Log-quadratic bounds for the Gaussian Q-function,” arXiv:1304.2488 [math.PR], Apr. 2013.
[Online]. Available: http://arxiv.org/abs/1304.2488
[28] S. Kim, N. Lee, and S.-N. Hong, “A low-complexity soft-output wMD decoding for uplink MIMO systems with one-bit
ADCs,” arXiv:1707.02868 [cs.IT], July 2017. [Online]. Available: http://arxiv.org/abs/1707.02868
[29] IEEE Approved Draft Standard for LAN - Specific Requirements - Part II: Wireless LAN Medium Access Control (MAC)
and Physical Layer (PHY) Specifications - Amendment 3: Enhancements for Very High Throughput in the 60GHz Band,
IEEE P802.11ad/D9.0 Std., Jul. 2012.
[30] K. D. Rao, Channel coding techniques for wireless communications, Springer, 2015.
[31] T. Richardson and R. Urbanke, Modern coding theory, Cambridge university press, 2008.
| 7cs.IT
|
1
Coded Computing for Distributed Graph Analytics
Saurav Prakash†‡ , Amirhossein Reisizadeh∗‡ , Ramtin Pedarsani∗ , Salman Avestimehr†
† University of Southern California ∗ University of California, Santa Barbara
‡ Authors have equal contribution
arXiv:1801.05522v1 [cs.DC] 17 Jan 2018
Abstract
Many distributed graph computing systems have been developed recently for efficient processing of massive graphs.
These systems require many messages to be exchanged among computing machines at each step of the computation, making
communication bandwidth a major performance bottleneck. We present a coded computing framework that systematically injects
redundancy in the computation phase to enable coding opportunities in the communication phase thus reducing the communication
load substantially. Specifically, we propose coded schemes that enable an inverse-linear trade-off (asymptotically) between
computation load and average communication load for three popular random graphs – Erdös-Rényi (ER), random bi-partite
(RB), stochastic block model (SBM). The proposed scheme for ER graph is shown to be optimal asymptotically as the graph size
n → ∞. For finite n, we demonstrate via numerical analysis that for a given computation load r, i.e. when each graph node is
carefully stored at r servers, the proposed scheme slashes the average communication load by (nearly) r.
I. I NTRODUCTION
Graphs are widely used to identify and incorporate the relationship patterns and anomalies inherent in real-life datasets. Their
growing scale and importance have prompted the development of various large-scale distributed graph processing frameworks,
such as Pregel [1] and GraphLab [2]. The underlying theme in these systems is the “think like a vertex” approach [3] where
the computation at each vertex requires only the data available in the neighbourhood of the vertex. This approach significantly
improves performance in comparison to general-purpose distributed processing systems (e.g., Dryad [4], MapReduce [5]),
which do not leverage the underlying structure of graphs.
These distributed graph processing systems, however, require many messages to be exchanged among computing machines
(servers) during job execution. As a result, communication bandwidth is a common bottleneck in parallel computations over
graphs [6], accounting for more than 50% of the overall execution time in representative cases [7].
We develop a new approach that leverages coding to reduce the communication load in distributed graph processing. Motivated
by the “think like a vertex” approach, we describe a mathematical model for MapReduce computations on graphs and show
how carefully injecting redundancy in Map phase results in significant reduction in the communication load during Shuffle
phase. The idea is to leverage the graph structure and create coded messages that simultaneously satisfy the data demand of
multiple computing machines in Reduce phase.
Our work is rooted in the recent development of a coding framework for general MapReduce computations that establishes
an inverse-linear trade-off between computation and communication – Coded Distributed Computing (CDC) [8]. CDC achieves
the communication bandwidth gain r, when each Map computation is carefully repeated at r servers.
As we move from the general MapReduce framework to graph analytics, the key challenge is that the computation associated
with each node is a function of graph structure. In particular, each computation needs data only from the neighbouring nodes,
while in the general MapReduce framework, each computation needs all the input files (corresponding to a complete graph).
This asymmetry in the data requirements of the computations is the main challenge in developing efficient subgraph and
computation allocation and Shuffling schemes.
As the main contribution of this paper, we solve the problem for random Erdös-Rényi (ER) graph. Specifically, for a
given computation load r, i.e. when each node is stored on average at r distinct servers, we show that the minimum average
r
normalized communication load is L∗ (r) = 1r p(1 − K
) + o(p), where K denotes the number of servers and p > Θ( n12 ) is
the edge probability in the ER graph of size n. To prove the achievability, we propose a coded scheme that creates coding
opportunities for communicating messages across machines by Mapping the same graph node at different machines, so that
each coded transmission satisfies the data demand of multiple machines. For converse, we use cut-set bounds and show that
the proposed scheme is asymptotically optimal as n → ∞.
We demonstrate via numerical results that our coded scheme for ER graph achieves near optimal average communication
load for finite n and provides a gain of (almost) r in comparison to an uncoded scheme described later. Furthermore, we
develop coded schemes for random bi-partite graphs and stochastic block models, and prove that our schemes asymptotically
enable an inverse-linear trade-off between computation and communication loads in distributed graph processing.
Related Works: Recently, there has been significant interest in the use of coding theoretic ideas for mitigating several
bottlenecks that arise in large scale distributed computation. These works can be divided into two categories: those that
tackle the communication bandwidth bottleneck in distributed computation, e.g. [8]–[14], and those who tackle the straggler
bottleneck in distributed computation, e.g. [9], [15]–[21].
Notation: We let [n] represent the set {1, 2, . . . , n} for n ∈ N. For functions of n, we denote f = Θ(g) if there are positive
constants c1 and c2 such that c1 ≤ |f /g| ≤ c2 , and f = o(g) if f /g converges to 0 as n goes to infinity.
2
II. P ROBLEM S ETTING
In this section, we describe the setting and formulate our problem.
A. Computation Model
Consider an undirected graph G = (V, E) where V = [n] denotes the set of graph nodes and E is the set of edges. A binary
file wi ∈ F2F of size F ∈ N is associated with each graph node i ∈ [n]. The set of all files is denoted by W = {wi : i ∈ V}.
The neighborhood of node i is denoted by N (i) = {j ∈ V : (j, i) ∈ E} and the set of files in N (i) is represented by
WN (i) = {wj : j ∈ N (i)}. In general, node i can be contained in N (i), i.e. G can have self-loops. For each node i ∈ V, a
computation is associated with it as follows
|N (i)|
φi : F2F
→ F2B ,
(1)
where φi (·) outputs the input files in WN (i) to a length B binary stream oi = φi (WN (i) ).
We assume the computation φi (·) can be decomposed as a MapReduce computation:
φi (WN (i) ) = hi ({gi,j (wj ) : wj ∈ WN (i) }),
(2)
where the Map function gi,j : F2F → F2T Maps file wj to a length T binary intermediate value vi,j = gi,j (wj ), ∀i ∈ N (j).
|N (i)|
The Reduce function hi : F2T
→ F2B Reduces the intermediate values associated with the output function φi (·) into the
final output value oi = hi ({vi,j : j ∈ N (i)}).
As an example, Fig. 1(a) illustrates a graph with n = 6 nodes, each associated with a file, and Fig. 1(b) illustrates the
corresponding MapReduce computations.
Popular iterative graph algorithms can be expressed in the computation framework described above. We now present two
popular graph algorithms that can be expressed in the computation framework described above [22].
Example 1. PageRank [23] is a popular algorithm to measure the quality of the nodes in a webgraph based on the underlying
hyperlink structure. In particular, the algorithm computes the likelihood that a random surfer would visit a page. Mathematically,
the rank of a node i satisfies the following relation:
X
1
Π(i) = (1 − d)
Π(j)P(j → i) + d ,
(3)
|V|
j∈N (i)
where (1 − d) is referred to as the damping factor, Π(i) denotes the likelihood that the random surfer will arrive at node
i, |V| is the total number of nodes in the webgraph, and P(j → i) is the transition probability from node j to node i. The
computation can be carried out iteratively as follows:
X
1
(4)
Πk (i) = (1 − d)
Πk−1 (j)P(j → i) + d ,
|V|
j∈N (i)
1
for all i ∈ V and k = 1, 2, · · · . The
where k and k − 1 are respectively the current and previous iterations and Π0 (i) = |V|
number of iterations depends on the stopping criterion for the algorithm. Usually, the algorithm is stopped when the change in
the PageRank mass of each node is less than a pre-defined tolerance. The PageRank update at each node can be decomposed
k
k−1
into Map and Reduce functions for each iteration k. For a given node i and iteration k, let WN
(j), j ∈ N (i)}, and
(i) = {Π
P
1
k
k
k−1
k
k
k
φi (WN (i) ) = (1 − d) j∈N (i) Π
(j)P(j → i) + d |V| . The Mapper gi,j (·) Maps the file wj to the intermediate values vi,j
=
k
k−1
gi,j (wj ) = Π
(j)P(j → i) for all neighboring nodes j ∈ N (i). Using the intermediate values from the
Map
computations,
P
1
k
k
the Reducer hki (·) computes node i’s new PageRank value as Πk (i) = hki ({vi,j
: j ∈ N (i)}) = (1 − d) j∈N (i) vi,j
+ d |V|
.
Example 2. Single-source shortest path is one of the most studied problems in graph theory. The task here is to find the
shortest path to each node i in the graph from a source node s. A sub-problem for this task is to compute the distance of
each node i from the source node s, where distance D(i) is the length of the shortest path from s to i. This can be carried
out iteratively in parallel. First, initialize D0 (s) = 0 and D0 (i) = +∞, ∀i ∈ V \ {s}. Subsequently, each node i is updated as
follows at each iteration k:
Dk (i) = min (Dk−1 (j) + t(j, i)),
j∈N (i)
(5)
where t(j, i) is the weight of the edge (j, i). The algorithm is stopped when the change in the distance value for each node
is within a pre-defined tolerance. The distance computation for each node at iteration k can be decomposed into Map and
k
k−1
k
Reduce computations. Particularly, for each node i and iteration k, let WN
(j), j ∈ N (i)}, and φki (WN
(i) = {D
(i) ) =
k−1
k
k
k
k
k−1
minj∈N (i) (D
(j) + t(j, i)). The Mapper gi,j (·) Maps the file wj to the intermediate values vi,j = gi,j (wj ) = D
(j) +
t(j, i) for all neighboring nodes j ∈ N (i). Using the intermediate values from the Map computations, the Reducer hki (·)
k
k
computes node i’s updated distance value as Dk (i) = hki ({vi,j
: j ∈ N (i)}) = minj∈N (i) vi,j
.
3
B. Distributed Implementation
We now describe our model for distributed implementation of the computations described above. We consider a network
consisting of K machines that are connected to each other. As the first step, a distributed implementation consists of allocating
a subgraph to each machine.
Subgraph Allocation: We denote the subgraph that is allocated to each machine k ∈ [K] by Mk ⊆ V. Machine k will
then store all files corresponding to that subgraph, and will be responsible for computing the Map functions on those files.
Note that each file (or node) should be Mapped by at least one machine, however, in order to help with reducing the overall
communication load, we allow each file to be Mapped by more than one machine (i.e, increasing the computation load). More
formally, the computation load is defined as follows.
Definition 1 (Computation Load). For a subgraph allocation, Mk , k ∈ [K], the computation load, r, is defined as
PK
|Mk |
r , k=1
.
(6)
n
n
Remark 1. For a desired computation load r, we assume that the subgraphs allocated to servers have all the same size of r K
nodes (i.e., equal memory is used at all servers).
Computation Allocation: Recall that a Reducer is associated with each node of the graph G. The set of nodes whose Reduce
computations are assigned to machine k is denoted by Rk ⊆ V. The Reduce computations are distributed disjointly and
n
uniformly across the K servers, i.e. |Rk | = K
, ∀k ∈ [K].
(a) An example of a graph with 6 nodes.
(b) MapReduce decomposition of computations for graph in (a).
(c) A subgraph and computation allocation for graph in (a).
Fig. 1: An illustrative example.
Given a subgraph and computation allocation to servers, denoted by A = (M, R) where M = (M1 , · · · , MK ) and
R = (R1 , · · · , RK ), the computation proceeds distributedly in the following three phases.
Map phase: Each server first Maps the files associated with the subgraph that is allocated to it. More specifically, for each
i ∈ Mk , machine k computes a vector of intermediate values corresponding to the nodes in N (i): ~gi = (vj,i : j ∈ N (i)).
Shuffle phase: To be able to do the final Reduce computations, each server needs the intermediate values corresponding to
the neighbors of each node that it is responsible for its Reduction. Servers exchange messages so that at the end of the Shuffle
phase, each server is able to recover its required set of intermediate values. More formally, the Shuffle phase proceeds as
follows. For each machine k ∈ [K],
(i) machine k creates a message Xk ∈ F2ck as a function of intermediate values computed locally at that server during the
Map phase, i.e. Xk = ψk ({~gi : i ∈ Mk }), where ck is the length of the binary message Xk ,
(ii) machine k multicasts Xk to all the remaining servers,
(iii) server k recovers the needed intermediate values {vi,j : i ∈ Rk , j ∈ N (i), j ∈
/ Mk } using locally computed intermediate
values {vi,j : i ∈ N (j), j ∈ Mk } and received messages {Xk0 : k 0 ∈ [K] \ {k}}.
4
We define the (normalized) communication load of the Shuffle phase as follows.
Definition 2 (Normalized Communication Load). The normalized communication load, denoted by L, is defined as the number
of bits communicated by K machines during the Shuffle phase, normalized by the maximum possible total number of bits in
the intermediate values associated with all the Reduce functions, i.e.
PK
ck
L , k=1
.
(7)
n2 T
Reduce phase: Server k uses its locally computed intermediate values and the messages received from other servers to first
construct the required intermediate values for Reduce functions that are allocated to it (i.e. for the nodes in Rk ), and then
calculates oi = hi ({vi,j : j ∈ N (i)}) for all i ∈ Rk .
To illustrate the above definitions, let us again consider the graph depicted in Fig. 1(a) where K = 3 machines are available
to carry out the computations. For the subgraph and computation allocation described in Fig. 1(c), each server k ∈ {1, 2, 3}
Maps the nodes in subgraph Mk and computes the Reductions associated with nodes in Rk . The computation load is r = 2
6
.
and the normalized (uncoded) communication load equals to L = 36
C. Problem Formulation
For an allowed computation load r, we aim to find the optimal allocation of subgraphs and computations to servers, and
the optimal coding for Shuffling in order to minimize the communication load. However, we note that this problem even in
the simplest case of r = 1 and uncoded transmission in the Shuffling phase is NP-hard for general graphs [24]. Hence, we
restrict our attention to random graphs and focus on the average communication load.
We consider a random undirected graph G = (V, E), where edges independently exist with probability P[(i, j) ∈ E] for all
i, j ∈ V. Let A(r) be the set of all possible subgraph and computation allocations for a given computation load r (as defined
in the previous subsection). For a graph realization G and an allocation A ∈ A(r), we denote by LA (r, G) the minimum
(normalized) communication load (as defined in Definition 2) over all possible Shuffling coding schemes that enable each
machine to compute the Reduce functions assigned to it.
We now formally define our problem as follows.
Problem: For a given random undirected graph G = (V, E) and a computation load r ∈ N, our goal is to characterize the
minimum average normalized communication load, i.e.
L∗ (r) , inf EG [LA (r, G)].
(8)
A∈A(r)
Remark 2. As defined above, L∗ (r) essentially reveals a fundamental trade-off between computation and communication in
distributed graph processing frameworks.
To solve the problem defined above, we need to establish the optimal subgraph and computation allocations for each server
along with the efficient Shuffle scheme.
III. M AIN R ESULTS
In this section, we present the main results of the paper. Our first result is the characterization of L∗ (r) (defined in (8)) for
Erdös-Rényi random graph that is defined below.
Erdös-Rényi: Denoted by ER(n, p), this model consists of graphs of size n in which each edge exists with probability p ∈ [0, 1],
independently of other edges (2(a)).
Theorem 1. For Erdös-Rényi graph ER(n, p) with p > Θ( n12 ), we have
L∗ (r)
1
r
lim
= (1 − ).
(9)
n→∞
p
r
K
Remark 3. Achievability of Theorem 1 is proved in Section IV, where we provide subgraph and computation allocations
followed by the code design for Shuffling. Proof of converse for Theorem 1 is provided in Section V.
Remark 4. Theorem 1 reveals an interesting inverse linear trade-off between computation and communication. In particular, the
scheme that we propose for achievability of Theorem 1 asymptotically gives a communication load gain of r in comparison
to the uncoded scheme that as we discuss later in Section IV only achieves an average normalized communication load
r
of p(1 − K
). This trade-off can be used to leverage additional computing resources and capabilities to alleviate the costly
communication bottleneck. Moreover, we numerically demonstrate that even for finite graphs, not only the proposed scheme
significantly reduces the communication load, but also has a small optimality gap (Fig. 3).
We also extend our achievability results to two more random graph models which are described below:
Random Bi-partite: Denoted by RB(n1 , n2 , q), this model consists of graphs with two disjoint clusters of sizes n1 and n2 in
which each inter-cluster edge exists with probability q ∈ [0, 1], independently of other edges (2(b)).
Stochastic Block Model: Denoted by SBM(n1 , n2 , p, q), this model consists of graphs with two disjoint clusters of sizes n1 and
n2 such that each intra-cluster edge exists with probability p and each inter-cluster edge exists with probability q, 0 ≤ q ≤ p ≤ 1,
all independent of each other (Fig. 2(c)).
5
(a) Erdös-Rényi model with n = 20
(b) Random bipartite model with n1 = 6
and n2 = 4
(c) Stochastic block model with n1 = 12
and n2 = 18
Fig. 2: Illustrative instances of the random graph models considered in the paper. In Fig. 2(a), each edge exists with a given probability p.
In Fig. 2(b), each cross-edge exists with a given probability q. In Fig. 2(c), each intra-cluster edge exists with a given probability p and each
cross-edge exists with a given probability q.
Fig. 3: Performance comparison of the coded scheme with uncoded scheme and the proposed lower bound. The experiments are carried out
for n = 300, p = 0.1 and K = 5.
The following theorems provide the achievability results for RB and SBM models and are proved in the appendices.
Theorem 2. For random bi-partite graph RB(n1 , n2 , q) with n = n1 + n2 , n1 = Θ(n), n2 = Θ(n) and q > Θ( n12 ), we have
L∗ (r)
1
r
lim sup 2qn1 n2 ≤ (1 − ).
(10)
r
K
n→∞ (n +n )2
1
2
Theorem 3. For stochastic block model graph SBM(n1 , n2 , p, q) with n = n1 +n2 , n1 = Θ(n), n2 = Θ(n), and p, q > Θ( n12 ),
we have
L∗ (r)
1
r
lim sup pn2 +pn2 +2qn n ≤ (1 − ).
(11)
1
2
1
2
r
K
n→∞
2
(n1 +n2 )
Remark 5. Theorems 2 and 3 demonstrate that the inverse linear trade-off between computation load and communication load
can also be achieved in random bi-partite and stochastic block models.
IV. ACHIEVABILITY FOR E RDÖS -R ÉNYI MODEL
We now propose our coded and uncoded schemes for Erdös-Rényi model, and prove the achievability of Theorem 1.
A. Proposed Scheme
As explained in Section II, a scheme for distributed implementation of the computation consists of subgraph allocation,
computation allocation, and Shuffling algorithm. We next precisely describe our proposed scheme.
6
K
Subgraph Allocation: The n nodes are partitioned into K
r batches of size g = n/ r , each corresponding to a set T ⊆ [K]
of size r, i.e. {1, · · · , n} = {BT : T ⊆ [K], |T | = r}. Server k ∈ [K] Maps the nodes in BT if k ∈ T . Equivalently,
n
BT ⊆ Mk if k ∈ T , i.e. each server Maps r K
nodes.
Computation Allocation: The n Reduce functions are uniformly partitioned into K subsets and each subset is assigned to
n
Reducers. We denote the proposed subgraph and computation
one machine. Thus, each server is responsible for computing K
allocation by Ã.
Uncoded Shuffle: Given the above subgraph and computation allocation, each node in V requires on average pn intermediate
r
values for the Reduce phase, where a fraction K
of them are locally available. The rest of the intermediate values should be
r
).
sent from the other servers. Therefore, the average normalized communication load for the uncoded scheme is p(1 − K
k
Coded Shuffle: Consider a set of servers S ⊆ [K], |S| = r + 1. For each server k ∈ S, let ZS\{k} be the set of all intermediate
values needed by Reduce functions in k, and are available exclusively at each server k 0 ∈ S \ {k}, i.e.
k
= {vi,j : (i, j) ∈ E, i ∈ Rk , j ∈ ∩k0 ∈S\{k} Mk0 }.
ZS\{k}
(1)
(r)
k
For each k ∈ S, each intermediate value vi,j ∈ ZS\{k}
is evenly split into r segments vi,j , · · · , vi,j , each of size Tr bits.
k
Each segment is associated with a distinct server in S \ {k}. Therefore, ZS\{k}
is evenly partitioned to r sets, which are
k
k
denoted by ZS\{k},s for s ∈ S \ {k}. Depending on the realization of the graph, the maximum possible size of ZS\{k}
is
n
=
g̃ = g K
n2
K (K
r)
= Θ(n2 ). Each server s ∈ S creates an r × g̃ table and fills that out with segments which are associated
k
with it. Each row of the table is filled from left by the segments in one of the sets ZS\{k},s
, where k ∈ S \ {s} (Fig. 4). Then,
server s broadcasts the XOR of all the segments in each (non-empty) column of the table (for each non-empty column, the
empty entries are zero padded). Clearly, there exist at most g̃ of such coded messages. The process is carried out similarly for
all remaining subsets S ⊆ [K] with |S| = r + 1.
After the Shuffle phase, each server can recover the intermediate values associated with its assigned set of Reduce functions
using the received coded messages and the locally computed intermediate values.
Remark 6. The proposed scheme carefully aligns and combines the existing intermediate values to benefit from the coding
opportunities. This resolves the issue posed by the asymmetry in the data requirements of the Reducers which is one of the
main challenges in moving from the general MapReduce framework to graph analytics.
As an example, consider a system of K = 3 servers and computation load r = 2. For the graph in Fig. 1(a), Fig. 1(c)
summarizes the subgraph and computation allocations followed from the proposed scheme. Consider the set S = [K] of
3
size r + 1 = 3. Every intermediate value in Z{1,2}
= {v5,1 , v6,2 } is split into r = 2 segments, each associated with a
distinct server in {1, 2}. This is done similarly for servers 1 and 2. Then, servers 1, 2, and 3 broadcast their coded messages
(1)
(1) (1)
(1)
(2)
(1) (2)
(1)
(2)
(2) (2)
(2)
X1 = {v5,1 ⊕ v4,3 , v3,4 ⊕ v6,2 }, X2 = {v5,1 ⊕ v1,5 , v6,2 ⊕ v2,6 }, and X3 = {v4,3 ⊕ v1,5 , v3,4 ⊕ v2,6 }, respectively.
All three servers can recover their needed intermediate values. For instance, server 3 needs v5,1 to carry out the Reduce
function associated with node 5. Since it has already Mapped nodes 3 and 5, intermediate values v4,3 and v1,5 are available
(1)
(2)
(1)
(1)
(2)
(1)
locally. Server 3 can recover v5,1 and v5,1 from v5,1 ⊕ v4,3 and v5,1 ⊕ v1,5 , respectively. Therefore, the overall uncoded
6
3
communication load 36 is reduced to coded load 36 .
B. Proof of Achievability
We first define the average normalized communication load for our scheme as follows. For a graph realization G, the proposed
allocation à ∈ A(r), and the proposed coded and uncoded Shuffling scheme, we denote the normalized coded and uncoded
communication loads by LC
(r, G) and LUC
(r, G), respectively. The average normalized coded and uncoded communication
Ã
Ã
C
UC
loads will then be L̄Ã , EG [LC
(r,
G)]
and
L̄
, EG [LUC
(r, G)], respectively.
Ã
Ã
Ã
We now apply the proposed coded scheme to graph G and compute the induced average coded load. WLOG, we analyze our
algorithm by a generic argument for servers S = {s1 , · · · , sr+1 } which can be similarly applied for other sets of servers due
to the symmetric structure induced by the graph model and allocations. Following the Shuffle phase of the proposed scheme,
consider r servers s2 , · · · , sr+1 and the (r + 1)’th server s1 . Server s1 broadcasts at most g̃ coded messages X 1 , · · · , X g̃
which are exclusively useful for severs s2 , · · · , sr+1 . Each X j , 1 ≤ j ≤ g̃, is XOR of at most r intermediate values (segments)
of size Tr bits, associated with server s1 . More formally, for all 1 ≤ j ≤ g̃,
r
M
(1)
Xj =
vα(i),β(j) ,
(12)
i=1
si+1
ZS\{s
i+1 },s1
(1)
{vα(i),β(j)
where
=
: 1 ≤ j ≤ g̃} for 1 ≤ i ≤ r.
Let Bern(p) random variable Eα,β indicate the existence of the edge (α, β) ∈ V × V, i.e.
1 if (α, β) ∈ E,
Eα,β =
0 if (α, β) ∈
/ E.
(13)
7
Fig. 4: Creating coded messages by aligning intermediate values.
Clearly, for all nodes α, β, u, v ∈ V, Eα,β is independent of Eu,v if {α, β} 6= {u, v}, and Eα,β = Eu,v , otherwise. For
i = 1, · · · , r, let us define the random variable Pi as
g̃
X
Pi =
Eα(i),β(j) ,
(14)
j=1
i.e. each Pi is sum of g̃ possibly dependent Bern(p) random variables. It is also worth noting that Pi ’s are not independent
in general. By careful alignment of present intermediate values (Fig. 4), server s1 broadcasts Q coded messages each of size
T
r bits, where
Q = max Pi .
i=1,··· ,r
(15)
Thus, the total coded communication load sent from server s1 exclusively for servers s2 , · · · , sr+1 is Tr Q bits. By similar
arguments for other sets of severs, we can characterize the average (normalized) coded communication load of the proposed
scheme as
K −1
1
C
L̄Ã = 2 K
E[Q].
(16)
r
rn
The following lemma asymptotically upper bounds E[Q] and the proof is provided in Section IV-C .
Lemma 1. For ER(n, p) graph with p > Θ( n12 ), we have
E[Q] ≤ pg̃ + o(pg̃).
(17)
Putting (16) and Lemma 1 together, we have
r
1
p(1 − ) + o(p),
(18)
r
K
hence the achievability claimed in Theorem 1 is proved. Finally, we note that as explained in the uncoded Shuffle algorithm,
r
the average normalized uncoded communication load of the proposed scheme is L̄UC
= p(1 − K
), which implies that our
Ã
scheme achieves an asymptotic gain r.
L∗ (r) ≤ L̄C
≤
Ã
C. Proof of Lemma 1
Before proving Lemma 1, we first present the following lemma that will be used in our proof.
Lemma 2. For random variables {Pi }ri=1 defined in (14), their moment generating functions can be bounded by
E[esPi ] ≤ (pe2s + 1 − p)g̃/2 .
(19)
Proof. Consider a generic random variable
P =
g̃
X
j=1
Ej ,
(20)
8
where Ej ’s are Bern(p) and possibly dependent. However, although Ej ’s may not be all independent, but dependency is
restricted to pairs of Ej ’s. In other words, for all 1 ≤ j ≤ g̃, Ej is either independent of all E[g̃]\{j} , or is equal to E` for
some ` ∈ [g̃] \ {j} and independent of all E[g̃]\{j,`} . By merging dependent pairs, we can write
P =
J
X
Fj ,
(21)
j=1
where
(i) Fj ’s are independent,
(ii) 2J − g̃ of Fj ’s are Bern(p),
(iii) g̃ − J of Fj ’s are 2 × Bern(p),
for some integer b g̃2 c ≤ J ≤ g̃. Now, we can bound the moment generating function of P . For s > 0,
E[esP ] = E[es
J
Y
=
PJ
j=1
Fj
]
(22)
E[esFj ]
(23)
j=1
2J−g̃
g̃−J
= pes + 1 − p
pe2s + 1 − p
h
g̃−J
2 iJ−g̃/2 2s
pe + 1 − p
= pes + 1 − p
(24)
(25)
(a)
J−g̃/2 2s
g̃−J
≤ pe2s + 1 − p
pe + 1 − p
g̃/2
= pe2s + 1 − p
,
where inequality (a) is obtained using Lemma 4 (proof available in Appendix C).
We now complete the proof of Lemma 1. For any s0 > 0, we can write
0
0
es E[Q] ≤ E[es Q ]
= E[ max e
i=1,··· ,r
r
X
s0 Pi
≤ E[
e
s0 Pi
]
(26)
(27)
(28)
(29)
]
(30)
E[es Pi ]
(31)
i=1
=
r
X
0
i=1
0
≤ r(pe2s + 1 − p)g̃/2 ,
where the last inequality follows from Lemma 2. Taking logarithm from both sides yields
0
1
g̃
E[Q] ≤ 0 log(r) + 0 log(pe2s + 1 − p).
s
2s
Let us substitute s = 2s0 in (33). Then,
1
g̃
E[Q] ≤ log(r2 ) + log(pes + 1 − p),
s
s
for any s > 0. Let p̄ = 1 − p and pick
s
log(r)
s∗ = 2
.
g̃pp̄
(32)
(33)
(34)
(35)
We proceed with evaluation of the right hand side (RHS) of (34) at s = s∗ . We first recall the following Taylor series
x3
x2
log(1 + x) = x −
+
− · · · , for x ∈ (−1, 1],
(36)
2
3
3
2
s
s
+
+ ··· ,
for s ∈ R.
(37)
es = 1 + s +
2
3!
Let x = p(es∗ − 1). It is easy to check that for p > Θ( n12 ), we have x → 0 and s∗ → 0 as n → ∞. Therefore, for n → ∞
we can write
log(pes∗ + 1 − p) = log(x + 1)
x2
x3
=x−
+
− ···
2
3
p2 (es∗ − 1)2
p3 (es∗ − 1)3
= p(es∗ − 1) −
+
− ···
2
3
9
p2
2 p3
3
s∗ 2
s∗ 3
s∗ 2
s∗ 3
s∗ 2
s∗ 3
+
+ ··· −
s∗ +
+
+ ··· +
s∗ +
+
+ ··· − ···
2
3!
2
2
3!
3
2
3!
pp̄ 2
2
= ps∗ + s∗ + o(ps∗ ).
(38)
2
Putting everything together, we have
1
g̃
E[Q] ≤
log(r2 ) +
log(pes∗ + 1 − p)
(39)
s∗
s∗
pp̄
1
g̃
log(r2 ) +
ps∗ + s2∗ + o(ps2∗ )
≤
(40)
s∗
s∗
2
g̃pp̄
1
log(r2 ) + g̃p +
s∗ + o(g̃ps∗ )
(41)
=
s∗
2
p
p
(42)
= g̃p + 2 g̃pp̄ log(r) + o( g̃p).
p
√
1
n2
2
Recall that g̃ = K K = Θ(n ). Therefore, for p > Θ( n2 ) we have g̃p > Θ(1) and thus g̃pp̄ log(r) = Θ( g̃p) = o(g̃p).
(r)
Therefore, E[Q] ≤ pg̃ + o(pg̃), as n → ∞.
= p s∗ +
V. C ONVERSE FOR E RDÖS -R ÉNYI MODEL
We complete the proof of Theorem 1 by deriving the lower bound on the best average communication load for Erdös-Rényi
graphs.
Let G be an ER(n, p) random graph and consider a subgraph and computation allocation A = (M, R) ∈ A(r), where
PK
n
k=1 |Mk | = rn and |Rk | = K , for all k ∈ [K]. We denote the number of files that are Mapped at j nodes under Map
j
assignment M, as aM , for all j ∈ [K]. The following lemma holds.
PK aj
Lemma 3. EG [LA (r, G)] ≥ p j=1 nM K−j
Kj .
Proof. We let intermediate values vi,j be realizations of random variables Vi,j , uniformly distributed over F2T . For a random
G
graph G = (V, E) and subsets I, J ⊆ V = [n], define VI,J
= {Vi,j : (i, j) ∈ E, i ∈ I, j ∈ J } as the set of present intermediate
values in graph G corresponding to Reduces in I and Maps in J . For a given allocation A = (M, R) ∈ A(r) and a subset
G
of servers S ⊆ [K], we define XS = {Xk : k ∈ S} and YSG = (VRGS ,: , V:,M
), where “:” denotes all possible indices (which
S
G
G
depend on both allocation and graph realization). For instance, H(Xk |V:,M
) = 0 and H(VRGk ,: |X[K] , V:,M
) = 0 for all
k
k
servers k ∈ [K] and graphs G. Moreover, we denote the number of nodes that are exclusively Mapped by j servers in S as
aj,S
M:
X
|(∩k∈S1 Mk ) \ (∪k0 ∈S
(43)
aj,S
/ 1 Mk0 )|
M ,
S1 ⊆S:|S1 |=j
Following the steps in proof of Claim 1 in [8], we prove by induction:
EG [H(XS |YSGc )] ≥ pT
|S|
X
j=1
aj,S
M
n |S| − j
,
K
j
(44)
for any subset S ⊆ [K].
(i) If S = {k}, for any k ∈ [K] and graph G we have H(XS |YSGc ) ≥ 0. Therefore,
1
X
n 1−1
.
EG [H(XS |YSGc )] ≥ 0 = pT
a1,S
M
K
1
j=1
(ii) Assume the claim (44) holds for all subsets of size S0 . For any subset S ⊆ [K] of size S0 + 1, we can write
1 X
G
H(XS |YSGc ) ≥
H(XS |V:,M
, YSGc ).
k
S0
(45)
(46)
k∈S
Moreover,
G
G
G
H(XS |V:,M
, YSGc ) = H(VRGk ,: |V:,M
, YSGc ) + H(XS |V:,M
, VRGk ,: , YSGc )
k
k
k
We can lower bound expectation of the first RHS term in (47) as follows
X
G
G
G
, YSGc )] = EG [
H(V{v},:
|V{v},M
)]
EG [H(VRGk ,: |V:,M
k
k ∪MS c
(47)
(48)
v∈Rk
= EG [
X
|N (v)| − |N (v) ∩ (Mk ∪ MS c )|]
(49)
v∈Rk
=
S0
X
n
j,S\{k}
pT
aM
K
j=0
(50)
10
≥
S0
X
n
j,S\{k}
pT
aM
.
K
j=1
(51)
Expected value of the second term in RHS of (47) can be lower bounded from the induction assumption:
G
G
)]
EG [H(XS |V:,M
, VRGk ,: , YSGc )] = EG [H(XS\{k} |YS\{k}
k
≥ pT
S0
X
j=1
j,S\{k}
aM
n S0 − j
.
K j
Putting (46), (47), (51), and (53) together, we have
1 X
G
EG [H(XS |YSGc )] ≥
EG [H(XS |V:,M
, YSGc )]
k
S0
k∈S
1 X
G
G
=
, VRGk ,: , YSGc )]
EG [H(VRGk ,: |V:,M
, YSGc )] + EG [H(XS |V:,M
k
k
S0
(52)
(53)
(54)
k∈S
≥
S0
S0
X
X
1 X n
i,S\{k}
j,S\{k} n S0 − j
pT
aM
+ pT
aM
S0
K
K j
i=1
j=1
(55)
k∈S
= pT
S0
X
n 1 X j,S\{k}
aM
Kj
j=1
(56)
k∈S
= pT
SX
0 +1
aj,S
M
j=1
n S0 + 1 − j
.
K
j
(57)
(iii) Therefore, for any subset S ⊆ [K], claim (44) holds.
Now, pick S = [K]. Then,
K
X
EG [H(XS |YSGc )]
ajM K − j
EG [LA (r, G)] ≥
≥
p
.
n2 T
n Kj
j=1
(58)
Converse proof of Theorem 1. Proof of Converse for Theorem 1. First, we can bound the best average communication load as
L∗ (r) ≥
inf
EG [LA (r, G)]
(59)
A:|M1 |+···+|MK |=rn
≥
inf
A:|M1 |+···+|MK |=rn
p
K
X
ajM K − j
.
n Kj
j=1
Additionally, for any Map allocation with computation load r, we have the following equations:
K
K
X
X
ajM = n,
jajM = rn.
j=1
Using convexity of
K−j
Kj
(60)
(61)
j=1
in j and (61), the converse is proved:
1
r
L∗ (r) ≥ p(1 − ).
r
K
(62)
VI. C ONCLUSION
We described a mathematical model for graph based MapReduce computations and demonstrated how coding theoretic
strategies can be employed to substantially reduce the communication load in distributed graph analytics. Our results reveal an
inverse linear trade-off exists between computation load and communication load in distributed graph anaytics. This trade-off
can be used to leverage additional computing resources and capabilities to alleviate the costly communication bottleneck in
distributed graph processing systems. In particular, we developed a coded scheme that asymptotically characterizes the minimum
average normalized communication load for random Erdös-Rényi graph. Moreover, we numerically demonstrate that even for
finite graphs, not only the proposed scheme significantly reduces the communication load, but also has a small optimality gap.
Furthermore, we extended the achievability of the minimum average normalized communication load for two more random
graphs – random bi-partite, and stochastic block model.
R EFERENCES
[1] G. Malewicz, M. H. Austern, A. J. Bik, J. C. Dehnert, I. Horn, N. Leiser, and G. Czajkowski, “Pregel: a system for large-scale graph processing,”
SIGMOD, 2010.
11
[2] Y. Low, D. Bickson, J. Gonzalez, C. Guestrin, A. Kyrola, and J. M. Hellerstein, “Distributed graphlab: a framework for machine learning and data
mining in the cloud,” VLDB, 2012.
[3] R. R. McCune, T. Weninger, and G. Madey, “Thinking like a vertex: a survey of vertex-centric frameworks for large-scale distributed graph processing,”
ACM Computing Surveys, 2015.
[4] M. Isard, M. Budiu, Y. Yu, A. Birrell, and D. Fetterly, “Dryad: distributed data-parallel programs from sequential building blocks,” EuroSys, 2007.
[5] J. Dean and S. Ghemawat, “Mapreduce: simplified data processing on large clusters,” Communications of the ACM, 2008.
[6] A. Lumsdaine, D. Gregor, B. Hendrickson, and J. Berry, “Challenges in parallel graph processing,” Parallel Processing Letters, 2007.
[7] R. Chen, X. Ding, P. Wang, H. Chen, B. Zang, and H. Guan, “Computation and communication efficient graph processing with distributed immutable
view,” HPDC, 2014.
[8] S. Li, M. A. Maddah-Ali, Q. Yu, and A. S. Avestimehr, “A fundamental tradeoff between computation and communication in distributed computing,”
IEEE Transactions on Information Theory, 2017.
[9] K. Lee, M. Lam, R. Pedarsani, D. Papailiopoulos, and K. Ramchandran, “Speeding up distributed machine learning using codes,” IEEE Transactions on
Information Theory, 2017.
[10] L. Song, C. Fragouli, and T. Zhao, “A pliable index coding approach to data shuffling,” ISIT, 2017.
[11] M. Kiamari, C. Wang, and A. S. Avestimehr, “On heterogeneous coded distributed computing,” arXiv preprint arXiv:1709.00196, 2017.
[12] M. A. Attia and R. Tandon, “Information theoretic limits of data shuffling for distributed learning,” GLOBECOM, 2016.
[13] Y. H. Ezzeldin, M. Karmoose, and C. Fragouli, “Communication vs distributed computation: an alternative trade-off curve,” arXiv preprint
arXiv:1705.08966, 2017.
[14] S. Li, M. A. Maddah-Ali, and A. S. Avestimehr, “Coding for distributed fog computing,” IEEE Communications Magazine, vol. 55, no. 4, pp. 34–40,
2017.
[15] S. Dutta, V. Cadambe, and P. Grover, “Short-dot: Computing large linear transforms distributedly using coded short dot products,” NIPS, 2016.
[16] R. Tandon, Q. Lei, A. G. Dimakis, and N. Karampatziakis, “Gradient coding: Avoiding stragglers in distributed learning,” ICML, 2017.
[17] Q. Yu, M. Maddah-Ali, and S. Avestimehr, “Polynomial codes: an optimal design for high-dimensional coded matrix multiplication,” NIPS, 2017.
[18] K. Lee, C. Suh, and K. Ramchandran, “High-dimensional coded matrix multiplication,” ISIT, 2017.
[19] A. Reisizadeh, S. Prakash, R. Pedarsani, and S. Avestimehr, “Coded computation over heterogeneous clusters,” ISIT, 2017.
[20] S. Li, M. A. Maddah-Ali, and A. S. Avestimehr, “A unified coding framework for distributed computing with straggling servers,” 2016.
[21] Q. Yu, M. A. Maddah-Ali, and A. S. Avestimehr, “Coded fourier transform,” arXiv preprint arXiv:1710.06471, 2017.
[22] J. Lin and M. Schatz, “Design patterns for efficient graph algorithms in mapreduce,” MLG Workshop, 2010.
[23] L. Page, S. Brin, R. Motwani, and T. Winograd, “The pagerank citation ranking: Bringing order to the web.,” 1999.
[24] E. Dahlhaus, D. S. Johnson, C. H. Papadimitriou, P. D. Seymour, and M. Yannakakis, “The complexity of multiway cuts,” STOC, 1992.
A PPENDIX A
ACHIEVABILITY FOR R ANDOM B I - PARTITE M ODEL
Consider RB(n1 , n2 , q) graph G = (V1 ∪ V2 , E) with n = n1 + n2 , |V1 | = n1 = Θ(n), and |V2 | = n2 = Θ(n). We first
provide a subgraph and computation allocation which is applied to both uncoded and coded schemes. Proposed allocation takes
place in two disjoint phases as follows.
Phase 1: Maps in V1 and Reduces in V2 are distributedly allocated to K servers according to the allocation proposed in
Section IV-A.
Phase 2: Maps in V2 and Reduces in V1 are distributedly allocated to K servers according to the allocation proposed in
Section IV-A.
We let à denote the subgraph and computation allocation induced by the two phases described above, i.e. à ∈ A(r).
Given the allocation à defined above, both uncoded and coded Shuffling schemes proceed in two steps. In uncoded Shuffling,
each server transmits the intermediate values required by other servers in each step corresponding to phases 1 and 2, without
coding. However, coded Shuffling applies the coded scheme proposed in Section IV-A in each step. Let L̄UC1
, L̄UC2
, L̄C1
and
Ã
Ã
Ã
C2
L̄à denote the average normalized communication load induced by Shuffling according to allocation à defined in Steps 1
and 2, respectively for uncoded and coded schemes. From the achievability of Theorem 1, for q > Θ( n12 ), we have
1
1
n1 n2
r
L̄C1
≤ L̄UC1
+ o(q) = q
(63)
(1 − ) + o(q),
Ã
r Ã
r (n1 + n2 )2
K
L̄C2
≤
Ã
1 UC2
1
n1 n2
r
L̄
+ o(q) = q
(1 − ) + o(q).
r Ã
r (n1 + n2 )2
K
(64)
Therefore, the total average normalized coded communication load of the scheme, L̄C
, is
Ã
L̄C
= L̄C1
+ L̄C2
(65)
Ã
Ã
Ã
1 UC1
≤ (L̄Ã + L̄UC2
) + o(q)
(66)
Ã
r
1
= L̄UC
+ o(q)
(67)
r Ã
1
2n1 n2
r
= q
(1 − ) + o(q),
(68)
r (n1 + n2 )2
K
r
∗
C
1 n2
where L̄UC
= q (n2n
2 (1 − K ) denotes the average normalized uncoded communication load. Given that L (r) ≤ L̄ , we
1 +n2 )
Ã
Ã
2n1 n2
1
r
∗
have L (r) ≤ r q (n1 +n2 )2 (1 − K ) + o(q) and Theorem 2 is proved.
12
A PPENDIX B
ACHIEVABILITY FOR S TOCHASTIC B LOCK M ODELS
Consider an SBM(n1 , n2 , p, q) graph G = (V1 ∪V2 , E1 ∪E2 ∪E3 ) with n = n1 +n2 , |V1 | = n1 = Θ(n), and |V2 | = n2 = Θ(n).
Edge subsets E1 , E2 and E3 repectively represent intra-cluster edges among nodes in V1 , intra-cluster edges among nodes in
V2 , and inter-cluster edges between nodes in V1 and V1 . Let G1 = (V1 , E1 ) and G2 = (V2 , E2 ) be graphs induced by V1 and
V2 , respectively, and denote the graph of inter-cluster connections by G3 = (V1 ∪ V2 , E3 ). Clearly, G1 and G2 are Erdös-Renyi
graphs RE(n1 , p) and RE(n2 , p), while G3 is a random bipartite graph RB(n1 , n2 , q).
Consider the two-step allocation, Ã ∈ A(r), proposed in Appendix A.
We propose to carry out the Shuffle phase in three steps in both uncoded and coded scenarios. In coded Shuffling, we first
apply the coded scheme described in Section IV-A to ER graph G1 ; second, the same scheme is applied to ER graph G2 ; and
third, the coded scheme proposed in Appendix A is applied to RB graph G3 . In uncoded scheme, required intermediate values
in each step are transmitted without coding.
For allocation Ã, let L̄UC1
, L̄UC2
, L̄UC3
, L̄C1
, L̄C2
and L̄C3
denote the average normalized communication loads induced by
Ã
Ã
Ã
Ã
Ã
Ã
the Shuffling described in each of the above three steps, respectively for uncoded and coded schemes. According to Theorems
1 and 2, for p ≥ q > Θ( n12 ), we have
r
1
n21
1 UC1
(1 − ) + o(p),
L̄Ã + o(p) = p
(69)
r
r (n1 + n2 )2
K
1
1
n22
r
L̄C2
≤ L̄UC2
+ o(p) = p
(1 − ) + o(p),
(70)
Ã
Ã
r
r (n1 + n2 )2
K
1
1
2n1 n2
r
L̄C3
≤ L̄UC3
+ o(q) = q
(1 − ) + o(q).
(71)
2
Ã
Ã
r
r (n1 + n2 )
K
and L̄UC
the total average normalized communication loads of the coded and uncoded schemes,
Ã
L̄C1
≤
Ã
Let us denote by L̄C
Ã
respectively. Therefore,
L∗ (r) ≤ L̄C
Ã
=
(72)
L̄C1
Ã
L̄C2
Ã
L̄C3
Ã
+
+
1 UC1
+ L̄UC3
) + o(p)
≤ (L̄Ã + L̄UC2
Ã
Ã
r
1
= L̄UC
+ o(p)
r Ã
2
r
pn1 + pn22 + 2qn1 n2
(1 − ) + o(p),
=
(n1 + n2 )2
K
(73)
(74)
(75)
(76)
and Theorem 3 is concluded.
A PPENDIX C
2
Lemma 4. For all p ∈ [0, 1] and s > 0, we have pe + 1 − p ≤ pe2s + 1 − p.
2
Proof. For given p ∈ [0, 1], define f (s) = pes + 1 − p − pe2s + 1 − p . Clearly f (0) = 0. Moreover,
f 0 (s) = 2pp̄(es − e2s ) < 0,
for s > 0. Therefore, f (s) ≤ 0 for all s > 0, concluding the claim of the lemma.
s
(77)
| 7cs.IT
|
Penalized Estimation in Additive Regression
with High-Dimensional Data
Zhiqiang Tan1 & Cun-Hui Zhang
1
arXiv:1704.07229v1 [math.ST] 24 Apr 2017
April 25, 2017
Abstract.
Additive regression provides an extension of linear regression by modeling the
signal of a response as a sum of functions of covariates of relatively low complexity. We study
penalized estimation in high-dimensional nonparametric additive regression where functional
semi-norms are used to induce smoothness of component functions and the empirical L2 norm
is used to induce sparsity. The functional semi-norms can be of Sobolev or bounded variation
types and are allowed to be different amongst individual component functions. We establish
new oracle inequalities for the predictive performance of such methods under three simple
technical conditions: a sub-gaussian condition on the noise, a compatibility condition on the
design and the functional classes under consideration, and an entropy condition on the functional classes. For random designs, the sample compatibility condition can be replaced by its
population version under an additional condition to ensure suitable convergence of empirical
norms. In homogeneous settings where the complexities of the component functions are of the
same order, our results provide a spectrum of explicit convergence rates, from the so-called
slow rate without requiring the compatibility condition to the fast rate under the hard sparsity
or certain Lq sparsity to allow many small components in the true regression function. These
results significantly broadens and sharpens existing ones in the literature.
Key words and phrases.
Additive model; Bounded variation space; ANOVA model; High-
dimensional data; Metric entropy; Penalized estimation; Reproducing kernel Hilbert space;
Sobolev space; Total variation; Trend filtering.
1
Department of Statistics & Biostatistics, Rutgers University. Address: 110 Frelinghuysen Road, Piscataway,
NJ 08854. E-mail: [email protected], [email protected]. The research of Z. Tan was supported in
part by PCORI grant ME-1511-32740. The research of C.-H. Zhang was supported in part by NSF grants
DMS-1513378, IIS-1250985, and IIS-1407939.
1
Introduction
Additive regression is an extension of linear regression where the signal of a response can
be written as a sum of functions of covariates of relatively low complexity. Let (Yi , Xi ),
i = 1, . . . , n, be a set of n independent (possibly non-identically distributed) observations,
where Yi ∈ R is a response variable and Xi ∈ Rd is a covariate (or design) vector. Consider an
additive regression model, Yi = g ∗ (Xi ) + εi with
g∗ (x) =
Pp
j=1
gj∗ (x(j) ),
(1)
where εi is a noise with mean 0 given Xi , x(j) is a vector composed of a small subset of the
components of x ∈ Rd , and gj∗ belongs to a certain functional class Gj . That is, g ∗ (x) lies in
P
the space of additive functions G = { pj=1 gj (x(j) ) : gj ∈ Gj , j = 1, . . . , p}. A function g ∈ G
Pp
(j)
may admit the decomposition g(x) =
j=1 gj (x ) for multiple choices of (g1 , . . . , gp ). In
what follows, such choices are considered equivalent but a favorite decomposition can be used
to evaluate properties of the components of g ∈ G.
In a classical setting (e.g., Stone 1985), each gj∗ is a univariate function and x(j) is the jth
component of x ∈ [0, 1]d , so that p = d. We take a broad view of additive regression and our
(j)
analysis will accommodate the general setting where gj∗ can be multivariate with Xi
being a
block of covariates, possibly overlapping across different j as in functional ANOVA (e.g., Gu
2002). However, most concrete examples will be given in the classical setting.
Additive modeling has been well studied in the setting where the number of components
p is fixed. See Hastie & Tibshirani (1990) and references therein. Recently, building upon
related works in penalized linear regression, there have been considerable progresses in the
development of theory and methods for sparse additive regression in high-dimensional settings
where p can be of greater order than the sample size n but the number of significant components
is still smaller than n. See, for example, Lin & Zhang (2006), Meier et al. (2009), Ravikumar
et al. (2009), Huang et al. (2010), Koltchinskii & Yuan (2010), Raskutti et al. (2012), Suzuki
& Sugiyama (2013), Petersen et al. (2016), and Yuan & Zhou (2016).
In this article, we study a penalized estimator ĝ with a specific associated decomposition
P
ĝ = pj=1 ĝj defined as a minimizer of a penalized loss
kY − gk2n /2 +
over g ∈ G and decompositions g =
Pp
Pp
j=1
j=1 gj ,
ρnj kgj kF,j + λnj kgj kn
where (λnj , ρnj ) are tuning parameters, k · kn is
1
the empirical L2 norm based on the data points, e.g. kY − gk2n = n−1
Pn
i=1 {Yi
− g(Xi )}2 , and
kgj kF,j is a semi-norm describing the complexity of gj ∈ Gj . For simplicity, the association of
(j)
kgj kn and kgj kF,j with Xi
is typically suppressed.
In the above penalty function, the primary role of the empirical norm k · kn is to induce
sparsity, whereas the primary role of the functional semi-norm k · kF,j is to induce smoothness
R 1 (m)
of the estimated regression function. For example, kgj kF,j = { 0 (gj )2 dz}1/2 when Gj is the
(m)
L2 -Sobolev space W2m on [0, 1], where gj
denotes the mth derivative of gj .
We consider both fixed and random designs and establish oracle inequalities for the predictive performance of ĝ under three simple technical conditions: a sub-gaussian condition on
noises, a compatibility condition on the design and the functional classes Gj , and an entropy
condition on Gj . The compatibility condition is similar to the restricted eigenvalue condition
used in analysis of Lasso, and for random designs, the empirical compatibility condition can be
replaced by its population version under an additional condition to ensue suitable convergence
of empirical norms. For the Sobolev and bounded variation classes, the entropy condition on
Gj follows from standard results in the literature (e.g., Lorentz et al. 1996).
The implications of our oracle inequalities can be highlighted in the classical homogeneous
(j)
setting where Xi
is the jth component of Xi and Gj = G0 for all j, where G0 is either an
Lr -Sobolev space Wrm or a bounded variation space V m of univariate functions on [0, 1], where
r ≥ 1 and m ≥ 1 are shape and smoothness indices of the space, and r = 1 for V m . In
this setting, it is natural to set (λnj , ρnj ) = (λn , ρn ) for all j. Consider random designs, and
suppose that for some choice of (g1∗ , . . . , gp∗ ) satisfying (1),
p
X
j=1
where kf k2Q = n−1
Pn
i=1 E{f
kgj∗ kF,j
2 (X )}
i
p
X
≤ C1 MF ,
j=1
kgj∗ kqQ ≤ C1q Mq ,
(2)
for a function f (x), C1 > 0 is a scaling constant depending
only on the moments of εi , and 0 ≤ q ≤ 1, Mq > 0 and MF > 0 are allowed to depend on
(n, p). In the case of hard sparsity, q = 0, M0 = #{j : gj∗ 6= 0}. As a summary, the following
result can be easily deduced from Proposition 3, 5, and 7.
Let β0 = 1/m and define
o
n
−1
1−q
wn∗ (q) = max n 2+β0 (1−q) , (log(p)/n) 2 ,
−1
−(1−q)β0
−1/2
∗
2+β0 (1−q)
4
,n
(log(p)/n)
γn (q) = min n
.
2
For simplicity, we restrict to the case where 1 ≤ r ≤ 2. For rm > 1, we assume that the
(j)
(j)
average marginal density of (X1 , . . . , Xn ) are uniformly bounded away from 0 and, if q 6= 1,
also uniformly bounded from above for all j = 1, . . . , p. The assumption of marginal densities
bounded from above, as well as the restriction 1 ≤ r ≤ 2, can be relaxed under slightly different
technical conditions (see Propositions 3, 4, and 6). For r = m = 1, neither the lower bound
nor the upper bound of marginal densities need to be assumed.
Proposition 1. Let G0 be a Sobolev space Wrm with 1 ≤ r ≤ 2 and m ≥ 1 or a bounded
variation space V m with r = 1 and m ≥ 1. Suppose that the noises are sub-gaussian and
√
log(p) = o(n). Let τ0 = 1/(2m + 1 − 2/r), Γn = 1 for rm > 1 and Γn = log n for r = m = 1.
(i) Let q = 1 and λn = ρn = A0 {log(p)/n}1/2 for a sufficiently large constant A0 . If p → ∞,
then
o
n
p
kĝ − g∗ k2Q = Op (1)C12 (MF2 + M12 ) n−1/2 Γn + log(p)/n .
(3)
(ii) Let q = 0, λn = A0 [γn∗ (0) + {log(p)/n}1/2 ] and ρn = λn wn∗ (0). Suppose that
n
wn∗ (0)−τ0
p
o
log(np)/n (1 + MF + M0 ) = o(1)
(4)
and a population compatibility condition (Assumption 5) holds. Then,
n −1
o2
p
kĝ − g∗ k2Q = Op (1)C12 (MF + M0 ) n 2+β0 + log(p)/n .
(5)
(iii) Let 0 < q < 1, λn = A0 [γn∗ (q) + {log(p)/n}1/2 ] and ρn = λn wn∗ (q). Suppose that
n
wn∗ (q)−τ0 (log(np)/n)
1−q
2
o
(1 + MF + Mq ) = O(1)
and a population compatibility condition (Assumption 7) holds. Then,
n
o2−q
p
−1
kĝ − g∗ k2Q = Op (1)C12 (MF + Mq ) n 2+β0 (1−q) + log(p)/n
.
(6)
There are several important features achieved by the foregoing result, distinct from existing
results. First, our results are established for additive regression with Sobolev spaces of general shape and bounded variation spaces. An important innovation in our proofs involves a
delicate application of maximal inequalities based on the metric entropy of a particular choice
of bounded subsets of G0 (see Lemma 1). All previous results seem to be limited to the L2 Sobolev spaces or similar reproducing kernel Hilbert spaces, except for Petersen et al. (2016),
who studied additive regression with the bounded variation space V 1 and obtained the rate
3
{log(np)/n}1/2 for in-sample prediction under assumption (2) with q = 1. In contrast, our
analysis in the case of q = 1 yields the sharper, yet standard, rate {log(p)/n}1/2 for in-sample
prediction (see Proposition 3), whereas {log(np)/n}1/2 for out-of-sample prediction by (3).
Second, the restricted parameter set (2) represents an L1 ball in k · kF semi-norm (inducing
smoothness) but an Lq ball in k · kQ norm (inducing sparsity) for the component functions
(g1∗ , . . . , gp∗ ). That is, the parameter set (2) decouples conditions for sparsity and smoothness
in additive regression: it can encourage sparsity at different levels 0 ≤ q ≤ 1 while enforcing
smoothness only to a limited extent. Accordingly, our result leads to a spectrum of convergence
rates, which are easily seen to slow down as q increases from 0 to 1, corresponding to weaker
sparsity assumptions. While most of previous results are obtained under exact sparsity (q =
0), Yuan & Zhou (2016) studied additive regression with reproducing kernel Hilbert spaces
Pp
∗ q
under an Lq ball in the Hilbert norm k · kH :
j=1 kgj kH ≤ Mq . This parameter set induces
smoothness and sparsity simultaneously and is in general more restrictive than (2). As a
result, the minimax rate of estimation obtained by Yuan & Zhou (2016), based on constrained
least squares with known Mq instead of penalized estimation, is faster than (6), in the form
n−2/(2+β0 ) + {log(p)/n}(2−q)/2 , unless q = 0 or 1.
Third, in the case of q = 1, our result (3) shows that the rate {log(p)/n}1/2 , with an
additional {log(n)/n}1/2 term for the bounded variation space V 1 , can be achieved via penalized
estimation without requiring a compatibility condition. This generalizes a slow-rate result for
constrained least-squares (instead of penalization) with known (M1 , MF ) in additive regression
with the Sobolev Hilbert space in Ravikumar et al. (2009). Both are related to earlier results
for linear regression (Greenhstein & Ritov 2004; Bunea et al. 2007).
Finally, compared with previous results giving the same rate of convergence (5) under exact
sparsity (q = 0) for Hilbert spaces, our results are stronger in requiring much weaker technical
conditions. The penalized estimation procedures in Koltchinskii & Yuan (2010) and Raskutti
et al. (2012), while minimizing a similar criterion as Kn (g), involve additional constraints:
Koltchinskii & Yuan (2010) assumed that the sup-norm of possible g∗ is bounded by a known
constant, where as Raskutti et al. (2012) assumed maxj kgj kH is bounded by a known constant.
(1)
(p)
Moreover, Raskutti et al. (2012) assumed that the covariates (Xi , . . . , Xi ) are independent
of each other. These restrictions were relaxed in Suzuki & Sugiyama (2013), but only explicitly
under the assumption that the noises εi are uniformly bounded by a constant. Moreover, our
rate condition (4) about the sizes of (M0 , MF ) is much weaker than in Suzuki & Sugiyama
4
(2013), due to improved analysis of convergence of empirical norms and the more careful
choices (λn , ρn ). For example, if (M0 , MF ) are bounded, then condition (4) holds whenever
log(p)/n = o(1) for Sobolev Hilbert spaces, but the condition previously required amounts to
log(p)n−1/2 = o(1). Finally, the seemingly faster rate in Suzuki & Sugiyama (2013) can be
deduced from our results when (λn , ρn ) is allowed to depend on (M0 , MF ). See Remarks 8 and
14–16 for relevant discussion.
The rest of the article is organized as follows. Section 2 gives a review of univariate functional
classes and entropies. Section 3 presents general results for fixed designs (Section 3.1) and
random designs (Section 3.2), and then provides specific results with Sobolev and bounded
variation spaces (Section 3.4) after a study of convergence of empirical norms (Section 3.3).
Section 4 concludes with a discussion. For space limitation, all proofs are collected in Section S1
and technical tools are stated in Section S2 of the Supplementary Material.
2
Functional classes and entropies
As a building block of additive regression, we discuss two broad choices for the function space Gj
and the associated semi-norm kgj kF,j in the context of univariate regression. For concreteness,
we consider a fixed function space, say G1 , although our discussion is applicable to Gj for j =
R1
1, . . . , p. For r ≥ 1, the Lr norm of a function f on [0, 1] is defined as kf kLr = { 0 |f (z)|r dz}1/r .
Example 1 (Sobolev spaces). For r ≥ 1 and m ≥ 1, let Wrm = Wrm ([0, 1]) be the Sobolev
(m−1)
space of all functions, g1 : [0, 1] → R, such that g1
(m)
(m)
kg1 kWrm = kg1 kLr + kg1 kLr is finite, where g1
is absolutely continuous and the norm
denotes the mth (weak) derivative of g1 . To
(m)
describe the smoothness, a semi-norm kg1 kF,1 = kg1 kLr is often used for g1 ∈ Wrm .
In the statistical literature, a major example of Sobolev spaces is W2m = {g1 : kg1 kL2 +
(m)
kg1 kL2 < ∞}, which is a reproducing kernel Hilbert space (e.g., Gu 2002). Consider a
univariate regression model
(1)
Yi = g1 (Xi ) + εi ,
i = 1, . . . , n.
(7)
The Sobolev space W2m is known to lead to polynomial smoothing splines through penalized
estimation: there exists a unique solution, in the form of a spline of order (2m − 1), when
minimizing over g1 ∈ W2m the following criterion
n
o2
1 Xn
(1)
Yi − g1 (Xi ) + ρn1 kg1 kF,1 .
2n
i=1
5
(8)
This solution can be made equivalent to the standard derivation of smoothing splines, where the
penalty in (8) is ρ′n1 kg1 k2F,1 for a different tuning parameter ρ′n1 . Particularly, cubic smoothing
splines are obtained with the choice m = 2.
Example 2 (Bounded variation spaces). For a function f on [0, 1], the total variation (TV)
of f is defined as
TV(f ) = sup
( k
X
i=1
)
|f (zi ) − f (zi−1 )| : z0 < z1 < . . . < zk is any partition of [0, 1] .
If f is differentiable, then TV(f ) =
R1
0
|f (1) (z)| dz. For m ≥ 1, let V m = V m ([0, 1]) be the
(m−2)
bounded variation space that consists of all functions, g1 : [0, 1] → R, such that g1
m ≥ 2, is absolutely continuous and the norm kg1 kV m = kg1 kL1 +
g1 ∈ V m , the semi-norm kg1 kF,1 =
(m−1)
TV(g1
)
(m−1)
TV(g1
)
, if
is finite. For
is often used to describe smoothness. The
bounded variation space V m includes as a strict subset the Sobolev space W1m , where the
(m−1)
semi-norms also agree: TV(g1
(m)
) = kg1 kL1 for g1 ∈ W1m .
For univariate regression (7) with bounded variation spaces, TV semi-norms can be used
as penalties in (8) for penalized estimation. This leads to a class of TV splines, which are
shown to adapt well to spatial inhomogeneous smoothness (Mammen & van de Geer 1997).
For m = 1 or 2, a minimizer of (8) over g1 ∈ V m can always be chosen as a spline of order m,
(1)
with the knots in the set of design points {Xi
: i = 1, . . . , n}. But, as a complication, this is
in general not true for m ≥ 3.
Recently, there is another smoothing method related to TV splines, called trend filtering
(1)
(Kim et al. 2009), where (8) is minimized over all possible values {g1 (Xi ) : i = 1, . . . , n}
with kg1 kF,1 replaced by L1 norm of mth-order differences of these values. This method is
equivalent to TV splines only for m = 1 or 2. But when the design points are evenly spaced,
it achieves the minimax rate of convergence over functions of bounded variation for general
m ≥ 1, similarly as TV splines (Tibshirani 2014).
The complexity of a functional class can be described by its metric entropy, which plays
an important role in the study of empirical processes (van der Vaart & Wellner 1996). For a
subset F in a metric space F endowed with norm k · k, the covering number N (δ, F, k · k) is
defined as the smallest number of balls of radius δ in the k·k-metric needed to cover F, i.e., the
smallest value of N such that there exist f1 , . . . , fN ∈ F , satisfying minj=1,...,N kf − fj k ≤ δ
for any f ∈ F. The entropy of (F, k · k) is defined as H(δ, F, k · k) = log N (δ, F, k · k).
6
For analysis of regression models, our approach involves using entropies of functional classes
(1)
for empirical norms based on design points, for example, {Xi
: i = 1, . . . , n} for subsets of G1 .
P
(1)
One type of such norms is the empirical L2 norm, kg1 kn = {n−1 ni=1 g12 (Xi )}1/2 . Another
(1)
is the empirical supremum norm, kg1 kn,∞ = maxi=1,...,n g1 (Xi ). If F is the unit ball in
the Sobolev space Wrm or the bounded variation space V m on [0, 1], the general picture is
H(δ, F, k · k) . δ−1/m for commonly used norms. See Section S2.5 for more.
3
Main results
As in Section 1, consider the estimator
Kn (g) = kY − gk2n /2 + A0 Rn (g),
ĝ = argming∈G Kn (g),
where A0 > 1 is a constant, G = {g =
Rn (g) =
p
X
j=1
for any decomposition g =
Pp
Pp
j=1 gj
(9)
: gj ∈ Gj } and the penalty is of the form
p
X
ρnj kgj kF,j + λnj kgj kn
Rnj (gj ) =
j=1
j=1 gj
with gj ∈ Gj , with certain functional penalties kfj kF,j and
the empirical L2 penalty kfj kn . Here the regularization parameters (λnj , ρnj ) are of the form
ρnj = λnj wnj ,
n
λnj = C1 γnj +
p
o
log(p/ǫ)/n ,
where C1 > 0 is a noise level depending only on parameters in Assumption 1 below, 0 < ǫ < 1
is a tail probability for the validity of error bounds, 0 < wnj ≤ 1 is a rate parameter, and
γnj = n−1/2 ψnj (wnj )/wnj
(10)
for a function ψnj (·) depending on the entropy of the unit ball of the space Gj under the
associated functional penalty. See Assumption 2 or 4 below.
Before theoretical analysis, we briefly comment on computation of ĝ. By standard properties
of norms and semi-norms, the objective function Kn (g) is convex in g. Moreover, there are
at least two situations where the infinitely-dimensional problem of minimizing Kn (g) can be
reduced to a finite-dimensional one. First, if each class Gj is a reproducing kernel Hilbert space
P
such as W2m , then a solution ĝ = pj=1 ĝj can be obtained such that each ĝj is a smoothing
(j)
spline with knots in the design points {Xi
: i = 1, . . . , n} (e.g., Meier et al. 2009). Second, by
the following proposition, the optimization problem can be also reduced to a finite-dimensional
7
one when each class Gj is the bounded variation space V 1 or V 2 . As a result, the algorithm in
Petersen et al. (2016) can be directly used to find ĝ when all classes (G1 , . . . , Gp ) are V 1 .
Proposition 2. Suppose that the functional class Gj is V m for some 1 ≤ j ≤ p and m = 1
P
or 2. Then a solution ĝ = pj=1 ĝj can be chosen such that ĝj is piecewise constant with jump
(j)
points only in {Xi
: i = 1, . . . , n} if m = 1, or ĝj is continuous and piecewise linear with
(j)
break points only in {Xi
: i = 1, . . . , n} if m = 2.
By Example 2, it can be challenging to compute ĝ when some classes Gj are V m with m ≥ 3.
However, this issue may be tackled using trend filtering (Kim et al. 2009) as an approximation.
3.1
Fixed designs
For fixed designs, the covariates (X1 , . . . , Xn ) are fixed as observed, whereas (ε1 , . . . , εn ) and
hence (Y1 , . . . , Yn ) are independent random variables. The responses are to be predicted when
new observations are drawn with covariates from the sample (X1 , . . . , Xn ). The predictive
performance of ĝ is measured by kĝ − g ∗ k2n .
Consider the following three assumptions. First, we assume sub-gaussian tails for the noises.
This condition can be relaxed, but with increasing technical complexity and possible modification of the estimators, which we will not pursue here.
Assumption 1 (Sub-gaussian noises). Assume that the noises (ε1 , . . . , εn ) are mutually independent and uniformly sub-Gaussian: For some constants D0 > 0 and D1 > 0,
max D0 E exp(ε2i /D0 ) ≤ D1 .
i=1,...,n
We will also impose this assumption for random designs with the interpretation that the above
probability and expectation are taken conditionally on (X1 , . . . , Xn ).
Second, we impose an entropy condition which describes the relationship between the function ψnj (·) in the definition of γnj and the complexity of bounded subsets in Gj . Although
entropy conditions are widely used to analyze nonparametric regression (e.g., Section 10.1, van
de Geer 2000), the subset Gj (δ) in our entropy condition below is carefully aligned with the
penalty Rnj (gj ) = λnj (wnj kgj kF,j + kgj kn ). This leads to a delicate use of maximal inequalities so as to relax and in some cased remove some restrictions in previous studies of additive
models. See Lemma 1 in the Supplement and Raskutti et al. (2012, Lemma 1).
8
Assumption 2 (Entropy condition for fixed designs). For j = 1, . . . , p, let Gj (δ) = {fj ∈ Gj :
kfj kF,j + kfj kn /δ ≤ 1} and ψnj (δ) be an upper bound of the entropy integral as follows:
Z δ
H 1/2 (u, Gj (δ), k · kn ) du, 0 < δ ≤ 1.
(11)
ψnj (δ) ≥
0
(j)
In general, Gj (δ) and the entropy H(·, Gj (δ), k · kn ) may depend on the design points {Xi }.
The third assumption is a compatibility condition, which resembles the restricted eigenvalue
condition used in high-dimensional analysis of Lasso in linear regression (Bickel et al. 2009).
Similar compatibility conditions were used by Meier et al. (2009) and Koltchinskii & Yuan
(2010) in their analysis of penalized estimation in high-dimensional additive regression.
Assumption 3 (Empirical compatibility condition). For certain subset S ⊂ {1, 2, . . . , p} and
constants κ0 > 0 and ξ0 > 1, assume that
2
X
X
λnj kfj kn ≤
κ20
λ2nj kf k2n
j∈S
j∈S
for any functions {fj ∈ Gj : j = 1, . . . , p} and f =
p
X
j=1
λnj wnj kfj kF,j +
X
j∈S c
Pp
j=1 fj
∈ G satisfying
λnj kfj kn ≤ ξ0
X
j∈S
λnj kfj kn .
Remark 1. The subset S can be different from {1 ≤ j ≤ p : gj∗ 6= 0}. In fact, S is arbitrary
in the sense that a larger S leads to a smaller compatibility coefficient κ0 which appears as a
factor in the denominator of the “noise” term in the prediction error bound below, whereas a
smaller S leads to a larger “bias” term. Assumption 3 is automatically satisfied for the choice
S = ∅. In this case, it is possible to take ξ0 = ∞ and any κ0 > 0, provided that we treat
summation over an empty set as 0 and ∞ × 0 as 0.
Our main result for fixed designs is an oracle inequality stated in Theorem 1 below, where
P
ḡ = pj=1 ḡj ∈ G as an estimation target is an additive function but the true regression function
g∗ may not be additive. Denote as a penalized prediction loss
Dn (ĝ, ḡ) =
1
1
kĝ − g∗ k2n + kĝ − ḡk2n + (A0 − 1)Rn (ĝ − ḡ).
2
2
For a subset S ⊂ {1, 2, . . . , p}, write as a bias term for the target ḡ
P
P
1
p
λ
kḡ
k
ρ
kḡ
k
+
∆n (ḡ, S) = kḡ − g∗ k2n + 2A0
j∈S c nj j n .
j=1 nj j F,j
2
The bias term is small when ḡ is smooth and sparse and predicts g ∗ well.
9
Theorem 1. Suppose that Assumptions 1, 2, and 3 hold. Then for any A0 > (ξ0 + 1)/(ξ0 − 1),
we have with probability at least 1 − ǫ,
Dn (ĝ, ḡ) ≤ ξ1−1 ∆n (ḡ, S) + 2ξ22 κ−2
0
P
2
λ
j∈S nj .
(12)
where ξ1 = 1 − 2A0 /{(ξ0 + 1)(A0 − 1)} ∈ (0, 1] and ξ2 = (ξ0 + 1)(A0 − 1).
Remark 2. As seen from our proofs, Theorem 1 and subsequent corollaries are directly ap(j)
plicable to functional ANOVA modeling, where each function gj may depend on Xi , a block
of covariates, and the variable blocks are allowed to overlap across different j. The entropy
associated with the functional class Gj need to be determined accordingly.
Remark 3. Using ideas from Bellec and Tsybakov (2016), it is possible to refine the oracle
inequality for ĝ, such that the scaling parameter ǫ is fixed, for example, ǫ = 1/2 in the definition
of ĝ in (9), but at any level 0 < ǫ̃ < 1, (12) holds with probability 1 − ǫ̃ when an additional
term of the form log(1/ǫ̃)/n on the right-hand side.
Taking S = ∅ and ξ0 = ∞ leads to the following corollary, which explicitly does not require
the compatibility condition (Assumption 3).
Corollary 1. Suppose that Assumptions 1 and 2 hold. Then for any A0 > 1, we have with
probability at least 1 − ǫ,
1
Dn (ĝ, ḡ) ≤ ∆n (ḡ, ∅) = kḡ − g∗ k2n + 2A0 Rn (ḡ).
2
(13)
The following result can be derived from Theorem 1 through the choice S = {1 ≤ j ≤ p :
kḡj kn > C0 λnj } for some constant C0 > 0.
Corollary 2. Suppose that Assumptions 1, 2, and 3 hold with S = {1 ≤ j ≤ p : kḡj kn >
C0 λnj } for some constant C0 > 0. Then for any 0 ≤ q ≤ 1 and A0 > (ξ0 + 1)/(ξ0 − 1), we
have with probability at least 1 − ǫ,
p
X
q
Dn (ĝ, ḡ) ≤ O(1) kḡ − g ∗ k2n +
ρnj kḡj kF,j + λ2−q
nj kḡj kn ,
j=1
where O(1) depends only on (q, A0 , C0 , ξ0 , κ0 ).
It is instructive to examine the implications of Corollary 2 in a homogenous situation where
for some constants B0 > 0 and 0 < β0 < 2,
Z δ
H 1/2 (u, Gj (δ), k · kn ) du ≤ B0 δ1−β0 /2 ,
max
j=1,...,p 0
10
0 < δ ≤ 1.
(14)
That is, we assume ψnj (δ) = B0 δ1−β0 /2 in (11). For j = 1, . . . , p, let
wnj = wn (q) = {γn (q)}1−q ,
2
2+β0 (1−q)
γnj = γn (q) = B0
−1
n 2+β0 (1−q) ,
(15)
1−q
which are determined by balancing the two rates ρnj = λ2−q
nj , that is, wnj = λnj , along with the
P
P
−β /2
definition γnj = B0 n−1/2 wnj 0 by (10). For g = pj=1 gj ∈ G, denote kgkF,1 = pj=1 kgj kF,j
P
and kgkn,q = pj=1 kgj kqn . For simplicity, we also assume that g∗ is an additive function and
set ḡ = g ∗ for Corollary 3.
Corollary 3. Assume that (1) holds and kg∗ kF,1 ≤ C1 MF and kg∗ kn,q ≤ C1q Mq for 0 ≤ q ≤ 1,
Mq > 0, and MF > 0, possibly depending on (n, p). In addition, suppose that (14) and (15)
hold, and Assumptions 1 and 3 are satisifed with S = {1 ≤ j ≤ p : kgj∗ kn > C0 λnj } for some
constant C0 > 0. If 0 < wn (q) ≤ 1 for sufficiently large n, then for any A0 > (ξ0 + 1)/(ξ0 − 1),
we have with probability at least 1 − ǫ,
Dn (ĝ, g∗ ) = kĝ − g ∗ k2n + (A0 − 1)Rn (ĝ − g∗ )
n
o2−q
p
≤ O(1)C12 (MF + Mq ) γn (q) + log(p/ǫ)/n
,
(16)
where O(1) depends only on (q, A0 , C0 , ξ0 , κ0 ).
Remark 4. There are several interesting features in the convergence rate (16). First, (16)
presents a spectrum of convergence rates in the form
n
−1
n 2+β0 (1−q) +
p
log(p)/n
o2−q
,
which are easily shown to become slower as q increases from 0 to 1, that is, the exponent
(2 − q)/{2 + β0 (1 − q)} is decreasing in q for 0 < β0 < 2. The rate (16) gives the slow rate
−2
{log(p)/n}1/2 for q = 1, or the fast rate n 2+β0 + log(p)/n for q = 0, as previously obtained
for additive regression with reproducing kernel Hilbert spaces. We defer to Section 3.4 the
comparison with existing results in random designs. Second, the rate (16) is in general at least
as fast as
n −1
o2−q
p
n 2+β0 + log(p)/n
.
Therefore, weaker sparsity (larger q) leads to a slower rate of convergence, but not as slow
−2
as the fast rate {n 2+β0 + log(p)/n} raised to the power of (2 − q)/2. This is in contrast
with previous results on penalized estimation over Lq sparsity balls, for example, the rate
{k/n + log(p)/n}(2−q)/2 obtained for group Lasso estimation in linear regression (Neghaban
11
et al. 2012), where k is the group size. Third, the rate (16) is in general not as fast as the
following rate (unless q = 0 or 1)
−2
n 2+β0 + {log(p)/n}(2−q)/2 ,
which was obtained by Yuan & Zhou (2016) using constrained least squares for additive
regression with reproducing kernel Hilbert spaces under an Lq ball in the Hilbert norm:
Pp
∗ q
j=1 kgj kH ≤ Mq . This difference can be explained by the fact that an Lq ball in k · kH
norm is more restrictive than in k · kn or k · kQ norm for our results.
3.2
Random designs
For random designs, prediction of the responses can be sought when new observations are randomly drawn with covariates from the distributions of (X1 , . . . , Xn ), instead of within the sample (X1 , . . . , Xn ) as in Section 3.1. For such out-of-sample prediction, the performance of ĝ is
P
measured by kĝ−g ∗ k2Q , where k·kQ denotes the theoretical norm: kf k2Q = n−1 ni=1 E{f 2 (Xi )}
for a function f (x).
Consider the following two extensions of Assumptions 2 and 3, such that dependency on
the empirical norm k · kn and hence on (X1 , . . . , Xn ) are removed.
Assumption 4 (Entropy condition for random designs). For some constant 0 < η0 < 1 and
j = 1, . . . , p, let ψnj (δ) be an upper bound of the entropy integral, independent of the realizations
(j)
{Xi
: i = 1, . . . , n}, as follows:
ψnj (δ) ≥
Z
δ
0
H ∗1/2 ((1 − η0 )u, Gj∗ (δ), k · kn ) du,
0 < δ ≤ 1,
(17)
where Gj∗ (δ) = {fj ∈ Gj : kfj kF,j + kfj kQ /δ ≤ 1} and
H ∗ (u, Gj∗ (δ), k · kn ) =
sup
(j)
(j)
(X1 ,...,Xn )
H(u, Gj∗ (δ), k · kn ).
Assumption 5 (Theoretical compatibility condition). For some subset S ⊂ {1, 2, . . . , p} and
constants κ∗0 > 0 and ξ0∗ > 1, assume that for any functions {fj ∈ Gj : j = 1, . . . , p} and
P
f = pj=1 fj ∈ G, if
p
X
j=1
λnj wnj kfj kF,j +
X
j∈S c
λnj kfj kQ ≤ ξ0∗
12
X
j∈S
λnj kfj kQ ,
(18)
then
κ∗2
0
X
j∈S
2
X
λnj kfj kQ ≤
λ2nj kf k2Q .
(19)
j∈S
Remark 5. Similarly as in Remark 1 about the empirical compatibility condition, Assumption 5 is also automatically satisfied for the choice S = ∅, in which case it is possible to take
ξ0∗ = ∞ and any κ∗0 > 0.
To tackle random designs, our approach relies on establishing appropriate convergence of
empirical norms k · kn to k · kQ uniformly over the space of additive functions G, similarly as
in Meier et al. (2009) and Koltchinskii & Yuan (2010). For clarity, we postulate the following
assumption on the rate of such convergence to develop general analysis of ĝ. We will study
convergence of empirical norms specifically for Sobolev and bounded variation spaces in Section 3.3, and then provide corresponding results on the performance of ĝ in Section 3.4. For
P
g = pj=1 gj ∈ G, denote
Rn∗ (g)
=
p
X
j=1
∗
Rnj
(gj ),
∗
Rnj
(gj ) = λnj (wnj kgj kF,j + kgj kQ ),
as the population version of the penalty Rn (g), with kgj kQ in place of kgj kn .
Assumption 6 (Convergence of empirical norms). Assume that
kgk2n − kgk2Q
P sup
>
φ
≤ π,
n
g∈G
Rn∗2 (g)
(20)
where 0 < π < 1 and φn > 0 such that for sufficiently large n, one or both of the following
conditions are valid.
(i) φn (maxj=1,...,p λ2nj ) ≤ η02 , where η0 is from Assumption 4.
(ii) For some constant 0 ≤ η1 < 1, we have
φn (ξ0∗ + 1)2 κ∗−2
0
X
j∈S
λ2nj ≤ η12 ,
(21)
where S is the subset of {1, 2, . . . , p} used in Assumption 5.
Our main result, Theorem 2, gives an oracle inequality for random designs, where the predicP
tive performance of ĝ is compared with that of an arbitrary additive function ḡ = pj=1 ḡj ∈ G,
13
but the true regression function g ∗ may not be additive, similarly as in Theorem 1 for fixed
designs. For a subset S ⊂ {1, 2, . . . , p}, denote
1
∆∗n (ḡ, S) = kḡ − g ∗ k2n + 2A0 (1 − η0 )
2
p
X
j=1
ρnj kḡj kF,j +
X
j∈S c
λnj kḡj kQ ,
which, unlike ∆n (ḡ, S), involves kḡj kQ and η0 from Assumptions 4 and 6(i).
Theorem 2. Suppose that Assumptions 1, 4, 5 and 6(i)–(ii) hold with 0 < η0 < (ξ0∗ − 1)/(ξ0∗ +
1). Let A(ξ0∗ , η0 ) = {ξ0∗ + 1 + η0 (ξ0∗ + 1)}/{ξ0∗ − 1 − η0 (ξ0∗ + 1)} > (1 + η0 )/(1 − η0 ). Then for
any A0 > A(ξ0∗ , η0 ), we have with probability at least 1 − ǫ − π,
1
1
kĝ − g ∗ k2n + kĝ − ḡk2n + (1 − η1 )A1 Rn∗ (ĝ − ḡ)
2
2
X
λ2nj ,
≤ ξ1∗−1 ∆∗n (ḡ, S) + 2ξ2∗2 κ∗−2
0
(22)
j∈S
where A1 = (A0 − 1) − η0 (A0 + 1) > 0, ξ1∗ = 1 − 2A0 /{(ξ0∗ + 1)A1 } ∈ (0, 1] and ξ2∗ = (ξ0∗ + 1)A1 .
Moreover, we have with probability at least 1 − ǫ − π,
1
1
kĝ − g ∗ k2n + kĝ − ḡk2Q + A2 Rn∗ (ĝ − ḡ)
2
2
X
φn
∗
≤ ξ3∗−1 ∆∗n (ḡ, S) + 2ξ4∗2 κ∗−2
λ2nj +
ξ ∗−2 ∆∗2
n (ḡ, g , S),
0
2A1 A2 3
Dn∗ (ĝ, ḡ) :=
(23)
j∈S
where A2 = A1 /(1 − η12 ), ξ3∗ = ξ1∗ (1 − η12 ), and ξ4∗ = ξ2∗ /(1 − η12 ).
Remark 6. Similarly as in Remark 2, we emphasize that Theorem 2 and subsequent corollaries
are also applicable to functional ANOVA modeling (e.g., Gu 2002). For example, consider
model (1) studied in Yang & Tokdar (2015), where each gj∗ is assumed to depend only on d0
of a total of d covariates and lie in a Hölder space with smoothness level α0 . Then p = dd0 ,
and the entropy condition (27) holds with β0 = d0 /α0 . Under certain additional conditions,
Corollary 6 with q = 0 shows that penalized estimation studied here achieves a convergence
−2
−2
rate M0 n 2+β0 + M0 log(p)/n under exact sparsity of size M0 , where n 2+β0 is the rate for
estimation of a single regression function in the Hölder class in dimension d0 with smoothness
β0−1 , and log(p)/n ≍ d0 log(d/d0 )/n is the term associated with handling p regressors. This
result agrees with the minimax rate derived in Yang & Tokdar (2015), but can be applied when
more general functional classes are used such as multi-dimensional Sobolev spaces. In addition,
14
Yang & Todkar (2015) considered adaptive Bayes estimators which are nearly minimax with
some extra logarithmic factor in n.
Taking S = ∅, ξ0∗ = ∞, and η1 = 0 leads to the following corollary, which explicitly
does not require the theoretical compatibility condition (Assumption 5) or the rate condition,
Assumption 6(ii), for convergence of empirical norms.
Corollary 4. Suppose that Assumptions 1, 4, and 6(i) hold. Then for any A0 > (1 + η0 )/(1 −
η0 ), we have with probability at least 1 − ǫ − π,
1
1
kĝ − g∗ k2n + kĝ − ḡk2n + A1 Rn∗ (ĝ − ḡ)
2
2
1
∗
≤ ∆n (ḡ, ∅) = λ2n0 + kḡ − g∗ k2n + 2A0 Rn∗ (ḡ).
2
(24)
Moreover, we have with probability at least 1 − ǫ − π,
1
1
kĝ − g∗ k2n + kĝ − ḡk2Q + A1 Rn∗ (ĝ − ḡ)
2
2
φn ∗2
∗
≤ ∆n (ḡ, ∅) +
∆ (ḡ, ∅).
2A21 n
(25)
The preceding results deal with both in-sample and out-of-sample prediction. For space limitation, except in Proposition 3, we hereafter focus on the more challenging out-of-sample prediction. Under some rate condition about φn in (20), the additional term involving φn ∆∗2
n (ḡ, S)
can be absorbed into the first term, as shown in the following corollary. Two possible scenarios
are accommodated. On one hand, taking ḡ = g ∗ directly gives high-probability bounds on the
prediction error kĝ − g∗ k2Q provided that g∗ is additive, that is, model (1) is correctly specified.
On the other hand, the error kĝ − g ∗ k2Q can also be bounded, albeit in probability, in terms of
an arbitrary additive function ḡ ∈ G, while allowing g ∗ to be non-additive.
Corollary 5. Suppose that the conditions of Theorem 2 hold with S = {1 ≤ j ≤ p : kḡj kQ >
C0∗ λnj } for some constant C0∗ > 0, and (20) holds with φn > 0 also satisfying
p
X
X
φn
ρnj kḡj kF,j +
λnj kḡj kQ ≤ η2 ,
(26)
j∈S c
j=1
for some constant η2 > 0. Then for any 0 ≤ q ≤ 1 and A0 > A(ξ0∗ , η0 ), we have with probability
at least 1 − ǫ − π,
p
X
2−q
ρnj kḡj kF,j + λnj
kḡj kqQ
Dn∗ (ĝ, ḡ) ≤ {O(1) + φn kḡ − g ∗ k2n } kḡ − g ∗ k2n +
,
j=1
15
where O(1) depends only on (q, A∗0 , C0∗ , ξ0∗ , κ∗0 , η0 , η1 , η2 ). In addition, suppose that φn kḡ − g∗ k2Q
is bounded by a constant and ǫ = ǫ(n, p) tends to 0 in the definition of ĝ in (9). Then for any
0 ≤ q ≤ 1, we have
p
X
q
kĝ − g∗ k2Q ≤ Op (1) kḡ − g ∗ k2Q +
ρnj kḡj kF,j + λ2−q
nj kḡj kQ .
j=1
Similarly as Corollary 3, it is useful to deduce the following result in a homogeneous situation
where for some constants B0∗ > 0 and 0 < β0 < 2,
Z δ
H ∗1/2 ((1 − η0 )u, Gj∗ (δ), k · kn ) du ≤ B0∗ δ1−β0 /2 ,
max
j=1,...,p 0
0 < δ ≤ 1.
(27)
That is, we assume ψnj (δ) = B0∗ δ1−β0 /2 in (17). For j = 1, . . . , p, let
wnj = wn∗ (q) = max{γn (q)1−q , νn1−q },
(28)
γnj = γn∗ (q) = min{γn (q), B0∗ n−1/2 νn−(1−q)β0 /2 },
(29)
where νn = {log(p/ǫ)/n}1/2 , and wn (q) = γn (q)1−q and
2
−1
−1
γn (q) = B0∗ 2+β0 (1−q) n 2+β0 (1−q) ≍ n 2+β0 (1−q)
are determined from the relationship (10), that is, γn (q) = B0∗ n−1/2 wn (q)−β0 /2 . The reason
why (wn∗ (q), γn∗ (q)) are used instead of the simpler choices (wn (q), γn (q)) is that the rate condition (30) needed below would become stronger if γn∗ (q) were replaced by γn (q). The rate
of convergence, however, remains the same even if γn∗ (q) is substituted for γn (q) in (31). See
P
P
Remark 16 for further discussion. For g = pj=1 gj ∈ G, denote kgkF,1 = pj=1 kgj kF,j and
P
kgkQ,q = pj=1 kgj kqQ .
Corollary 6. Assume that (1) holds and kg∗ kF,1 ≤ C1 MF and kg ∗ kQ,q ≤ C1q Mq for 0 ≤ q ≤ 1,
Mq > 0, and MF > 0, possibly depending on (n, p). In addition, suppose that (27), (28),
and (29) hold, Assumptions 1, 5, and 6(i) are satisfied with 0 < η0 < (ξ0∗ − 1)/(ξ0∗ + 1) and
S = {1 ≤ j ≤ p : kgj∗ kQ > C0∗ λnj } for some constant C0∗ > 0, and (20) holds with φn > 0
satisfying
o2−q
n
p
= o(1).
φn C12 (MF + Mq ) γn∗ (q) + log(p/ǫ)/n
(30)
Then for sufficiently large n, depending on (MF , Mq ) only through the convergence rate in (30),
and any A0 > A(ξ0∗ , η0 ), we have with probability at least 1 − ǫ − π,
o2−q
n
p
Dn∗ (ĝ, g∗ ) ≤ O(1)C12 (MF + Mq ) γn (q) + log(p/ǫ)/n
,
16
(31)
where O(1) depends only on (q, A∗0 , C0∗ , ξ0∗ , κ∗0 , η0 ).
In the case of q 6= 0, Corollary 6 can be improved by relaxing the rate condition (30) but
requiring the following compatibility condition, which is seemingly stronger than Assumption 5,
and also more aligned with those used in related works on additive regression (Meier et al. 2009;
Koltchinskii & Yuan 2010).
Assumption 7 (Monotone compatibility condition). For some subset S ⊂ {1, 2, . . . , p} and
constants κ∗0 > 0 and ξ0∗ > 1, assume that for any functions {fj ∈ Gj : j = 1, . . . , p} and
P
f = pj=1 fj ∈ G, if (18) holds then
κ∗2
0
X
j∈S
kfj k2Q ≤ kf k2Q .
(32)
Remark 7. By the Cauchy–Schwartz inequality, (32) implies (19), and hence Asssumption 7
is stronger than Assumption 5. However, there is a monotonicity in S for the validity of
Assumption 7 with (32) used. In fact, for any subset S ′ ⊂ S and any functions {fj′ ∈ Gj : j =
P
1, . . . , p} and f ′ = pj=1 fj′ ∈ G, if
p
X
j=1
λnj wnj kfj′ kF,j +
X
j∈S ′c
λnj kfj′ kQ ≤ ξ0∗
X
j∈S ′
λnj kfj′ kQ ,
then (18) holds with fj = fj′ , j = 1, . . . , p, and hence, via (32), implies
kf ′ k2Q ≥ κ∗2
0
X
j∈S
kfj′ k2Q ≥ κ∗2
0
X
j∈S ′
kfj′ k2Q .
Therefore, if Assumption 7 holds for a subset S, then it also holds for any subset S ′ ⊂ S with
the same constants (ξ0∗ , κ∗ ).
Corollary 7. Suppose that the conditions of Corollary 6 are satisfied with 0 < q ≤ 1 (excluding
q = 0), Assumption 7 holds instead of Assumption 5, and the following condition holds instead
of (30),
o2−q
n
p
≤ η3 ,
φn C12 (MF + Mq ) γn∗ (q) + log(p/ǫ)/n
(33)
for some constant η3 > 0. If 0 < wn∗ (q) ≤ 1 for sufficiently large n, then for any A0 >
A(ξ0∗ , η0 ), inequality (31) holds with probability at least 1 − ǫ − π, where O(1) depends only on
(q, A∗0 , C0∗ , ξ0∗ , κ∗0 , η0 , η3 ).
17
To demonstrate the flexibility of our approach and compare with related results, notably
Suzuki & Sugiyama (2013), we provide another result in the context of Corolloary 6 with
(wnj , γnj ) allowed to depend on (MF , Mq ), in contrast with the choices (28)–(29) independent
of (MF , Mq ). For j = 1, . . . , p, let
wnj = wn† (q) = max{wn′ (q), νn1−q (Mq /MF )},
(34)
γnj = γn† (q) = min{γn′ (q), B0∗ n−1/2 νn−(1−q)β0 /2 (Mq /MF )−β0 /2 },
(35)
2
−β0
−1
where wn′ (q) = γn′ (q)1−q (Mq /MF ) and γn′ (q) = B0∗ 2+β0 (1−q) n 2+β0 (1−q) (Mq /MF ) 2+β0 (1−q) are determined along with the relationship γn′ (q) = B0∗ n−1/2 wn′ (q)−β0 /2 by (10). These choices are
picked to balance the two rates: λn wn MF and λn2−q Mq , where wn and λn denote the common
values of wnj and λnj for j = 1, . . . , p.
Corollary 8. Suppose that the conditions of Corollary 6 are satisfied except that (wnj , γnj )
are defined by (34)–(35), and the following condition holds instead of (30),
n
o2−q
p
φn C12 Mq γn† (q) + log(p/ǫ)/n
= o(1).
(36)
Then for sufficiently large n, depending on (MF , Mq ) only through the convergence rate in (36),
and any A0 > A(ξ0∗ , η0 ), we have with probability at least 1 − ǫ − π,
(
Dn∗ (ĝ, g ∗ ) ≤ O(1)C12
2−β0
2+β0 (1−q)
Mq
(2−q)β0
2+β0 (1−q)
MF
n
−(2−q)
2+β0 (1−q)
+ Mq νn2−q
)
,
(37)
where O(1) depends only on (q, B0∗ , A∗0 , C0∗ , ξ0∗ , κ∗0 , η0 ).
Remark 8. In the special case of q = 0 (exact sparsity), the convergence rate (37) reduces to
2−β0
2+β0
M0
2β0
2+β0
MF
−2
n 2+β0 + M0 νn2 . The same rate was obtained in Suzuki & Sugiyama (2013) under
p
X
j=1
kgj∗ k0Q
≤ M0 ,
p
X
j=1
kgj∗ kH ≤ MF ≤ cM0 ,
(38)
with a constant c for additive regression with reproducing kernel Hilbert spaces, where kgj∗ kH
is the Hilbert norm. As one of their main points, this rate was argued to be faster than
−2
(M0 + MF )n 2+β0 + M0 νn2 , that is, the rate (31) with q = 0 under (38). Our analysis sheds
new light on the relationship between the rates (31) and (37): their difference mainly lies in
whether the tuning parameters (wnj , γnj ) are chosen independently of (MF , M0 ) or depending
on (MF , M0 ).
18
3.3
Convergence of empirical norms
We provide two explicit results on the convergence of empirical norms as needed for Assumption 6. These results can also be useful for other applications.
Our first result, Theorem 3, is applicable (but not limited) to Sobolev and bounded variation
spaces in general. For clarity, we postulate another entropy condition, similar to Assumption 4
but with the empirical supremum norms.
Assumption 8 (Entropy condition in supremum norms). For j = 1, . . . , p, let ψnj,∞ (·, δ) be
(j)
an upper envelope of the entropy integral, independent of the realizations {Xi
: i = 1, . . . , n},
as follows:
ψnj,∞(z, δ) ≥
Z
0
z
H ∗1/2 (u/2, Gj∗ (δ), k · kn,∞ ) du,
z > 0, 0 < δ ≤ 1,
where Gj∗ (δ) = {fj ∈ Gj : kfj kF,j + kfj kQ /δ ≤ 1} as in Assumption 4 and
H ∗ (u, Gj∗ (δ), k · kn,∞ ) =
sup
(j)
(j)
(X1 ,...,Xn )
H(u, Gj∗ (δ), k · kn,∞ ).
We also make use of the following two conditions about metric entropies and sup-norms.
Suppose that for j = 1, . . . , p, ψnj (δ) and ψnj,∞ (z, δ) in Assumptions 4 and 8 are in the
polynomial forms
ψnj (δ) = Bnj δ1−βj /2 ,
0 < δ ≤ 1,
ψnj,∞ (z, δ) = Bnj,∞z 1−βj /2 ,
(39)
z > 0, 0 < δ ≤ 1,
(40)
where 0 < βj < 2 is a constant, and Bnj > 0 and Bnj,∞ > 0 are constants, possibly depending
on n. Denote Γn = maxj=1,...,p (Bnj,∞ /Bnj ). In addition, suppose that for j = 1, . . . , p,
kgj k∞ ≤ (C4,j /2) kgj kF,j + kgj kQ
τj
1−τj
kgj kQ
,
gj ∈ Gj ,
(41)
where C4,j ≥ 1 and 0 < τj ≤ (2/βj − 1)−1 are constants. Let γnj = n−1/2 ψnj (wnj )/wnj =
−β /2
n−1/2 Bnj wnj j
−τ
by (10) and γ̃nj = n−1/2 wnj j for j = 1, . . . , p. As a function of wnj , the
quantity γ̃nj in general differs from γnj even up to a multiplicative constant unless τj = βj /2
as in the case where Gj is an L2 -Sobolev space; see (43) below.
Theorem 3. Suppose that Assumptions 4 and 8 hold with ψnj (δ) and ψnj,∞(z, δ) in the forms
(39) and (40), and condition (41) holds. In addition, suppose that for sufficiently large n,
19
1−βj /2
γnj ≤ wnj ≤ 1 and Γn γnj
≤ 1 for j = 1, . . . , p. Then for any 0 < ǫ′ < 1 (for example,
ǫ′ = ǫ), inequality (20) holds with π = ǫ′ 2 and φn > 0 such that
β
τ /2
γ̃nj wnjp+1 j
γnj
1/2
max
φn = O(1) n Γn max
j
j λnj
λnj
p
2 log(p/ǫ′ )
γ̃nj
log(p/ǫ′ )
γ̃nj
+ max
max
+ max
,
j λnj
j
j
λnj
λ2nj
(42)
where βp+1 = minj=1,...,p βj , and O(1) depends only on (C2 , C3 ) from Lemmas 13 and 14 and
C4 = maxj=1,...,p C4,j from condition (41).
To facilitate justification of conditions (39), (40), and (41), consider the following assumption on the marginal densities of the covariates, as commonly imposed when handling random
designs (e.g., Stone 1982).
Assumption 9 (Non-vanishing marginal densities). For j = 1, . . . , p, denote by qj (x(j) ) the
(j)
(j)
average marginal density function of (X1 , . . . , Xn ), that is, the density function associated
P
(j)
with the probability measure n−1 ni=1 QX (j) , where QX (j) is the marginal distribution of Xi .
i
i
For some constant 0 < ̺0 ≤ 1, assume that qj (x(j) ) is bounded from below by ̺0 simultaneously
for j = 1, . . . , p.
Remark 9. Conditions (39), (40), and (41) are satisfied under Assumption 9, when each Gj
m
is a Sobolev space Wrj j for rj ≥ 1 and mj ≥ 1, or a bounded variation space V mj for rj = 1
and mj ≥ 1, on [0, 1]. Let βj = 1/mj . First, (41) is implied by the interpolation inequalities
for Sobolev spaces (Nirenberg 1966) with
τj = (2/βj + 1 − 2/rj )−1
(43)
∗
and C4,j = ̺−1
0 C4 (mj , rj ) as stated in Lemma 21 of the Supplement. Moreover, if fj ∈ Gj (δ)
with 0 < δ ≤ 1, then kfj kF,j ≤ 1 and kfj kQ ≤ δ, and hence kfj kLrj ≤ kfj k∞ ≤ C4,j by
(41). By rescaling the entropy estimates for Sobolev and bounded variation spaces (Lorentz
et al. 1996) as in Lemmas 19 and 20 of the Supplement, Assumptions 4 and 8 are satisfied
such that (39) and (40) hold with Bnj independent of n, and Bnj,∞ = O(1)Bnj if rj > βj or
Bnj,∞ = O(log1/2 (n))Bnj if rj = βj = 1.
Remark 10. Assumption 9 is not needed for justification of (39), (40), and (41), when each
class Gj is W11 or V 1 on [0, 1], that is, rj = mj = 1. In this case, condition (41) directly holds
with τj = 1, because kgj k∞ ≤ TV(gj ) + kgj kQ . Then (39) and (40) easily follow from the
entropy estimates in Lemmas 19 and 20.
20
Our second result provides a sharper rate than in Theorem 3, applicable (but not limited)
to Sobolev and bounded variation spaces, provided that the following conditions hold. For
P∞
gj ∈ Gj , assume that gj (·) can be written as
ℓ=1 θjℓ ujℓ (·) for certain coefficients θjℓ and
basis functions ujℓ (·) on a set Ω. In addition, for certain positive constants C5,1 , C5,2 , C5,3 ,
0 < τj < 1, and 0 < wnj ≤ 1, assume that for all 1 ≤ j ≤ p,
sup
P k
max
k≥1
Pℓj0
2
ℓ=1 ujℓ (x)/k
P
: x ∈ Ω, k ≥ ℓj0 ≤ C5,1 ,
2 1/τj
ℓj,k−1 <ℓ≤ℓjk θjℓ ℓjk
2 −2
ℓ=1 θjℓ wnj
(44)
−1
≤ C5,2 (kgj kF,j + wnj
kgj kQ )2 ,
(45)
−1
kgj kQ )2 ,
≤ C5,2 (kgj kF,j + wnj
(46)
with ℓjk = ⌈(2k /wnj )2τj ⌉ for k ≥ 0 and ℓj,−1 = 0, and for all 1 ≤ j ≤ p and k ≥ 0,
n P
P
2
= 1 ≤ C5,3 .
sup k ℓj,k−1 <ℓ≤ℓjk θjℓ ujℓ k2Q : ℓj,k−1 <ℓ≤ℓjk θjℓ
(47)
Theorem 4. Suppose that (44), (45), (46) and (47) hold as above, and maxj=1,...,p {e2/(1−τj ) +
−τ
2wnj j } ≤ n. Then for any 0 < ǫ′ < 1 (for example, ǫ′ = ǫ), inequality (20) holds with π = ǫ′ 2
and φn > 0 such that
(
γ̃nj
φn = O(1) max
max
j (1 − τj )λnj
j
p
2 log(np/ǫ′ )
γ̃nj
log(np/ǫ′ )
+ max
j
λnj
(1 − τj )2 λ2nj
)
,
−τ
where γ̃nj = n−1/2 wnj j and O(1) depends only on {C5,1 , C5,2 , C5,3 }.
m
Remark 11. Let Gj be a Sobolev space Wrj j with rj ≥ 1, mj ≥ 1, and (rj ∧ 2)mj > 1 or a
bounded variation space V mj with rj = 1 and mj > 1 (excluding mj = 1) on [0, 1]. Condition
(44) holds for commonly used Fourier, wavelet and spline bases in L2 . For any L2 orthonormal
bases {ujℓ , ℓ ≥ 1}, condition (46) follows from Assumption 9 when C5,2 ≥ ̺−1
0 , and condition
(47) is also satisfied under an additional assumption that the average marginal density of
(j)
{Xi
: i = 1, . . . , n} is bounded from above by C5,3 for all j. In the proof of Proposition 5
we verify (44) and (45) for suitable wavelet bases with τj = 1/{2mj + 1 − 2/(rj ∧ 2)}, which
m
j
satisfies τj < 1 because (rj ∧ 2)mj > 1. In fact, Gj is allowed to be a Besov space Brj ,∞
, which
m
contains Wrj j for rj ≥ 1 and V mj for rj = 1 (e.g., DeVore & Lorentz 1993).
Remark 12. The convergence rate of φn in Theorem 4 is no slower than (42) in Theorem 3
if 1 ≤ rj ≤ 2 and (1 − τj )−1 {log(n)/n}1/2 = O(γ̃nj ), the latter of which is valid whenever τj
−τ /2
is bounded away from 1 and γ̃nj = n−1/2 wnj j
21
is of a slower polynomial order than n−1/2 .
However, Theorem 4 requires an additional side condition (47) along with the requirement of
τj < 1, which excludes for example the bounded variation space V 1 on [0, 1]. See Equations
(52) and (53) for implications of these rates when used in Assumption 6.
3.4
Results with Sobolev and bounded variation spaces
We combine the results in Section 3.2 and 3.3 (with ǫ′ = ǫ) to deduce a number of concrete
results on the performance of ĝ. For simplicity, consider a fully homogeneous situation where
each class Gj is a Sobolev space Wrm0 0 for some constants r0 ≥ 1 and m0 ≥ 1 or a bounded
variation space V m0 for r0 = 1 and m0 ≥ 1 on [0, 1]. Let β0 = 1/m0 . By Remark 9, if
r0 > β0 , then Assumptions 4 and 8 are satisfied such that ψnj (δ) = B0∗ δ1−β0 /2 and ψnj,∞ (z, δ) =
O(1)B0∗ z 1−β0 /2 for z > 0 and 0 < δ ≤ 1 under Assumption 9 (non-vanishing marginal densities),
where B0∗ > 0 is a constant depending on ̺0 among others. On the other hand, by Remark 10,
if r0 = β0 = 1, then Assumptions 4 and 8 are satisfied such that ψnj (δ) = B0∗ δ1/2 and
ψnj,∞ (z, δ) = O(log1/2 (n))B0∗ z 1/2 for z > 0 and 0 < δ ≤ 1, even when Assumption 9 does not
hold. That is, Γn in Theorem 3 reduces to
Γn = O(1) if r0 > β0 or O(log1/2 (n)) if r0 = β0 = 1.
We present our results in three cases, where the underlying function g ∗ =
(48)
Pp
∗
j=1 gj
is assumed
to satisfy (2) with q = 1, q = 0, or 0 < q < 1. As discussed in Section 1, the parameter set (2)
decouples sparsity and smoothness, inducing sparsity at different levels through an Lq ball in
k · kQ norm for 0 ≤ q ≤ 1, while only enforcing smoothness through an L1 ball in k · kF norm
on the components (g1∗ , . . . , gp∗ ).
The first result deals with the case q = 1 for the parameter set (2).
Proposition 3. Assume that (1) holds and kg ∗ kF,1 ≤ C1 MF and kg ∗ kQ,1 ≤ C1 M1 for MF > 0
and M1 > 0, possibly depending on (n, p). Let wnj = 1 and γnj = γn∗ (1) ≍ n−1/2 by (28)–(29).
Suppose that Assumptions 1 and 9 hold, and log(p/ǫ) = o(n). Then for sufficiently large n,
independently of (MF , M1 ), and any A0 > (1 + η0 )/(1 − η0 ), we have with probability at least
1 − 2ǫ,
p
kĝ − g ∗ k2n + A1 Rn∗ (ĝ − g ∗ ) ≤ O(1)C12 (MF + M1 ) log(p/ǫ)/n,
22
where O(1) depends only on (B0∗ , A0 , η0 , ̺0 ). Moreover, we have
1
kĝ − g∗ k2Q + A1 Rn∗ (ĝ − g ∗ )
2
n
o
p
≤ O(1)C12 (MF2 + M12 ) n−1/2 Γn + log(p/ǫ)/n ,
with probability at least 1 − 2ǫ, where Γn is from (48) and O(1) depends only on (B0∗ , A0 , η0 , ̺0 )
and (C2 , C3 , C4 ) as in Theorem 3. If r0 = β0 = 1, then the results are valid even when
Assumption 9 and hence ̺0 are removed.
Remark 13 (Comparison with existing results). Proposition 3 leads to the slow rate {log(p)/n}1/2
under L1 -ball sparsity in k·kQ norm, as previously obtained for additive regression with Sobolev
Hilbert spaces in Ravikumar et al. (2009), except in the case where r0 = β0 = 1, that is, each
class Gj is W11 or V 1 . In the latter case, Proposition 3 shows that the convergence rate is
{log(np)/n}1/2 for out-of-sample prediction, but remains {log(p)/n}1/2 for in-sample predic-
tion. Previously, only the slower rate, {log(np)/n}1/2 , was obtained for in-sample prediction
in additive regression with the bounded variation space V 1 by Petersen et al. (2016).
The second result deals with the case q = 0 for the parameter set (2).
Proposition 4. Assume that (1) holds and kg∗ kF,1 ≤ C1 MF and kg∗ kQ,0 ≤ M0 for MF > 0
and M0 > 0, possibly depending on (n, p). By (28)–(29), let
(
1/2 )
−1
2
log(p/ǫ)
,
wnj = wn∗ (0) = max B0∗ 2+β0 n 2+β0 ,
n
− β 0
4
−1
2
log(p/ǫ)
γnj = γn∗ (0) = min B0∗ 2+β0 n 2+β0 , B0∗ n−1/2
.
n
Suppose that Assumptions 1, 5, and 9 hold with 0 < η0 < (ξ0∗ − 1)/(ξ0∗ + 1) and S = {1 ≤ j ≤
p : kgj∗ kQ > C0∗ λnj } for some constant C0∗ > 0, and
n
Γn wn∗ (0)−(1−β0 /2)τ0 γn∗ (0) + wn∗ (0)−τ0
p
o
log(p/ǫ)/n (1 + MF + M0 ) = o(1),
(49)
where τ0 = 1/(2/β0 + 1 − 2/r0 ). Then for sufficiently large n, depending on (MF , M0 ) only
through the convergence rate in (49), and any A0 > A(ξ0∗ , η0 ), we have
n −1
o2
p
Dn∗ (ĝ, g∗ ) ≤ O(1)C12 (MF + M0 ) n 2+β0 + log(p/ǫ)/n ,
(50)
with probability at least 1 − 2ǫ, where O(1) depends only on (B0∗ , A∗0 , C0∗ , ξ0∗ , κ∗0 , η0 , ̺0 ). If
r0 = β0 = 1, then the results are valid even when Assumption 9 and hence ̺0 are removed.
23
Condition (49) is based on Theorem 3 for convergence of empirical norms. By Remark 12,
a weaker condition can be obtained using Theorem 4 when 1 ≤ r0 ≤ 2 and τ0 < 1 (that is,
r0 > β0 ). It is interesting to note that (49) reduces to (51) below in the case r0 = β0 = 1.
Proposition 5. Proposition 4 is also valid with (49) replaced by the weaker condition
o
n
p
wn∗ (0)−τ0 log(np/ǫ)/n (1 + MF + M0 ) = o(1),
(j)
(51)
(j)
in the case where 1 ≤ r0 ≤ 2, r0 > β0 , and the average marginal density of (X1 , . . . , Xn ) is
bounded from above for all j.
Remark 14 (Comparison with existing results). Propositions 4 and 5 yield the fast rate
−2
n 2+β0 + log(p)/n under L0 -ball sparsity in k · kQ norm. Previously, the same rate was obtained
for high-dimensional additive regression only with reproducing kernel Hilbert spaces (including
the Sobolev space W2m ) by Koltchinskii & Yuan (2010) and Raskutti et al. (2012), but under
more restrictive conditions. They studied hybrid penalized estimation procedures, which involve additional constraints such that the Hilbert norms of (g1 , . . . , gp ) are bounded by known
constants when minimizing a penalized criterion. Moreover, Koltchinskii & Yuan (2010) assumed a constant bound on the sup-norm of possible g∗ , whereas Raskutti et al. (2012) assumed
(1)
(p)
the independence of the covariates (Xi , . . . , Xi ) for each i. These restrictions were relaxed
in subsequent work by Suzuki & Sugiyama (2013), but only explicitly under the assumption
that the noises εi are uniformly bounded by a constant. Moreover, our condition (49) is much
weaker than related ones in Suzuki & Sugiyama (2013), as discussed in Remarks 15 and 16
below. See also Remark 8 for a discussion about the relationship between our results and the
seemingly faster rate in Suzuki & Sugiyama (2013).
Remark 15. To justify Assumptions 6(i)–(ii) on convergence of empirical norms, our rate
condition (49) is much weaker than previous ones used. If each class Gj is a Sobolev Hilbert
space (r0 = 2), then τ0 = β0 /2 and (49) becomes
o
n
p
2
n1/2 wn∗ (0)β0 /4 γn∗ (0)2 + γn∗ (0) log(p/ǫ) (1 + MF + M0 ) = o(1).
(52)
Moreover, by Proposition 5, condition (49) can be weakened to (51), that is,
p
γn∗ (0) log(np/ǫ)(1 + MF + M0 ) = o(1),
(53)
(j)
(j)
under an additional condition that the average marginal density of (X1 , . . . , Xn ) is bounded
from above for all j. Either condition (52) or (53) is much weaker than those in related analysis
24
with reproducing kernel Hilbert spaces. In fact, techniques based on the contraction inequality
(Ledoux & Talagrand 1991) as used in Meier et al. (2009) and Koltchinskii & Yuan (2010),
lead to a rate condition such as
n1/2 {γn2 (0) + νn2 }(1 + MF + M0 ) = o(1),
2
(54)
−1
where γn (0) = B0∗ 2+β0 n 2+β0 and νn = {log(p/ǫ)/n}1/2 . This amounts to condition (6) assumed
in Suzuki & Sugiyama (2013), in addition to the requirement n−1/2 (log p) ≤ 1. But condition
(54) is even stronger than the following condition:
o
n
2
n1/2 γn (0)2+β0 /4 + γn (0)νn (1 + MF + M0 ) = o(1),
(55)
2
because Γn γn (0)2+β0 /4 + γn (0)νn ≪ γn2 (0) + νn2 if either γn (0) ≫ νn or γn (0) ≪ νn . Condition
(55) implies (52) and (53), as we explain in the next remark.
Remark 16. Our rate condition (49) is in general weaker than the corresponding condition
with (wn∗ (0), γn∗ (0)) replaced by (wn (0), γn (0)), that is,
o
n
Γn γn (0)1−(1−β0 /2)τ0 + γn (0)−τ0 νn (1 + MF + M0 ) = o(1),
(56)
This demonstrates the advantage of using the more careful choices (wn∗ (0), γn∗ (0)) and also
explains why (55) implies (52) in Remark 15. In fact, if γn (0) ≥ νn then (49) and (56) are
identical to each other. On the other hand, if γn (0) < νn , then wn∗ (0) = νn > γn (0) and
wn∗ (0)−(1−β0 /2)τ0 γn∗ (0) = B0∗ n−1/2 wn∗ (0)−(1−β0 /2)τ0 −β0 /2 < γn (0)1−(1−β0 /2)τ0 . This also shows
that if γn (0) ≪ νn , then (49) is much weaker than (56). For illustration, if r0 = 2 and
hence τ0 = β0 /2, then (56) or equivalently (55) requires at least γn (0)−β0 /2 νn = o(1), that is,
−2
(log p)n 2+β0 = o(1), and (54) requires at least n1/2 νn2 = o(1), that is, log(p)n−1/2 = o(1). In
contrast, the corresponding requirement for (49), wn∗ (0)−β0 /2 νn = o(1), is automatically valid
as long as νn = o(1), that is, log(p)n−1 = o(1).
The following result deals with the case 0 < q < 1 for the parameter set (2).
Proposition 6. Assume that (1) holds and kg∗ kF,1 ≤ C1 MF and kg∗ kQ,q ≤ C1q Mq for 0 <
q < 1, Mq > 0, and MF > 0, possibly depending on (n, p). Let wnj = wn∗ (q) and γnj = γn∗ (q)
by (28)–(29). Suppose that Assumptions 1, 7, and 9 hold with 0 < η0 < (ξ0∗ − 1)/(ξ0∗ + 1) and
S = {1 ≤ j ≤ p : kgj∗ kQ > C0∗ λnj } for some constant C0∗ > 0, log(p/ǫ) = o(n), and
n
o
Γn wn∗ (q)−(1−β0 /2)τ0 γn∗ (q)1−q + wn∗ (q)−τ0 νn1−q (1 + MF + Mq ) ≤ η4 ,
25
(57)
for some constant η4 > 0, where νn = {log(p/ǫ)/n}1/2 . Then for sufficiently large n, independently of (MF , Mq ), and any A0 > A(ξ0∗ , η0 ), we have
n
o2−q
p
−1
Dn∗ (ĝ, g ∗ ) ≤ O(1)C12 (MF + Mq ) n 2+β0 (1−q) + log(p/ǫ)/n
,
with probability at least 1 − 2ǫ, where O(1) depends only on (q, B0∗ , A∗0 , C0∗ , ξ0∗ , κ∗0 , η0 , ̺0 , η4 ) and
(C2 , C3 , C4 ) as in Theorem 3. If r0 = β0 = 1, then the results are valid even when Assumption 9
and hence ̺0 are removed.
Similarly as in Propositions 4 and 5, condition (57) can be weakened as follows when
1 ≤ r0 ≤ 2 and τ0 < 1 (that is, r0 > β0 ). It should also be noted that (57) is equivalent to
(58) below (with different η4 in the two equations) in the case r0 = β0 = 1, because γn∗ (q) with
q < 1 is of a slower polynomial order than n−1/2 and hence {log(n)/n}1/2 γn∗ (q)−1 = o(1).
Proposition 7. Proposition 6 is also valid with (49) replaced by the weaker condition
n
o
wn∗ (q)−τ0 (log(np/ǫ)/n)(1−q)/2 (1 + MF + M0 ) ≤ η4 ,
(58)
for some constant η4 > 0, in the case where 1 ≤ r0 ≤ 2, r0 > β0 , and the average marginal
(j)
(j)
density of (X1 , . . . , Xn ) is bounded from above for all j.
Remark 17. Propositions 6 and 7 yield, under Lq -ball sparsity in k · kQ norm, a convergence
rate interpolating the slow and fast rates smoothly from q = 1 to q = 0, similarly as in fixed
designs (Section 3.1). However, the rate condition (57) involved does always exhibit a smooth
transition to those for the slow and fast rates. In the extreme case q = 1, condition (57) with
q = 1 cannot be satisfied when M1 is unbounded or when M1 is bounded but Γn is unbounded
with r0 = β0 = 1. In contrast, Proposition 3 allows for unbounded M1 and the case r0 = β0 = 1.
This difference is caused by the need to justify Assumption 6(ii) with q 6= 1. In the extreme
case q = 0, condition (57) with q = 0 also differ drastically from (49) in Proposition 4. As seen
from the proof of Corollary 7, this difference arises because Assumption 6(ii) can be justified
by exploiting the fact that z q → ∞ as z → ∞ for q > 0 (but not q = 0).
For illustration, Table 1 gives the convergence rates from Propositions 3–6 in the simple
situation where (MF , Mq ) are bounded from above, independently of (n, p). The rate conditions
(49) and (57) are easily seen to hold in all cases except that (49) is not satisfied for q = 0
when r0 = β0 = 1 but νn 6= o(γn (0)). In this case, we show in the following result that the
convergence rate {γn (0) + νn }2 can still be achieved, but with the tuning parameters (wnj , γnj )
26
Table 1: Convergence rates for out-of-sample prediction under parameter set (2) with (MF , Mq )
bounded from above
r0 > β0
0≤q≤1
scale
adaptive
rate
yes
{γn (q) + νn }2−q
−1
r0 = β0 = 1
q=1
0<q<1
yes
p
log(n)/n + νn
yes
q=0
νn = o(γn (0))
otherwise
yes
no
{γn (q) + νn }2−q
Note: γn (q) ≍ n 2+β0 (1−q) and νn = {log(p/ǫ)/n}1/2 . Scale-adaptiveness means the convergence rate is
achieved with (wnj , γnj ) chosen independently of (MF , Mq ).
chosen suitably depending on the upper bound of (MF , Mq ). This is in contrast with the other
cases in Table 1 where the convergence rates are achieved by our penalized estimators in a
scale-adaptive manner: (wnj , γnj ) = (wn∗ (q), γn∗ (q)) are chosen independently of (MF , Mq ) or
their upper bound.
Proposition 8. Assume that r0 = β0 = 1, and MF and M0 are bounded from above by a
constant M > 0. Suppose that the conditions of Proposition 4 are satisfied except with (49)
and Assumption 9 removed, and Assumption 7 holds instead of Assumption 5. Let ĝ ′ be the
−β0 /2 ∗
γn (0)
′ = K
′
estimator with (wnj , γnj ) replaced by wnj
= K0 wn∗ (0) and γnj
0
for K0 > 0.
Then K0 can be chosen, depending on M but independently of (n, p), such that for sufficiently
large n, depending on M , and any A0 > A(ξ0∗ , η0 ), we have
n −1
o2
p
Dn∗ (ĝ ′ , g ∗ ) ≤ O(1)C12 (MF + M0 ) n 2+β0 + + log(p/ǫ)/n ,
with probability at least 1 − 2ǫ, where O(1) depends only on (M , B0∗ , A∗0 , C0∗ , ξ0∗ , κ∗0 , η0 ) and
(C2 , C3 , C4 ) as in Theorem 3.
4
Discussion
For additive regression with high-dimensional data, we have established new convergence results on the predictive performance of penalized estimation when each component function can
be a Sobolev space or a bounded variation space. A number of open problems remain to be
fully investigated. First, our results provide minimax upper bounds for estimation when the
27
component functions are restricted within an L1 ball in k · kF,j semi-norm and an Lq ball in
k · kQ norm. It is desirable to study whether these rates would match minimax lower bounds.
Second, while the penalized estimators have been shown under certain conditions to be adaptive to the sizes of L1 (k · kF ) and Lq (k · kQ ) balls for fixed q, we are currently investigating
adaptive estimation over such balls with varying q simultaneously. Finally, it is interesting to
study variable selection and inference about component functions for high-dimensional additive
regression, in addition to predictive performance studied here.
References
Bellec, P.C. and Tsybakov, A.B. (2016) Bounds on the prediction error of penalized least
squares estimators with convex penalty. Festschrift in Honor of Valentin Konakov, to
appear.
Bickel, P., Ritov, Y., and Tsybakov, A.B. (2009) Simultaneous analysis of Lasso and Dantzig
selector, Annals of Statistics, 37, 1705–1732.
Bunea, F., Tsybakov, A.B., and Wegkamp, M. (2007) Sparsity oracle inequalities for the
Lasso, Electronic Journal of Statistics, 1, 169–194.
DeVore, R. A. and Lorentz, G.G. (1993) Constructive Approximation, Springer: New York,
NY.
Greenshtein, E. and Ritov, Y. (2004) Persistency in high-dimensional linear predictor selection
and the virtue of over-parametrization, Bernoulli 10, 971–988.
Gu, C. (2002) Smoothing Spline ANOVA Models, Springer: New York, NY.
Hastie, T. and Tibshirani, R. (1990) Generalized Additive Models, Chapman & Hall: New
York, NY.
Huang, J., Horowitz, J.L., and Wei, P. (2010) Variable selection in nonparametric additive
models, Annals of Statistics, 38, 2282–2313.
Kim, S.-J., Koh, K., Boyd, S., and Gorinevsky, D. (2009) ℓ1 trend filtering, SIAM Review 51,
339–360.
Koltchinskii, V. and Yuan, M. (2010) Sparsity in multiple kernel learning, Annals of Statistics,
38, 3660–3695.
28
Ledoux, M. and Talagrand, M. (1991) Probability in Banach Spaces: Isoperimetry and Processes, Springer: Berlin.
Lin, Y. and Zhang, H.H. (2006) Component selection and smoothing in multivariate nonparametric regression, Annals of Statistics, 34, 2272–2297
Lorentz, G.G., Golitschek, M.v. and Makovoz, Y. (1996) Constructive Approximation: Advanced Problems, Springer: New York, NY.
Mammen, E. and van de Geer, S. (1997) Locally adaptive regression splines, Annals of Statistics, 25, 387–413.
Meier, L., van de Geer, S., and Buhlmann, P. (2009) High-dimensional additive modeling,
Annals of Statistics, 37, 3779–3821.
Negahban, S.N., Ravikumar, P., Wainwright, M.J., and Yu, B. (2012) A unified framework
for high-dimensional analysis of M-estimators with decomposable regularizers, Statistical
Science, 27, 538–557.
Nirenberg, L. (1966) An extended interpolation inequality, Annali della Scuola Normale Superiore di Pisa, Classe di Scienze, 20, 733–737.
Petersen, A., Witten, D., and Simon, N. (2016) Fused Lasso additive model, Journal of
Computational and Graphical Statistics, 25, 1005–1025.
Raskutti, G., Wainwright, M.J., and Yu, B. (2012) Minimax-optimal rates for sparse additive models over kernel classes via convex programming, Journal of Machine Learning
Research 13, 389–427.
Ravikumar, P., Liu, H., Lafferty, J., and Wasserman, L. (2009) SPAM: Sparse additive models,
Journal of the Royal Statistical Society, Series B, 71, 1009–1030.
Stone, C.J. (1982) Optimal global rates of convergence for nonparametric regression. Annals
of Statistics, 10, 1040–1053.
Stone, C.J. (1985) Additive regression and other nonparametric models. Annals of Statistics,
13, 689–705.
29
Suzuki, T. and Sugiyama, M. (2013) Fast learning rate of multiple kernel learning: Trade-off
between sparsity and smoothness, Annals of Statistics, 41, 1381–1405.
Tibshirani, R.J. (2014) Adaptive piecewise polynomial estimation via trend filtering, Annals
of Statistics, 42, 285–323.
van de Geer, S. (2000) Empirical Processes in M-Estimation, Cambridge University Press.
van der Vaart, A.W. and Wellner, J. (1996) Weak Convergence and Empirical Processes,
Springer: New York, NY.
Yuan, M. and Zhou, D.-X. (2016) Minimax optimal rates of estimation in high dimensional
additive models, Annals of Statistics, 44, 2564–2593.
Yang, Y. and Tokdar, S.T. (2015) Minimax-optimal nonparametric regression in high dimensions, Annals of Statistics, 43, 652–674.
30
Supplementary Material for “Penalized Estimation in
Additive Regression with High-Dimensional Data”
Zhiqiang Tan & Cun-Hui Zhang
S1
Proofs
S1.1
Proof of Proposition 2
(1)
(1)
Without loss of generality, assume that k = 1 and 0 ≤ X1 < . . . < Xn ≤ 1.
P
Consider the case m = 1. For any g = pj=1 gj with g1 ∈ V 1 , define g̃1 as a piecewise
(1)
(1)
(1)
(1)
(1)
constant function: g̃1 (z) = g1 (X1 ) for 0 ≤ z < X2 , g̃1 (z) = g1 (Xi ) for Xi ≤ z < Xi+1 ,
P
(1)
(1)
i = 2, . . . , n − 1, and g̃1 (z) = g1 (Xn ) for Xn ≤ z ≤ 1. Let g̃ = g̃1 + pj=2 gj . Then
g̃(Xi ) = g(Xi ) for i = 1, . . . , n, but TV(g̃1 ) ≤ TV(g1 ) and hence Rn (g̃) ≤ Rn (g), which implies
the desired result for m = 1.
Consider the case m = 2. For any g =
(1)
Pp
j=1 gj
(1)
with g1 ∈ V 2 , define g̃1 such that g̃1 (Xi ) =
(1)
(1)
(1)
g1 (Xi ), i = 1, . . . , n, and g̃1 (z) is linear in the intervals [0, X2 ], [Xi , Xi+1 ], i = 2, . . . , n −
Pn−1
(1)
(1)
2, and [Xn−1 , 1]. Then TV(g̃1 ) =
i=1 |bi+1 − bi |, where bi is the slope of g̃1 between
(1)
(1)
(1)
(1)
[Xi , Xi+1 ]. On the other hand, by the mean-value theorem, there exists zi ∈ [Xi , Xi+1 ] such
P
Pn−1
(1)
(1)
that g1 (zi ) = bi for i = 1, . . . , n − 1. Then TV(g1 ) ≥ i=1
|bi+1 − bi |. Let g̃ = g̃1 + pj=2 gj .
Then g̃(Xi ) = g(Xi ) for i = 1, . . . , n, but Rn (g̃) ≤ Rn (g), which implies the desired result for
m = 2.
S1.2
Proofs of Theorem 1 and corollaries
We split the proof of Theorem 1 and Corollary 1 into five lemmas. The first one provides a
probability inequality controlling the magnitude of hε, hj in , in terms of the semi-norm khj kF,j
and the norm khj kn for all hj ∈ Gj with a single j.
Lemma 1. For fixed j ∈ {1, . . . , p}, let
Anj = ∪hj ∈Gj {|hε, hj in |/C1 > γnj,twnj khj kF,j + γnj,tkhj kn } ,
where γnj,t = γnj +
p
t/n for t > 0, γnj = n−1/2 ψnj (wnj )/wnj , and wnj ∈ (0, 1]. Under
Assumptions 1 and 2, we have
P (Anj ) ≤ exp(−t).
1
Proof. In the event Anj , we renormalize hj by letting fj = hj /(khj kF,j + khj kn /wnj ). Then
kfj kF,j + kfj kn /wnj = 1 and hence fj ∈ Gj (wnj ). By Lemma 12 with F1 = Gj (wnj ) and
δ = wnj , we have for t > 0,
(
P (Anj ) ≤ P
=P
(
sup
fj ∈Gj (wnj )
sup
fj ∈Gj (wnj )
|hε, fj in |/C1 > γnj,t wnj
)
|hε, fj in |/C1 > n−1/2 ψnj (wnj ) + wnj
p
)
t/n
≤ exp(−t).
By Lemma 1 and the union bound, we obtain a probability inequality controlling the magnitude of hε, hj in for hj ∈ Gj simultaneously over j = 1, . . . , p.
Lemma 2. For each j ∈ {1, . . . , p}, let
Anj = ∪hj ∈Gj {|hε, hj in | > λnj wnj khj kF,j + λnj khj kn } ,
where λnj /C1 = γnj + {log(p/ǫ)/n}1/2 . Under Assumptions 1 and 2, we have
P (∪pj=1 Anj ) ≤ ǫ.
Proof. By Lemma 1 with t = log(p/ǫ), we have for j = 1, . . . , p,
P (Anj ) ≤ exp(−t) =
ǫ
.
p
Applying the union bound yields the desired inequality.
If g ∗ ∈ G, then Kn (ĝ) ≤ Kn (g∗ ) directly gives the basic inequality:
1
kĝ − g ∗ k2n + A0 Rn (ĝ) ≤ hε, ĝ − g ∗ in + A0 Rn (g∗ ).
2
(S1)
By exploiting the convexity of the regularizer Rn (·), we provide a refinement of the basic
inequality (S1), which relates the estimation error of ĝ to that of any additive function ḡ ∈ G
and the corresponding regularization Rn (ḡ).
Lemma 3. The fact that ĝ is a minimizer of Kn (g) implies that for any function ḡ(x) =
Pp
(j)
j=1 ḡj (x ) ∈ G,
1
1
kĝ − g ∗ k2n + kĝ − ḡk2n + A0 Rn (ĝ)
2
2
1
≤ kḡ − g ∗ k2n + hε, ĝ − ḡin + A0 Rn (ḡ).
2
2
(S2)
Proof. For any t ∈ (0, 1], the fact that Kn (ĝ) ≤ Kn ((1 − t)ĝ + tḡ) implies
t2
kĝ − ḡk2n + Rn (ĝ) ≤ hY − ((1 − t)ĝ + tḡ), t(ĝ − ḡ)in + Rn ((1 − t)ĝ + tḡ)
2
≤ hy − ((1 − t)ĝ + tḡ), t(ĝ − ḡ)in + (1 − t)Rn (ĝ) + tRn (ḡ),
by similar calculation leading to the basic inequality (S1) and by the convexity of Rn (·):
Rn ((1 − t)ĝ + tḡ) ≤ (1 − t)Rn (ĝ) + tRn (ḡ). Using Y = g∗ + ε, simple manipulation of the
preceding inequality shows that for any t ∈ (0, 1],
t
hĝ − g∗ , ĝ − ḡin − kĝ − ḡk2n + Rn (ĝ) ≤ hε, ĝ − ḡin + Rn (ḡ),
2
which reduces to
1
1−t
1
kĝ − g ∗ k2n +
kĝ − ḡk2n + Rn (ĝ) ≤ kḡ − g∗ k2n + hε, ĝ − ḡin + Rn (ḡ)
2
2
2
by the fact that 2hĝ − g∗ , ĝ − ḡin = kĝ − g ∗ k2n + kĝ − ḡk2n − kḡ − g∗ k2n . Letting t ց 0 yields the
desired inequality (S2).
From Lemma 3, we obtain an upper bound of the estimation error of ĝ when the magnitudes
of hε, ĝj − ḡj in , j = 1, . . . , p, are controlled by Lemma 2.
Lemma 4. Let An = ∪pj=1 Anj with hj = ĝj − ḡj in Lemma 2. In the event Acn , we have for
any subset S ⊂ {1, 2, . . . , p},
1
1
kĝ − g ∗ k2n + kĝ − ḡk2n + (A0 − 1)Rn (ĝ − ḡ)
2
2
X
λnj kĝj − ḡj kn ,
≤ ∆n (ḡ, S) + 2A0
j∈S
where
p
X
X
1
∆n (ḡ, S) = kḡ − g∗ k2n + 2A0
ρnj kḡj kF,j +
λnj kḡj kn .
2
c
j=1
j∈S
Proof. By the refined basic inequality (S2), we have in the event Acn ,
1
1
kĝ − g ∗ k2n + kĝ − ḡk2n + A0 Rn (ĝ)
2
2
1
∗ 2
≤ kḡ − g kn + Rn (ĝ − ḡ) + A0 Rn (ḡ).
2
Applying to the preceding inequality the triangle inequalities,
kĝj kF,j ≥ kĝj − ḡj kF,j − kḡj kF,j ,
j = 1, . . . , p,
kĝj kn ≥ kĝj − ḡj kn − kḡj kn ,
j ∈ S c,
kĝj kn ≥ kḡj kn − kĝj − ḡj kn ,
j ∈ S,
3
(S3)
and rearranging the result leads directly to (S3).
Taking S = ∅ in (S3) yields (13) in Corollary 1. In general, we derive implications of (S3)
by invoking the compatibility condition (Assumption 3).
Lemma 5. Suppose that Assumption 3 holds. If A0 > (ξ0 + 1)/(ξ0 − 1), then (S3) implies
(12) in Theorem 1.
Proof. For the subset S used in Assumption 3, write
1
1
Zn = kĝ − g ∗ k2n + kĝ − ḡk2n ,
2
2
p
X
X
ρnj kĝj − ḡj kF,j +
λnj kĝj − ḡj kn ,
Tn1 =
Tn2 =
j∈S c
j=1
X
j∈S
λnj kĝj − ḡj kn .
Inequality (S3) can be expressed as
Zn + (A0 − 1)(Tn1 + Tn2 ) ≤ ∆n (ḡ, S) + 2A0 Tn2 ,
which leads to two possible cases: either
ξ1 {Zn + (A0 − 1)(Tn1 + Tn2 )} ≤ ∆n (ḡ, S),
(S4)
or (1 − ξ1 ){Zn + (A0 − 1)(Tn1 + Tn2 )} ≤ 2A0 Tn2 , that is,
Zn + (A0 − 1)(Tn1 + Tn2 ) ≤
2A0
Tn2 = (ξ0 + 1)(A0 − 1)Tn2 = ξ2 Tn2 ,
1 − ξ1
(S5)
where ξ1 = 1 − 2A0 /{(ξ0 + 1)(A0 − 1)} ∈ (0, 1] because A0 > (ξ0 + 1)/(ξ0 − 1). If (S5) holds,
then Tn1 ≤ ξ0 Tn2 , which, by Assumption 3 with fj = ĝj − ḡj , implies
Tn2
1/2
X
λ2nj kĝ − ḡkn .
≤ κ−1
0
(S6)
j∈S
Combining (S5) and (S6) and using kĝ − ḡk2n /2 ≤ Zn yields
X
λ2nj .
Zn + (A0 − 1)(Tn1 + Tn2 ) ≤ 2ξ22 κ−2
0
(S7)
j∈S
Therefore, inequality (S3), through (S4) and (S7), implies
Zn + (A0 − 1)(Tn1 + Tn2 ) ≤ ξ1−1 ∆n (ḡ, S) + 2ξ22 κ−2
0
4
X
j∈S
λ2nj .
Finally, combining Lemmas 2, 4 and 5 completes the proof of Theorem 1.
P
λ2nj and j∈S c λnj kḡj kn
Pp
2−q
q
> C0 λnj }. First, because
j=1 λnj kḡj kn ≥
Proof of Corollary 2. The result follows from upper bounds of
by the definition S = {1 ≤ j ≤ p : kḡj kn
P
2−q
+ q q
j∈S λnj (C0 ) λnj , we have
X
λ2nj
j∈S
≤
(C0+ )−q
p
X
j=1
P
j∈S
2−q
λnj
kḡj kqn ,
(S8)
where for z ≥ 0, (z + )q = z q if q > 0 or = 1 if q = 0. Second, because
Pp
1−q kḡ kq , we have
j n
j=1 λnj (C0 λnj )
X
j∈S c
λnj kḡj kn ≤ C01−q
p
X
j=1
P
j∈S c
2−q
λnj
kḡj kqn .
λnj kḡj kn ≤
(S9)
Inserting (S8) and (S9) into (12) yields the desired inequality.
Proof of Corollary 3. The result follows directly from Corollary 2, because λ2−q
=
nj
C12−q {γn (q) + νn }2−q and ρnj = C1 {γn (q) + νn }γn1−q (q) ≤ C1 {γn (q) + νn }2−q , where νn =
{log(p/ǫ)/n}1/2 .
S1.3
Proofs of Theorem 2 and corollaries
Write hj = ĝj − ḡj and h = ĝ − ḡ and, for the subset S used in Assumption 5,
1
1
Zn = kĝ − g ∗ k2n + khk2n ,
2
2
p
X
X
∗
ρnj khj kF,j +
λnj khj kQ ,
Tn1
=
∗
Tn2
=
j∈S c
j=1
X
j∈S
λnj khj kQ .
∗ and T ∗ are
Compared with the definitions in Section S1.2, Zn is the same as before, and Tn1
n2
similar to Tn1 and Tn2 , but with khj kQ used instead of khj kn .
Let
Ωn1 =
sup
g∈G
kgk2n − kgk2Q
Rn∗2 (g)
≤ φn
.
Then P (Ωn1 ) ≥ 1 − π. In the event Ωn1 , we have by Assumption 6(i),
max sup
j=1,...,p gj ∈Gj
|kgj kn − kgj kQ |
≤ λn,p+1 φn1/2 ≤ η0 .
wnj kgj kF,j + kgj kQ
5
(S10)
By direct calculation, (S10) implies that if kgj kF,j + kgj kn /wnj ≤ 1 then kgj kF,j + kgj kQ /wnj ≤
(1 − η0 )−1 and hence (S10) implies that
H(u, Gj (wnj ), k · kn ) ≤ H((1 − η0 )u, Gj∗ (wnj ), k · kn ),
and ψnj (wnj ) satisfying (17) also satisfies (11) for δ = wnj . Let Ωn2 = Acn in Lemma 4. Then
conditionally on X1:n = (X1 , . . . , Xn ) for which Ωn1 occurs, we have P (Ωn2 |X1:n ) ≥ 1 − ǫ by
Lemma 2. Therefore, P (Ωn1 ∩ Ωn2 ) ≥ (1 − ǫ)(1 − π) ≥ 1 − ǫ − π.
In the event Ωn2 , recall that (S3) holds, that is,
Zn + (A0 − 1)Rn (h) ≤ ∆n (ḡ, S) + 2A0
X
j∈S
λnj khj kn .
(S11)
In the event Ωn1 ∩ Ωn2 , simple manipulation of (S11) using (S10) shows that
Zn + A1 Rn∗ (h) ≤ ∆∗n (ḡ, S) + 2A0
X
j∈S
λnj khj kQ ,
(S12)
where A1 = (A0 − 1) − η0 (A0 + 1) > 0 because A0 > (1 + η0 )/(1 − η0 ). In the following, we
restrict to the event Ωn1 ∩ Ωn2 with probability at least 1 − ǫ − π.
Proof of Corollary 4. Taking S = ∅ in (S12) yields (24), that is,
Zn + A1 Rn∗ (h) ≤ ∆∗n (ḡ, g∗ , ∅).
−2 ∗2
∗
∗
2
2
∗2
2
∗
As a result, Rn∗ (h) ≤ A−1
1 ∆n (ḡ, g , ∅) and hence khkQ ≤ khkn +φn Rn (h) ≤ khkn +φn A1 ∆n (ḡ, g , ∅).
Inequality (25) then follows from (24).
Proof of Theorem 2. Inequality (S12) can be expressed as
∗
∗
∗
Zn + A1 (Tn1
+ Tn2
) ≤ ∆∗n (ḡ, g ∗ , S) + 2A0 Tn2
,
which leads to two possible cases: either
∗
∗
ξ1∗ {Zn + A1 (Tn1
+ Tn2
)} ≤ ∆∗n (ḡ, S),
(S13)
∗ + T ∗ )} ≤ 2A T ∗ , that is,
or (1 − ξ1∗ ){Zn + A1 (Tn1
0 n2
n2
∗
∗
Zn + A1 (Tn1
+ Tn2
)≤
2A0 ∗
∗
∗
T = (ξ0∗ + 1)A1 Tn2
= ξ2∗ Tn2
,
1 − ξ1∗ n2
(S14)
where ξ1∗ = 1 − 2A0 /{(ξ0∗ + 1)A1 } ∈ (0, 1] because A0 > {ξ0∗ + 1 + η0 (ξ0∗ + 1)}/{ξ0∗ − 1 −
∗ ≤ ξ ∗ T ∗ , which, by the theoretical compatibility condition
η0 (ξ0∗ + 1)}. If (S14) holds, then Tn1
0 n2
6
(Assumption 5) with fj = ĝj − ḡj , implies
1/2
X
∗
λ2nj khkQ
Tn2
≤ κ∗−1
0
(S15)
j∈S
1/2
n
o
X
∗
∗
λ2nj
≤ κ∗−1
khkn + φ1/2
n (Tn1 + Tn2 )
0
(S16)
j∈S
P
2
2 2
By Assumption 6(ii), we have φn ξ2∗2 κ∗−2
0 ( j∈S λnj ) ≤ η1 A1 . Combining this fact, (S14) and
(S16) and simple manipulation yields
1/2
X
∗
∗
λ2nj khkn ,
Zn + (1 − η1 )A1 (Tn1
+ Tn2
) ≤ ξ2∗ κ0∗−1
j∈S
which, by the fact that khk2n /2 ≤ Zn , implies
∗
∗
Zn + (1 − η1 )A1 (Tn1
+ Tn2
) ≤ 2ξ2∗2 κ∗−2
0
X
j∈S
λ2nj .
(S17)
Therefore, inequality (S12), through (S13) and (S17), implies (22):
∗
∗
Zn + (1 − η1 )A1 (Tn1
+ Tn2
) ≤ ξ1∗−1 ∆∗n (ḡ, S) + 2ξ2∗2 κ∗−2
0
X
j∈S
λ2nj .
To demonstrate (23), we return to the two possible cases, (S13) or (S14). On one hand, if
∗ + T ∗ ) is also bounded from above by the right hand side
(S13) holds, then A1 Rn∗ (h) = A1 (Tn1
n2
of (S13) and hence
khk2Q ≤ khk2n + φn Rn∗2 (h) ≤ khk2n +
φn ∗−2 ∗2
ξ ∆n (ḡ, g∗ , S).
A21 1
(S18)
Simple manipulation of (S13) using (S18) yields
1
φn ∗−2 ∗2
1
∗
∗
kĝ − g∗ k2n + khk2Q + A1 (Tn1
+ Tn2
) ≤ ξ1∗−1 ∆∗n (ḡ, S) +
ξ ∆n (ḡ, g ∗ , S).
2
2
2A21 1
On the other hand, combining (S14) and (S15) yields
∗
∗
Zn + A1 (Tn1
+ Tn2
) ≤ ξ2∗ κ∗−1
0
X
j∈S
1/2
λ2nj
khkQ .
(S19)
(S20)
∗ + T ∗ ) is also bounded from above by the right hand side of
As a result, A1 Rn∗ (h) = A1 (Tn1
n2
(S20) and hence by Assumption 6(ii),
khk2Q ≤ khk2n + φn Rn∗2 (h)
X
1
φ
n
λ2nj khk2Q ≤ khk2n + η12 khk2Q .
≤ khk2n + 2 ξ2∗2 κ∗−2
0
2
A1
j∈S
7
(S21)
Simple manipulation of (S20) using (S21) yields
1
1−
kĝ − g ∗ k2n +
2
2
η12
1/2
X
∗
∗
λ2nj khkQ ,
khk2Q + A1 (Tn1
+ Tn2
) ≤ ξ2∗ κ∗−1
0
j∈S
which, when squared on both sides, implies
− η12
1
1
kĝ − g ∗ k2n +
2
2
∗
∗
khk2Q + A1 (Tn1
+ Tn2
)≤
2
1 − η12
X
λ2nj .
ξ2∗2 κ∗−2
0
(S22)
j∈S
Therefore, inequality (S12), through (S19) and (S22), implies
1 − η12
1
∗
∗
kĝ − g∗ k2n +
khk2Q + A1 (Tn1
+ Tn2
)
2
2
X
2
φn ∗−2 ∗2
≤ ξ1∗−1 ∆∗n (ḡ, S) +
λ2nj +
ξ2∗2 κ∗−2
ξ1 ∆n (ḡ, g ∗ , S),
0
2
2
1 − η1
2A1
j∈S
which yields (23) after divided by 1 − η12 on both sides.
Proof of Corollary 5. We use the following upper bounds, obtained from (S8) and (S9)
with S = {1 ≤ j ≤ p : kḡj k > C0∗ λnj },
X
j∈S
λ2nj
≤
(C0∗+ )−q
p
X
j=1
q
λ2−q
nj kḡj kQ ,
(S23)
and
X
j∈S c
λnj kḡj kQ ≤ C0∗1−q
p
X
j=1
q
λ2−q
nj kḡj kQ .
(S24)
Equations (21) and (26) together imply φn ∆∗n (ḡ, , S) = O(1) + φn kḡ − g∗ k2n /2. Inserting this
into (23) and applying (S23) and (S24) yields the high-probability result about Dn∗ (ĝ, ḡ). The
in-probability result follows by combining the facts that ǫ → 0, kḡ − g∗ k2n = Op (1)kḡ − g∗ k2Q by
the Markov inequality, and kĝ − g ∗ k2Q ≤ 2(kĝ − ḡk2Q + kḡ − g ∗ k2Q ) by the triangle inequality.
Proof of Corollary 6. First, we show
wn∗ (q) ≤ {γn∗ (q) + νn }1−q .
(S25)
In fact, if γn (q) ≥ νn , then γn∗ (q) = γn (q) and wn∗ (q) = γn (q)1−q ≤ {γn∗ (q) + νn }1−q . If
γn (q) < νn , then wn∗ (q) = νn1−q ≤ {γn∗ (q) + νn }1−q . By (S23), (S24), and (S25), inequality (30)
implies that for any constants 0 < η1 < 1 and η2 > 0, (21) and (26) are satisfied for sufficiently
2−q
large n. The desired result follows from Corollary 5 with ḡ = g ∗ , because λ2−q
{γn∗ (q) +
nj = C1
8
νn }2−q ≤ C12−q {γn (q)+ νn }2−q and, by (S25), ρnj = C1 wn∗ (q){γn∗ (q)+ νn } ≤ C1 {γn (q)+ νn }2−q .
Proof of Corollary 7. For a constant 0 < η1 < 1, we choose and fix C0∗ ′ ≥ C0∗ sufficiently
large, depending on q > 0, such that
∗′ q 2
(ξ0∗ + 1)2 κ∗−2
0 η3 ≤ (C0 ) η1 .
Let S ′ = {1 ≤ j ≤ p : kgj∗ kQ > C0∗ ′ λnj }. Then (21) is satisfied with S replaced by S ′ , due
to (33), (S23), and the definition λnj = C1 {γn∗ (q) + νn }. Similarly, (26) is satisfied with S
replaced by S ′ for η2 = M q + (C0∗ ′ )1−q M q , by (S24) and simple manipulation. By Remark 7,
Assumption 7 implies Assumption 5 and remains valid when S is replaced by S ′ ⊂ S. The
desired result follows from Corollary 5 with ḡ = g ∗ .
Proof of Corollary 8. The proof is similar to that of Corollary 6. First, we show
wn† (q)MF ≤ {γn† (q) + νn }1−q Mq .
(S26)
In fact, if γn′ (q) ≥ νn , then γn† (q) = γn′ (q) and wn† (q)MF = γn′ (q)1−q Mq ≤ {γn† (q) + νn }1−q Mq .
If γn (q) < νn , then wn† (q)MF = νn1−q Mq ≤ {γn† (q) + νn }1−q Mq . Then (36) implies that for
any constants 0 < η1 < 1 and η2 > 0, (21) and (26) are satisfied for sufficiently large n.
2−q
†
The desired result follows from Corollary 5 with ḡ = g ∗ , because λ2−q
nj Mq = C1 {γn (q) +
νn }2−q Mq ≤ C12−q {γn′ (q) + νn }2−q Mq and, by (S26), ρnj MF = C1 wn† (q){γn† (q) + νn }MF ≤
C1 {γn′ (q) + νn }2−q Mq .
S1.4
Proof of Theorem 3
We split the proof into three lemmas. First, we provide maximal inequalities on convergence
of empirical inner products in functional classes with polynomial entropies.
Lemma 6. Let F1 and F2 be two functional classes such that
sup kfj kQ ≤ δj ,
fj ∈Fj
sup kfj k∞ ≤ bj ,
j = 1, 2.
fj ∈Fj
Suppose that for some 0 < βj < 2 and Bnj,∞ > 0, condition (S45) holds with
ψn,∞ (z, Fj ) = Bnj,∞ z 1−βj /2 ,
9
j = 1, 2.
(S27)
Then we have
(
E
sup
f1 ∈F1 ,f2 ∈F2
|hf1 , f2 in − hf1 , f2 iQ | /C2
)
2C2 ψn,∞ (b1 , F1 ) 1−β1 /2
2C2 ψn,∞ (b2 , F2 ) β1 /2 ψn,∞ (b2 , F1 )
√
√
√
≤ 2 δ1 +
δ2 +
n
n
n
1−β2 /2
β2 /2
2C2 ψn,∞ (b2 , F2 )
ψn,∞ (b1 , F2 )
2C2 ψn,∞ (b1 , F1 )
√
√
√
+ 2 δ2 +
.
δ1 +
n
n
n
(S28)
Moreover, we have for any t > 0,
sup
f1 ∈F1 ,f2 ∈F2
≤E
(
|hf1 , f2 in − hf1 , f2 iQ | /C3
sup
f1 ∈F1 ,f2 ∈F2
)
|hf1 , f2 in − hf1 , f2 iQ |
+ δ1 b2
r
t
t
+ b1 b2 ,
n
n
(S29)
with probability at least 1 − e−t .
Proof. For any function f1 , f1′ ∈ F1 and f2 , f2′ ∈ F2 , we have by triangle inequalities,
kf1 f2 − f1′ f2′ kn ≤ δ̂2 kf1 − f1′ kn,∞ + δ̂1 kf2 − f2′ kn,∞ .
As a result, we have for u > 0,
H(u, F1 × F2 , k · kn ) ≤ H{u/(2δ̂2 ), F1 , k · kn,∞ } + H{u/(2δ̂1 ), F2 , k · kn,∞ },
(S30)
where F1 × F2 = {f1 f2 : f1 ∈ F1 , f2 ∈ F2 }.
By symmetrization inequality (van der Vaart & Wellner 1996),
)
(
(
E
sup
f1 ∈F1 ,f2 ∈F2
|hf1 , f2 in − hf1 , f2 iQ |
≤ 2E
sup
f1 ∈F1 ,f2 ∈F2
)
|hσ, f1 f2 in | .
Let δ̂12 = supf1 ∈F1 ,f2 ∈F2 kf1 f2 kn ≤ min(δ̂1 b2 , δ̂2 b1 ). By Dudley’s inequality (Lemma 13) conditionally on X1:n = (X1 , . . . , Xn ), we have
)
(
E
sup
f1 ∈F1 ,f2 ∈F2
|hσ, f1 f2 in | X1:n
/C2 ≤ E
(Z
δ̂12
0
H 1/2 (u, F1 × F2 , k · kn ) du X1:n
)
.
Taking expectations over X1:n , we have by (S30), (S45), and definition of H ∗ (),
)
(
E
sup
f1 ∈F1 ,f2 ∈F2
≤E
"Z
δ̂1 b2
0
|hf1 , f2 in − hf1 , f2 iQ | /C2
H ∗1/2 {u/(2δ̂2 ), F1 , k · kn,∞ } du +
Z
δ̂2 b1
0
h
i
≤ E δ̂2 ψn,∞ (δ̂1 b2 /δ̂2 , F1 ) + δ̂1 ψn,∞ (δ̂2 b1 /δ̂1 , F2 ) .
10
H ∗1/2 {u/(2δ̂1 ), F2 , k · kn,∞ } du
#
(S31)
By (S27) and the Hölder inequality, we have
n
o
1−β /2
β /2 1−β /2
E δ̂2 ψn,∞ (δ̂1 b2 /δ̂2 , F1 ) ≤ Bn1,∞ b2 1 E δ̂2 1 δ̂1 1
1−β1 /2
≤ Bn1,∞ b2
1−β1 /2
E β1 /2 (δ̂2 )E 1−β1 /2 (δ̂1 ) ≤ Bn1,∞ b2
E β1 /4 (δ̂22 )E (2−β1 )/4 (δ̂12 ),
and similarly
n
o
1−β /2
E δ̂1 ψn,∞ (δ̂2 b1 /δ̂1 , F2 ) ≤ Bn2,∞ b1 2 E β2 /4 (δ̂12 )E (2−β2 )/4 (δ̂22 ).
Then inequality (S28) follows from (S31) and Lemma 16. Moreover, inequality (S29) follows
from Talagrand’s inequality (Lemma 14) because kf1 f2 kQ ≤ δ1 b2 and kf1 f2 k∞ ≤ b1 b2 for
f1 ∈ F1 and f2 ∈ F2 .
By application of Lemma 6, we obtain the following result on uniform convergence of empirical inner products under conditions (39), (40), and (41).
Lemma 7. Suppose the conditions of Theorem 3 are satisfied for j = 1, 2 and p = 2. Let
Fj = Gj∗ (wnj ) for j = 1, 2. Then we have
(
E
sup
f1 ∈F1 ,f2 ∈F2
|hf1 , f2 in − hf1 , f2 iQ | /C2
)
β τ /2
β τ /2
≤ 2(1 + 2C2 C4 )C4 n1/2 Γn wn1 wn2 γn1 γ̃n2 wn21 2 + γn2 γ̃n1 wn12 1
.
−τ
where 0 < τj ≤ (2/βj −1)−1 and C4 = maxj=1,2 C4,j from condition (41), and γ̃nj = n−1/2 wnj j .
Moreover, we have for any t > 0,
sup
f1 ∈F1 ,f2 ∈F2
≤E
(
|hf1 , f2 in − hf1 , f2 iQ | /C3
sup
f1 ∈F1 ,f2 ∈F2
)
|hf1 , f2 in − hf1 , f2 iQ |
+ wn1 wn2 C4 t1/2 γ̃n2 + C42 tγ̃n1 γ̃n2 ,
with probability at least 1 − e−t .
Proof. For fj ∈ Fj with wnj ≤ 1, we have kfj kF,j ≤ 1 and kfj kQ ≤ wnj , and hence
1−τj
kfj k∞ ≤ C4 wnj
by (41). Let ψn,∞ (·, Fj ) = ψnj,∞ (·, wnj ) from (40), that is, in the form (S27)
1−τ
such that (S45) is satisfied. We apply Lemma 7 with δj = wnj and bj = C4 wnj j . By simple
manipulation, we have
1−τ
n−1/2 ψn,∞ (bj , Fj ) = n−1/2 ψnj,∞ (C4 wnj j , wnj )
−β /2
1−(1−βj /2)τj
≤ C4 Bnj,∞n−1/2 wnj j wnj
11
1−βj /2
≤ C4 Γn γnj wnj
≤ C4 wnj ,
where C4 ≥ 1 is used in the second step, Bnj,∞ ≤ Γn Bnj and (1 − βj /2)τj ≤ βj /2 in the
−β /2
third step, and γnj ≤ wnj and Γn γnj wnj j
inequality (S28) yields
(
E
sup
f1 ∈F1 ,f2 ∈F2
1−βj /2
≤ Γn γnj
|hf1 , f2 in − hf1 , f2 iQ | /C2
1−β1 /2
≤ 2(1 + 2C2 C4 )n−1/2 wn1
≤ 1 in the fourth step. Therefore,
)
β /2
1−τ2
wn21 ψn1,∞ (C4 wn2
, wn1 )
1−β2 /2
+ 2(1 + 2C2 C4 )n−1/2 wn2
1−β1 /2
≤ 2(1 + 2C2 C4 )C4 n−1/2 wn1
β /2
1−τ1
wn12 ψn2,∞ (C4 wn1
, wn2 )
1−β2 /2
+ 2(1 + 2C2 C4 )C4 n−1/2 wn2
−τ +β1 τ2 /2
Bn1,∞ wn2 wn2 2
−τ +β2 τ1 /2
Bn2,∞ wn1 wn1 1
,
which leads to the first desired inequality because Bnj,∞ ≤ Γn Bnj . Moreover, simple manipulation gives
δ1 b2
b1 b2
r
t
1−τ2
= C4 wn1 wn2
n
r
t
= C4 t1/2 wn1 wn2 γ̃n2 ,
n
t
1−τ1 1−τ2 t
= C42 wn1
wn2
= C42 twn1 γ̃n1 wn2 γ̃n2 .
n
n
The second desired inequality follows from (S29).
The following result concludes the proof of Theorem 3.
Lemma 8. In the setting of Theorem 3, let
βp+1 τj /2
γ̃nj wnj
γnj
max
φn = 4C2 C3 (1 + 2C2 C4 )C4 n Γn max
j
j λnj
λnj
p
2 log(p/ǫ′ )
′
√
γ̃nj
log(p/ǫ )
γ̃nj
+ 2C3 C4 max
max
+ 2C3 C42 max
,
j λnj
j
j
λnj
λ2nj
1/2
−τ
where γ̃nj = n−1/2 wnj j and βp+1 = minj=1,...,p βj . Then
kgk2n − kgk2Q
2
>
φ
≤ ǫ′ .
P sup
n
∗2
g∈G
Rn (g)
∗ (g ) = kg k
∗
Proof. For j = 1, . . . , p, let rnj
j
j F,j + kgj kQ /wnj and fj = gj /rj (gj ). Then kfj kF,j +
P
kfj kQ /wnj = 1 and hence fj ∈ Gj∗ (wnj ). By the decomposition kgk2n = j,k hgj , gk in , kgk2Q =
P
j,k hgj , gk iQ , and the triangle inequality, we have
kgk2n − kgk2Q ≤
=
X
j,k
X
j,k
|hgj , gk in − hgj , gk iQ |
∗
∗
(gk ) |hfj , fk in − hfj , fk iQ | .
rnj
(gj )rnk
12
Because R∗2 (g) =
sup
P
⊂
(
g=
P
∗
∗
j,k rnj (gj )rnk (gk )wnj λnj wnk λnk ,
kgk2n − kgk2Q
R∗2 (g)
p
j=1 gj
[
j,k
> φn
sup
fj ∈G ∗ (wnj ),fk ∈G ∗ (wnk )
=
g=
we have
[
Pp
j=1
gj
n
o
kgk2n − kgk2Q > φn R∗2 (g)
|hfj , fk in − hfj , fk iQ | > φn wnj λnj wnk λnk
)
By Lemma 7 with F1 = Gj∗ (wnj ), F2 = Gk∗ (wnk ), and t = log(p2 /ǫ′ 2 ), we have with probability
no greater than ǫ′ 2 /p2 ,
sup
fj ∈G ∗ (wnj ),fk ∈G ∗ (wnk )
|hfj , fk in − hfj , fk iQ | /C3
> 4C2 (1 + 2C2 C4 )C4 n1/2 Γn Wn wnj γnj wnk γnk
q
2
1/2
+ C4 n Vn wnj wnk γnk log(p2 /ǫ′ 2 )/n + C42 Vn2 log(p2 /ǫ′ )wnj γnj wnk γnk .
Therefore, we have by the definition of φn ,
P
sup
fj ∈G ∗ (wnj ),fk ∈G ∗ (wnk )
|hfj , fk in − hfj , fk iQ | > φn wnj λnj wnk λnk
The desired result follows from the union bound.
S1.5
!
≤
ǫ′ 2
.
p2
Proofs of Propositions 3, 4, 6, and 8
Denote wnj = wn,p+1 and γnj = γn,p+1 for j = 1, . . . , p. By direct calculation, (42) implies
that for any 0 ≤ q ≤ 1,
n
2−q
1−q
φn (γn,p+1 + νn )2−q ≤ O(1) n1/2 Γn Wn γn,p+1
+ n1/2 Vn min γn,p+1 νn1−q , γn,p+1
νn
o
2−q
2
(S32)
+ nVn2 min γn,p+1
νn2−q , γn,p+1
νn2 ,
where
β /2−τ
0
0
Vn = wn,p+1
,
β /2−τ0 +β0 τ0 /2
0
Wn = wn,p+1
.
(S33)
We verify that the technical conditions hold as needed for Theorem 3, with wnj = wn∗ (q)
and γnj = γn∗ (q) for 0 ≤ q ≤ 1. First, we verify γnj ≤ wnj for sufficiently large n. It suffices
to show that γn∗ (q) ≤ wn∗ (q) whenever γn (q) ≤ 1 and νn ≤ 1. In fact, if γn (q) ≥ νn , then
wn∗ (q) = γn (q)1−q and γn∗ (q) = γn (q) ≤ γn (q)1−q provided γn (q) ≤ 1. If γn (q) < νn , then
−(1−q)β0 /2
wn∗ (q) = νn1−q and γn∗ (q) = B0∗ n−1/2 νn
≤ νn ≤ νn1−q provided νn ≤ 1. Moreover, we
13
have Γn γn∗ (q)1−β0 /2 ≤ 1 for sufficiently large n, because Γn is no greater than O(log1/2 (n)) and
γn∗ (q)1−β0 /2 ≤ γn (q)1−β0 /2 decreases polynomially in n−1 for 0 < β0 < 2.
Proof of Proposition 3. For wnj = 1 and γnj = γn∗ (1) ≍ n−1/2 , inequality (S32) with
q = 0 and νn = o(1) gives
o
n
φn {γn∗ (1) + νn }2 ≤ O(1) n1/2 Γn γn∗2 (1) + n1/2 γn∗ (1)νn + nγn∗2 (1)νn2
= O(1) n−1/2 Γn + νn ,
Assumption 6(i) holds because Γn is no greater than O(log1/2 (n)). Inserting the above inequality into (25) in Corollary 4 yields the out-of-sample prediction result. The in-sample prediction
result follows directly from Corollary 4.
Proof of Proposition 4. For γnj = γn∗ (0), inequality (S32) with q = 0 gives
o
n
φn {γn∗ (0) + νn }2 ≤ O(1) n1/2 Γn Wn γn∗2 (0) + n1/2 Vn γn∗ (0)νn + nVn2 γn∗2 (0)νn2 .
(S34)
By (S33) and γn∗ (0) = B0∗ n−1/2 wn∗ (0)−β0 /2 , simple manipulation gives
n1/2 Vn γn∗ (0)νn = B0∗ wn∗ (0)−τ0 νn ,
(S35)
n1/2 Wn γn∗ (0)2 = B0∗ wn∗ (0)−(1−β0 /2)τ0 γn∗ (0).
Then (49) and (S34) directly imply that Assumption 6(i) holds for sufficiently large n and also
(30) holds. The desired result follows from Corollary 6 with q = 0.
Proof of Proposition 6. For γnj = γn∗ (q), inequality (S32) with q = 0 gives
o
n
φn {γn∗ (q) + νn }2 ≤ O(1) n1/2 Γn Wn γn∗2 (q) + n1/2 Vn γn∗ (q)νn + nVn2 γn∗2 (q)νn2 .
(S36)
By (S33) and γn∗ (q) = B0∗ n−1/2 wn∗ (q)−β0 /2 , simple manipulation gives
n1/2 Vn γn∗ (q)νn1−q = B0∗ wn∗ (q)−τ0 νn1−q ,
n1/2 Wn γn∗ (q)2−q = B0∗ wn∗ (q)−(1−β0 /2)τ0 γn∗ (q)1−q .
Then (57) and (S36) imply that Assumption 6(i) holds for sufficiently large n, along with the
fact that νn = o(1), γn (q) = o(1), and q > 0. Moreover, (57) and (S32) with γnj = γn∗ (q)
directly yield (33). The desired result follows from Corollary 7.
Proof of Proposition 8.
′
, Vn′ , Wn′ , etc., the corresponding quanDenote by γn,p+1
′ , γ ′ ). By (S33) and (S35) with τ = 1, we have n1/2 V ′ γ ′
tities based on (wnj
0
n n,p+1 νn =
nj
14
K0−1 (n1/2 Vn γn,p+1 νn ) and n1/2 Vn γn,p+1 νn = B0∗ min{νn γn−1 (0), 1} ≤ B0∗ . Moreover, we have
n1/2 Γn Wn′ γ ′ 2n,p+1 = o(1) for a constant K0 , because Wn′ = 1, Γn is no greater than O(log1/2 (n)),
and n1/2 γn2 (0) decreases polynomially in n−1 . For a constant 0 < η1 < 1, we choose and fix
K0 ≥ 1 sufficiently large, depending on M but independently of (n, p), such that Assump-
′ , γ ′ ), for sufficiently large n, due
tions 6(i)–(ii) are satisfied, with (wnj , γnj ) replaced by (wnj
nj
′ + ν ). Moreover, by (S25), ρ′ = λ′ w ′ ≤
to (S23), (S34), and the definition λ′nj = C1 (γnj
n
nj
nj nj
1−β0 /2
K0
1−β0 /2
λnj wnj ≤ K0
C1 {γn∗ (0)+νn }2 , which together with (S24) implies that (26) is satisβ /2
fied for some constant η2 > 0. Assumption 7 is also satisfied with C0∗ replaced by C0∗′ = C0∗ K0 0
and S replaced by {1 ≤ j ≤ p : kgj∗ kQ > C0∗′ λ′nj } ⊂ S for K0 ≥ 1 due to monotonicity in S for
′ , γ ′ ) because
the validity of Assumption 7 by Remark 7, and with (wnj , γnj ) replaced by (wnj
nj
′ ≥w
(18) after the modification implies (18) itself, with wnj
nj for K0 ≥ 1 and λnj constant in
j. The desired result follows from Corollary 5 with ḡ = g ∗ .
S1.6
Proof of Theorem 4
We use the non-commutative Bernstein inequality (Lemma 15) to prove Theorem 4. Suppose
that (X1 , . . . , Xn ) are independent variables in a set Ω. First, consider finite-dimensional
functional classes Fj with elements of the form
fj (x) = uTj (x)θj ,
∀ θj ∈ Rdj , j = 1, 2,
(S37)
where uj (x) is a vector of basis functions from Ω to Rdj , and θj is a coefficient vector. Let
Uj = {uj (X1 ), . . . , uj (Xn )}T , and Σjj ′ = E UjT Uj ′ /n ∈ Rdj ×dj′ . The population inner product
is hfj , fj′ iQ = θjT Σjj ′ θj ′ , j, j ′ = 1, 2. The difference between the sample and population inner
products can be written as
sup
fj , fj ′
kθj k=kθj ′ k=1
n
− fj , fj ′
Q
=
sup
kθj k=kθj ′ k=1
|θjT (UjT Uj ′ /n − Σjj ′ )θj′ |
= kUjT Uj ′ /n − Σjj ′ kS .
Lemma 9. Let fj be as in (S37). Assume that for a constant C5,1 ,
sup kuj (x)k2 ≤ C5,1 ℓj ,
x∈Ω
∀j = 1, 2.
Then for all t > 0,
kUjT Uj ′ /n − Σjj ′ kS >
q
r
(ℓj kΣj ′ j ′ kS ) ∨ (ℓj ′ kΣjj kS )
with probability at least 1 − (dj + dj ′ )e−t .
15
p
2C5,1 t
4t
+ C5,1 ℓj ℓj ′
n
3n
Let Mi = uj (Xi )uTj ′ (Xi ) − E{uj (Xi )uTj ′ (Xi )}. Because uj (Xi )uTj ′ (Xi ) is of rank
p
p
1, kMi kS ≤ 2 supx∈Ω {kuj (x)kkuTj ′ (x)k} ≤ 2C5,1 ℓj ℓj ′ . Hence we set s0 = 2C5,1 ℓj ℓj ′ in
Proof.
Lemma 15. Similarly, Wcol ≤ C5,1 ℓj ′ kΣjj kS because
E(Mi MiT ) ≤ E{uj (Xi )uTj ′ (Xi )uj ′ (Xi )uTj (Xi )} ≤ C5,1 ℓj ′ E{uj (Xi )uTj (Xi )},
and Wrow ≤ C5,1 ℓj kΣj ′ j ′ kS . Thus, (S44) gives the desired result.
Now consider functional classes Fj such that fj ∈ Fj admits an expansion
fj (·) =
∞
X
θjℓ ujℓ (·),
ℓ=1
where {ujℓ (·) : ℓ = 1, 2, . . .} are basis functions and {θjℓ : ℓ = 1, 2, . . .} are the associated
coefficients.
Lemma 10. Let 0 < τj < 1, 0 < wnj ≤ 1 and
n
P
P
2
≤ k−1/τj ∀ k ≥ (1/wnj )2τj ,
Bj = fj : k/4<ℓ≤k θjℓ
0≤ℓ1/τj w 2
nj
2
≤ wnj
θ2
<1 j,ℓ+1
o
Suppose that (44) and (47) hold with certain positive constants C5,1 , C5,3 . Then, for a certain
constant C5,4 depending on {C5,1 , C5,3 } only,
sup
fj ∈Bj ,fj ′ ∈Bj ′
hfj , fj ′ in − hfj , fj ′ iQ
q
h
−τ ′
−τ ′
−τ
−τ
≤ C5,4 wnj wnj ′ (µj wnj j + µj ′ wnj ′j )
µj + µj ′ + log(wnj j + wnj ′j ) + t /n
i
−τ ′
−τ ′
−τ
−τ
+ µj + µj ′ + log(wnj j + wnj ′j ) + t (µj wnj j )(µj ′ wnj ′j )/n
with at least probability 1 − e−t for all t > 0, where µj = 1/(1 − τj ).
Proof. Let ℓjk = ⌈(2k /wnj )2τj ⌉. We group the basis and coefficients as follows:
uj,Gjk (x) = (ujℓ (x), ℓ ∈ Gjk )T ,
θj,Gjk = (θjℓ , ℓ ∈ Gjk )T ,
k = 0, 1, . . .
where Gj0 = {1, . . . , ℓj0 } of size |Gj0 | = ℓj0 and Gjk = {ℓj,k−1 + 1, . . . , ℓjk } of size |Gjk | =
ℓjk − ℓj,k−1 ≤ (2k /wnj )2τj for k ≥ 1. Define θ̃j , a rescaled version of θj , by
−1
θj,Gjk .
θ̃j,Gjk = (θ̃jℓ , ℓ ∈ Gjk ) = 2k wnj
It follows directly from (45) and (46) that
kθ̃j,Gj,0 k2 ≤ 1,
−1/(2τj )
kθ̃j,Gjk krj ≤ (2k /wnj )/ℓjk
16
≤ 1 ∀ k ≥ 1, ∀ fj ∈ Bj .
Let Ujk = {uj,Gjk (X1 ), . . . , uj,Gjk (Xn )}T ∈ Rn×|Gjk | . We have
sup
fj ∈Bj ,fj ′ ∈Bj ′
=
hfj , fj ′ in − hfj , fj ′ iL2
sup
fj ∈Bj ,fj ′ ∈Bj ′
≤
max
k=0 ℓ=0
∞
∞ X
X
T
T
T
′
′
θj,G
U
U
/n
−
E
U
U
/n
θj ′ ,Gj′ ,ℓ
jk j ,ℓ
jk j ,ℓ
jk
T
T
T
Uj ′ ,ℓ /n
Uj ′ ,ℓ /n − E Ujk
Ujk
θ̃j,Gjk
k=0 ℓ=0
∞
∞ X
T
X
Uj ′ ,ℓ /n
Ujk
kθ̃j k∨kθ̃j ′ k≤1
≤ wnj wnj ′
∞ X
∞
X
k=0 ℓ=0
−1 ℓ −1
2k wnj
2 wnj ′
T
Uj ′ ,ℓ /n
− E Ujk
2k 2ℓ
θ̃j ′ ,Gj′ ,ℓ
.
(S38)
S
Let ak = 1/{(k + 1)(k + 2)}. By (44), supx∈Ω kuj,Gjk (x)k2 ≤ supx∈Ω
Pℓjk
2
ℓ=1 ujℓ (x)
≤ C5,1 ℓjk for
T
Ujk /nkS ≤ C5,3 . Because |Gjk | ≤ ℓj,k , it follows from Lemma 9 that
k ≥ 0. By (47), kEUjk
T
T
Uj ′ ,ℓ /nkS
Uj ′ ,ℓ /n − E Ujk
kUjk
q
log(ℓjk + ℓj ′ ,ℓ ) − log(ak aℓ ) + t 2C5,1 C5,3 (ℓjk ∨ ℓj ′ ,ℓ )/n
≤
p
+ log(ℓjk + ℓj ′ ,ℓ ) − log(ak aℓ ) + t (4/3)C5,1 ℓjk ℓj ′ ,ℓ /n
(S39)
with probability at least 1 − ak aℓ e−t for any fixed k ≥ 0 and ℓ ≥ 0. By the union bound and
P
the fact that ∞
k=0 ak = 1, inequality (S39) holds simultaneously for all k ≥ 0 and ℓ ≥ 0 with
probability at least 1 − e−t . Because ℓjk = ⌈(2k /wnj )2τj ⌉, we rewrite (S39) as
T
T
kUjk
Uj ′ ,ℓ /n − E Ujk
Uj ′ ,ℓ /nkS
q
h
−τ ′
−τ ′
−τ
−τ
k + ℓ + log(wnj j + wnj ′j ) + t /n
≤C5,4 (2τj k wnj j + 2τj′ ℓ wnj ′j )
i
−τ ′
−τ ′
−τ
−τ
+ k + ℓ + log(wnj j + wnj ′j ) + t (2τj k wnj j )(2τj′ ℓ wnj ′j )/n .
where C5,4 is a constant depending only on {C5,1 , C5,3 }. For any α ≥ 0,
P∞
k=0 k
(S40)
α 2−k(1−τj )
≤
Cα µα+1
, where Cα is a numerical constant and µj = 1/(1 − τj ). Using this fact and inserting
j
(S40) into (S38) yields the desired result.
Finally, the following result concludes the proof of Theorem 4.
Lemma 11. In the setting of Theorem 4, let
)
(
p
2
2 log(np/ǫ′ )µ2j γ̃nj
2 log(np/ǫ′ )
µj γ̃nj
max
+ max
,
φn = C5,2 C5,4 max
j
j
j
λnj
λnj
λ2nj
17
−τ
where γ̃nj = n−1/2 wnj j , µj = 1/(1−τj )−1 , and C5,4 is a constant depending only on {C5,1 , C5,3 }
as in Lemma 10. Then
P
sup
g∈G
kgk2n − kgk2Q
Proof. Recall that ℓjk = ⌈(2k /wnj )2τj ⌉. For gj =
2
rnj
(gj )
X
ℓj0
=
> φn
Rn∗2 (g)
2
2
θjℓ
/wnj
ℓ=1
∨
P∞
ℓ=1 θjℓ ujℓ ,
max
k≥1
2
≤ ǫ′ .
X
define rnj (gj ) by
2 1/τj
θjℓ
ℓjk
ℓj,k−1 <ℓ≤ℓjk
.
Let fj = gj /rnj (gj ) and µj = 1/(1 − τj ). Then fj ∈ Bj as in Lemma 10 and
kgk2n
=
−
p X
p
X
j=1
Because
Pp
i=1 wnj λnj rnj (gj )
sup
g∈G
⊂
[
j,j ′
kgk2Q
j ′ =1
≤
Rn∗2 (g)
sup
fj ∈Bj ,fj ′ ∈Bj ′
j=1 j ′ =1
hgj , gj ′ in − hgj , gj ′ iQ
rnj (gj )rnj ′ (gj ′ ) hfj , fj ′ in − hfj , fj ′ iQ .
1/2
j=1 C5,2 λnj (wnj kgj kF,j
Pp
kgk2n − kgk2Q
(
≤
p X
p
X
> φn
1/2
+ kgj kQ ) = C5,2 Rn∗ (g) by (45),
−1
φn wnj λnj wnj ′ λnj ′
|hfj , fk in − hfj , fk iQ | > C5,2
)
.
(S41)
−τ
By Lemma 10 with t = log(p2 /ǫ′ 2 ) and e2µj + 2wnj j ≤ n, we have
sup
fj ∈Bj ,fj ′ ∈Bj ′
hfj , fj ′ in − hfj , fj ′ iQ
q
h
−τ ′
−τ ′
−τ
−τ
≤ C5,4 wnj wnj ′ (µj wnj j + µj ′ wnj ′j )
µj + µj ′ + log(wnj j + wnj ′j ) + log(p2 /ǫ′ 2 ) /n
i
−τ ′
−τ ′
−τ
−τ
2
+ µj + µj ′ + log(wnj j + wnj ′j ) + log(p2 /ǫ′ ) (µj wnj j )(µj ′ wnj ′j )/n
n
o
−τ ′
−τ ′ p
−τ
−τ
≤ C5,4 wnj wnj ′ (µj wnj j + µj ′ wnj ′j ) 2 log(np/ǫ′ )/n + 2 log(p/ǫ′ )(µj wnj j )(µj ′ wnj ′j )/n ,
with probability at least 1 − ǫ′ 2 /p2 . By the definition of φn , we have
(
P
sup
fj ∈Bj ,fj ′ ∈Bj ′
hfj , fj ′ in − hfj , fj ′ iQ ≤
−1
C5,2
φn wnj wnj ′ λnj λnj ′
The conclusion follows from the union bound using (S41).
18
)
≥1−
ǫ′ 2
.
p2
S1.7
Proof of Proposition 5
Here we verify explicitly conditions of Theorem 4 for Sobolev spaces Wrmi i and bounded varia-
tion spaces V mi with ri = 1 on [0, 1] in the case of τj < 1, where τi = 1/(2mi +1−2/(ri ∧2)). Be-
cause conditions (44), (45), (46) and (47) depend on (mj , rj ) only through τj , we assume with(j)
out loss of generality 1 ≤ rj ≤ 2. When the average marginal density of {Xi
: i = 1, . . . , n}
is uniformly bounded away from 0 and ∞, the norms kgj kQ and kgj kL2 are equivalent, so that
condition (46) and (47) hold for any L2 -orthonormal bases {ujℓ : ℓ ≥ 1}. Let u0 (x) be a mother
R
R
wavelet with m vanishing moments, e.g., u0 (x) = 0 for |x| > c0 , u20 (x)dx = 1, xm u0 (x)dx =
√
0 for m = 0, . . . , maxj mj , and {u0,kℓ (x) = 2k u0 (2k (x − j)) : ℓ = 1, . . . , 2k , k = 0, 1, . . .} is
L2 -orthonormal. We shall identify {ujℓ : ℓ ≥ 1} as {u0,11 , u0,21 , u0,22 , u0,31 , . . .}. Because
#{ℓ : u0,kℓ (x) 6= 0} ≤ 2c0 k ∀x,
2k+1
X−1
k
u2jℓ (x)
=
ℓ=1
ℓ=2k
cause u0 has
u20,kℓ (x) ≤ 2c0 2k ku0 k∞ , ∀x,
P∞ P2k
(−m)
(x)
ℓ=1 θjkℓ u0,kℓ (x). Define u0
ℓ=1 θjℓ ujℓ (x) =
k=0
R
(m)
(−m+1)
(−m)
x
(x) = −∞ u0
(t)dt, and gj (x) = (d/dx)m gj (x). Beintegral of u0 , u0
R (m)
(m)
vanishing moments, u0 (x)dx = 0 for m = 0, . . . , maxj mj , so that u0 (x) = 0
so that (44) holds. Suppose gj (x) =
as the m-th
2
X
P∞
for |x| > c0 . Due to the orthonormality of the basis functions, for 1 ≤ ℓ ≤ 2k , we have
Z
Z
(m )
2mj k θjkℓ = 2mj k gj (x)u0,kℓ (x)dx = (−1)m gj j (x)u0,mj kℓ (x)dx
with u0,mkℓ (x) =
k −1
2X
mj k
2
θjℓ
rj
ℓ=2k−1
√
(−m)
2k u0
≤
k −1
2X
ℓ=2k−1
(2k (x − j)). By the Hölder inequality,
Z
(mj )
gj
(mj ) rj
≤ gj
Lr j
(x)
rj
u0,mj kℓ (x)
rj (1−2(1−/rj ))
dx u0,mj kℓ
(mj ) rj (1−2(1−/rj )) (k/2)rj (1−2(1−/rj ))
2c0 u0
2
∞
rj (2(1−1/rj ))
L2
(mj )
u0
L2
2rj (2(1−1/rj )) .
Because 2mj k−(k/2)(1−2(1−/rj )) = 2k(mj +1/2−1/rj ) = 2k/(2τj ) and 1 ≤ rj ≤ 2, we have
k −1
2X
ℓ=2k−1
2
2k/τj θjℓ
1/2
≤
k −1
2X
k/(2τj )
2
ℓ=2k−1
1/(2τj )
Because ℓjk
1/rj
(mj ) 2/rj −1
(mj )
≤ gj
θjℓ
rj
Lr j
(2c0 )1/rj u0
∞
(mj ) 2−2/rj
u0
L2
.
≥ 2k /wnj and ℓjk ≤ 1 + 22τj ℓj,k−1 with τj < 1, we have ℓjk ≤ 4ℓj,k−1 , so that
{ℓj,k−1 + 1, . . . , ℓj,k } involves at most three resolution levels. Thus, condition (45) follows from
19
the above inequality. For the bounded variation class, we have
Z
Z
(m −1)
mj k
mj k
m
2
θjkℓ = 2
gj (x)u0,kℓ (x)dx = (−1)
u0,mj kℓ (x)dgj j (x),
so that (45) follows from the same proof with rj = 1.
S2
S2.1
Technical tools
Sub-gaussian maximal inequalities
The following maximal inequality can be obtained from van de Geer (2000, Corollary 8.3), or
directly derived using Dudley’s inequality for sub-gaussian variables and Chernoff’s tail bound
(see Proposition 9.2, Bellec et al. 2016).
Lemma 12. For δ > 0, let F1 be a functional class such that supf1 ∈F1 kf1 kn ≤ δ, and
Z δ
H 1/2 (u, F1 , k · kn ) du.
(S42)
ψn (δ, F1 ) ≥
0
Let (ε1 , . . . , εn ) be independent variables. Under Assumption 1, we have for any t > 0,
)
(
p
−1/2
ψn (δ, F1 ) + δ t/n ≤ exp(−t),
P sup |hε, f1 in |/C1 > n
f1 ∈F1
where C1 = C1 (D0 , D1 ) > 0 is a constant, depending only on (D0 , D1 ).
S2.2
Dudley and Talagrand inequalities
The following inequalities are due to Dudley (1967) and Talagrand (1996).
Lemma 13. For δ > 0, let F1 be a functional class such that supf1 ∈F1 kf1 kn ≤ δ and (S42)
holds. Let (σ1 , . . . , σn ) be independent Rademacher variables, that is, P (σi = 1) = P (σi =
−1) = 1/2. Then for a universal constant C2 > 0,
)
(
E
sup |hσ, f1 in |/C2
f1 ∈F1
≤ n−1/2 ψ(δ, F1 ).
Lemma 14. For δ > 0 and b > 0, let (X1 , . . . , Xn ) be independent variables, and F be a
functional class such that supf ∈F kf kQ ≤ δ and supf ∈F kf k∞ ≤ b. Define
n
Zn = sup
f ∈F
1X
{f (Xi ) − Ef (Xi )} .
n
i=1
Then for a universal constant C3 > 0, we have
)
(
r
t
t
≤ exp(−t),
+b
P Zn /C3 > E(Zn ) + δ
n
n
20
t > 0.
S2.3
Non-commutative Bernstein inequality
We state the non-commutative Bernstein inequality (Troop, 2011) as follows.
Lemma 15. Let {Mi : i = 1, . . . , n} be independent random matrices in Rd1 ×d2 such that
E(Mi ) = 0 and P {kMi kS ≤ s0 } = 1, i = 1, . . . , n, for a constant s0 > 0, where k · kS denotes
P
P
the spectrum norm of a matrix. Let Σcol = ni=1 E(Mi MiT )/n and Σrow = ni=1 E(MiT Mi )/n.
Then, for all t > 0,
n
1X
P
Mi
n
i=1
>t
S
≤ (d1 + d2 ) exp
−nt2 /2
.
kΣcol kS ∨ kΣrow kS + s0 t/3
Consequently, for all t > 0,
n
p
p
1X
Mi > kΣcol kS ∨ kΣrow kS 2t/n + (s0 /3)2t/n ≤ (d1 + d2 )e−t .
P
n
S
(S43)
(S44)
i=1
S2.4
Convergence of empirical norms
For δ > 0 and b > 0, let F1 be a functional class such that
sup kf1 kQ ≤ δ,
f1 ∈F1
sup kf1 k∞ ≤ b,
f1 ∈F1
and let ψn,∞ (·, F1 ) be an upper envelope of the entropy integral:
Z z
H ∗ 1/2 (u/2, F1 , k · kn,∞ ) du,
ψn,∞ (z, F1 ) ≥
z > 0,
(S45)
0
where H ∗ (u, F1 , k · kn,∞ ) = sup(X (1) ,...,X (1) ) H(u, F1 , k · kn,∞ ). Let δ̂ = supf1 ∈F1 kf1 kn . The
1
n
following result can be obtained from Guedon et al. (2007) and, in its present form, van de
Geer (2014), Theorem 2.1.
Lemma 16. For the universal constant C2 in Lemma 13, we have
)
(
2
(b, F1 )
2δC2 ψn,∞ (b, F1 ) 4C22 ψn,∞
2
2
√
E sup kf1 kn − kf1 kQ ≤
.
+
n
n
f1 ∈F1
Moreover, we have
q
S2.5
E(δ̂2 ) ≤ δ +
2C2 ψn,∞ (b, F1 )
√
.
n
Metric entropies
m
For r ≥ 1 and m > 0 (possibly non-integral), let W r = {f : kf kLr + kf (m) kLr ≤ 1}. The
following result is taken from Theorem 5.2, Birman & Solomjak (1967).
21
Lemma 17. If rm > 1 and 1 ≤ q ≤ ∞, then
m
H(u, W r , k · kLq ) ≤ B1 u−1/m ,
u > 0,
where B1 = B1 (m, r) > 0 is a constant depending only on (m, r). If rm ≤ 1 and 1 ≤ q <
r/(1 − rm), then
m
H(u, W r , k · kLq ) ≤ B2 u−1/m ,
u > 0,
where B2 = B2 (m, r, q) > 0 is a constant depending only on (m, r, q).
For m ≥ 1, let V
m
= {f : kf kL1 + TV(f (m−1) ) ≤ 1}. The following result can be obtained
from Theorem 15.6.1, Lorentz et al. (1996), on the metric entropy of the ball {f : kf kLr +
[f ]Lip(m,Lr ) ≤ 1}, where [f ]Lip(m,Lr ) is a semi-norm in the Lipschitz space Lip(m, Lr ). By
Theorem 9.9.3, DeVore & Lorentz (1993), the space Lip(m, L1 ) is equivalent to V m , with
the semi-norm [f ]Lip(m,Lr ) equal to TV(f ), up to suitable modification of function values at
1
(countable) discontinuity points. However, it should be noted that the entropy of V endowed
with the norm k · kL∞ is infinite.
Lemma 18. If m ≥ 2 and 1 ≤ q ≤ ∞, then
m
H(u, V , k · kLq ) ≤ B3 u−1/m ,
u > 0,
where B3 = B3 (m) > 0 is a constant depending only on m. If 1 ≤ q < ∞, then
1
H(u, V , k · kLq ) ≤ B4 u−1 ,
u > 0,
where B4 = B4 (r) > 0 is a constant depending only on r.
By the continuity of functions in Wrm for m ≥ 1 and V m for m ≥ 2, the maximum entropies
of these spaces in k · kn,∞ and k · kn norms over all possible design points can be derived from
Lemmas 17 and 18.
Lemma 19. If rm > 1, then for B1 = B1 (m, r),
m
m
H ∗ (u, W r , k · kn ) ≤ H ∗ (u, W r , k · kn,∞ ) ≤ B1 u−1/m ,
m
u > 0,
m
and hence (S42) and (S45) hold with ψn (z, W r ) ≍ ψn,∞ (z, W r ) ≍ z 1−1/(2m) . If m ≥ 2, then
for B3 = B3 (m),
m
m
H ∗ (u, V , k · kn ) ≤ H ∗ (u, V , k · kn,∞ ) ≤ B3 u−1/m ,
m
m
u > 0,
and hence (S42) and (S45) hold with ψn (z, V ) ≍ ψn,∞ (z, V ) ≍ z 1−1/(2m) .
22
1
The maximum entropies of V over all possible design points can be obtained from Section
5, Mammen (1991) for the norm k · kn and Lemma 2.2, van de Geer (2000) for the norm
k · kn,∞ . In fact, the proof of van de Geer shows that for F the class of nondecreasing functions
f : [0, 1] → [0, 1], H ∗ (u, F, k · kn,∞ ) ≤ n log(n + u−1 ) if u ≤ n−1 or ≤ u−1 log(n + u−1 ) if
u > n−1 . But if u ≤ n−1 , then n log(n + u−1 ) ≤ n(log n + n−1 u−1 ) ≤ (1 + log n)u−1 . If
u > n−1 , then u−1 log(n + u−1 ) ≤ u−1 log(2n). Combining the two cases gives the stated result
1
1
about H ∗ (u, V , k · kn,∞ ), because each function in V can be expressed as a difference two
nondecreasing functions.
Lemma 20. For a universal constant B5 > 0, we have
1
1
H ∗ (u, W 1 , k · kn ) ≤ H ∗ (u, V , k · kn ) ≤ B5 u−1 ,
1
u > 0,
1
and hence (S42) holds with ψn (z, W 1 ) ≍ ψn (z, V ) ≍ z 1/2 . Moreover, for a universal constant
B6 > 0, we have
1
1
H ∗ (u, W 1 , k · kn,∞ ) ≤ H ∗ (u, V , k · kn,∞ ) ≤ B6
1
1 + log n
,
u
u > 0,
1
and hence (S45) holds with ψn,∞ (z, W 1 ) ≍ ψn,∞ (z, V ) ≍ (1 + log n)1/2 (z/2)1/2 .
S2.6
Interpolation inequalities
The following inequality (S46) can be derived from the Gagliardo-Nirenberg inequality for
Sobolev spaces (Theorem 1, Nirenberg 1966). Inequality (S47) can be shown by approximating
f ∈ V m by functions in W1m .
Lemma 21. For r ≥ 1 and m ≥ 1, we have for any f ∈ Wrm ,
oτ
n
kf k∞ ≤ (C4 /2) kf (m) kLr + kf kL2 kf k1−τ
L2 ,
(S46)
where τ = (2m + 1 − 2/r)−1 ≤ 1 and C4 = C4 (m, r) ≥ 1 is a constant depending only on (m, r).
In addition, we have for any f ∈ V m ,
oτ
n
kf k∞ ≤ (C4 /2) TV(f (m−1) ) + kf kL2 kf k1−τ
L2 .
(S47)
From this result, kf k∞ can be bounded in terms of kf kL2 and kf (m) kLr or TV(f (m−1) )
in a convenient manner. For f ∈ Wrm and 0 < δ ≤ 1, if kf kL2 ≤ δ and kf (m) kLr ≤ 1,
then kf k∞ ≤ C4 δ1−1/(2m+1−2/r) . Similarly, for f ∈ V m and 0 < δ ≤ 1, if kf kL2 ≤ δ and
TV(f (m−1) ) ≤ 1, then kf k∞ ≤ C4 δ1−1/(2m−1) .
23
References
Bellec, P.C., Lecue, G., Tsybakov, A.B. (2016) Slope meets Lasso: Improved oracle bounds
and optimality, arXiv:1605.08651.
Birman, M.S. and Solomjak, M.Z. (1967) Piecewise-polynomial approximations of functions
of the classes Wpα , Mathematics of the USSR–Sbornik, 2, 295–317.
Dudley, R.M. (1967) The sizes of compact subsets of Hilbert space and continuity of Gaussian
processes, Journal of Functional Analysis, 1, 290–330.
Guedon, O., Mendelson, S., Pajor, A., and Tomczak-Jaegermann, N. (2007) Subspaces and
orthogonal decompositions generated by bounded orthogonal systems, Positivity, 11,
269–283.
Mammen, E. (1991) Nonparametric regression under qualitative smoothness assumptions,
Annals of Statistics, 19, 741–759.
Talagrand, M. (1996) New concentration inequalities in product spaces. Inventiones Mathematicae 126, 505–563.
Tropp, J.A. (2011) Freedman’s inequality for matrix martingales, Electronic Communications
in Probability, 16, 262–270.
van de Geer, S. (2014) On the uniform convergence of empirical norms and inner products,
with application to causal inference, Electronic Journal of Statistics, 8, 543–574.
24
| 10math.ST
|
EXTREMES AND LIMIT THEOREMS FOR DIFFERENCE OF CHI-TYPE PROCESSES
PATRIK ALBIN, ENKELEJD HASHORVA, LANPENG JI, AND CHENGXIU LING
(κ)
Abstract. Let {ζm,k (t), t ≥ 0}, κ > 0 be random processes defined as the differences of two independent stationary
n
o
(κ)
chi-type processes with m and k degrees of freedom. In this paper we derive the asymptotics of P supt∈[0,T ] ζm,k (t) > u , u → ∞
under some assumptions on the covariance structures of the underlying Gaussian processes. Further, we establish a
arXiv:1508.02758v2 [math.PR] 15 Jul 2016
Berman sojourn limit theorem and a Gumbel limit result.
1. Introduction
Let X(t) = (X1 (t), . . . , Xm+k (t)), t ≥ 0, m ≥ 1, k ≥ 0 be a vector process with independent components which are
centered stationary Gaussian processes with almost surely (a.s.) continuous sample paths, and unit variance. Set
ri (t) = E {Xi (t)Xi (0)} , t ≥ 0 and suppose that
(1)
α
α
∀t6=0, 1 ≤ i ≤ m + k,
n
o
(κ)
where α ∈ (0, 2] and C := (C1 , . . . , Cm+k ) ∈ (0, ∞)m+k . Define the random processes ζm,k (t), t ≥ 0 , κ > 0 by
(2)
ri (t) = 1 − Ci |t| + o(|t| ),
(κ)
ζm,k (t)
:=
m
X
i=1
!κ/2
Xi2 (t)
−
t→0
m+k
X
and ri (t) < 1,
!κ/2
Xi2 (t)
i=m+1
In this paper we shall investigate the asymptotics of
(
(3)
P
sup
t∈[0,T ]
(κ)
ζm,k (t)
=: |X (1) (t)|κ − |X (2) (t)|κ ,
)
>u ,
t ≥ 0.
u→∞
for given positive T . The study of the asymptotics of (3) is of interest in engineering applications concerned with
the safety of structures; see, e.g., [17–19] and the references therein. More precisely, of interest is the calculation of
the probability that the Gaussian vector process exits a predefined safety region S u ⊂ Rm+k up to the time T , i.e.,
P {X(t) 6∈ S u , for some t ∈ [0, T ]} .
In the aforementioned papers, various types of safety regions S u have been considered for smooth Gaussian vector
processes. Particularly, a safety region given by a ball centered at 0 with radius u > 0
!1/2
)
(
m+k
X
2
m+k
xi
≤u
B u = (x1 , . . . , xm+k ) ∈ R
:
i=1
has been extensively studied; see, e.g., [2, 5, 14, 22]. Referring to [1, 2], we know that for k = 0
(
)
P {X(t) 6∈ B u , for some t ∈ [0, T ]} =
=
sup |X(t)| > u
P
t∈[0,T ]
2
m,0
(C)u α P {|X(0)| > u} (1 + o(1)),
T Hα,1
u → ∞,
m,0
where Hα,1
(C) is a positive constant (see (6) below for a precise definition). Very recently [23] obtained the tail
(2)
asymptotics of the product of two Gaussian processes which has the same tail asymptotic behavior as supt∈[0,T ] ζ1,1 (t).
1991 Mathematics Subject Classification. 60G15, 60G70.
Key words and phrases. Stationary Gaussian process; stationary chi-type process; extremes; Berman sojourn limit theorem; Gumbel
limit theorem; Berman’s condition.
E. Hashorva, L. Ji and C. Ling were partially supported by the Swiss National Science Foundation (SNSF) grant 200021-140633/1,
E. Hashorva is partially supported from SNSF grant 200021-166274, C. Ling also acknowledges Fundamental Research Funds for Central
Universities (XDJK2016C118), and Fundamental Science and Advanced Technology Funds of Chongqing, China (cstc2016jcyjA0036).
1
2
PATRIK ALBIN, ENKELEJD HASHORVA, LANPENG JI, AND CHENGXIU LING
Our first result extends the findings of [2, 23] and suggests an asymptotic approximation for the exit probability of
X from the safety regions S u(κ) defined by
n
o
S u(κ) = (x1 , . . . , xm+k ) ∈ Rm+k : |x(1) |κ − |x(2) |κ ≤ u .
Since chi-type processes appear naturally as limiting processes (see, e.g., [3, 4, 20]), when one considers two independent asymptotic models, the study of the supremum of the difference of the two chi-type processes is also of some
interest in mathematical statistics and its applications.
(κ)
Although for k ≥ 1 the random process ζm,k is not Gaussian and the analysis of the supremum cannot be directly
transformed into the study of the supremum of a related Gaussian random field (which is the case for chi-type
processes; see, e.g., [10, 17–19, 21, 22, 25]), it turns out that it is possible to apply the techniques for dealing with
extremes of stationary processes developed mainly in [2, 5, 6]. In the second part of Section 2 we derive a sojourn
(κ)
(κ)
limit theorem for ζm,k . Further, we show a Gumbel limit theorem for the supremum of ζm,k over an increasing
infinite interval. We refer to [2–4, 15, 22, 26] for results on the Gumbel limit theorem for Gaussian processes and
chi-type processes.
Brief outline of the paper: our main results are stated in Section 2. In Section 3 we present proofs of Theorem 2.1,
Theorem 2.2 and Theorem 2.3 followed then by an appendix.
2. Main Results
We first introduce some notation. Let {Z(t), t ≥ 0} be a standard fractional Brownian motion (fBm) with Hurst
index α/2 ∈ (0, 1], i.e., it is a centered Gaussian process with a.s. continuous sample paths and covariance function
1 α
α
s + tα − |s − t| , s, t ≥ 0.
Cov(Z(s), Z(t)) =
2
In the following, let {Zi (t), t ≥ 0}, 1 ≤ i ≤ m + k be independent copies of Z and define Wκ to be a Gamma
distributed random variable with parameter (k/κ, 1). Further let O 1 = (O1 , . . . , Om ), O 2 = (Om+1 , . . . , Om+k )
denote two random vectors uniformly distributed on the unit sphere of Rm and Rk , respectively. Hereafter we shall
suppose that O 1 , O 2 , Wκ and Zi ’s are mutually independent. Define for m ≥ 1, k ≥ 0, κ > 0
(κ)
e(κ) (t) + E, t ≥ 0,
ηm,k (t) = Z
m,k
(4)
where E is a unit mean exponential random variable being independent of all the other random elements involved,
and (recall C = (C1 , . . . , Cm+k ) given in (1))
Pm √
Pm
2 α
L
(t)
=
t ,
C
O
2C
O
Z
(t)
−
1
i
i
i
i
i
i=1
i=1
κ > 1,
L1 (t),
Pm+k √
2/κ
1/κ
L2 (t) = Wκ − Wκ + 2(Wκ /κ)
e(κ) (t) =
i=m+1 2Ci Oi Zi (t)
(5)Z
L1 (t) + L2 (t), κ = 1, with
m,k
κ/2
P
L2 (t),
κ < 1,
+2κ−2/κ m+k Ci Zi2 (t)
,
i=m+1
Pm
Here we set i=m+1 ci =: 0.
We state next our main result.
(κ)
Theorem 2.1. If {ζm,k (t), t ≥ 0} is given by (2) with the involved Gaussian processes Xi ’s satisfying (1), then, for
any T > 0
(
)
n
o
2τ
(κ)
(κ)
m,k
P sup ζm,k (t) > u = T Hα,κ
(C)u ακ P ζm,k (0) > u (1 + o(1))
t∈[0,T ]
(κ)
holds as u → ∞, where τ = 2/κ − 1 for κ ∈ (0, 1), and 1 otherwise, with ηm,k given by (4),
1
(κ)
m,k
(6)
Hα,κ (C) = lim P sup ηm,k (aj) ≤ 0 ∈ (0, ∞).
a↓0 a
j≥1
(κ)
Remarks: a) The tail asymptotics of the Gaussian chaos ζm,k (0) is discussed in Lemma 3.1 below.
b) The most obvious choice of κ is 1, which corresponds to the difference of L2 -norm of two independent multivariate
Gaussian processes. For the case κ = 2 and m = k = 1 the problem was (implicitly) investigated by considering the
product of two independent Gaussian processes in the recent contribution [23].
EXTREMES AND LIMIT THEOREMS FOR DIFFERENCE OF CHI-TYPE PROCESSES
3
(κ)
d
c) Since O 1 is uniformly distributed on the unit sphere of Rm , we have, for κ > 1 and C = 1, that ηm,k (t) =
√
m,k
2Z(t) − tα + E. In such a case, the constant Hα,κ
(1) coincides with the classical Pickands constant Hα ; see, e.g.,
[22]. Approximation of Pickands constant Hα has been considered by a number of authors; see the recent contribution
m,k
(C) seems to
[8] which gives some simulation algorithms. Precise estimation of the general Pickands constant Hα,κ
(κ)
be hard to find, due to the complexity of the process ηm,k .
d) We see from Theorem 2.1 and Lemma 3.1 that, if κ > 2, then, for any m, k ≥ 1
(
)
(
)
P
(κ)
sup ζm,k (t) > u
(κ)
=P
t∈[0,T ]
sup ζm,0 (t) > u (1 + o(1))
t∈[0,T ]
(κ)
holds as u → ∞, which means that Xm+1 , . . . , Xm+k do not influence the tail asymptotic of supt∈[0,T ] ζm,k (t). This
(κ)
is not so surprising as the tail asymptotic behavior of ζm,k (0) is subexponential.
(κ)
Next, we consider the sojourn time of ζm,k above a threshold u > 0 in the time interval [0, t] defined by
(κ)
Lm,k,t (u)
=
Z
0
t
(κ)
I{ζm,k (s) > u} ds,
t > 0,
where I{·} stands for the indicator function. Our second result below establishes a Berman sojourn limit theorem
(κ)
for ζm,k . See [5] for related discussions on sojourn times of Gaussian processes and related processes.
Theorem 2.2. Under the assumptions and notation of Theorem 2.1, we have, for any t > 0
Z ∞ n
o
n
o
2τ
2τ
(κ)
(κ)
P u ακ Lm,k,t (u) > y dy = u ακ E Lm,k,t (u) Υκ (x)(1 + o(1))
x
holds as u → ∞ for all continuity point x > 0 of Υκ (x) := P
nR
∞
0
o
(κ)
I{ηm,k (s) > 0} ds > x .
(κ)
In the following, we derive a Gumbel limit theorem for supt∈[0,T ] ζm,k (t) under a linear normalization, which is also
of interest in extreme value analysis and statistical tests. We refer to [4, 6, 13, 15] for its applications in deriving
approximations of the critical values of the proposed test statistics.
Theorem 2.3. Under the assumptions and notation of Theorem 2.1, if further
2/κ − 1,
(7)
lim max |rl (t)|(ln t)c = 0, with c =
1,
t→∞ 1≤l≤m+k
k + 1 − 2k/κ,
the following Berman-type condition
0 < κ < 1,
1 ≤ κ ≤ 2,
κ>2
holds, then
lim sup P
T →∞ x∈R
(
(κ)
aT
sup
t∈[0,T ]
(κ)
ζm,k (t)
−
(κ)
bT
where, for all T large
(8)
(κ)
aT =
(2 ln T )1−κ/2
,
κ
(κ)
bT = (2 ln T )κ/2 +
)
≤x
− exp −e−x
= 0,
κ
(K0 ln ln T + ln D0 ) ,
2(2 ln T )1−κ/2
with (below Γ(·) denotes the Euler Gamma function)
2 2
k
( −1)+2(1− κ
) Γ k κ(k/κ−1) 2 ,
2α κ
κ
!2
2 α2 +2(1− κk ) Γ k κ(k/κ−1) 2 ,
m,k
Hα,κ
(C)
×
D0 =
κ
2
k 2
Γ(m/2)Γ(k/2)
α −2 Γ
2
,
2
α2
k 2
,
2 Γ 2
m − 2 + (2/α)(2/κ − 1) + k(1 − 2/κ), 0 < κ ≤ 1,
K0 =
m − 2 + 2/α + k(1 − 2/κ),
1 < κ < 2,
m − 2 + 2/α,
κ ≥ 2.
0 < κ ≤ 1,
1 < κ < 2,
κ = 2,
κ>2
4
PATRIK ALBIN, ENKELEJD HASHORVA, LANPENG JI, AND CHENGXIU LING
p
Under the assumptions of Theorem 2.3, we have the following convergence in probability (denoted by →)
(κ)
supt∈[0,T ] ζm,k (t)
(2 ln T )κ/2
p
→ 1,
T → ∞,
(κ)
(κ)
which follows from the fact that limT →∞ bT /(2 ln T )κ/2 = 1 and that aT is bounded away from zero, together
with elementary considerations. In several cases such a convergence in probability can be strengthened to the pth
mean convergence which is referred to as the Seleznjev pth mean convergence since the idea was first suggested by
Seleznjev in [24], see also [12]. In order to show the Seleznjev pth mean convergence of crucial importance is the
Piterbarg inequality (see [22], Theorem 8.1). Since the Piterbarg inequality holds also for chi-square processes (see
[25], Proposition 3.2), using further the fact that
(κ)
ζm,k (t) ≤ |X (1) (t)|κ ,
t ≥ 0,
we immediately get the Piterbarg inequality for the difference of chi-type processes by simply applying the aforementioned proposition. Specifically, under the assumptions of Theorem 2.3 for any T > 0 and all large u
(
)
1 2/κ
(κ)
β
P sup ζm,k (t) > u ≤ KT u exp − u
,
2
t∈[0,T ]
where K and β are two positive constants not depending on T and u. Note that the above result also follows
immediately from Theorem 2.1 combined with Lemma 3.1 below. Hence utilizing Lemma 4.5 in [25] we arrive at
our last result.
Corollary 2.4. (Seleznjev pth mean theorem) Under the assumptions of Theorem 2.3, we have, for any p > 0
!p
(κ)
sup
ζ
(t)
t∈[0,T ] m,k
lim E
= 1.
T →∞
(2 ln T )κ/2
3. Further Results and Proofs
Before presenting the proof of Theorem 2.1 we first give some preliminary lemmas. Hereafter we use the same
d
d
notation and assumptions as in Section 1. By → and = we shall denote the convergence in distribution (or the
convergence of finite dimensional distributions if both sides of it are random processes) and equality in distribution
function, respectively. Further, we write fξ (·) for the pdf of a random variable ξ and write h1 ∼ h2 if two functions
hi (·), i = 1, 2 are such that h1 /h2 goes to 1 as the argument tends to some limit. For simplicity we shall denote,
with κ > 0 and τ = max(2/κ − 1, 1),
1 2/κ−1
u
, u > 0.
κ
= u + x/wκ (u) for all u, x > 0.
qκ = qκ (u) = u−2τ /(ακ) ,
In the proofs of Lemmas 3.1–3.3, we denote uκ,x
wκ (u) =
(κ)
Lemma 3.1. Let {ζm,k (t), t ≥ 0} be given by (2). For all integers m ≥ 1, k ≥ 0 we have as u → ∞
Γ(k/κ)
κ < 2,
(wκ (u))k/κ ,
n
o fζ (κ) (0) (u)
m/κ−1
2−(m+k)/2
u
2
1
(κ)
∼ 2
exp − u2/κ ×
P ζm,k (0) > u ∼ m,k
Γ(k/2),
κ = 2,
wκ (u)
κ Γ(k/2)Γ(m/2) wκ (u)
2
κ2k/2−1 Γ(k/2), κ > 2,
where Γ(k/κ)/Γ(k/2) := 1 for k = 0 and all κ > 0.
Proof. The claim follows from Theorem 1 in [11].
(κ)
Lemma 3.2. If {ζm,k (t), t ≥ 0} is as in Theorem 2.1, then
(κ)
n
o
n
o
d
(κ)
(κ)
(κ)
wκ (u)(ζm,k (qκ t) − u)|{ζm,k (0) > u}, t ≥ 0 → ηm,k (t), t ≥ 0 ,
d
u → ∞,
with ηm,k given by (4). Recall that → stands for the convergence of finite dimensional distributions.
EXTREMES AND LIMIT THEOREMS FOR DIFFERENCE OF CHI-TYPE PROCESSES
5
Proof. We henceforth adopt the notation introduced in Section 2. By Lemma 3.1, we have
(κ)
d
(κ)
wκ (u)(ζm,k (0) − u) {ζm,k (0) > u} → E,
u → ∞.
Thus, in view of Theorem 5.1 in [5], it suffices to show that, for any 0 < t1 < · · · < tn < ∞, n ∈ N
o
n
(κ)
(κ)
pk (u) := P ∩nj=1 {ζm,k (qκ tj ) ≤ uκ,zj } ζm,k (0) = uκ,x
n
o
e(κ) (tj )+x ≤ zj } , u → ∞
(9)
→ P ∩nj=1 {Z
m,k
holds for all x > 0 and zj ∈ R, 1 ≤ j ≤ n. Define below
∆iu (tj ) = Xi (qκ tj ) − ri (qκ tj )Xi (0),
1 ≤ i ≤ m + k, 1 ≤ j ≤ n.
By (1) we have
α
u2τ /κ Cov(∆iu (s), ∆iu (t)) → Ci (sα + tα − |s − t| )
= 2Ci Cov(Zi (s), Zi (t)),
Therefore,
u → ∞, s, t > 0, 1 ≤ i ≤ m + k.
p
d
{uτ /κ∆iu (t), t ≥ 0} → { 2Ci Zi (t), t ≥ 0}, u → ∞, 1 ≤ i ≤ m + k.
Furthermore, by the independence of ∆iu (t)’s and Xi (0)’s, the random processes Zi ’s can be chosen such that they
d
(κ)
are independent of ζm,k (0). Note that X (1) (0) = R1 O 1 holds for some R1 > 0 which is independent of O 1 . Then,
using the Taylor’s expansion of (1 + x)κ/2 = 1 + κx/2 + o(x), x → 0 for any zj ∈ R, 1 ≤ j ≤ n we have
n n
\
o
|X (1) (qκ tj )|κ ≤ uκ,zj |X (1) (0)|κ = uκ,x
p0 (u) = P
j=1
)
!
)
(
κ/2
n
\
1
− R1κ ≤ zj − x R1κ = uκ,x
wκ (u) R1κ 1 + 2 Vu (tj )
=P
R
1
j=1
)
n n
\
o
κ
κ−2
κ
=P
wκ (u)R1 Vu (tj )(1 + op (1)) ≤ zj − x R1 = uκ,x
2
j=1
)
(m √
!
m
n
\
X 2Ci Oi Zi (tj )
X
Ci Oi2
α
(10)
, u → ∞,
t
(1
+
o
(1))
+
x
≤
z
(1
+
o
(1))
−
=P
p
j
p
u(τ −1)/κ
u2(τ −1)/κ j
j=1
i=1
i=1
Pm
Pm
Pm
where Vu (tj ) := i=1 ∆2iu (tj ) + 2 i=1 ∆iu (tj )ri (qκ tj )Xi (0) − i=1 (1 − ri2 (qκ tj ))Xi2 (0). Consequently, the claim
for k = 0 follows. Next, for k ≥ 1, we rewrite pk (u) as (recall that uκ,x = u + x/wκ (u))
Z ∞ \
n n
o
y
(κ)
ζm,k (qκ tj ) ≤ uκ,zj |X (1) (0)|κ = uκ,x+y , |X (2) (0)|κ =
pk (u) =
P
wκ (u)
0
j=1
×
f|X (1) (0)|κ (uκ,x+y )f|X (2) (0)|κ (y/wκ (u))
wκ (u)fζ (κ) (0) (uκ,x )
(11)
=
Z
0
where
dy
m,k
∞
n n
\
o
y
|X (1) (qκ tj )|κ ≤ uκ,zj +wκ (u)·|X (2) (qκ tj )|κ |X (1) (0)|κ = uκ,x+y , |X (2) (0)|κ =
P
× hκ,u (y) dy,
wκ (u)
j=1
hκ,u (y) :=
f|X (1) (0)|κ (uκ,x+y )f|X (2) (0)|κ (y/wκ (u))
wκ (u)fζ (κ) (0) (uκ,x )
m,k
=
(u) f (1) κ (uκ,y )f (2) κ (y/wκ (u))
(κ)
f|X (1) (0)|κ (uκ,x+y ) fζm,k
(0)
|X (0)|
|X (0)|
f|X (1) (0)|κ (uκ,y ) fζ (κ) (0) (uκ,x )
m,k
wκ (u)fζ (κ) (0) (u)
m,k
6
PATRIK ALBIN, ENKELEJD HASHORVA, LANPENG JI, AND CHENGXIU LING
f|X (1) (0)|κ (uκ,y )f|X (2) (0)|κ (y/wκ (u))
R∞
,
0 f|X (1) (0)|κ (uκ,y )f|X (2) (0)|κ (y/wκ (u)) dy
∼
(12)
u → ∞.
Here the last step follows by Lemma 3.1.
d
Noting that X (2) (0) = R2 O 2 holds for some R2 > 0 which is independent of O 2 , we have by similar arguments as
in (10) that, for any t ≥ 0
2/κ
{wκ (u)|X (2) (0)|κ = y}
wκ (u)|X (2) (qκ t)|κ
m+k
X
= (wκ (u))2/κ
Xi2 (0) + 2
−
ri (qκ t)Xi (0)∆iu (t) +
!
(1 − ri (qκ t)2 )Xi2 (0)
R2κ =
i=m+1
m+k
X
∆2iu (t)
i=m+1
i=m+1
i=m+1
m+k
X
m+k
X
y
(wκ (u))1/κ
m+k
m+k
X
R2 X p
2
2C
Ci Zi2 (t)(1 + op (1))
O
Z
(t)(1
+
o
(1))
+
i
i
i
p
uτ /κ i=m+1
u2τ /κ i=m+1
!
2 m+k
X
R2
y
2 α
κ
−2
Ci Oi t (1 + op (1))
R2 =
uτ /κ
(wκ (u))1/κ
i=m+1
= (wκ (u))2/κ R22 + 2
=y
(13)
2/κ
+ 2y
1/κ
wκ (u)
uτ
=: θκ,u (y, t).
1/κ m+k
2/κ m+k
X p
X
wκ (u)
2Ci Oi Zi (t)(1 + op (1)) + 2
Ci Zi2 (t)(1 + op (1))
τ
u
i=m+1
i=m+1
This together with (10) and (11) implies that
(m √
!
Z ∞ \
n
m
X 2Ci Oi Zi (tj )
X
Ci Oi2
pk (u) =
P
tα
(1 + op (1)) −
j (1 + op (1)) + x + y
(τ −1)/κ
2(τ −1)/κ
u
u
0
i=1
j=1
i=1
)
)
y
(2)
(2)
κ
κ
≤ zj + wκ (u)|X (qκ tj )|
|X (0)| =
hκ,u (y) dy
wκ (u)
(m √
!
Z ∞ \
m
n
X 2Ci Oi Zi (tj )
X
Ci Oi2
tα
(1 + op (1)) −
=
P
j (1 + op (1)) + x + y
(τ −1)/κ
2(τ −1)/κ
u
u
0
i=1
j=1
i=1
oo
κ
≤ zj + (θκ,u (y, tj )) 2
(14)
hκ,u (y) dy.
Recalling that τ = max(2/κ − 1, 1) and wκ (u) = (1/κ)u2/κ−1 , we have by (13) that (θκ,u (y, tj ))κ/2 = y + op (1) for
κ > 1. While for κ ∈ (0, 1], it follows by (12) and Lemma 3.1 that,
(15)
hκ,∞ (y) := lim hκ,u (y) =
u→∞
1
y k/κ−1 e−y ,
Γ(k/κ)
y > 0,
which is the pdf of a Gamma distributed random variable with parameter (k/κ, 1). Hence, combining (12)–(15) and
e(κ) (t), the claim in (9) follows. Consequently, the proof of Lemma 3.2 is complete.
(5) for the definition of Z
m,k
The next lemma corresponds to Condition B in [2]; see also [1, 3]. As shown in Chapter 5 in [1] this condition is
crucial in ensuring that the double sum part is asymptotically negligible with respect to the principal sum. Denote
in the following by [x] the integer part of x ∈ R.
(κ)
Lemma 3.3. If {ζm,k (t), t ≥ 0} is as in Theorem 2.1, then for any T, a > 0
[T /(aqκ )]
lim sup
u→∞
X
j=N
n
o
(κ)
(κ)
P ζm,k (aqκ j) > u ζm,k (0) > u → 0,
N → ∞.
Proof. Note first that the case k = 0 is treated in [2], p.119. Using the fact that the standard bivariate Gaussian
distribution is exchangeable for u > 0 we have
n
o
n
o
(κ)
(κ)
(κ)
(κ)
P ζm,k (qκ t) > u ζm,k (0) > u = 2P ζm,k (qκ t) > u, |X (1) (qκ t)| > |X (1) (0)| ζm,k (0) > u =: 2Θ(u).
EXTREMES AND LIMIT THEOREMS FOR DIFFERENCE OF CHI-TYPE PROCESSES
7
Further, it follows from Lemma 3.1 that, for any k ≥ 1
Z ∞Z ∞
y
(κ)
(1)
(1)
(1)
(2)
κ
κ
Θ(u) =
P ζm,k (qκ t) > u, |X (qκ t)| > |X (0)| |X (0)| = uκ,x+y , |X (0)| =
wκ (u)
0
0
f|X (1) (0)|κ (uκ,x+y )f|X (2) (0)|κ wκy(u)
n
o
×
dxdy
(κ)
wκ2 (u)P ζm,k (0) > u
Z ∞Z ∞ n
o f|X (1) (0)|κ (uκ,x+y ) f|X (2) (0)|κ w y(u)
κ
n
o
dxdy
P |X (1) (qκ t)|κ > uκ,y |X (1) (0)|κ = uκ,x+y
≤
(κ)
2
0
0
wκ (u)P ζm,k (0) > u
o
n
Z ∞ n
o P |X (1) (0)|κ > uκ,y
y
n
o f|X (2) (0)|κ
=
P |X (1) (qκ t)|κ > uκ,y |X (1) (0)|κ > uκ,y
dy.
(κ)
wκ (u)
0
wκ (u)P ζ (0) > u
m,k
Moreover, in view of the treatment of the case k = 0 in [2], p.119 we readily see that, for any p ≥ 1, with
R(t) := max1≤i≤m ri (t), r(t) := min1≤i≤m ri (t) and Φ(·) denoting the N (0, 1) distribution function
!!
o
n
(1 − R(qκ t))u1/κ
(1)
(1)
κ
κ
≤ 4m 1 − Φ p
P |X (qκ t)| > uκ,y |X (0)| > uκ,y
m(1 − r2 (qκ t))
≤ Kp t−αp/2 ,
∀qκ t ∈ (0, T ]
holds for some Kp > 0 not depending on u, t and y. Consequently,
o
n
Z ∞ P |X (1) (0)|κ > uκ,y
n
o
y
(κ)
(κ)
−αp/2
n
o f|X (2) (0)|κ
dy
P ζm,k (qκ t) > u ζm,k (0) > u
≤ 2Kp t
(κ)
wκ (u)
0
wκ (u)P ζm,k (0) > u
(16)
2Kp t−αp/2 ,
=
∀qκ t ∈ (0, T ].
Therefore, with p = 4/α
[T /(aqκ )]
lim sup
u→∞
X
j=N
≤ 2Kp
Z
n
o
(κ)
(κ)
P ζm,k (aqκ j) > u ζm,k (0) > u
∞
aN
x−2 dx =
2Kp
→ 0,
aN
N →∞
establishing the proof.
The lemma below concerns the accuracy of the discrete approximation to the continuous process, which is related
to Condition C in [2]. As shown in [3] (see Eq. (7) therein), in order to verify Condition C the following lemma is
sufficient. Its proof is relegated to the appendix.
(κ)
Lemma 3.4. If {ζm,k (t), t ≥ 0} is as in Theorem 2.1, then there exist some constants C, p > 0, d > 1 and λ0 , u0 > 0
such that
n
o
λ
(κ)
(κ)
(κ)
P ζm,k (qκ t) > u +
, ζm,k (0) ≤ u ≤ Ctd λ−p P ζm,k (0) > u
wκ (u)
for 0 < t̟ < λ < λ0 and u > u0 . Here ̟ is α/2 for κ ≥ 1, and (α/2) min(κ/(4(1 − κ)), 1) otherwise.
Proof of Theorem 2.1: It follows from Lemmas 3.1–3.4 that all the assumptions of Theorem 1 in [2] are satisfied
(κ)
by the process ζm,k , which immediately establishes the proof.
2τ /(ακ)
Proof of Theorem 2.2: In view of (16) with p = 4/α and letting vκ = vκ (u) = 1/qκ (u) = u
, we obtain
Z T
Z
vκ T
n
o
n
o
(κ)
(κ)
(κ)
(κ)
vκ
P ζm,k (s) > u ζm,k (0) > u ds =
P ζm,k (s/vκ ) > u ζm,k (0) > u ds
N/vκ
N
≤ K4/α
Z
vκ T
N
s−2 ds ≤
K4/α
,
N
u → ∞.
8
PATRIK ALBIN, ENKELEJD HASHORVA, LANPENG JI, AND CHENGXIU LING
Hence,
lim lim sup vκ
N →∞ u→∞
Z
T
N/vκ
n
o
(κ)
(κ)
P ζm,k (s) > u ζm,k (0) > u ds = 0.
Since further Lemma 3.2 holds, the claim follows by Theorem 3.1 in [5].
(κ)
As shown by Theorem 10 in [2], in order to derive the Gumbel limit theorem for the random process ζm,k two
additional conditions, which were first addressed by the seminal contributions [15, 16], need to be checked, namely
the mixing Condition D and the Condition D′ therein. These two conditions will follow from Lemma 3.5 and Lemma
3.6 below; their proofs are displayed in the appendix.
(κ)
Lemma 3.5. Let T and a be any given positive constants and M ∈ (0, T ). If {ζm,k (t), t ≥ 0} is as in Theorem 2.1,
then for any 0 ≤ s1 < · · · < sp < t1 < · · · < tp′ in {aqκ j : j ∈ Z, 0 ≤ aqκ j ≤ T } such that t1 − sp ≥ M
n
o
n
o n ′
o
′
(κ)
(κ)
(κ)
(κ)
P ∩pi=1 {ζm,k (si ) ≤ u}, ∩pj=1 {ζm,k (tj ) ≤ u} − P ∩pi=1 {ζm,k (si ) ≤ u} P ∩pj=1 {ζm,k (tj ) ≤ u}
X
u2/κ
(17)
≤ Kuς
re(tj − si ) exp −
1 + re(tj − si )
′
1≤i≤p,1≤j≤p
and
(18)
n
o
n
o n ′
o
′
(κ)
(κ)
(κ)
(κ)
P ∩pi=1 {ζm,k (si ) > u}, ∩pj=1 {ζm,k (tj ) > u} − P ∩pi=1 {ζm,k (si ) > u} P ∩pj=1 {ζm,k (tj ) > u}
X
u2/κ
≤ Kuς
re(tj − si ) exp −
1 + re(tj − si )
′
1≤i≤p,1≤j≤p
hold for all u > 0 and some K > 0 not depending on u. Here ς = 2/κ m − k(2/κ − 1) − 1 + max(0, 2(1/κ − 1) and
re(t) := max1≤l≤m+k |rl (t)|, t > 0.
Lemma 3.6. Under the assumptions of Theorem 2.3, for ς, re(·) as in Lemma 3.5 and Tκ given by
(19)
Tκ = Tκ (u) =
1
m,k
Hα,κ
(C)
we have, for any given constant ε ∈ (0, Tκ )
(20)
uς
Tκ
qκ
X
ε≤aqκ j≤Tκ
re(aqκ j) exp −
q (u)
n κ
o,
(κ)
P ζm,k (0) > u
u2/κ
1 + re(aqκ j)
→ 0,
u → ∞.
Proof of Theorem 2.3: To establish Conditions D and D′ in [2], we shall make use of Lemma 3.5 with T = Tκ
given by (19) and M = ε ∈ (0, Tκ ), and Lemma 3.6. First note that the right-hand side of (17) is bounded from
above by
X
Tκ
u2/κ
Kuς
,
re(aqκ j) exp −
aqκ
1 + re(aqκ j)
ε≤aqκ j≤Tκ
(κ)
which by an application of (20) implies that the mixing Condition D in [2] holds for the random process ζm,k .
Next, we prove Condition D′ in [2], i.e., for any given positive constants a and Te
h
n
oi
(κ)
ε/P ζm,k (0)>u
(21)
lim sup
u→∞
X
j=[Te/(aqκ )]
n
o
(κ)
(κ)
P ζm,k (aqκ j) > u ζm,k (0) > u → 0,
ε ↓ 0.
f > Te and a positive constant K
Indeed, by (18) for some M
n
o
n
o
u2/κ
(κ)
(κ)
(κ)
ς Tκ
P ζm,k (aqκ j) > u ζm,k (0) > u ≤ P ζm,k (0) > u + Ku
re(aqκ j) exp −
qκ
1 + re(aqκ j)
f. Consequently,
holds for u > 0 and aqκ j > M
h
n
oi
(κ)
ε/P ζm,k (0)>u
lim sup
u→∞
X
j=[Te/(aqκ )]
n
o
(κ)
(κ)
P ζm,k (aqκ j) > u ζm,k (0) > u
EXTREMES AND LIMIT THEOREMS FOR DIFFERENCE OF CHI-TYPE PROCESSES
≤ lim sup
u→∞
f/(aqκ )]
[M
X
n
o
(κ)
(κ)
P ζm,k (aqκ j) > u ζm,k (0) > u + ε
j=[Te/(aqκ )]
h
Tκ
+ lim sup Ku
qκ
u→∞
ς
9
n
oi
(κ)
ε/P ζm,k (0)>u
X
f/(aqκ )]
j=[M
u2/κ
re(aqκ j) exp −
1 + re(aqκ j)
,
which equals ǫ by an application of Lemma 3.3 and (20), respectively. It follows then that (21) holds. Consequently,
in view of Theorem 10 in [2] we have, for Tκ given by (19)
(
)
x
(κ)
lim P
sup ζm,k (t) ≤ u +
= exp −e−x , x ∈ R.
u→∞
w
(u)
κ
t∈[0,Tκ ]
(κ)
(κ)
Expressing u in terms of Tκ using (19) (see also (32)) we obtain the required claim with aT , bT given by (8) for
any x ∈ R; the uniform convergence in x follows since all functions (with respect to x) are continuous, bounded and
increasing.
4. Appendix
Proof of Lemma 3.4: By (1), for any small ǫ ∈ (0, 1) there exists some positive constant B such that
1
and 1 − ri (t) ≤ Btα , ∀t ∈ (0, ǫ], 1 ≤ i ≤ m + k.
2
Furthermore, for any positive t satisfying (recall ̟ = α/2I{κ ≥ 1} + α/2 min(κ/(4(1 − κ)), 1)I{0 < κ < 1})
κ
1
̟
,
ǫ
,
0 < t̟ < λ < λ0 := min
2κ+4 B 2κ+2
ri (t) ≥
and any u > 2
κtα/2
16
1
− 1, r(t) := min ri (t).
1≤i≤m+k
(r(qκ t))κ
(2)
(1)
−1
−1
Let (X 1/r (t), X 1/r (t)) := X1 (t) − r1 (t)X1 (0), . . . , Xm+k (t) − rm+k (t)Xm+k (0) which by definition is independent
u2τ /κ θκ (t) ≤ 2κ κBtα ≤
(22)
with θκ (t) :=
(κ)
of {ζm,k (t), t ≥ 0}. For j = 1, 2
o
n
λ
x
(j)
(j)
(23)
.
, uθκ (t) ≤
P |X 1/r (qκ t)| > x ≤ P |X (0)| > √
−2τ
/κ
α
2wκ (u)
2 2Bu
t
In the following, the cases κ = 1, κ ∈ (1, ∞) and κ ∈ (0, 1) will be considered in turn.
Case κ = 1: Note by the triangular inequality
1
(2)
(1)
(1)
(1)
ζm,k (q1 t) ≤ |X 1/r (q1 t)| + |X 1/r (q1 t)| +
ζ (0) + θ1 (t)|X (2) (0)|.
r(q1 t) m,k
Consequently, from (23) we get
λ (1)
(1)
P ζm,k (q1 t) > u + , ζm,k (0) ≤ u
u
λ (1)
(2)
(1)
(2)
, ζ (q1 t) > u
≤ P |X 1/r (q1 t)| + |X 1/r (q1 t)| + θ1 (t)|X (0)| >
2u m,k
n
o
λ
λ
(2)
(1)
(1)
(2)
≤ P |X 1/r (q1 t)| + |X 1/r (q1 t)| >
P ζm,k (q1 t) > u + P θ1 (t)|X (0)| >
3u
6u
=: I1u + I2u .
By (22) and (23), we have, for any p > 1
−p
λ
λ
λ
(1)
≤ P |X (1) (0)| > √
P |X 1/r (q1 t)| >
≤ K α/2
6u
t
12 2Btα/2
holds with some K > 0 (the values of p and K might change from line to line below). Similarly,
−p
λ
λ
(2)
≤ K α/2
P |X 1/r (q1 t)| >
6u
t
10
PATRIK ALBIN, ENKELEJD HASHORVA, LANPENG JI, AND CHENGXIU LING
and hence
(24)
I1u ≤ K
λ
tα/2
−p
n
o
(1)
P ζm,k (0) > u .
Moreover, in view of Lemma 3.1 and (22) we have for sufficiently large u that
o
n
−(p−k+2)
n
o
n
o
P |X (2) (0)| > t2λu
α/2
λ
(1)
(1)
n
o
(25)
u−(p+m−2k) P ζm,k (0) > u .
I2u ≤
P ζm,k (0) > u ≤ K α/2
(1)
t
P ζm,k (0) > u
Hence, the claim for κ = 1 follows from (24) and (25) by choosing p > max(4/α + k, 2k).
−1
Case κ ∈ (1, ∞): Denote below by (Y (1) (t), Y (2) (t)) := r1−1 (t)X1 (0), . . . , rm+k
(t)Xm+k (0) . Note that |Y (1) (t)| ≤
|X (1) (0)|/r(t) and |X (2) (0)| ≤ |Y (2) (t)| ≤ |X (2) (0)|/r(t) for all t < ε, and for some constants K1 , K2 > 0 whose
values might change from line to line below
|1 + x|κ ≥ 1 + κx,
x∈R
and (1 + x)κ ≤ 1 + K1 x + K2 xκ ,
x ≥ 0.
We have further by the triangle inequality
κ
(κ)
(1)
(2)
ζm,k (qκ t) ≤
|Y (1) (qκ t)| + |X 1/r (qκ t)| − |Y (2) (qκ t)| − |X 1/r (qκ t)|
≤
κ
(1)
(1)
|Y (1) (qκ t)|κ + K1 |X 1/r (qκ t)||Y (1) (qκ t)|κ−1 + K2 |X 1/r (qκ t)|κ
(2)
−|Y (2) (qκ t)|κ + κ|X 1/r (qκ t)||Y (2) (qκ t)|κ−1
≤
(1)
(1)
K1 |X 1/r (qκ t)||X (1) (0)|κ−1 + K2 |X 1/r (qκ t)|κ
(κ)
(2)
+K3 |X 1/r (qκ t)||X (2) (0)|κ−1
+
ζm,k (0)
(r(qκ t))κ
+ θκ (t)|X (2) (0)|κ
holds for qκ t ≤ ǫ and some constant K3 > 0. Therefore, with µ = 1/(2(κ − 1)) and ϕ = α/(4(κ − 1)),
λ
(κ)
(κ)
P ζm,k (qκ t) > u +
, ζ (0) ≤ u
wκ (u) m,k
λµ u1/κ
λµ u1/κ
(2)
+
P
|X
(0)|
>
≤ P |X (1) (0)| >
tϕ
tϕ
(
µ 1/κ κ−1
µ 1/κ κ−1
λ u
λ u
(1)
(2)
(1)
κ
+P K1 |X 1/r (qκ t)|
(q
t)|
(q
t)|
+
K
|X
+
K
|X
κ
κ
3
2
1/r
1/r
ϕ
t
tϕ
λ
(κ)
+θκ (t)|X (2) (0)|κ ≥
, ζm,k (qκ t) > u
2wκ (u)
˜
˜
˜
(26)
=: I1u + I2u + I3u .
Note by (22) that λµ /tϕ > 1. Similar arguments as in (25) yield that
n
o
µ −(p−m+2)
(κ)
u−(p−k(2/κ−1)I{κ≤2})/κ P ζm,k (0) > u
I˜1u ≤ K λtϕ
n
o
µ −(p−k+2)
(κ)
I˜2u ≤ K λtϕ
u−(p−k+m−k(2/κ−1)I{κ≤2})/κ P ζm,k (0) > u
and
I˜3u
≤
P
+P
(
(1)
K1 |X 1/r (qκ t)|
(
(2)
K3 |X 1/r (qκ t)|
λµ u1/κ
tϕ
λµ u1/κ
tϕ
κ−1
κ−1
λ
>
8wκ (u)
λ
>
8wκ (u)
λ
(2)
κ
+P θκ (t)|X (0)| >
8wκ (u)
n
o
(κ)
=: (II1u + II2u + II3u )P ζm,k (0) > u + II4u .
)
(1)
+ P K2 |X 1/r (qκ t)|κ >
)!
n
o
(κ)
P ζm,k (qκ t) > u
λ
8wκ (u)
EXTREMES AND LIMIT THEOREMS FOR DIFFERENCE OF CHI-TYPE PROCESSES
11
Furthermore,
II1u
≤
≤
P |X (1) (0)| > K1
λ1/2 u−1/κ
t−α/4 (r−2 (qκ t) − 1)1/2
−p/2
λ
λ1/2
.
P |X (1) (0)| > K1 α/4 ≤ K α/2
t
t
Similarly,
II2u ≤ K
λu2(1−1/κ)
tακ/2
−p/κ
,
II3u ≤ K
λ
tα/2
−p/2
.
Next, we deal with II4u . We have by (22) that 2κ+4 Btα/2 ≤ 1. Hence as in the proof of (25), we have
1
2λu
II4u ≤ P |X (2) (0)|κ > α/2 κ+4 α/2
t
2
Bt
−(p−k+2)
n
o
λ
(κ)
≤ K α/2
(27)
u−(p−k+m−k(2/κ−1)I{κ≤2})/κ P ζm,k (0) > u .
t
Therefore, the claim for κ ∈ (1, ∞) follows from (26) and the inequalities for I˜1u , I˜2u and II1u – II4u by choosing
p > max(8(κ − 1)/α + k + m, 2k).
Case κ ∈ (0, 1): Note that
(1 + x)κ ≤ 1 + x,
x≥0
and
− |1 − x|κ ≤ −(1 − x),
x ∈ [0, ∞).
We have further by the triangle inequality
κ
(κ)
(1)
(2)
ζm,k (qκ t) ≤
|Y (1) (qκ t)| + |X 1/r (qκ t)| − |Y (2) (qκ t)| − |X 1/r (qκ t)|
(1)
κ
(2)
≤ |Y (1) (qκ t)|κ + |X 1/r (qκ t)||X (1) (0)|κ−1 − |Y (2) (qκ t)|κ + |X 1/r (qκ t)||Y (2) (qκ t)|κ−1
≤
|X (1) (0)|κ
(2)
(1)
+ |X 1/r (qκ t)||X (1) (0)|κ−1 − |X (2) (0)|κ + |X 1/r (qκ t)||X (2) (0)|κ−1
(r(qκ t))κ
(κ)
=
ζm,k (0)
(r(qκ t))κ
(1)
(2)
+ θκ (t)|X (2) (0)|κ + |X 1/r (qκ t)||X (1) (0)|κ−1 + |X 1/r (qκ t)||X (2) (0)|κ−1 .
Therefore, we have by (23), with ψ = α/(4(1 − κ))
λ
(κ)
(κ)
P ζm,k (qκ t) > u +
, ζ (0) ≤ u
wκ (u) m,k
(2)
(1)
(q
t)|
|X
(q
t)|
|X
λ
1/r κ
1/r κ
(κ)
, ζm,k (qκ t) > u
≤ P θκ (t)|X (2) (0)|κ + −τ /κ ψ 1−κ + −τ /κ ψ 1−κ >
2wκ (u)
(u
t )
(u
t )
n
o
τ
τ
(κ)
(κ)
(κ)
+P |X (1) (0)| ≤ u− κ tψ , ζm,k (qκ t) > u + P |X (2) (0)| ≤ u− κ tψ , ζm,k (0) ≤ u, ζm,k (qκ t) > u +
∗
∗
∗
=: I1u
+ I2u
+ I3u
.
Now we deal with the three terms one by one. Clearly, for any u > 2
n
o
λκtα/4
λ
(1)
(κ)
∗
+ P |X 1/r (qκ t)| >
P
ζ
(q
t)
>
u
I1u
≤ P θκ (t)|X (2) (0)|κ >
m,k κ
6wκ (u)
6uτ /κ
n
o
λκtα/4
(κ)
(2)
P
ζ
(q
t)
>
u
,
+P |X 1/r (qκ t)| >
κ
m,k
6uτ /κ
where the first term can be treated as for II4u , see (27). For the rest two terms, we have, by using (23)
−p
λκtα/4
λ
κ
λ
(j)
(j)
P |X 1/r (qκ t)| >
≤ P |X (0)| > √
≤ K α/4
(28)
, j = 1, 2.
6uτ /κ
t
12 2B tα/4
λ
wκ (u)
12
PATRIK ALBIN, ENKELEJD HASHORVA, LANPENG JI, AND CHENGXIU LING
(2)
∗
∗
In order to deal with I2u
and I3u
, set below (X (1)
r (t), X r (t)) := X1 (0)−r1 (t)X1 (t), . . . , Xm+k (0)−rm+k (t)Xm+k (t)
(κ)
which by definition is independent of {ζm,k (t), t ≥ 0}. For j = 1, 2
)
(
√
n
o
2 λx
(j)
(j)
(29)
.
P |X r (qκ t)| > x ≤ P |X (0)| > √
u−2τ /κtα
(1)
κ
κ
Using further the triangle inequality |X (1)
(qκ t)|κ −|X (1) (0)|κ and (22) (recalling |X (1) (qκ t)|κ ≥
r (qκ t)| ≥ (r(qκ t)) |X
(κ)
ζm,k (qκ t) > u), we have
n
o
tψκ
(κ)
∗
κ
κ
P
ζ
(q
t)
>
u
I2u
≤ P |X (1)
(q
t)|
>
u
(r(q
t))
−
κ
κ
κ
r
m,k
u1+τ
n
o
(1 − 2−κ )u
(κ)
κ
≤ P |X (1)
P ζm,k (0) > u
r (qκ t)| >
κ
2
(
√ ) n
o
λ
(κ)
(1)
−κ 1/κ
≤ P |X (0)| > (1 − 2 )
P
ζ
(0)
>
u
m,k
tα/2
−p/2 n
o
λ
(κ)
(30)
P ζm,k (0) > u .
≤ K α
t
∗
For I3u
, using |X (1) (qκ t)|κ > u + λ/wκ (u) and
|X
we have
∗
I3u
where by (22)
(1)
κ
(0)| =
(κ)
ζm,k (0)
+ |X
(2)
tψκ
(0)| ≤ u 1 + 1+τ
u
κ
n
o
λ
tψκ
(κ)
κ
κ
1
+
≤ P |X (1)
(q
t)|
>
u
(r(q
t))
−
1
+
P ζm,k (qκ t) > u
κ
κ
r
1+τ
uwκ (u)
u
n
o
o n (κ)
κ
−τ
= P |X (1)
λκ(r(qκ t))κ − u1+τ (1 − (r(qκ t))κ ) − tψκ P ζm,k (0) > u ,
r (qκ t)| > u
λκ(r(qκ t))κ − u1+τ (1 − (r(qκ t))κ ) − tψκ ≥
λκ
λκ
− tψκ ≥ κ+2 .
κ+1
2
2
Consequently, it follows further by (29) that
n
o
1/κ+1/2
(κ)
(1)
∗
−2/κ 1/κ λ
I3u ≤ P |X (0)| > 2
κ
P
ζ
(0)
>
u
m,k
tα/2
1/κ+1/2 −p n
o
λ
(κ)
≤ K
P
ζ
(0)
>
u
,
m,k
tα/2
which together with (27), (28) and (30) completes the proof for κ ∈ (0, 1) by taking p > 4/α + k. Consequently, the
desired claim of Lemma 3.4 follows. This completes the proof.
Proof of Lemma 3.5: We give only the proof for (17) since (18) follows by similar arguments. Since the claims
consider that k ≥ 1 below.
Define, for j = 1, 2, independent random
for k = 0 are already shown in [1],
we only
(j)
(j)
(κ)
(j)
(j)
e
e
vectors |Y (s1 )|, . . . , |Y (sp )| and |Y (t1 )|, . . . , |Y (tp′ )| , which are independent of the process ζm,k and
have the same distributions as those of |X (j) (s1 )|, . . . , |X (j) (sp )| and |X (j) (t1 )|, . . . , |X (j) (tp′ )| , respectively.
Note that, for any u > 0, the left-hand side of (17) is clearly bounded from above by
p′ n
p n
\
o
o \
|X (2) (tj )|κ ≥ |X (1) (tj )|κ − u
|X (2) (si )|κ ≥ |X (1) (si )|κ − u ,
P
j=1
i=1
p′
p n
\
o \
(2)
|Ye (tj )|κ ≥ |X (1) (tj )|κ − u
|Y (2) (si )|κ ≥ |X (1) (si )|κ − u ,
−P
j=1
i=1
p′
p n
\
o \
(2)
(1)
(2)
κ
κ
|X (1) (tj )|κ ≤ |Ye (tj )|κ + u
|X (si )| ≤ |Y (si )| + u ,
+ P
i=1
j=1
EXTREMES AND LIMIT THEOREMS FOR DIFFERENCE OF CHI-TYPE PROCESSES
−P
(31)
p n
\
i=1
13
p′
o \
(1)
(2)
|Ye (tj )|κ ≤ |Ye (tj )|κ + u
|Y (1) (si )|κ ≤ |Y (2) (si )|κ + u ,
.
j=1
Next, note by Cauchy-Schwarz inequality that u2 + v2 ≤ (u2 − 2ρuv + v 2 )/(1
− |ρ|) for all ρ ∈ (−1, 1) and u, v ∈ R.
(1)
(1)
It follows that, fij (·, ·), the joint density function of |X (si )|, |X (tj )| , satisfies
fi,j (x, y)
=
2
xl − 2rl (tj − si )xl yl + yl2
1
p
exp
−
dxdy
2
2(1 − rl2 (tj − si ))
|x|=x,|y|=y l=1 2π 1 − rl (tj − si )
Z
m
Y
x2l + yl2
1
exp −
dxdy
2(1 + |rl (tj − si )|)
(2π)m (1 − (e
r (tj − si ))2 )m/2 |x|=x,|y|=y
l=1
Z
x2 + y 2
1
dxdy
exp −
2(1 + re(tj − si ))
(2π)m (1 − (e
r (tj − si ))2 )m/2
|x|=x,|y|=y
(xy)m−1
x2 + y 2
, x, y > 0.
exp −
2(1 + re(tj − si ))
2m−2 (Γ(m/2))2 (1 − (e
r (tj − si ))2 )m/2
Z
≤
≤
=
m
Y
Therefore, in view of Lemma 2 in [1], with K a constant whose value might change from line to line, the first
absolute value in (31) is bounded from above by
(k−1)/κ
(xκ − u)2/κ + (y κ − u)2/κ
re(tj − si ) (xκ − u)(y κ − u)
exp −
fij (x, y) dxdy
2(1 + re(tj − si ))
κ
κ
i=1 j=1 x >u y >u
!2
Z ∞
p′
p X
X
x2/κ
(k−1)/κ m/κ−1
re(tj − si )
(x − u)
x
exp −
≤K
dx
2(1 + re(tj − si ))
u
i=1 j=1
p Z
p X
X
′
K
Z
′
≤ Ku
(2/κ)(m−(k−1)(2/κ−1)−2)
p
p X
X
i=1 j=1
re(tj − si ) exp −
u2/κ
1 + re(tj − si )
,
where in the first inequality, we use first the bound e−x ≤ 1, x ≥ 0 and then a change of variable x′ = xκ , while the
second inequality follows by a change of variable x′ = u2/κ−1 (x − u) and Taylor’s expansion of (u + x′ /u2/κ−1 )2/κ =
u2/κ + (2/κ)x′ + O(u−2/κ ) for large u and x′ ≥ 0. Similarly, denoting by g(·) the pdf of |X (2) (0)|, we obtain that
the second absolute value in (31) is bounded from above by
(m−1)/κ
(xκ + u)2/κ + (y κ + u)2/κ
re(tj − si )
(x + u)(y + u)
exp −
g(x)g(y) dxdy
K
2(1 + re(tj − si ))
0
0
i=1 j=1
!2
Z ∞
p′
p X
X
(xκ + u)2/κ
κ
(m−1)/κ k−1
re(tj − si )
≤K
dx
(x + u)
x
exp −
2(1 + re(tj − si ))
0
i=1 j=1
′
p
p X
X
Z
∞
Z
∞
κ
κ
′
≤ Ku
(2/κ)(m−k(2/κ−1)−1)
p
p X
X
i=1 j=1
re(tj − si ) exp −
u2/κ
1 + re(tj − si )
,
where the last step follows by a change of variable x′ = u2/κ−1 xκ . Hence the proof of (17) is established since
m − k(2/κ − 1) − 1 − m − (k − 1)(2/κ − 1) − 2 = −2(1/κ − 1).
The desired result in Lemma 3.5 follows.
Proof of Lemma 3.6: The proof follows by the same arguments as for Lemma 12.3.1 in [15], using alternatively
the following asymptotic relation (recall (19) and Lemma 3.1)
(32)
u2/κ = 2 ln Tκ + K0 ln ln Tκ + ln D0 (1 + o(1)),
Tκ → ∞
with D0 , K0 defined in Theorem 2.3. We split the sum in (20) at Tκβ , where β is a constant such that 0 < β <
(1 − δ)/(1 + δ) and δ = sup{e
r (t) : t ≥ ǫ} < 1 (see, e.g., Lemma 8.1.1 (i) in [15]). Below K is again a positive
14
PATRIK ALBIN, ENKELEJD HASHORVA, LANPENG JI, AND CHENGXIU LING
constant which value might change from line to line. From (32) we conclude that exp −u2/κ /2 ≤ K/Tκ and
u2/κ = 2 ln Tκ (1 + o(1)). Further,
X
Tκ
u2/κ
uς
re(aqκ j) exp −
qκ
1 + re(aqκ j)
β
ε≤aqκ j≤Tκ
2
4τ
κς
u2/κ
2τ
β+1− 1+δ
≤ uς+ ακ Tκβ+1 exp −
≤ K(ln Tκ ) 2 + α Tκ
,
1+δ
which tends to 0 as Tκ → ∞ since β+1−2/(1+δ) < 0. For the remaining sum, denoting δ(t) = sup{|e
r(s) ln s| : s ≥ t},
β
t > 0, we have re(t) ≤ δ(t)/ ln t as t → ∞, and thus in view of (32) for aqκ j ≥ Tκ
δ(Tκβ )
u2/κ
≤ exp −u2/κ 1 −
exp −
1 + re(aqκ j)
ln Tκβ
≤ K exp(−u2/κ ) ≤ KTκ−2(ln Tκ )−K0 .
Consequently, with c given by Theorem 2.3 (recall τ = 2 max(1/κ − 1, 0) + 1),
X
u2/κ
ς Tκ
re(aqκ j) exp −
u
qκ β
1 + re(aqκ j)
Tκ ≤aqκ j≤Tκ
≤ Kuς
(33)
Tκ
qκ
≤ K(ln Tκ )
2
κς
2
Tκ−2 (ln Tκ )−K0
+ 2τ
α −K0 −c
1
Tκ /qκ
1
c
ln Tκβ
X
1
Tκ /qκ
Tκβ ≤aqκ j≤Tκ
X
Tκβ ≤aqκ j≤Tκ
c
re(aqκ j) ln(aqκ j)
c
re(aqκ j) ln(aqκ j) .
Since K0 = m − 2 + 2τ /α + k min(1 − 2/κ, 0) and ς := 2/κ(m − k(2/κ − 1) − 1 + max(0, 2(1/κ − 1))), we have
c
κς/2 + 2τ/α − K0 − c = 0 for all κ > 0. Noting further that the Berman-type condition limt→∞ re(t) ln t = 0 holds
and β < 1, the right-hand side of (33) tends to 0 as u → ∞. Thus the proof is complete.
Acknowledgments. We would like to thank the associate editor and the two anonymous referees for their constructive suggestions and corrections which greatly improved the paper.
References
[1] J.M.P. Albin. On extremal theory for non differentiable stationary processes. PhD Thesis, University of Lund,
Sweden, 1987.
[2] J.M.P. Albin. On extremal theory for stationary processes. Ann. Probab., 18(1):92–128, 1990.
[3] J.M.P. Albin and D. Jarušková. On a test statistic for linear trend. Extremes, 6(3):247–258, 2003.
[4] A. Aue, L. Horváth, and M. Hušková. Extreme value theory for stochastic integrals of Legendre polynomials.
J. Multivariate Anal., 100(5):1029–1043, 2009.
[5] S.M. Berman. Sojourns and extremes of stationary processes. Ann. Probab., 10(1):1–46, 1982.
[6] S.M. Berman. Sojourns and extremes of stochastic processes. The Wadsworth & Brooks/Cole Statistics/Probability Series. Wadsworth & Brooks/Cole Advanced Books & Software, Pacific Grove, CA, 1992.
[7] K. Dȩbicki, E. Hashorva, and L. Ji. Gaussian approximation of perturbed chi-square risks. Statistics and Its
Interface, 7:363–373, 2014.
[8] A.B. Dieker and B. Yakir. On asymptotic constants in the theory of Gaussian processes. Bernoulli, 20(3):1600–
1619, 2014.
[9] P. Embrechts, C. Klüppelberg, and T. Mikosch. Modelling extremal events for insurance and finance. SpringerVerlag, Berlin, 1997.
[10] E. Hashorva and L. Ji. Piterbarg theorems for chi-processes with trend. Extremes, 18(1):37–64, 2015.
[11] E. Hashorva, D. Korshunov, and V.I. Piterbarg. Asymptotic expansion of Gaussian chaos via probabilistic
approach. Extremes, 18(3):315–347, 2015.
[12] J. Hüsler, V.I. Piterbarg, and O. Seleznjev. On convergence of the uniform norms for Gaussian processes and
linear approximation problems. Ann. Appl. Probab., 13(4):1615–1653, 2003.
EXTREMES AND LIMIT THEOREMS FOR DIFFERENCE OF CHI-TYPE PROCESSES
15
[13] D. Jarušková. Detecting non-simultaneous changes in means of vectors. TEST, 24(4):681–700, 2015.
[14] C. Klüppelberg and M.G. Rasmussen. Outcrossings of safe regions by generalized hyperbolic processes. Stat.
Prob. Letters., 83(10):2197–2204, 2013.
[15] M.R. Leadbetter, G. Lindgren, and H. Rootzén. Extremes and related properties of random sequences and
processes, volume 11. Springer Verlag, 1983.
[16] M.R. Leadbetter and H. Rootzén. Extreme value theory for continuous parameter stationary processes. Z.
Wahrsch. Verw. Gebiete, 60(1):1–20, 1982.
[17] G. Lindgren. Extreme values and crossings for the χ2 -process and other functions of multidimensional Gaussian
processes, with reliability applications. Adv. in Appl. Probab., 12(3):746–774, 1980.
[18] G. Lindgren. Extremal ranks and transformation of variables for extremes of functions of multivariate Gaussian
processes. Stochastic Process. Appl., 17(2):285–312, 1984.
[19] G. Lindgren. Slepian models for χ2 -processes with dependent components with application to envelope upcrossings. J. Appl. Probab., 26(1):36–49, 1989.
[20] C. Ling and Z. Tan. On maxima of chi-processes over threshold dependent grids. Statistics, 50(3):579–595,
2016.
[21] C. Ling and Z. Peng. Extremes of order statistics of self-similar processes (in Chinese). Sci Sin Math, 46: 1–10,
doi: 10.1360/012016-15, 2016.
[22] V.I. Piterbarg. Asymptotic methods in the theory of Gaussian processes and fields, volume 148. American
Mathematical Society, Providence, RI, 1996.
[23] V.I. Piterbarg and A. Zhdanov. On probability of high extremes for product of two independent Gaussian
stationary processes. Extremes, 18(1):99–108, 2015.
[24] O. Seleznjev. Asymptotic behavior of mean uniform norms for sequences of Gaussian processes and fields.
Extremes, 8(3):161–169 (2006), 2005.
[25] Z. Tan and E. Hashorva. Exact asymptotics and limit theorems for supremum of stationary χ-processes over a
random interval. Stochastic Process. Appl., 123(8):1983–2998, 2013.
[26] Z. Tan and E. Hashorva. Limit theorems for extremes of strongly dependent cyclo-stationary χ-processes.
Extremes, 16(2):241–254, 2013.
Patrik Albin, Department of Mathematical Sciences, Chalmers University of Technology, SE-412, 96 Gothenburg, Sweden
Enkelejd Hashorva, Department of Actuarial Science, University of Lausanne, UNIL-Dorigny, 1015 Lausanne, Switzerland
Lanpeng Ji, Department of Actuarial Science, University of Lausanne, UNIL-Dorigny, 1015 Lausanne, Switzerland
School of Mathematics and Statistics, Southwest University, Beibei District, 400715 Chongqing, China
| 10math.ST
|
STRUCTURE AND MINIMAL GENERATING SETS OF SYLOW
2-SUBGROUPS OF ALTERNATING GROUPS, PROPERTIES OF ITS
COMMUTATOR SUBGROUP
arXiv:1702.05784v4 [math.GR] 10 Nov 2017
RUSLAN SKURATOVSKII
Abstract. In this article the research of Sylows p-subgroups of An and Sn , which
was started in [2, 3, 4] is continued. Let Syl2 A2k and Syl2 An be Sylow 2-subgroups
of corresponding alternating groups A2k and An . We find a least generating set and a
structure for such subgroups Syl2 A2k and Syl2 An and commutator width of Syl2 A2k
[1].
The purpose of this paper is to research the structure of a Sylow 2-subgroups and to
construct a minimal generating set for such subgroups. The main result is to prove minimality of this generating set for the above indicated subgroups and also the description
of their structure.
Key words: minimal set of generators; wreath product of group; Sylow subgroups;
commutator subgroup, semidirect product.
1. Introduction
The aim of this paper is to research the structure of Sylow 2-subgroups of A2k , An and
to construct a minimal generating set for Syl2A2k . The case of Sylow subgroup where
p = 2 is very special because group C2 ≀ C2 ≀ C2 . . . ≀ C2 admits odd permutations, this
case was not fully investigated in [2, 3]. The authors of [2, 4] didn’t prove minimality of
found by them system of generators for such Sylow 2-subgroups of An and structure of
a Sylow 2-subgroups was found by them not fully, for case n = 2k structure of Syl2An
was not found. This question is up till now under consideration. There was a mistake
in a statement about irreducibility of a set of k + 1 elements for Syl2(A2k ) that was in
abstract [5] on Four ukraine conference of young scientists in 2015 year. These groups
have applications in the automaton theory, because if all states of automaton A have
output function that can be presented as cycle (1, 2, ..., p) then group GA (X) of this
automaton, where X is finite alphabet, is Sylows p-subgroup of the group of all automaton
transformations GA(X) [7]. Also in this case holds Sylp (AutX) > F GA(X) [7], where
| X |∈ N. Thus, finding the minimum size of generating set number is important. Let X ∗
be the free monoid freely generated by X. All undeclared terms are from [8, 9].
2. Preliminaries
The set X ∗ is naturally a vertex set of a regular rooted tree, i.e. a connected graph
without cycles and a designated vertex v0 called the root, in which two words are connected
by an edge if and only if they are of form v and vx, where v ∈ X ∗ , x ∈ X. The set X n ⊂ X ∗
is called the n-th level of the tree X ∗ and X 0 = {v0 }. We denote by vj,i the vertex of
1
2
X j , which has the number i. The subtree of X ∗ induced by the set of vertices ∪ki=0 X i
is denoted by X [k]. Note that the unique vertex vk,i corresponds to the unique word v
in alphabet X. For every automorphism g ∈ AutX ∗ and every word v ∈ X ∗ define the
section (state) g(v) ∈ AutX ∗ of g at v by the rule: g(v) (x) = y for x, y ∈ X ∗ if and only
if g(vx) = g(v)y. The restriction of the action of an automorphism g ∈ AutX ∗ to the
subtree X [l] is denoted by g(v) |X [l] . A restriction g(v) |X [1] is called the vertex permutation
(v.p.) of g in a vertex v. Let us introduce conventional signs for a v.p. state value of α
in vki as ski (α) we put that ski (α) = 1 if α(vki ) |X [1] (x) = y, x 6= y such state of v.p. is
active, and ski (α) = 0 if α(vki ) |X [1] (x) = x such state of v.p. is trivial. Let us label every
vertex of X l , 0 ≤ l < k by sign 0 or 1 in relation to state of v.p. in it. Obtained by such
way a vertex-labeled regular tree is element of AutX [k] .
We denote by vj,i X [k−j] subtree of X [k] with a root in vj,i . An automorphism of X [k] with
non-trivial states of v.p. in some of v1,1 , v1,2 , v2,1 ,..., v2,4 , ... ,vm,1 , ... ,vm,j , m < k, j ≤ 2m
is denoted by β1,(i11 ,i12 );...;l,(il1,...,il2l );...;m,(im1 ,...,im2m ) where the index that stands straight
before parentheses are number of level in parentheses we write a tuple of states of v.p.
of this level. In other words we set imj = 0 if v.p. in vmj is trivial, imj = 1 in other
case, i.e. imj = smj (β), where β ∈ AutX [k] , m < k. If for some l all ilj = 0 then 2l -tuple
l, (il1 , ..., il2l ) does not figure in indexes of β. But if numbers of active vertices are certain,
for example vj,1 and vj,s , we can use more easy notation βj,(1,s); , where in parentheses
numbers of vertices with active state of v.p. from level j. If in parentheses only one
index then parentheses can be omitted for instance βj,(s); = βj,s; . Denote by τi,...,j the
automorphism of X [k] , which has a non-trivial v.p. only in vertices vk−1,i , ... ,vk−1,j ,
j ≤ 2k−1 of the level X k−1. Denote by τ the automorphism τ1,2k−1 . Let us consider special
elements such that: α0 = β0 = β0,(1,0,...,0), α1 = β1 = β1,(1,0,...,0) , . . . , αl = βl = βl,(1,0,...,0).
3. Main result
Recall that a wreath product of permutation groups is associative construction. We
consider C2 as additive group with two elements 0, 1. For constructing a wreath product
we define an action of C2 by shift on X = {0, 1}. As well known that AutX [k−1] ≃
C2 ≀ ... ≀ C2 [7].
| {z }
k−1
Lemma 1. Every automorphism that has active v.p. only on X l , l < k − 1 acts by even
permutation on X k .
Proof. Actually every transposition in vertex from X l , l < k − 1 acts on even number of
pair of vertexes because of binary tree structure. More precisely it realize even permuk−1
k−l−l
k−l−l
tation on the set X k with cyclic structure [11] (12 −2
, 22
) because it formed by
the structure of binary tree.
Corollary 1. Due to Lemma 1 automorphisms from AutX [k−1] = hα0 ...., αk−2 i form a
group Bk−1 = C2 ≀ ... ≀ C2 which acts on X k−1 by even permutations. Size of Bk−1 equal
| {z }
k−1
to 2
2k−1 −1
.
3
Let us denote by Wk−1 the subgroup of AutX [k] such that has active states only on
X k−1 and number of such states is even, i.e. Wk−1 ⊳ StGk (k − 1) [8].
k−1 −1
Proposition 1. The size of Wk−1 is equal to 22
k−1 −1
, k > 1 and its structure is (C2 )2
.
Proof. On X k−1 we have 2k−1 vertices where can be elements of a group Vk−1 ≃ C2 × C2 ×
... × C2 ≃ (C2 )k−1 , but as a result of the fact that X k−1 contains only even number of non
trivial v.p. from X k−1, there are only half of all permutations from Vk−1 ≃ StGk (k − 1) on
2k−1
X k−1 . So it is subgroup Wk−1 ≃ C2 /C2 of Vk−1. So we can state that |Wk−1 | = 2k−1 − 1,
Wk−1 has k−1 generators and we can consider Wk−1 as vector space of dimension k−1.
4−1
For example let us consider the subgroup W4−1 of A24 its size is 22 −1 = 27 and
|A24 | = 214 . Let us denote by Gk the subgroup of AutX [k] such that Gk ≃ Bk−1 ⋉ Wk−1.
Lemma 2. The elements τ and α0 , ..., αk−1 generate arbitrary element τij .
Proof. According to [9, 3] the set α0 , ..., αk−2 is minimal generating set for group
AutX [k−1] . Firstly, we shall prove the possibility of generating an arbitrary τij , from
set v(k−1,i) , 1 ≤ i ≤ 2k−2. Since Autv1,1 X [k−2] ≃ hα1 , ..., αk−2 i acts on X k−1 transitively
from it follows existing of an ability to permute vertex with a transposition from automorphism τ and stands in vk−1,1 in arbitrary vertex vk−1,j , j ≤ 2k−2 of v1,1 X [k−1] . For
this goal we act by αk−j at τ : αk−j τ αk−j = τj,2k−2 . Similarly we act on τ by corespon−1
dent αk−i to get τi,2k−2 from τ : αk−i τ αk−i
= τi,2k−2 . Note that automorphisms αk−j and
αk−i , 1 < i, j < k − 1 acts only on subtree v1,1 X [k−1] that’s why they fix v.p. in vk−1,2k−1 .
Now we see that τi,2k−2 τj,2k−2 = τi,j , where 1 ≤ i, j < 2k−2 . To get τm,l from v1,2 X [k−1],
i.e. 2k−2 < m, l ≤ 2k−1 we use α0 to map τi,j in τi+2k−2 ,j+2k−2 ∈ v1,2 AutX [k−1]. To
construct arbitrary transposition τi,m from Wk−1 we have to multiply τ1,i τ τm,2k−1 = τi,m .
Let us realize a natural number of vk,l , 1 < l < 2k in 2-adic set of presentation (binary arithmetic). Then l = δl1 2ml + δl2 2ml −1 + ... + δlml +1 , δli ∈ {0, 1} where is a
correspondence between δli that from such presentation and expressing of automorQml δli
Q l δli
α
τ
phisms: τl,2k−1 = m
i=1 αk−2−(ml −i) , 1 ≤ ml ≤ k − 2. In other words
i=1 k−2−(ml −i)
hα0 , ..., αk−2 , τ i ≃ Gk .
Corollary 2. The elements from condition of Lemma 2 are enough to generate a basis
of Wk−1 .
k
Lemma 3. Sizes of groups Gk = hα0 , α1 , α2 , ..., αk−2 , τ i and Syl2(A2k ) are equal to 22 −2 .
h ki h ki
k
Proof. In accordance with Legendre’s formula, the power of 2 in 2 ! is 22 + 222 +
h ki
h ki
k −1
2
2
+
...
+
= 22−1
. We need to subtract 1 from it because we have only n!2 of all
23
2k
k
k
−1
permutations as a result: 22−1
− 1 = 2k − 2. So |Syl(A2k )| = 22 −2 . The same size
has group Gk = Bk−1 ⋉ Wk−1 and |Gk | = |Bk−1 | · |Wk−1 | = |Syl2A2k |. Since size of
k−1
k
groups Gk according to Proposition 1 and the fact that |Bk−1 | = 22 −1 is 22 −2 . For
3−1
instance the sizes of Syl2(A8 ), B3−1 and W3−1 are such |W3−1 | = 22 −1 = 23 = 8,
|B3−1 | = |C2 ≀ C2 | = 2 · 22 = 23 and according to Legendre’s formula, the power of 2 in
3
3
3
k
2k ! is 22 + 222 + 223 − 1 = 6 so Syl2(A8 ) = 26 = 22 −2 , where k = 3. Next example for
4
4
4−1
A16 : Syl2(A16 ) = 22 −2 = 214 , k = 4, |W4−1 | = 22 −1 = 27 , |B4−1 | = |C2 ≀ C2 ≀ C2 | =
2 · 22 · 24 = 27 . So we have the |A16 | = |W3 ||B3 | equality which endorse the condition of
this lemma.
Theorem 1. A maximal 2-subgroup of AutX [k] that acts by even permutations on X k
has the structure of the semidirect product Gk ≃ Bk−1 ⋉ Wk−1 and isomorphic to Syl2A2k .
Proof. A maximal 2-subgroup of AutX [k−1] is isomorphic to Bk−1 ≃ C2 ≀ C2 ≀ ... ≀ C2 (this
|
{z
}
k−1
group acts on X k−1). A maximal 2-subgroup which has elements with active states only
on X k−1 corresponds subgroup Wk−1 . Since subgroups Bk−1 and Wk−1 are embedded
in AutX [k] , then define an action of Bk−1 on elements of Wk−1 as τ σ = στ σ −1 , σ ∈
Bk−1 , τ ∈ Wk−1 , i.e. action by inner automorphism (inner action) from AutX [k] . Note
that Wk−1 is subgroup of stabilizer of X k−1 i.e. Wk−1 < StAutX [k] (k − 1) ⊳ AutX [k] and is
normal too Wk−1 ⊳ AutX [k], because conjugation keeps a cyclic structure of permutation
so even permutation maps in even. Therefore such conjugation induce automorphism of
k−1
Wk−1 and Gk ≃ Bk−1 ⋉ Wk−1 . Since at X k−1 is 22
vertexes and half of combinations of
k−1
k−1
active states from X
can form even permutation thus |Wk−1 | = 22 −1 that is proved
in Proposition 1. Using the Corollary 1 about Bk−1 we get size of Gk ≃ Bk−1 ⋉ Wk−1 is
k−1
k−1
k
22 −1 · 22 −1 = 22 −2 . Since Gk is the maximal 2-subgroup then Gk ≃ Syl2 A2k .
Theorem 2. The set Sα = {α0 , α1 , α2 , ..., αk−2 , τ } of elements from subgroup of AutX [k]
generates a group Gk which isomorphic to Syl2 (A2k ).
Proof. As we see from Corollary 1, Lemma 2 and Corollary 2 group Gk are generated by
Sα and their sizes according to Lemma 3 are equal. So according to Sylow’s theorems
2-subgroup Gk < A2k is Syl2(A2k ).
Consequently, we construct a generating set, which contains k elements, that is less
than in [5].
k−1
Q k−1
The structure of Sylow 2-subgroup of A2k is the following: ≀ C2 ⋉ 2i=1 −1 C2 , where
i=1
we take C2 as group of action on two elements and this action is faithful. It adjusts
with construction of normalizer for Sylp (Sn ) from [14], where it was said that Syl2(A2l )
is self-normalized in S2l .
Definition 1. Let us call the index of automorphism β on X l as a number of active v.p.
of β on X l .
Definition 2. Define an element of type T as an automorphism τi0 ,...,i2k−1 ;j2k−1 ,...,j2k , that
has even index at X k−1 and it has exactly m1 active states, m1 ≡ 1(mod2), in vertexes
vk−1,j , with number 1 ≤ j ≤ 2k−2 and m2 , m2 ≡ 1(mod2) active states in vertices vk−1,l ,
2k−2 < l ≤ 2k−1 . Set of such elements is denoted by T.
Definition 3. A combined element is such an automorphism β1,i1 ;2,i2 ;...;k−1,ik−1;τ̃ , that it’s
restriction β1,i1 ;2,i2 ;...;k−1,ik−1;τ̃ Xk−1 coincide with one of elements that can be generated
by Sα and Rist<β1,i1 ;2,i2 ;...;k−1,ik−1 ;τ̃ > (k − 1) = hτ ′ i [8] where τ ′ ∈T. Set of such elements is
denoted by C.
5
In other word elements g ∈C on level X k−1 have such structure as element and generator
of type T. As well τi0 ,...,i2k−1 ;j2k−1 ,...,j2k ∈ StAutX k (k − 1).
The minimum size of a generating set S of G we denote by rkG and call the rank of G
[15]. By distance between vertices we shall understand usual distance at graph between its
vertexes. By distance of automorphism g (element) we shall understand maximal distance
between two vertexes with active states of g.
Lemma 4. A vertices permutations on X k that has distance d0 can not be generated by
vertex permutations with distance d1 such that d1 < d0 .
Proof. The element τij with distance ρ(τij ) = d0 , d0 < d1 can be mapped by automorphic
mapping only in automorphism with distance d0 because automorphic mapping keep
incidence relation and so it possess property of isometry. Also multiplication of portrait
(labeled graph) of automorphism τij : ρ(τij ) = d1 give us portrait of element with
distance no greater than d1 , it follows from properties of group operation. For instance
τ1i τ1j = τii , where i, j > 2k−2 ρ(τ1i ) = ρ(τ1j ) = 2k − 2 but ρ(τij ) < 2k − 2.
Lemma 5. An arbitrary automorphism τ ′ ∈T (or in particular τ ) can be generated only
with using odd number of automorphisms from C or T.
Proof. Let us assume that there is no such element τij which has distance 2k − 2 then
accord to Lemma 4 it is imposable to generate are pair of transpositions τ ′ with distance
ρ(τij ) = 2k since such transpositions can be generated only by τij that described in the
conditions of this Lemma: i ≤ 2k−2 , j > 2k−2. Combined element can be decomposed in
product τ β̇il = βil ;τ so we can express by using τ or using a product where odd number
an elements from T or C. If we consider product P of even number elements from T then
automorphism P has even number of active states in vertexes vk−1,i with number i ≤ 2k−2
so P does not satisfy the definition of generator of type T.
Corollary 3. Any element of type T can not be generated by τij ∈ Autv1,1 X [k−1] and
τml ∈ Autv1,2 X [k−1] . The same corollary is true for a combined element.
Proof. It can be obtained from the Lemma 4 because such τij ∈ Autv1,1 X [k−1] has distance
less then 2k − 2 so it does not satisfy conditions of the Lemma 4. I.e. τ ′ can not be
generated by vertices automorphisms with distance between vertices less than 2k − 2 such
distance has only automorphisms of type T and C. But elements from Autv1,1 X [k−1] do
not belongs to type T or C.
Lemma 6. Sets of elements of types T, C are not closed by multiplication and raising to
even power.
Proof. Let ̺, ρ ∈ T (or C) and ̺, ρ = η. The numbers of active states from ̺ and ρ in tuple
of vertices vk−1,i , 1 ≤ i ≤ 2k−2 sums by mod2, numbers of active states from ̺ and ρ in
vertices on vertices vk−1,i , 2k−2 < i ≤ 2k−1 sums by mod2 too. Thus η has even numbers
of active states on these tuples. Hence RiSthηi (k − 1) doesn’t contain elements of type τ
so η ∈T.
/ Really if we raise the element β1,i1 ;2,i2 ;...;k−1,ik−1;τ ∈T to even power or we evaluate
a product of even number of multipliers from C corteges µ0 and µ1 permutes with whole
6
subtrees v1,1 X [k−1] and v1,2 X [k−1] , then we get an element g with even indices of X k in
v1,1 X k−1 and v1,2 X k−1 . Thus g ∈T.
/ Consequently elements of C do not form a group, and
the set T as a subset of C is not closed too.
′
′
Let Sα = hα0 , α1 , ..., αk−2i so as well known [9] Sα = AutX [k−1] . The cardinality of
′
a generating set S is denoted by | S | so | Sα |= k − 1. Recall that rk (G) is the rank of a
group G [15].
′
′
Let Sβ = Sα ∪ τi...j , where τi...j ∈T and Sβ is generating system which contains combine
′
′
elements, | Sβ |= k. It’s known that rk(AutX [k−1] ) = k − 1 and | Sα |= k − 1 [9]. So
′
if we complete Sα by τ or element of type T we obtain set Sβ such that Gk ≃ hSβ i and
′
|Sβ | = k. Hence to construct combined element β we multiply generator αi of Sα or
′
arbitrary element that can be express from Sα on the element of type T, i.e., we take τ ′ · βi
instead of βi and denote it βi;τ ′ . It’s equivalent that Ristβi;τ ′ (k − 1) = hτ ′ i, where τ ′ ∈T.
′
′
Let us assume that Sβ has a cardinality k − 1. If in this case Sβ is generating system
again, then element τ can be expressed from it. There exist too ways to express element
′
′
of type T from Sβ . To express element of type T from Sβ we can use a word βi,τ βi−1 =
′
′
′
/ Sβ in contrary case | Sβ |= k. So we can not express
τ but if βi,τ ∈ Sβ then βi ∈
word βi,τ βi−1 |X [k−1] = e to get βi,τ βi−1 = τ . For this goal we have to find relation in a
group that is a restriction of the group Gk on X [k−1] . We have to take in consideration
(j)
that Gk |X [k−1] = Bk−1 . Really in wreath product ≀kj=1 C2 ≃ Bk−1 holds a constitutive
m
′
′
−i
j
−j
i
relations αi2 = e and [αm
, αm
αik α m
] = e, i 6= j, where αm ∈ Sα , αik ∈ Sα
α in α m
(j)
are generators of factors of ≀kj=1C2 (m < n, m < k) [3, 6]. Such relations are a words
m
−j
j
−i
i
] , i 6= j or βi2 = e, βin , βm , βik , βin ,π are generators of Sβ∗ (k −1),
βik ,π βm
, βm
βin ,π βm
[βm
−j
j
−i
i
] , i 6= j does not belongs
βik ,π βm
, βm
βin ,π βm
that could be an automorphism θ. But [βm
to T because this word has logarithm 0 by every element [12]. According to Lemma 5 and
Lemma 6 product of even number element of type C doesn’t equal to element of C or T.
Let us assume an existence of generating set of cardinality k − 1 for Syl2 (A2k ) that in
general case has form Sβ∗ (k − 1) = β0 , β0;1,(i11 ,i12 );π1 , ..., β0;...;k−1,(ik−11,...,ik−1j );πk−1 ,
0 < i < k −1, j ≤ 2i , π1 ∈ T , where πi is the cortege of vertices from X k−1 with non trivial
v.p. which realize permutation with distance 2(k − 1). In other word if element πi ∈ T
then β0;...;i,(ii1,...,iij );πi = πi α0;...;i,(ii1,...,iij ) . From it follows β0;1,(i11 ,i12 );...;m,(im1 ,...,imj ;πm ) |X [k−1] =
′
α0;1,(i11 ,i12 );...,m,(im1 ,...,imj ) ∈ Sα .
Note, that automorphisms from set Sβ∗ (k − 1) generate on truncated rooted tree [8]
k−1 (i)
X [k−1] group hSβ∗ (k − 1)i |X [k−1] ≃ ≀i=1
C2 ≃ Bk .
Theorem 3. Any element of type T can not be expressed by elements of Sβ∗ (k − 1).
Proof. It is necessary to express an automorphism θ of type T express such automorphism
which has zero indexes of X 0 , ..., X k−2, this conclusion follows from structure of elements
from T. It means that word w from letters of Sβ∗ (k − 1) such that w = θ is trivial in group
Bk , that arise on restriction of hSβ∗ (k − 1)i on X [k−1] , as well restriction Gk |X [k−1] ≃ Bk .
Every relation from Bk can be expressed as a product of words from the normal closure
Bk
R of the set of constitutive relations of the group Bk [15]. But defined relations ri of Bk
have form of commutators [6, 3] so the number of inclusions of every multiplier is even and
7
(j)
as follows from lemma 6 that ri |[X k ] ∈
/ T. Really in wreath product ≀kj=1C2 ≃ Bk−1 holds
−i
j
−j
2m
i
, αm
αik α m
] = e, i 6= j, where n, k, m
a constitutive relations αi = e and [αm
α in αm
(j)
k
are number of groups in ≀j=1Cl (m < n, m < k) [3, 6], where αin , αin , αm , αik are
′
i
−i
j
−j
βin ,π βm
, βm
βik ,π βm
] , i 6= j that
generators of the Bk−1 from Sα . So it give us a word [βm
could be an automorphism θ but does not belongs to T because the word has structure
of commutator or belongs to normal closure RBk−1 so has logarithm 0 by every element,
where βin , βm , βik , βin ,π are generators of Gk from Sβ∗ (k − 1).
Let β1,1,(i11 ,i12 ),...,i,(ii1 ,...,ii2i );πi be an arbitrary element of type C, where πi is cortege of
vertexes from X k−1 having non trivial states which realize permutation with distance
2(k − 1). The case where θ = βk−1;πi ∈ T can be expressed by multiplying arbitrary
−1
β1,(i11,12 ),...,i,(ii1 ,...,ii2i );πi on β1,(i
means that such set has size more then k − 1
11,12 ),...,i,(ii1 ,...,ii2i )
what is contradiction. Really if β1,(i11,12 ),...,i,(ii1 ,...,ii2i ) and β1,(i11,12 ),...,i,(ii1 ,...,ii2i );πi ∈ Sβ∗ then
β1,(i11,12 ),...,i,(ii1 ,...,ii2i ) |X [k−1] = β1,(i11,12 ),...,i,(ii1 ,...,ii2i );πi |X [k−1] i.e. these elements are mutually
inverse at restriction on X [k−1] , but it means that in restriction of Sβ∗ (k − 1) to X [k−1]
′
that corresponds to the generating set Sα (k − 1) for AutX [k−1] ≃ Bk−1 we use two equal
generators. So it has at least k generators, because rk(AutX [k−1] ) = k − 1 according to
lemma 1.
−1
The subcase of this case where β1,(i
can be expressed from Sβ∗ (k − 1) as
11 ,i12 ),...,l,(il1 ,...,il2i )
a product of its generators has the same conclusion. Really if we can generate arbitrary
element from Bk−1 by generators from Sβ∗ (k − 1) then k − 1 generators is contained in set
Sβ∗ (k − 1) but we have a further β1,(i11 ,i12 ),...,i,(ii1 ,...,ii2i );π . In other words if arbitrary element
β1,(i11 ,i12 ),...,l,(il1 ,...,il2i ) of the Bk−1 does not contains in Sβ∗ (k − 1) but can be expressed from
it then Sβ∗ (k − 1) has at least k − 1 elements exclusive of β1,(i11 ,i12 ),...,i,(ii1 ,...,ii2i );π then
Sβ∗ (k − 1) ≥ k.
′
Corollary 4. A necessary and sufficient condition of expressing an element τ from Sβ is
′
′
existing of pair: βim ;τ , βim in hSβ i. So size of a generating set of Gk which contains Sβ
is at least k.
Proof. Proof can be obtained from Theorem 3 and Lemma 3 from which we have that
i
−i
j
−j
element of type T cannot be expressed from [βm
βin .τ βm
, βm
βik .τ βm
] = e, i 6= j because
such word has even number of elements from C. Sufficient condition follows from formula
βim ;τ βi−1
= τ.
m
′
Lemma 7. A generating set of Gk contains Sα and has at least k − 1 generators.
Proof. The subgroup Bk−1 < Gk is isomorphic to AutX k−1 that has a minimal set of
generators of k − 1 elements [9]. Moreover, the subgroup Bk−1 ≃ Gk /Wk−1 , because Gk ≃
Bk−1 ⋉ Wk−1 , where Wk−1 ⊲ Gk . As it is well known that if H ⊳ G then rk(G) ≥ rk(H),
because all generators of Gk may belongs to different quotient classes [10].
As a corollary of last lemma and Theorem 3 we see that generating set of size k − 1
does not exist.
Let us sharpen and reformulate following theorem which is in [20].
8
Theorem 4. An element of form (r1 , . . . , rp−1 , rp ) ∈ W ′ = (Cp ≀ B)′ iff product of all ri
(in any order) belongs to B ′ and cw(B) = 1.
We deduce here a special form of commutator elements. If we multiply elements having
−1
form of a tuple (r1 , . . . , rp−1 , rp ), where ri = hi ga(i) h−1
ab(i) gaba−1 (i) , h, g ∈ B and a, b ∈ Cp ,
then we obtain a product
(1)
p
Y
ri =
i=1
k
p
Y
−1
′
hi ga(i) h−1
ab(i) gaba−1 (i) ∈ B .
i=1
Recall that ≀ Cp ≃ Sylp Spk .
i=1
k
k
i=1
i=1
Lemma 8. An element g ∈ Bk ≃ ≀ Cp belongs to commutator subgroup Bk′ ≃ ( ≀ Cp )′
iff g has all possible even indexes on X l , l < k.
Proof. Let us prove the ampleness by induction by a number of level l. We first show that
our statement for base of the induction is true. Actually, if α, β ∈ B0 then (αβα−1 )β −1
determine a trivial v.p. on X 0 . If α, β ∈ B1 and β has an odd index on X 1 , then (αβα−1 )
and β −1 have the same index on X 1 . Consequently, in this case an index of the product
(αβα−1)β −1 can be 0 or 2. Case where α, β ∈ B1 and has even index on X 1 , needs no
proof, because the product and the sum of even numbers is an even number.
To finish the proof it suffices to assume that for Bl−1 statement holds and prove that
it holds for Bl . Let α, β are an arbitrary automorphisms from AutX [k] and β has index
x on X l , l < k, where 0 ≤ x ≤ 2l .A conjugation of an automorphism β by arbitrary
α ∈ AutX [k] gives us arbitrary permutations of X l where β has active v.p.
Thus following product (αβα−1 )β −1 admits all possible even indexes on X l , l < k from 0
to 2x. In addition [α, β] can has arbitrary assignment (arrangement) of v.p. on X l . Let us
present Bk as Bk = Bl ≀Bk−l , so elements α, β can be presented in form of wreath recursion
α = (h1 , ..., h2l )π1 , β = (f1 , ..., f2l )π2 , hi , fi ∈ Bk−l , 0 < i ≤ 2l and hi , fj corresponds to
sections of automorphism in vertices of X l of isomorphic subgroup to Bl in AutX [k]. Actually, the parity of this index are formed independently of the action of AutX [l] on X l . So
this index forms as a result of multiplying of elements of commutator presented as wreath
recursion (αβα−1) · β −1 = (h1 , ..., h2l )π1 · (f1 , ..., f2l )π2 = (h1 , ..., h2l )(fπ1 (1) , ..., fπ1(2l ) )π1 π2 ,
where hi , fj ∈ Bk−l , l < k and besides automorphisms corresponding to hi are x automorphisms which has active v.p. on X l . Analogous automorphisms hi has number of
active v.p. equal to x. As a result of multiplication we have automorphism with index
2i : 0 ≤ 2i ≤ 2x. Consequently, commutator [α, β] has arbitrary even indexes on X m ,
m < l and we showed by induction that it has even index on X l .
Let us prove the necessity by induction by number of level l. Let g ∈ Bk and g has all
even indexes on X j 0 ≤ j < k we need to show that g ∈ Bk′ . According to condition of this
Lemma g1 g2 has even indexes. An element g has form g = (g1 , g2 ), where g1 , g2 ∈ Bk−1 ,
′
and products g1 g2 = h, g2 g1 = h′ ∈ Bk−1
because h, h′ ∈ Bk−1 and for Bk−1 induction
assumption holds. Therefore all products of form g1 g2 indicated in formula 1 belongs to
′
Bk−1
. Hence, from Theorem 4 follows that g = (g1 , g2 ) ∈ Bk′ .
9
An even easier Proposition, that needs no proof, is the following.
k
Proposition 1. An element (g1 , g2 )σ i , i ∈ {0, 1} of wreath power ≀ C2 belongs to its
i=1
subgroup Gk , iff g1 g2 ∈ Gk−1 .
k
Proposition 2. If g is an element of wreath power ≀ C2 ≃ Bk then g 2 ∈ Bk′ .
i=1
Proof. As it was proved in Lemma 8 commutator [α, β] from Bk has arbitrary even indexes
on X m , m < k. Let us show that elements of Bk2 have the same structure.
Let α, β ∈ Bk an index of the automorphisms α2 , (αβ)2 and α, β ∈ Gk on X l , l < k − 1
are always even. In more detail the indexes of α2 , (αβ)2 and α−2 on X l are determined
exceptionally by the parity of indexes of α and β on X l . Actually, the parity of this index
are formed independently of the action of AutX l on X l . So this index forms as a result
of multiplying of elements α ∈ Bk presented as wreath recursion α2 = (h1 , ..., h2l )π1 ·
(f1 , ..., f2l )π1 = (h1 , ..., h2l )(fπ1 (1) , ..., fπ1 (2l ) )π1 π1 , where hi , fj ∈ Bk−l , π1 ∈ Bl , l < k and
besides automorphisms corresponding to hi are x automorphisms which has active v.p.
on X l . Analogous automorphisms hi has number of active v.p. equal to x. As a result of
multiplication we have automorphism with index 2i : 0 ≤ 2i ≤ 2x.
Since g 2 admits only an even index on X l of AutX [k] , 0 < l < k, then g 2 ∈ Bk′ according
to lemma 8 about structure of a commutator subgroup.
′
Lemma 9. An element (g1 , g2)σ i ∈ G′k iff g1 , g2 ∈ Gk−1 and g1 g2 ∈ Bk−1
.
Proof. Indeed, if (g1 , g2) ∈ G′k then indexes of g1 and g2 on X k−1 are even according to
Lemma 1 thus, g1 , g2 ∈ Gk−1. A sum of indexes of g1 and g2 on X l , l < k − 1 are even
′
.
according to Lemma 1 too, so index of product g1 g2 on X l is even. Thus, g1 g2 ∈ Bk−1
Let us prove the sufficiency via Lemma 1. Wise versa, if g1 , g2 ∈ Gk−1 then indexes of
these automorphisms on X k−2 of subtrees v11 X [k−1] and v12 X [k−1] are even as elements
′
from G′k have. The product g1 g2 belongs to Bk−1
by condition of this Lemma so sum of
indexes of g1 , g2 on any level X l , 0 ≤ l < k − 1 is even. Thus, the characteristic properties
of G′k described in this Lemma 1 holds.
Let X1 = {vk−1,1, vk−1,2 , ..., vk−1,2k−2 } and X2 = {vk−1,2k−2 +1 , ..., vk−1,2k−1 }. Let group
Syl2A2k acts on X [k] .
Lemma 10. An element g belongs to G′k ≃ (Syl2A2k )′ iff g is arbitrary element from Gk
which has all even indexes.
Proof. Let us prove the ampleness by induction by a number of level l. Recall that any
authomorphism θ ∈ Syl2 An has an even index on X k−1 so number parities of active v. p.
on X1 and on X2 are the same. Conjugation by automorphism α from Autv11 X [k−1] of
automorphism θ, that has some number x : 1 ≤ x ≤ 2k−2 of active v. p. on X1 does not
change x. Also automorphism θ−1 has the same number x of v. p. on Xk−1 as θ has. If
α from Autv11 X [k−1] and α ∈
/ AutX [k] then conjugation (αθα−1) permutes vertices only
inside X1 (X2 ).
10
Thus, αθα−1 and θ have the same parities of number of active v.p. on X1 (X2 ). Hence,
a product αθα−1 θ−1 has an even number of active v.p. on X1 (X2 ) in this case. More
over a coordinate-wise sum by mod2 of active v. p. from (αθα−1 ) and θ−1 on X1 (X2 ) is
even and equal to y : 0 ≤ y ≤ 2x.
If conjugation by α permutes sets X1 and X2 then there are coordinate-wise sums of
no trivial v.p. from αθα−1 θ−1 on X1 (analogously on X2 ) have form:
(sk−1,1 (αθα−1), ..., sk−1,2k−2 (αθα−1 )) ⊕ (sk−1,1 (θ−1 ), ..., sk−1,2k−2 (θ−1 )). This sum has even
number of v.p. on X1 and X2 because (αθα−1 ) and θ−1 have a same parity of no trivial
v.p. on X1 (X2 ). Hence, (αθα−1 )θ−1 has even number of v.p. on X1 as well as on X2 .
An authomorphism θ from Gk was arbitrary so number of active v.p. x on X1 is
arbitrary. And α is arbitrary from AutX [k−1] so vertices can be permuted in such way
that the commutator [α, θ] has arbitrary even number y of active v.p. on X1 , 0 ≤ y ≤ 2x.
A conjugation of an automorphism θ having arbitrary index x, 1 ≤ x ≤ 2l on X l by
different α ∈ AutX [k] gives us all permutations of active v.p. that θ has on X l . So
multiplication (αθα−1 )θ generates a commutator having index y equal to coordinate-wise
sum by mod2 of no trivial v.p. from vectors (sl1 (αθα−1), sl2 (αθα−1 ), ..., sl2l (αθα−1 )) ⊕
(sl1 (θ), sl2 (θ), ..., sl2l (θ)) on X l . A indexes parities of αθα−1 and θ−1 are same so their
sum by mod2 are even. Choosing θ we can choose an arbitrary index x of θ also we can
choose arbitrary α to make a permutation of active v.p. on X l . Thus, we obtain an
element with arbitrary even index on X l and arbitrary location of active v.p. on X l .
Check that property of number parity of v.p. on X1 and on X2 is closed with respect to
conjugation. We know that numbers of active v. p. on X1 as well as on X2 have the same
parities. So action by conjugation only can permutes it, hence, we again get the same
structure of element. Conjugation by automorphism α from Autv11 X [k−1] automorphism
θ, that has odd number of active v. p. on X1 does not change its parity. Choosing the θ
we can choose arbitrary index x of θ on X k−1 and number of active v.p. on X1 and X2
also we can choose arbitrary α to make a permutation active v.p. on X1 and X2 . Thus,
we can generate all possible elements from a commutant. Also this result follows from
Lemmas 9 and 8.
Let us check that the set of all commutators K from Syl2 A2k is closed with respect to
multiplication of commutators. Let κ1 , κ2 ∈ K then κ1 κ2 has an even index on X l , l < k−1
because coordinate-wise sum (sl,1 (κ1 ), ..., sk−1,2l (κ1 )) ⊕ (sl,κ1 (1) (κ2 ), ..., sl,κ1(2l ) (κ2 )). of two
2l -tuples of v.p. with an even number of no trivial coordinate has even number of such
coordinate. Note that conjugation of κ can permute sets X1 and X2 so parities of x1 and
X2 coincide. It is obviously index of ακα−1 is even as well as index of κ.
Check that a set K is a set closed with respect to conjugation.
Let κ ∈ K, then ακα−1 also belongs to K, it is so because conjugation does not change
index of an automorphism on a level. Conjugation only permutes vertices on level because
elements of AutX [l−1] acts on vertices of X l . But as it was proved above elements of K
have all possible indexes on X l , so as a result of conjugation ακα−1 we obtain an element
from K.
11
Check that the set of commutators is closed with respect to multiplication of commutators. Let κ1 , κ2 be an arbitrary commutators of Gk . The parity of the number of vertex
permutations on X l in the product κ1 κ2 is determined exceptionally by the parity of the
numbers of active v.p. on X l in κ1 and κ2 (independently from the action of v.p. from
the higher levels). Thus κ1 κ2 has an even index on X l .
Hence, normal closure of the set K coincides with K.
Statement 1. Frattini subgroup φ(Gk ) = Gk 2 · [Gk , Gk ] = Gk 2 acts by all even permutations on X l , 0 ≤ l < k − 1 and by all even permutations on X k except for those from
T.
Proof. Index of the automorphism α2 , α ∈ Sβ on X l is always even. Really the parity
of the number of vertex permutations at X l in the product (αi αj )2 , αi , αj ∈ Sα , i, j < k
is determined exceptionally by the parity of the numbers of active states of v.p. on X l
in α and β (independently of the action of v.p. from the higher levels). On X k−1 group
G2 contains all automorphisms of form τ1i , i ≤ 2k−1 which can be generated in such way
(αk−2 τ12 )2 = τ1234 , τ12 τ1234 = τ34 , (αk−i τ12 )2 = τ1,2,1+2k−i ,2+2k−i then τ1,2,1+2k−i ,2+2k−i τ12 =
τ1+2k−i ,2+2k−i . In such way we get set of form τ12 , τ23 , , τ34 , ..., τ2k−1 −1,2k−1 . This set is the
base for Wk−1 .
The parity of the number of vertex permutations at X l in the product αi or αi αj ,
αi , αj ∈ Sα ) is determined exceptionally by the parity of the numbers of active v.p.
on X l in α and β (independently of the action of v.p. from the higher levels). Thus
[α, β] = αβα−1 β −1 has an even number of v. p. at each level. Therefore, the commutators
of the generators from Sα and elements from G2 generate only the permutations with even
number of v. p. at each X l , (0 ≤ l ≤ k − 2).
Let us consider (α0 αl )2 = βl(1,2l−1 +1) . Conjugation by the element β1(1,2) (or βi(1,2) , 0 <
i < l) give us ability to express arbitrary coordinate x : 1 ≤ x ≤ 2l−1 where x = 2k−1 − i,
i.e. from the element βl(1,2l−1 +1) we can express βl(x,2l−1 +1) . For instance x = 2j−1 + 1,
j < l: βl−j(1,2) βl(1,2l−1 +1) βl−j(1,2) = βl(2j−1 +1,2l−1 +1) . If x = 2l−j + 2 than to realize every
shift on x on set X l the element βl(1,2l−1 +1) should to be conjugated by such elements
βl−j(1,2) βl−1(1,2) . So in such way can be realized every βl(x,2l−1 +1) and analogously every
βl(2l−1 ,y) and βl(x,y) . Hence we can express from elements of G2 every even number of
active states of v.p. on X l .
Define the subgroup G(l) < AutX [k] , where l < k, as StabAutX [k] (l) |X l . It is plain,
that G(l) ≃StabAutX [k] (l) /StabAutX [k] (l+1) because StabAutX [k] (l + 1) is normal subgroup of
finite index in AutX [k] [8]. Let us construct a homomorphism from G(l) onto C2 in the
2l
2l
P
P
following way: ϕ(α) =
sli (α) mod 2. Note that ϕ(α · β) = ϕ(α) ◦ ϕ(β) = ( sli (α) +
i=1
2l
P
i=1
sli (β))mod2.
i=1
k
Structure of subgroup G2k Gk ′ ⊳ ≀ S2 ≃ AutX [k] can be described in next way. This
1
subgroup contains the commutant G′k . So it has on each X l , 0 ≤ l < k − 1 all even
12
indexes that can exists there. On X k−1 it does not exist v.p. of type T, which has the
distance 2k − 2, rest of even the indexes are present on X k−1 . It’s so, because the sets of
elements of types T and C are not closed under operation of calculating the even power as
it proved in Lemma 6. Thus, the squares of the elements don’t belong to T and C (because
they have the distance, which is less than 2k − 2). This implies the following corollary.
Corollary 5. A quotient group
Gk
/G2 G′ is isomorphic to C2 × C2 × ... × C2 .
{z
}
|
k k
k
′
′
Proof. The proof is based on two facts G2k Gk ≃ G2k ⊳ Gk and G : G2k Gk = 2k . Construct
2l
P
a homomorphism from Gk (l) onto C2 in the following way: ϕ(α) =
sli (α) mod 2. Note
i=1
that ϕ(α · β) = ϕ(α) ◦ ϕ(β) = (
2l
P
i=1
sli (α) +
2l
P
sli (β))mod2, where α, β ∈ AutX [n] . Index
i=1
of α ∈ G2k on X l , l < k − 1 is even but index of β ∈ Gk on X l can be both even and odd.
Note that Gk (l) is abelian group and G2k (l) E Gk . Since words with equal logarithms to
all bases [12] belong to distinct cosets of the commutator, the subgroup G2k (l) is the kernel
of this mapping. Also we can use homomorphism ϕ which is described above and denote
it as ϕl , to map Gk (l) onto C2 the ker ϕl = G2k (l). Really if α from Gk (l) has odd number
of active states of v.p. on X l , l < k − 1 than ϕl (α) = 1 in C2 otherwise if this number is
even than α from ker ϕi so ϕl (α) = 0 hence ker ϕl = G2k (l). So Gk (l) /G2k (l) = C2 analogously
Bk (l)
/Bk2 (l) = C2 . Let us check that mapping (ϕ0 , ϕ1 , ..., ϕk−2, φk−1 ) is the homomorphism
from Gk to C2 × C2 × ... × C2 . By virtue of the fact that we can construct homomorphism
|
{z
}
k
ϕi from every factor Gk (i) of this direct product to C2 . The group Gk /G2k is elementary
abelian 2-group because g 2 = e, g ∈ G.
Parity of index of α · β on X l is equal to sum by mod2 of indexes of α and β hence
ϕl (α · β) = (ϕl (α) + ϕl (β)) because multiplication α · β in Gk does not change a parity of
index of β, β ∈ Gk on X l .
Really action of element of active group A = C2 ≀ C2 ≀ ... ≀ C2 from wreath power
{z
}
|
l−1
(C2 ≀ C2 ≀ ... ≀ C2 ) ≀ C2 on element from passive subgroup C2 of second multiplier from prod{z
}
|
l−1
uct gf, g, f ∈ (C2 ≀ C2 ≀ ... ≀ C2 )≀C2 does not change a parity of index of β on X l , if index of
|
{z
}
l−1
β was even then under action it stands to be even and the sum ϕ(α) mod 2 + ϕ(β) mod 2
will be equal to (ϕ(α) + ϕ(β)) mod 2, hence it does not change a ϕ(β). Since words with
equal logarithms to all bases [12] belong to distinct cosets of the commutator, the subgroup
G2k (l) is the kernel of this mapping. Let us define the permutations of the type 2 that act
on X1 and X2 , where X1 = {vk,1 , ..., vk,2k−1 }, X2 = {vk,2k−1 +1 , ..., vk,2k }, X1 ∪ X2 = X k
only by even permutations. Subgroup G2k G′ k acts only by permutations of type 2 on X1 ,
X2 , according to Statement 1.
The restriction G2k |X [k−1] acts only by permutations of the second type (elements of
it form a normal subgroup in Gk ) by parity of permutation on sets X1 and X2 . A
13
permutation of Type 1, where on X1 and X2 the group Gk can acts by odd as well as
by even permutations but in such way to resulting permutation on X k is always even.
The number of active states from subgroup Gk (k − 1) on X k−1 can be even as well as
odd. It means that on set of vertices of X k−1 over X1 i.e. vertices that are connected
by edges with vertices of X k−1 over X1 automorphism of Gk can contains odd number
of active states (and X2 analogously). Hence for a subgroup G(k − 1) ≃ Wk−1 such
that has the normal subgroup Gk 2 (k − 1) ⊳ Gk (k − 1) we construct a homomorphism:
φk−1 (Gk (Xk )) → C2 ≃ Gk (k − 1)/G2k (k−1) as product of sum by mod2 of active states
(sk−1,i ∈ {0, 1}, 0 < j ≤ 2k−2 if vij ∈ X1 and X2 corespondently) on each set X1 and X2 :
k−1
k−2
2P
2P
sk−1,i (α)(mod2). Where sk−1,i (α) = 1
sk−1,i (α)(mod2) ·
φα (X1 ) · φα (X2 ) =
i=1
i=2k−2 +1
k−1
if there is active state in vk−1,i , i < 2
+ 1 and sk−1,i (α) = 0 if there is no active state.
′
It follows from structure of Gk that φα (X1 ) = φα (X2 ) so it is 0 or 1. But G2k Gk admits
only permutations of Type 2 on X k so G2k Gk ′ (Xk ) ⊳ Gk (Xk ) because it holds a conjugacy
and it is a kernel of mapping from Gk (Xk ) onto C2 .
Hence for a subgroup Wk−1 such that has the normal subgroup Gk 2 (k − 1) ⊳ Wk−1
it was constructed a homomorphism: φk−1 (Wk−1 ) → C2 ≃ Gk (k − 1)/G2k (k−1) as
product of sum by mod2 of active states from X1 and X2 . As the result we have
Gk
/G2k ≃ C2 × C2 × ... × C2 .
{z
}
|
k
Considering that it was proved in Theorem 1 and Theorem 2 that Gk ≃ A2k we can
formulate next Corollary.
Corollary 6. The group Syl2 A2k has a minimal generating set with k generators.
′
Proof. Since quotient group of Gk by subgroup of Frattini G2k Gk has minimal set of generators from k elements because Gk /G2 G′ is isomorphic to linear p-space (p = 2) of dimension
k k
k (or elementary abelian group). Then according to theorems from [16] rk(Gk ) = k. It
means that A2k is a group with fixed size of minimal generating set.
Main Theorem. The set Sβ = {β0 , β1 , β2 , . . . , βk−2, τ }, where βi = αi , is a minimal
generating set for a group Gk that is isomorphic to Sylow 2-subgroup of A2k .
We have isomorphism of Gk and Syl2 (A2k ) from Theorem 2, the minimality of Sβ
following from Theorem 3 which said that Sβ has to contain an element of type T, Theorem
3 and Lemma 7 about minimal rank. Another way to prove the minimality of Sβ is given
to us by Corollary 5 about quotient by Frattini subgroup.
For example a minimal set of generators for Syl2 (A8 ) can be constructed by following
way, for convenience let us consider the next set:
Consequently, in such way we construct second generating set for A2k of k elements
that is less than in [5], and this set is minimal.
We will call diagonal base (Sd ) for Syl2 S2k ≃ AutX [k] such generating set that
has the property sjx (αi ) = 0 iff i 6= j, (for 1 ≤ x ≤ 2j ) and every αi , i < k has
odd number of active v.p. A number of no trivial v.p. that can be on X j is odd the
14
Picture 1.
number of ways to chose tuple of no trivial v.p. on X j for generator from Sd and equal
j
j−1
k
to 22 : 2 = 22 . Thus, general cardinality of Sd for Syl2S2k is 22 −k−1 . There is
minimum one generator of type T in Sd for Syl2 A2k . This generator can be chosen not
2
less than in C21k−2 C21k−2 = (2k−2) = 22k−4 ways. Thus, total cardinality of Sd for Syl2 A2k
k−1
2
is 22 −k−2(2k−2 ) .
And there are k generators in a minimal set of generators, therefore |φ(Gk )| = |Gk | : 2k
k
k
k
k
should be raised to the power of k. It equals to (|Gk | : 2k ) = (22 −2 : 2k ) = 2k(2 −k−2) .
k
As a result, we have 2k(2 −k−1) · (2k − 1)(2k − 2)(2k − 22 )...(2k − 2k−1 ).
Let us consider an examples of Syl2 An for a cases n = 4k + r, where r ≤ 3. The
structure of Syl2A12 is the same as of the subgroup H12 < Syl2 (S8 ) × Syl2(S4 ), for that
[Syl2 (S8 ) × Syl2(S4 ) : H12 ] = 2, |Syl2(A12 )| = 2[12/2]+[12/4]+[12/8]−1 = 29 . Also |Syl2(S8 )| =
27 , |Syl2(S4 )| = 23 , so |Syl2 (S8 ) × Syl2(S4 )| = 210 and |H12 | = 29 , because its index in
Syl2(S8 )×Syl2(S4 ) is 2. The structure of Syl2 (A6 ) is the same as of H6 < Syl2(S4 )×(C2 ).
Here H6 = {(g, hg )|g ∈ Syl2 (S4 ), hg ∈ C2 }, where
(
hg = e, if g|X2 ∈ Syl2 (A6 ),
(2)
hg = (5, 6), if g|X 2 ∈ Syl2(S6 ) \ Syl2A6 .
The structure of Syl2(A6 ) is the same as subgroup H6 : H6 < Syl2(S4 ) × (C2 ) where H6 =
{(g, h)|g ∈ Syl2 (S4 ), h ∈ AutX}. So last bijection determined by (2) giving us Syl2A6 ≃
Syl2S4 . As a corollary we have Syl2 A2k +2 ≃ Syl2S2k . The structure of Syl2(A7 ) is the
same as of the subgroup H7 : H7 < Syl2 (S4 )×S2 where H6 = {(g, h)|g ∈ Syl2(S4 ), h ∈ S2 }
and h depends of g:
(
hg = e, if g|X 2 ∈ Syl2A7 ,
(3)
hg = (i, j), i, j ∈ {5, 6, 7}, if g|X 2 ∈ Syl2S7 \ Syl2A7 .
The generators of the group H7 have the form (g, h), g ∈ Syl2(S4 ), h ∈ C2 , namely:
{β0 ; β1 , τ } ∪ {(5, 6)}. An element hg can’t be a product of two transpositions of the set:
(i, j), (j, k), (i, k), where i, j, k ∈ {5, 6, 7}, because (i, j)(j, k) = (i, k, j) but ord(i, k, j) = 3,
so such element doesn’t belong to 2-subgroup. In general elements of Syl2 A4k+3 have the
structure (3), where hg = (i, j), i, j ∈ {4k + 1, 4k + 2, 4k + 3} and g ∈ Syl2 S4k .
15
Also |Syl2(S8 )| = 27 , |Syl2(S4 )| = 23 , so |Syl2(S8 ) × Syl2(S4 )| = 210 and |H12 | = 29 ,
because its index in Syl2(S8 ) × Syl2 (S4 ) is 2. The structure of Syl2(A6 ) is the same as of
H6 < Syl2 (S4 ) × (C2 ). Here H6 = {(g, hg )|g ∈ Syl2(S4 ), hg ∈ C2 }.
The sizes of this groups are equal, really |Syl2(A7 )| = 2[7/2]+[7/4]−1 = 23 = |H7 |. In case
g|L2 ∈ S7 \ A7 we have C32 ways to construct one transposition that is direct factor in H
which complete Syl2S4 to H7 by one transposition : {(5, 6); (6, 7); (5, 7)}.
The structure of Syl2(A2k +2l ) (k > l) is the same as of the subgroup H2k +2l <
Syl2(S2k ) × Syl2(S2l ), for that [Syl2(S2k ) × Syl2(S2l ) : H] = 2. |Syl2(A2k +2l )| =
k
l
k
l
2[(2 +2 )!/2]+[(2 +2 )!/4]+....−1. Here H = {(g, hg )|g ∈ Syl2 (S2k ), hg ∈ Syl2(S2l }, where
(4)
(
h ∈ A2l , if g|X k−1 ∈ A2k ,
h : h|X 2 ∈ Syl2(S2l ) \ Syl2A2l , if g|X k ∈ Syl2S2k \ Syl2A2k .
The generators of the group H7 have the form (g, h), g ∈ Syl2(S4 ), h ∈ C2 , namely:
β0 ; β1 , τ ∪ (5, 6).
I.e. for element βσ (2i − 1) = 2σ(i) − 1, βσ (2i) = 2σ(i), σi ∈ 1, 2, ..., 2k−1 .
Let us present new operation ⊠ (similar to that is in [2]) as a even subdirect product
of SylS2i , n = 2k0 + 2k1 + ... + 2km , 0 ≤ k0 < k1 < ... < km and m ≥ 0, i.e. SylS2k1 ⊠
SylS2k2 ⊠ ... ⊠ SylS2kl = P ar(SylS2k1 × SylS2k2 × ... × SylS2kl ), where P ar(G) – set
of all even permutations of G. Note, that ⊠ is not associated operation, for instance
ord(G1 ⊠ G2 ⊠ G3 ) = |G1 × G2 × G3 | : 2 but ord((G1 ⊠ G2 ) ⊠ G3 ) = |G1 × G2 × G3 | : 4.
For cases n = 4k + 1, n = 4k + 3 it follows from formula of Legendre.
i
It is well known that the Syl2S2ki ≃ ≀kj=1
C2 . Since Sylow p-subgroup of direct product is
direct product of Sylow p-subgroups and fact that automorphism of rooted tree keeps an
vertex-edge incidence relation then we have AutX [k0] ×AutX [k1 ] ×...×AutX [km ] ≃ Syl2Sn ,
n = 2k1 + 2k2 + ... + 2kl , ki ≥ 0, ki < ki−1 . Let us denote a subgroup, that consists of all
even substitutions from Syl2Sn as AutX [k0] ⊠ AutX [k1 ] ⊠ ... ⊠ AutX [km ] , where a states of
v.p. on X k0 −1 ⊔ X k1 −1 ⊔ ... ⊔ X km −1 are related by congruence:
k −1
(5)
i
m 2X
X
ski −1,j (αi ) ≡ 0 (mod2).
i=0 j=1
Lemma 11. If number of active states on a last level of AutX [ki ] from AutX [k0 ] ⊠ ... ⊠
AutX [km ] is odd, then it is subdirect product of groups AutX [k0 ] , ... , AutX [km ] .
Proof. It is a quotient group which is a homomorphic image obtained by a mapping from
AutX [k0 ] ×AutX [k1 ] ×...×AutX [km ] ≃ Syl2 Sn to this quotient group. A kernel of ϕ consists
i −1
m 2kP
P
of all automorphisms which satisfy a congruence
ski −1,j (αi ) ≡ 1 (mod2). At once
i=0 j=1
from definition follows, that if number of states on last level of AutX [ki] is odd, then
a subgroup from the condition is subdirect product of groups AutX [k0] , ... , AutX [km ] .
Actually for every state of automorphism α from AutX [ki] on X l , l < ki − 1 we have
that (e, ..., e, αi , e, ..., e) belongs to AutX [k0 ] × AutX [k1] × ... × AutX [km] . An arbitrary
16
state from X ki−1 is included in AutX [k0] ⊠ AutX [k1] ⊠ ... ⊠ AutX [km] together with even
number of states from last levels of X [k0 ] , ... , X [km] . Analogous fact was proved in [1] for
a direct sum of permutations groups and for their subgroups which consists of all even
permutations. Our statement is a restiction on a Syl2Sn .
The Sylow subgroup Syl2(An ) has index 2 in Syl2(Sn ) and it’s structure: Syl2 S2k1 ⊠
Syl2S2k2 ⊠ ... ⊠ Syl2 S2kl .
Lemma 12. If n = 4k + 2, then the subgroup Syl2 An is isomorphic to Syl2S4k , where
k ∈ N.
Proof. Let us consider the subgroup H4k+2 = {(g, hg )|g ∈ Syl2 (S4k ), hg ∈ S2 }, where
(
hg = e, if g|X k ∈ Syl2(A4k+2 ),
(6)
hg = (4k + 1, 4k + 2), if g|X k ∈ Syl2 (S4k+2 ) \ Syl2(A4k+2 ).
For instance the structure of Syl2(A6 ) is the same as subgroup H6 : H6 < Syl2(S4 ) × (C2 ),
where H6 = {(g, h)|g ∈ Syl2 (S4 ), h ∈ AutX}. So last bijection determined by (6) give us
Syl2A6 ≃ Syl2S4 . As a corollary we have Syl2A2k +2 ≃ Syl2S2k .
Bijection correspondence between set of elements of Syl2(An ) and Syl2(S4k ) we have
from (6). Let’s consider a mapping φ : Syl2 (S4k ) → Syl2(A4k+2 ) if σ ∈ Syl2 (S4k )
then φ(σ) = σ ◦ (4k + 1, 4k + 2)χ(σ) = (σ, (4k + 1, 4k + 2)χ(σ) ), where χ(σ) is number of transposition in σ by module 2. So φ(σ) ∈ Syl2(A4k+2 ). If φ(σ) ∈ An then
χ(σ) = 0, so φ(σ) ∈ Syl2 (An−1 ). Check that φ is homomorphism. Assume that
σ1 ∈ Syl2(S4k \A4k ), σ2 ∈ Syl2(A4k ), then φ(σ1 )φ(σ2 ) = (σ1 , hχ(σ1 ) )(σ2 , e) = (σ1 σ2 , h) =
σ1 σ2 ◦ (4k + 1, 4k + 2), where (σi , h) = σi ◦ hχ(σi ) ∈ Syl2(A4k+2 ). If σ1 , σ2 ∈ S2k \A2k , then
φ(σ1 )φ(σ2 ) = (σ1 , hχ(σ1 ) )(σ2 , hχ(σ2 ) ) = (σ1 σ2 , e) = (a, e), where σ1 σ2 = a ∈ A4k+2 . So it
is isomorphism.
Let nm = 2k0 + 2k1 + ... + 2km , where 0 ≤ k0 < k1 < ... < km and m ≥ 0.
Theorem 5. If nm = 4k + 2, then the minimal set of generators for Syl2Anm has
m
P
ki
i=1
elements.
Proof. Actually, according to Lemma 12, Syl2A4k+2 is isomorphic to Syl2S4k . In Statement 2 it was proved that Syl2 S4k ≃ Syl2 S2k1 × ... × Syl2S2km , where 4k = 2k1 + ... + 2km ,
k1 < ... < km . On the other hand, Syl2S2ki ≃ AutX [ki ] , so there exists the homomorphism ϕ from every factor Syl2 S2ki onto C2ki . Such homomorphism was defined in Corollary 5 and in [9]. And what is more it is known that AutX [ki ] has a minimal generating set
of ki generators [9]. Thus, there exists the homomorphism from AutX [k1 ] × ... ×AutX [km ]
m
P
ki , where k1 = 1.
onto C2k1 × ... × C2km , so the rank of Syl2A4k+2 is
i=1
This result was confirmed by the algebraic system GAP. Actually, it was founded by
GAP that the minimal generating set for Syl2 A14 , Syl2A14 ≃ Syl2 S12 ≃ Syl2 S22 ×Syl2S23 ,
of 5 elements: (11, 12)(13, 14), (9, 11)(10, 12), (7, 8)(9, 10), (1, 5)(2, 6)(3, 7)(4, 8), (1, 3)(2, 4).
17
Lemma 13. If nm ≡ 1(mod2), then there exists a point n from tuple M of nm points
indexed by numbers from 1 to nm , such that StSyl2 Snm (n) is isomorphic to Syl2Snm acting
on a tuple M.
Proof. If Snm acts on M, then one of a Sylow 2-subgroups H < Snm is isomorphic to
AutX [k0 ] × AutX [k1] × ... × AutX [km] , that acts on tuple of nm points, wherenm = 2k0 +
2k1 + ... + 2km , k0 < k1 < ... < km . By virtue of the fact that nm ≡ 1(mod2), then
k0 = 0. Thus point n, that is in X [k0 ] has a stabilizer StH (n) ≃ Syl2Snm . It is so, because
group of automorphism of such group keeps an vertex-edge incidence relation of X [ki ] , i ∈
{0, ..., m}. Thus, action of every Sylow 2-subgroup of Snm , where nm ≡ 1(mod2), fix one
element from {1, 2, ..., nm}.
According to the Sylow theorem all Sylows p-subgroups are conjugated so their actions
are isomorphic. In particular, a Sylow 2-subgroup of S2r is self-normalizing. The number
of Sylow 2-subgroups of S2r is 2r ! : 2e where e = 1 + 2 + ... + 2r−1 [14].
Remark 1. The mentioned isomorphism may be extended to Syl2 A4k+3 ≃ Syl2A4k+2 ≃
Syl2S4k+1 ≃ Syl2 S4k .
Proof. Since in accordance with Lemma 13 an action of Syl2A4k+3 on the set of 4k + 3
elements fixes one point, then this group as group of action is isomorphic to Syl2A4k+2 .
For a similar reason Syl2A4k+1 ≃ Syl2 A4k . As well as it was proved in Lemma 12 that
Syl2A4k+2 ≃ Syl2S4k .
The number of generating sets for AutX ki is not less then Nki = 1 · 2 · 22 · ... · 2ki −1 =
ki (ki −1)
k1 (k1 −1)
2 2 . Hence, the number of generating sets for Syl2A4k+2 is not less than 2 2 · ... ·
km (km −1)
. Thus, it can be applied in cryptography [18].
2 2
Property 1. Relation between sizes of the Sylows subgroup for n = 4k − 2 and n = 4k
is given by |Syl2 (A4k−2 )| = 2i |Syl2 (A4k )|, where value i depends only of power of 2 in
decomposition of prime number of k.
, therefore |A4k | =
Proof. Really |A4k−2 | = (4k−2)!
2
mines only by k and is not bounded.
(4k−2)!
(4k
2
− 1)4k, it means that i deter
Proposition 2. If n = 4k, then index Syl2(An+3 ) in An+3 is equal to [S4k+1 :
Syl2(A4k+1 )](2k + 1)(4k + 3), index Syl2(An+1 ) in An+1 as a subgroup of index 2m−1 ,
where m is the maximal natural number, for which 4k! is divisible by 2m .
Proof. For Syl2(An+3 ) its cardinality equal to maximal power of 2 which divide (4k + 3)!
this power on 1 grater then correspondent power in (4k + 1)! because (4k + 3)! = (4k +
1)!(4k + 2)(4k + 3) = (4k + 1)!2(2k + 1)(4k + 3) so | Syl2An+3 |= 2m · 2 = 2m+1 . As a
and
result of it indexes of An+3 and An+1 are following: [S4k+1 : Syl2(A4k+1 )] = (4k+1)!
2m
(4k+1)!
[S4k+3 : Syl2(A4k+3 )] = [S4k+1 : Syl2 (A4k+1 )](2k + 1)(4k + 3) = 2m (2k + 1)(4k + 3).
Proposition 3. If n = 2k then [Syl2(An ) : Syl2(Sn−1 )] = 2m−1 , where m is the maximal
power of 2 in factorization of n.
18
Proof. |Syl2(Sn−1 )| is equal to t that is a maximal power of 2 in (n − 1)!. |Syl2(An )|
is equal to maximal power of 2 in (n!/2). Since n = 2k then (n/2)! = (n − 1)! n2
and 2f is equal to product maximal power of 2 in (n − 1)! on maximal power of 2 in
m−1
|Syl2 (An )|
n
. Therefore |Syl
= 2 2t 2t = 2m−1 . Note that for odd m = n − 1 the group
2
2 (Sn−1 )|
Syl2(Sm ) ≃ Syl2 (Sm−1 ) i.e. Syl2(Sn−1 ) ≃ Syl2(Sn−2 ). The group Syl2(Sn−2 ) contains
the automorphism of correspondent binary subtree with last level X n−2 and this automorphism realizes the permutation σ on X n−2 . For every σ ∈ Syl2(Sn−2 ) let us set in
correspondence a permutation σ(n − 1, n)χ(σ) ∈ Syl2(An ), where χ(σ) – number of transposition in σ by mod 2, so it is bijection φ(σ) 7−→ σ(n − 1, n)χ(σ) that has property of
homomorphism, see Lemma 12. Thus, we prove that Syl2(Sn−1 ) ֒→ Syl2 (An ) and its
index is 2d−1 .
Proposition 4. The ratio of |Syl2(A4k+3 )| and |Syl2(A4k+1 )| is equal to 2 and ratio of
indexes [A4k+3 : Syl2(A4k+3 )] and [A4k+1 : Syl2 (A4k+1 )] is equal (2k + 1)(4k + 3).
Proof. The ratio |Syl2 (A4k+3 )| : |Syl2(A4k+1 )| = 2 holds because formula of Legendre gives
us new one power of 2 in (4k + 3)! in compering with (4k + 1)!. Second part of statement
follows from theorem about p-subgroup of H, [G : H] 6= kp then one of p-subgroups of H
is Sylow p-group of G. In this case p = 2 but |Syl2(A4k+3 )| : |Syl2(A4k+1 )| = 2 so we have
to divide ratio of indexes on 2.
Proposition 5. If n = 2k + 1 then Syl2 (An ) ∼
= Syl2(An−1 ) and Syl2 (Sn ) ∼
= Syl2 (Sn−1 ).
Proof. Sizes of these subgroups are equal to each other according to Legender’s formula
which counts power of 2 in (2k + 1)! and (2k)! we obtain that these powers are equal.
So these maximal 2-subgroups are isomorphic. From Statment 1 can be obtained that
vertex with number 2k + 1 will be fixed to hold even number of transpositions on X k1
from decomposition of n which is in Statement 1. For instance Syl2 (A7 ) ≃ Syl2(A6 ) and
by the way Syl2 (A6 ) ≃ C2 ≀ C2 ≃ D4 , Syl2 (A11 ) ≃ Syl2 (A10 ) ≃ C2 ≀ C2 ≀ C2 .
Proof. Sizes of these subgroups are equal to each other according to Legender’s formula
which counts power of 2 in (2k + 1)! and (2k)! we obtain that these powers are equal.
So these maximal 2-subgroups are isomorphic. From Statment 1 can be obtained that
vertex with number 2k + 1 will be fixed to hold even number of transpositions on X k1
from decomposition of n which is in Statement 1. For instance Syl2 (A7 ) ≃ Syl2(A6 ) and
by the way Syl2 (A6 ) ≃ C2 ≀ C2 ≃ D4 , Syl2 (A11 ) ≃ Syl2 (A10 ) ≃ C2 ≀ C2 ≀ C2 .
Let us denote by S(n) and S[n] a minimal generating system of Syl2 (An ), Syl2 (Sn )
correspondently. Let nm = 2k0 + 2k1 + ... + 2km , where 0 ≤ k0 < k1 < ... < km and m ≥ 0.
Definition 4. We shall call the top of system of generators a portion of generators from
subgroup with the largest degree km , which belong to tuples, that represent elements from
subdirect product, where there are not other non-trivial elements.
Theorem 6. Any minimal set of generators for Syl2An has
and it has k0 generators, if m = 0.
m
P
i=0
ki −1 generators, if m > 0,
19
Proof. Construction of generating set S(nm ) is such that it contains on the first coordinate
all generators from Syl2S2k0 , it contains all generators of Syl2S2k1 on the second one,
analogously it contains on i-th coordinate all generators of Syl2S2ki , 0 < i ≤ m. Hence
we can generate on i-th coordinate an arbitrary element from Syl2 S2ki . States on X k0 −1 ⊔
X k1 −1 ⊔ ... ⊔ X km −1 are related by congruence:
k −1
(7)
i
m 2X
X
ski −1j (αi ) ≡ 0 (mod2)
i=0 j=1
This congruence holds due to structure of constructed by us system of generators.
Let us consider a construction of a system of generators for Syl2An and prove
its minimality. Consider an induction base on the example of the group A28 , since
A28 = Syl2A4 × Syl2A8 × Syl2 A16 , according to the proved fact it is equivalent to a
construction of minimal system of generators for Syl2 A28 = Syl2S4 ⊠ Syl2S8 ⊠ Syl2 S16 .
All elements of this system define odd substitutions on the respective sets X 2 and X 3
by the generators of the subgroups AutX [1] = Syl2 S4 = α0,1(1) , α1 , and Syl2S8 where
α0,2(1) , α1(2),2(1) , α2 = AutX [3] = Syl2 S8 , α2 = α2(1) . Last subgroup of the subdirect product has the next generating system Syl2S16 = α0 , α1(1) , α2(1) , α3(1) hence
here only the last element determines odd permutation on X 4 . Other generators for
Syl2S8 can be decomposed in a product of two special elements α2 αi = αi(1);2(1) . We
call such structure of element as odd structure. Generators of Syl2S4 , 0 ≤ i < 2
have the same structure α1 α0 = α0;1(1) , where 0 ≤ i < 1. Hence a product of two
arbitrary elements from the set α0,2(1) , α1(2),2(1) , α2 is automorphism that
has even
2
number of active states of v.p. on X . Analogously statement is true for α0,1(1) , α1
on X 1 and for α1(1) , α2(1) , α3(1) on X 3 . Now rename generators of group Syl2 S4 as
β0,1(1) = t0 , β1 = t1 , for the second group as α0,1(1),2(1) = s0 , α1,2(1) = s1 , α2 = s2 ,
rename generators for Syl2 S16 as f0 = α0 , ... , f3 = α3(1) . A characteristic feature of
the built elements is, that those of them, which are generators of the same subgroup
Syl2S2i and implement odd substitutions, contains the unique common 2-cycle, that
implements odd substitution on the set, where the correspondent group acts. Thus,
for AutX [2] it is the cycle (1, 2) on set with 4 elements. Therefore, according to
the given by us structure of generators of groups Syl2 S2i , those of them, that implement odd substitutions, their products and their squares, are already even substitutions. The system of generators for Syl2 A28 , Syl2A28 = Syl2 S4 ⊠ Syl2S8 ⊠ Syl2S16 is
h(s0 , t0 , e), (t0 , e, f3 ), (e, s1 , f3 ), (s1 , t0 , e), (e, t1 , f3 ), (e, t2 , f3 ), (e, e, f2 ), (e, e, f1 ), (e, e, f0 )i
that is why substitutions, that are defined by generators f2 , f1 , f0 must be even (we
shall call just this set as top) and f3 must has odd structure.
Since for the group Syl2 S2ki there exists a surjective homomorphism (ϕ0 , ϕ1 , ϕ2 , ..., ϕki −1 )
onto C2ki , where ϕi is homomorphism from Gki (l) onto C2 , so there exists a surjective
homomorphism from Syl2 S4 × Syl2S8 × Syl2S16 onto C29 [9]. We apply it to every Gki (l),
which are the subgroups of Syl2S4 × Syl2 S8 × Syl2 S16 , where ki ∈ {2, 3, 4} and l changes
from 0 to 2ki − 1 for each ki .
20
Show, that there is enough
m
P
ki − 1 generators in general case, when m > 0. Arbitrary
i=0
element from can be generated on the first coordinate as follows (s0 , β1 , e)(s1 , β1 , e) =
(α0 , e, e) and (α0 , e, e)(s0 , β1 , e) = (α1(2) , β1 , e) then (α1 , β1 , e)(α0,1(1) , β1 , e) = (α1(12) , e, e).
Hence, we have derived on the first coordinate all non-trivial elements from Syl2 A4 .
All automorphisms, which has odd structure, provides an odd permutations on X ki ,
which contain in its cyclic decomposition the unique cycle of length 2 (in common case it
is odd number of such cycles) that rearranges neighboring elements from X ki and arises
under the action of active state from X ki −1 , all other cycles in substitutions are even,
because they are defined by states from X k−j , 1 < j < ki . Thus, in the multiplication of
these automorphisms, reduction of such cycle is obtained or one more conjugated to it 2
cycle appears in the correspondent substitution.
The considered by us group Syl2S4 ⊠ Syl2 S8 ⊠ Syl2S16 is a quotient group of Syl2S4 ×
Syl2S8 × Syl2S16 by normal closure RG , where R is the relation that can be derived from
the congruence 7. Thus from the group Syl2S4 ⊠ Syl2S8 ⊠ Syl2 S16 has on 1 generator less
than Syl2 S4 × Syl2S8 × Syl2S16 because we have new relation from which this generator
can be expressed from form t−1 ω, ω ∈ Syl2A28 and t is one of generators for Syl2 S4 ×
Syl2S8 × Syl2S16 . Hence now t can be express due to a new relation similar as it was in
transformation of Tietce. Thus we have on one generator less because t is depends from
rest of generators for Syl2 S4 × Syl2S8 × Syl2S16 .
For the group Syl2A28 there exists a surjective homomorphism (ϕ0 , ϕ1 , ϕ2 , ϕk3 ) onto
8
C2 , which is built previously. We apply it to every Gki (l), which are the subgroups
of Syl2S4 ⊠ Syl2S8 ⊠ Syl2 S16 , where ki ∈ {2, 3, 4} and l changes from 0 to 2ki − 1
for each ki . The main property of a homomorphism holds for all Gki (l). Actually, a
parity of index of α · β on X l is equal to sum by mod 2 of indexes of α and β hence
ϕ(αl · βl ) = (ϕ(αl ) + ϕ(βl )) because multiplication α · β in Gk does not change a parity of
index of β, β ∈ Gk on X l . This parity depends only from parities of α and β on X l . That’s
why the described mapping is a homomorphism onto C2 . The same is true for every l ∈
{0, ..., 3} and for every ki ∈ {2, 3, 4} from Gki (l). A group Anm has the similar structure
Syl2S2k0 ⊠ . . . ⊠ Syl2 S2km−1 ⊠ Syl2 S2km and main property of constructed homomorphism
ϕ(αl · βl ) = (ϕ(αl ) + ϕ(βl )) holds by the same reason for every, l ∈ {0, ..., km − 1}.
Let us separate out a subgroup which corresponds to the last level, at which the active
states exist. The parity of the index of this subgroup, let it be G4 (3), depends also on
the parities of all other level subgroups of AutX [2] ⊠ AutX [3] ⊠ AutX [4] because relation
7 holds. Thus Syl2A28 has minimal generating system from 8 elements.
Let us present S(28) in form useful for generalization and for counting rank of S(28):
h(s0 , e, f3 ), (s1 , e, f3 ), (e, t0 , f3 ), (e, t1 , f3 ), (e, t2 , f3 ), (e, e, f2 ), (e, e, f1 ), (e, e, f0 )i.
So rk(S(28))=8. Obviously that generator f3 from generating system of Syl2S16 occurs
only in combination with generators of other subgroups and any other of generators occurs
m
P
only one time. Thus formula
ki −1 holds. This result was confirmed by algebraic system
i=0
GAP by which it was founded the minimal generating set for Syl2A28 from 8 elements:
(25, 27)(26, 28), (23, 24)(25, 26), (17, 21)(18, 22)(19, 23)(20, 24), (17, 19)(18, 20),
21
(15, 16)(17, 18), (1, 9)(2, 10)(3, 11)(4, 12)(5, 13)(6, 14)(7, 15)(8, 16), (1, 5)(2, 6)(3, 7)(4, 8),
(1, 3)(2, 4). Thus, the base of induction is checked.
The construction of homomorphism and minimal generating set for case n = 2k was
fully investigated in Corollary 5.
m
P
Show, that there is enough
ki − 1 generators in general case, when m > 0. There
i=0
are all ki generators of Syl2 A2ki on i-th coordinate. Consider the example from the
base of induction arbitrary element from Syl2 A4 can be generated on the first coordinate as follows (s0 , β1 , e)(s1 , β1 , e) = (α0 , e, e) and (α0 , e, e)(s0 , β1 , e) = (α1(2) , β1 , e) then
(α1 , β1 , e)(α0,1(1) , β1 , e) = (α1(12) , e, e). Hence, we have derived on the first coordinate all
non-trivial elements from Syl2A4 . All automorphisms, which has odd structure, provides
an odd permutations on X ki , which contain in its cyclic decomposition the unique cycle
of length 2 (in common case it is odd number of such cycles) that rearranges neighboring
elements from X ki and arises under the action of active state from X ki −1 , all other cycles
in substitutions are even, because they are defined by states from X k−j , 1 < j < ki .
Thus, in the multiplication of these automorphisms, reduction of such cycle is obtained
or one more conjugated to it 2 cycle appears in the correspondent substitution.
So, closure holds with respect to parity of obtained tuples of elements as a result of
multiplication in the whole subdirect product, because in the product automorphism
is reduced on the second coordinate, which is determined by the same state, that is
determined by β3 . It is so because it has the order 2 therefore e are on other coordinates,
and automorphism on the first coordinate becomes an even substitution already after the
first multiplication by (s0 , β1 , e). Indeed, system of generators for ht0 , t1 i has such cyclic
structure, that the same cycle implements the unique odd substitution - this is the cycle
of length 2, therefore it is reduced as a result of multiplication of two such generators and
even substitution is obtained, the generators s0 , s1 , s2 , f0 , f1 , f2 , f3 for Syl2A8 and Syl2 A16
correspondently have the same structure, so this is the system of generators, that has only
7 elements, while the direct product of these groups should have 9 elements. Thus, for
this example the statement is true. Let us show, that minimal system of generators can
be constructed by the same method for Anm that contains in subdirect product the factor
Syl2S2km .
All generator except those from the top of generating system for Syl2 S2km has odd
structure. For Syl2S2km we choose one generator that realize odd permutation on X km
analogously as αkm −1 and other generators αk0 , ..., αkm −2 belong to top of system of generators of Syl2S2km , so they define even substitutions on X km . Here every separately
taken generator for the separated subgroup Syl2S2ki , ki < km defines odd substitution
and has odd structure, only generators from the top of system of generators of the subgroup Syl2S2km implement even substitutions. Only one generator from S(2km ) has odd
structure and only this generator stands in pair with generator from S(2km −1 ) in tuple
of generators of subdirect product for Syl2S2k0 ⊠ Syl2S2k1 ⊠ ... ⊠ Syl2S2km . Thus rank
of S(nm ) in comparing to S(nm−1 ) will grows on km . Thus rank of S(nm ) became to be
m
P
ki − 1 as it states in formula from condition.
i=1
22
For the subgroup Syl2S2ki , i < m, distinguish its subgroups of the form Gki (l), l <
ki − 1 elements from this subgroup act on X ki by even substitutions and therefore are
not linked with elements Gkj (s), s < kj − 1, where i 6= j, ki − s = kj − l by parity
relation and they forming a direct product. States from last levels X k0 , ..., X km are linked
by parity relation and form the subdirect product, in which there is only even number of
states in every tuple with m + 1 coordinates. Construction of generating system S(nm ) is
such that it presents on the first coordinate all generators from Syl2 S2k0 , it presents all
generators of Syl2S2k1 on the second one, analogously it presents on i-th coordinate all
generators of Syl2 S2ki , 0 < i ≤ m. Hence we can generate on i-th coordinate an arbitrary
element from Syl2 S2ki . A states on last level of X [k0 −1] ⊔ X [k1−1] ⊔ ... ⊔ X [km−1] are related
by congruence 7. And form the subgroup of H = Syl2S2k0 ⊠ ... ⊠ Syl2 S2km that is a
quotient group of G = Syl2 S2k0 × ... × Syl2S2km by normal c closure RG , where R is
the relation equivalent to relation 7. Since as in proof of induction base we have that
the group H has on 1 generator less than G because we have new relation 7 that does
not belongs to normal closure of relations from the group G and R does not contains
new generators than those for G, therefore one of old generators can be expressed. This
congruence holds due to the structure of constructed by us system of generators.
To prove that proposed by us system generates whole group Syl2An let us prove that
k
km
there is equality of orders of hS(n)i and Syl2 An . Order of Syl2Sn is 22 0 +...+2 −(m+1) =
k
k
2n−m−1 because SylS2k ≃ AutX [2 ] and its order is equal to 22 −1 , where n = 2k0 +
2k1 + ... + 2km . Thus |Syl2 An | = 2n−m−2 . But order of constructed by us quotient group
Syl2S2k0 ⊠Syl2 S2k1 ⊠ ... ⊠Syl2S2km is also equal to 2n−m−2 because the relation 7 restricts
class of states from last level of X [k0 −1] ⊔ X [k1 −1] ⊔ ... ⊔ X [km −1] to the subclass that realize
even permutations on X n . The same order for n!2 give us Legender’s formula for finding
n!
maximal
power
:
2
n
nof2 which kdivides
k1
0
+ ... + 2km − 1 = 2 + 2 + ... + 2km − 1 = (n − m − 1) − 1 = n − m − 2.
2
Also from Statement 1 and corollary from it about (AutX [k] )′ can be deduced that
derived length of Syl2Ak2 is not always equal to k as it was said in Lemma 3 of [2] because
in case A2k if k = 2 its Syl2A4 ≃ K4 but K4 is abelian group so its derived length is 1.
4. Some applications of constructed generating systems
As it was calculated in previous paragraph a total cardinality of Sd for Syl2A2k is
2
2
(2k−2 ) . Thus, if we associate generating set with alphabet and choice of generating set will be a private key, then it can be applied in cryptography [18]. More over
our group Gk has exponential grows of different generating sets and diagonal bases that
can be used for extention of key space. Diagonal bases are useful for easy constructing of
normal form [23] of an element g ∈ Gk .
Let us consider a function of Morse [19] f : D 2 → R that painted at pict. 2 and
graph of Kronrod-Reeb [21] that obtained by contraction every set’s component of level
of f −1 (c) in point. Group of automorphism of this graph is isomorphic to Syl2S2k , where
k = 2 in general case we have regular binary rooted tree for arbitrary k ∈ N.
2k−1 −k−2
23
Picture 2.
According to investigations of [22] for D 2 we have that Syl2S2k > Gk ≃ Syl2A2k is
quotient group of diffeomorphism group that stabilize function and isotopic to identity.
Analogously to investigations of [21, 22, 24] there is short exact sequence 0 → Zm →
π1 Of (f ) → G → 0, where G-group of automorphisms Reeb’s (Kronrod-Reeb) graph [21]
and Of (f ) is orbit under action of diffeomorphism group, so it could be way to transfer
it for a group Syl2 (S2k ), where m in Zm is number of inner edges or vertices in Reeb’s
graph, in case for Syl2S4 we have m = 3.
Higher half of projection of manifold from pic. 2 can be determed by product of the
quadratic forms −((x + 4)2 + y 2)((x + 3)2 + y 2 )((x − 3)2 + y 2)((x − 4)2 + y 2) = z in points
(−4, 0)(−3, 0)(3, 0)(4, 0) it reach a maximum value 0. Generally there is −d21 d22 d23 d24 = z.
5. Conclusion
The proof of minimality of constructed generating sets was done, also the description
of the structure Syl2A2k , Syl2An and its property was founded.
References
[1] Alexey Muranov, Finitely generated infinite simple groups of infinite commutator width.
arXiv:math/0608688v4 [math.GR] 12 Sep 2009.
[2] U. Dmitruk, V. Suschansky, Structure of 2-sylow subgroup of symmetric and alternating group. UMJ.
(1981), N. 3, pp. 304-312.
[3] R. Skuratovskii "Corepresentation of a Sylow p-subgroup of a group Sn". Cybernetics and systems
analysis, (2009), N. 1, pp. 27-41.
[4] B. Pawlik The action of Sylow 2-subgroups of symmetric groups on the set of bases and the problem
of isomorphism of their Cayley graphs. Algebra and Discrete Mathematics. (2016), Vol. 21, N. 2, pp.
264-281.
[5] V. Ivanchenko, System of generators for 2-sylow subgroup alternating group, Four ukraine conference
of young scientists. Kiev: KPI (2015, (http://matan.kpi.ua/uk/ysmp4conf.html)), pp. 60.
[6] R.V. Skuratovskii, Y.A. Drozd Generators and and relations for wreath products of groups. Ukr
Math J. (2008), vol. 60. Issue 7, pp. 1168–1171.
[7] R. Grigorchuk, V. Nekrashevich, V. Sushchanskii, Automata, Dynamical Systems, and
Groups, Trudy mat. inst. imeny Steklova. №231. (2000) 134–214.
[8] V. Nekrashevych, Self-similar groups. International University Bremen. American Mathematical Society. 2005. Monographs, Vol. 117, 230 p.
[9] R.I. Grigorchuk, Solved and unsolved problems around one group. Infinite Groups:
Geometric, Combinatorial and Dynamical Aspects. Вasel, (2005). Progress Math., vol 248.
pp. 117-218.
24
[10] V. Magnus, A. Karras, D. Soliter, Combinatorial Group Theory: Presentations of Groups
in Terms of Generators and Relations. – M.: Science, 1974, 453 p.
[11] V.S. Sikora, V. I. Suschanskii, Operations on groups of permutations. Chernivci: Ruta,
2003, 256 p.
[12] M. I. Kargapolov, J. I. Merzljakov, Fundamentals of the Theory of Groups Springer.
Softcover reprint of the original 1st ed. 1979 edition. Springer. 1st ed. 1979. 312 p.
[13] R. Skuratovskii. Generators and relations for sylows р-subgroup of group Sn . Naukovi Visti
KPI. 4 (2013)
[14] Louis Weisner, On the Sylow Subgroups of the Symmetric and Alternating Groups. American Journal of Mathematics, (Apr., 1925), vol. 47, Issue 2, pp. 121-124.
[15] Oleg Bogopolski, An Introduction to the Theory of Groups. European Mathematical Society.
2008. — 189 p.
[16] J. J. Rotman, An introduction to the Theory of Groups. New Yourk. Springer, 1995, XV.
513 p.
[17] V. V. Sharko Smooth topological equivalence of functions of surfaces. Ukrainian Mathematical Journal, 5 (2003), 687-700
[18] A.G.Myasnikov, V. Shpilrain, A.Ushakov, A practical attack on some braid group based
cryptographic protocols, Crypto 2005, Springer Lect., Notes Comp. (2005), Sc. 3621, pp.
86-96.
[19] V. V. Sharko, Smooth topological equivalence of functions of surfaces. Ukrainian Mathematical Journal, May (2003), Volume 55, Issue 5, pp. 832–846.
[20] J.D.P. Meldrum, Wreath Products of Groups and Semigroups. Pitman Monographs and
Surveys in Pure and Applied Mathematic English. 1st Edition. Jun (1995). 425 p.
[21] S. I. Maksymenko, Homotopy types of stabilizers and orbits of Morse functions on surfaces.
Annals of Global Analysis and Geometry. (2006), May vol. 29, issue 3, pp. 241–285.
[22] S. I. Maksymenko, Path-components of Morse mappings space Some problems of
contemporary mathematics. Proceedings of Intsitute of Mathematics of Ukrainian NAS. 1998,
vol. 25, pp. 135-153.
[23] V. Shpilrain, A. Ushakov. A new key exchange protocol on the decomposition problem.
Contemp. Math. Volume 418. (2006), pp. 161-167.
[24] R. V. Skuratovskii, Minimal generating systems and properties of Syl2 A2k and Syl2 An . X
International Algebraic Conference in Odessa dedicated to the 70th anniversary of Yu. A.
Drozd. (2015), pp. 104.
[25] R.
V.
Skuratovskii,
Minimal
generating
systems
and
structure
of
and
Syl2 An .
International
Conference
and
PhD-Master
Summer
Syl2 A2k
School on Graphs and Groups,
Spectra and Symmetries. (2016),
source:
http://math.nsc.ru/conference/g2/g2s2/exptext/Skuratovskii-abstract-G2S2+.pdf.
[26] R. V. Skuratovskii, Structure of commutant and centralizer of Sylow 2-subgroups of
alternating and symmetric groups, minimal generating sets of Syl2 An , its applications in
cryptography. CAIM 2017, ROMAI. Algebra, Logic and Geometry (2017), pp. 74-75, source:
http : //www.romai.ro/documentepoze/Conf erinte/Caim17/CAIM 2017 2.pdf
[27] R. V. Skuratovskii, Commutant of Sylow 2-subgroups of alternating group and
minimal generating sets of Syl2 An . Mal’tsev Meeting Novosibirsk, (2017), source:
http://math.nsc.ru/conference/malmeet/17/Main.htm
| 4math.GR
|
arXiv:1711.06221v1 [stat.ML] 16 Nov 2017
A Forward-Backward Approach for Visualizing
Information Flow in Deep Networks
Aditya Balu, Thanh V. Nguyen, Apurva Kokate, Chinmay Hegde, Soumik Sarkar∗
Iowa State University
Ames, IA 50010
Abstract
We introduce a new, systematic framework for visualizing information flow in
deep networks. Specifically, given any trained deep convolutional network model
and a given test image, our method produces a compact support in the image
domain that corresponds to a (high-resolution) feature that contributes to the given
explanation. Our method is both computationally efficient as well as numerically
robust. We present several preliminary numerical results that support the benefits
of our framework over existing methods.
1
Introduction
Deep neural networks have resulted in widespread and compelling advances in a variety of machine learning tasks such as object recognition, image segmentation, anomaly detection, machine
translation, and synthesis. However, these advances have often been accompanied by a significant
reduction in interpretability, or the ability to visualize the flow of information being extracted at
various layers of abstraction. In contrast to traditional rule-based learning methods (which search for
specific, semantically hand-crafted features or patterns), deep networks often produce decisions that
are seemingly hard to decipher or justify for a given test data sample, even though their aggregate
generalizability measured with respect to a hold-out test dataset is excellent. This issue of unpacking the “black-box” nature of deep networks has been identified as a key issue by several recent
works Springenberg et al. [2014], Selvaraju et al. [2016], Shrikumar et al. [2017], Sundararajan et al.
[2016].
In this work, we focus on the task of object detection in images. Broadly, algorithms for interpreting
the action of deep networks for this task can be grouped as follows: Class-discriminative approaches,
such as Class Activation Mappings (CAM) Zhou et al. [2016], or its gradient-based variant Selvaraju
et al. [2016], produce a support in the original image domain that approximately corresponds to
a given object class detected in that image. However, such methods are coarse and only produce
low-resolution visualizations, and as such cannot be directly applied to very high resolution images.
On the other hand, pixel-space gradient-based methods such as deconvolution networks Zeiler and
Fergus [2014] and guided back-propagation Springenberg et al. [2014] produce fine-grained features
in a given image. However, gradient based methods suffer from either significant computational
efficiency concerns, or are susceptible to saturation phenomena due to vanishing/exploding gradients,
or both. In Shrikumar et al. [2017], this issue is alleviated by suitably using a second reference
input to stabilize the estimates. However, choosing this reference image is qualitative and can be
challenging. Finally, model-agnostic approaches such as LIME Ribeiro et al. [2016] are theoretically
sound and can be applied for interpreting deep convolution networks, but involve solving challenging
optimization problems.
∗
{baditya,thanhng,akokate,chinmay,soumik*@iastate.edu}.
Interpretable ML Symposium, 31st Conference on Neural Information Processing Systems (NIPS 2017), Long
Beach, CA, USA.
Dense Layer
Input Image
Conv. Filters
Dropout
Output Layer
(Classification)
Max.pooling
Relu (Feature maps)
F
B
F
B
F
F
B
Output Image
B
:Dropout zero
:Masked zero
:Selected Filters
Figure 1: Overview of the forward-backward scheme for visualizing the information flow. Forward information
(F) of the model is combined with the backward information (B) and flow throughout the network.
In this short paper, we outline a systematic framework for visualizing information flow in deep
convolutional networks that resolves both the computational efficiency as well as the numerical
robustness issues described above. We present several preliminary numerical results that support the
benefits of our framework over existing methods.
At a high level, our approach is based on a novel forward-backward scheme which operates as follows.
Consider a trained deep convolutional network model and a given test image for which our model is
able to identify the existence of a given target class. Then, our algorithm produces as output, a support
(i.e., a subset of pixel locations) corresponding to the class predicted by our model in a manner similar
to pixel-space gradient methods. However, in contrast with gradient-based approaches, our algorithm
not only leverages the backward (class) information flow from the output layer(s) to the input, but
also the the forward (image) information extracted at various layers of abstraction. See Figure 1.
More specifically, our method has the following distinguishing characteristics:
1. We propose a mathematically principled approach to achieve “backward information flow” within
a deep convolutional network, leveraging the ideas proposed in the deconvolutional networks
approach of Zeiler and Fergus [2014]. However, this approach is computationally very expensive
since it requires solving a sparse recovery problem for each layer of the network, and this limits the
depth of a network on which this method is applicable. On the other hand, our approach only needs
simple application of matrix adjoints and (element-wise) nonlinearities for each convolutional
layer and can be easily implementable on very deep networks.
2. We propose a systematic way of using the forward information to guide the backward-traversal. In
particular, we use the forward information to extract a support within a given layer of representation
that best corresponds to a specific feature map. We achieve this using a novel masking scheme
which transparently combines both forward and backward information flows through the network.
3. As opposed to gradient-based schemes (such as Selvaraju et al. [2016], Springenberg et al. [2014])
that aggregate the information from all feature maps, our algorithm produces binary support
estimates layer by layer. In that sense, our method avoids any numerical stability and robustness
issues that may arise via the well-known problem of exploding/vanishing gradients that can
potentially affect the interpretability. In particular, in contrast with Shrikumar et al. [2017], we
remove the need for any separate reference image, and our method only involves making two
passes through the network for a given image.
We present preliminary numerical evidence supporting our method, and demonstrate advantages over
gradient-based methods such as guided backpropagation Springenberg et al. [2014].
2
2
Proposed Approach: Forward-Backward Interpretability
We now describe our scheme for visualizing a convolutional neural network. We term our method
Forward-Backward Interpretability (or FBI for short). Given a test image, the goal is to identify
important regions that explain the prediction of the learned network. To do this, we propagate the
class-probed information back to the image pixel space through the complete network, using the
guidance of learned model weights as well as the forward activations of each neuron in the network.
Our approach shares several similarities with the deconvolutional networks (DeconvNet) approach
introduced in Zeiler and Fergus [2014]. However, instead of reconstructing lower layer feature maps
from higher layer activations as in DeconvNet, we merely try to identify important regions (supports)
preserved in the forward activations in each layer from the backward (class-specific) information
flow.
Suppose that we have already trained the network to an optimal state. In the forward pass, an image
is presented to the network, and the activations in the entire network are computed. To explain the
classification, we consider the class indicator vector ŷ where ŷc = 1 for the predicted class c and
zero otherwise, and back-propagate this information to the input space. We use ŷ as the input of the
backward pass to approximately “invert” each layer, while iteratively filtering these inverses using
the forward activations. The process is repeated until the input layer is reached.
Dense layers. For each fully-connected layer (indexed by l), denote its activation as:
z (l) = W (l) a(l−1) + b(l) , a(l) = ReLU (z (l) )
and the softmax activation y = σ(a(L) ) is achieved at the final (Lth ) layer. Our goal is to traverse
each of these layers backwards. In order to achieve this, we define the “adjoint” of each operation2 as
follows. The adjoint of the softmax layer, ẑ (L) , is defined point-wise such that
max(z (L) ) i = c
ẑi =
min(z (L) ) i 6= c.
The adjoint of the ReLU activation function is the ReLU itself. Overall, the “adjoint” of each fully
connected layer is â(l−1) = (W (l) )T (ẑ (l) − b(l) ).
Foward-Backward masking. The backward information flow is now filtered using the forward
activations. Specifically, we only keep entries of â(l) such that their entry-wise product with respective
entries in a(l) are above some threshold parameter τ . The other entries are set to zero otherwise. This
enables us to identify a candidate support corresponding to an interpretable feature in the input of a
given layer.
Contributing feature maps. Among many backward feature maps at the top convolutional layer, we
keep only k of them in the backward information flow. The contribution of each map is determined as
the total activation of the entire map. Hence, the features irrelevant to the probed class are removed.
Unpooling. To perform an adjoint of the max pooling layer, we reshape the obtained pooled map
(l)
âp from the backward pass and replicate (copy) its values across the domain of the max operator.
Then, we evaluate entry-wise:
(l) (l)
â(l)
up = min(ap , âp )
This is similar to an analogous unpooling operation in DeconvNet; however, that approach only copies
(l)
the value of âp in a single location via switches that are stored in memory. In contrast, our new
scheme allows the backward feature maps after unpooling to be not overly sparse, and retains enough
spatial information about the interpretation. We note that the replication step is suitable for any
downsampling filter of size 2 × 2 and stride 2 (wherein the receptive fields are non-overlapping). For
other filter sizes and stride lengths, the replicated values are averaged over the overlapping locations.
Deconvolution. The deconvolution step is similar to DeconvNet, where we compute the adjoint by
convolving the backward activation with the flipped filter weights of a corresponding filter.
As we traverse backwards through the network using the above operations, we successively retain
a subset of pixel indices (or support) at the input of each layer that plausibly corresponds to the
2
The term “adjoint” is only loosely defined here due to the nonlinearities involved.
3
Table 1: Illustrative examples of the forward-backward inversion visualizations.
Input
FBI
Interpretations
{our method}
tabby cat
boxer
jaguar
fire hydrant
tabby cat
boxer
jaguar
fire hydrant
Guided Backprop
Visualizations
“interpretable” portion of the given image. In the end, we display the locations of these indices
together with the values produced by the adjoint. The selectivity achieved by successive masking
means that we always obtain a fairly sparse support in our final estimate; the sparsity can be controlled
via appropriate choice of the threshold parameter τ .
3
Results
We visualize the interpretations provided by the proposed algorithm in Table 1. We use VGG-16
model pretrained with ImageNet dataset Chollet et al. [2015]. The visualizations are generated
for the top 1 predictions of the image. The choice of top k filters for computing the inverse is an
important factor contributing to the interpretation obtained. If the value of k is very less (around
10% of the filters), then the interpretations obtained looses a lot of important features and when k is
too high (close to 100%) then we see that the interpretations are too noisy. We have found that the
visualizations obtained while using 100% of the filters, and no pointwise thresholding based on the
forward function value, produces an output that is close to the DeconvNet algorithm. Thus, using the
forward function and masking the inverse computed as explained above, we achieve better output
than the DeconvNet as well as the guided backpropagation algorithms.
For the experiments shown in the Table 1, we use the top 50% filters to propagate the inverse of each
layer. We also use a thresholding value of 10.0 for the pointwise mask between the forward and
backward values. We compare our method with the guided backpropagation algorithm. We notice
that the resulting visualizations have lesser noise compared to that of the guided backpropagation
algorithm.
4
Conclusions
In this work, we introduce a novel Forward-Backward approach for visualizing the interpretations
that correspond to a particular class. However, we see that the choice of k filters for computing
the interpretations has become a hyper-parameter to ensure that the interpretations are good. It is
also seen that the filters which contribute to one particular class might have similar activations to
the activations pertaining to some other class. Hence, decoupling the activations of these filters to
4
choose which filters to use for computing the inverse, so that we maintain class discriminativeness, is
something of much interest to the authors.
References
Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity:
The all convolutional net. arXiv preprint arXiv:1412.6806, 2014.
Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv
Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. See https://arxiv.
org/abs/1610.02391 v3, 2016.
Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating
activation differences. arXiv preprint arXiv:1704.02685, 2017.
Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Gradients of counterfactuals. arXiv preprint arXiv:1611.02639,
2016.
Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features
for discriminative localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 2921–2929, 2016.
Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European
Conference on Computer Vision (ECCV), pages 818–833. Springer, 2014.
Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Why should i trust you?: Explaining the predictions of
any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, pages 1135–1144. ACM, 2016.
François Chollet et al. Keras. https://github.com/fchollet/keras, 2015.
5
| 1cs.CV
|
Unsupervised Regenerative Learning of Hierarchical
Features in Spiking Deep Networks for Object
Recognition
Priyadarshini Panda*, and Kaushik Roy
School of Electrical and Computer Engineering, Purdue University
[email protected]*
Abstract—We
present
a
spike-based
unsupervised
regenerative learning scheme to train Spiking Deep Networks
(SpikeCNN) for object recognition problems using biologically
realistic leaky integrate-and-fire neurons. The training
methodology is based on the Auto-Encoder learning model
wherein the hierarchical network is trained layer wise using the
encoder-decoder principle. Regenerative learning uses spiketiming information and inherent latencies to update the weights
and learn representative levels for each convolutional layer in an
unsupervised manner. The features learnt from the final layer in
the hierarchy are then fed to an output layer. The output layer is
trained with supervision by showing a fraction of the labeled
training dataset and performs the overall classification of the
input. Our proposed methodology yields 0.92%/29.84%
classification error on MNIST/CIFAR10 datasets which is
comparable with state-of-the-art results. The proposed
methodology also introduces sparsity in the hierarchical feature
representations on account of event-based coding resulting in
computationally efficient learning.
Keywords— Spiking Neural Networks; Deep Learning; AutoEncoder; Leaky Integrate-and-Fire (LIF) Neuron; Unsupervised
Hierarchical Learning; Regenerative learning; Object Recognition.
I. INTRODUCTION
“Can machines think?”, the question brought up by Turing
in his paper, [1], has led to the development of the field of
brain-inspired machine learning wherein researchers have put
substantial effort in building smarter, more aware devices and
technology that have the potential of having human-like
understanding. In fact, large scale deep neural network
architectures, such as Convolutional Neural Nets (CNNs), have
demonstrated unprecedented performance (in terms of
classification and recognition accuracy) in a wide range of
computer vision and related applications. Such deep networks
inspired by the cortical visual processing systems have seen
increasing success in recent years due to the availability of
more powerful computing hardware (GPU accelerators) and
massive datasets for training. Regardless of their success, the
substantial computational cost of training and testing such
large-scale networks has limited their implementation to clouds
and servers. In order to build devices with cognitive abilities,
there is a need for specialized hardware with new
computational theories. Spiking Neural Networks (SNNs) are a
prime candidate for enabling such on-chip intelligence.
Driven by brain-like asynchronous event based
computations, SNNs focus their computational effort on
currently active parts of the network, effectively saving power
on the remaining part, thereby achieving orders of magnitude
lesser power consumption in comparison to their Artificial
Neural Network (ANN) counterparts [2, 3]. In 2014, IBM
research demonstrated a large-scale (>1Million neurons & 256
Million synapses) digital CMOS neurosynaptic chip,
TrueNorth [4], which implements a network of integrate-andfire spiking neurons. However, TrueNorth does not incorporate
any information pertaining to the learning mechanisms, which
is at present a major constraint for realizing SNNs for realworld practical applications like visual and speech recognition
among others. Thus, there is a need to develop efficient
learning algorithms that might take the advantage of the
specific features of SNNs (event-driven, low power, on-chip
learning) while keeping the properties (general-purpose,
scalable to larger problems with higher accuracy) of
conventional ANN models.
Recent efforts on training of deep spiking networks do not
use spike-based learning rules, but instead start from a
conventional ANN fully trained using labeled training data,
followed by a conversion of the same into a model consisting
of simple spiking neurons [5-7]. However, in order to extend
the applicability of learning methods, the use of unlabeled data
for machine learning is imperative. In the non-spiking domain,
unsupervised learning of hierarchical regenerative models such
as Auto-Encoders [8] have been successfully used to learn
high-level features. The learnt features are then used as inputs
to a supervised classification task [9] or to initialize a CNN
[10] to avoid local minima. In this paper we develop upon
Auto-Encoders where we build a spiking deep CNN by training
each layer in the hierarchy in a purely unsupervised manner
using the regenerative model and the temporal spike
information to update the weights.
SNNs are equipped with unsupervised weight-modification
rules like Spike Timing Dependent Plasticity (STDP) that learn
the structure of input examples without using labels [11].
However, the network structure that has been successful in
achieving competitive classification accuracy for pattern
recognition problems is a single-layer SNN, which does not
scale well to realistic sized high dimensional images in terms
of computational complexity [12]. Moreover, STDP does not
support learning hierarchical models that simultaneously
represent multiple levels like edges or object parts in a visual
context that is fundamental to a deep learning model. We
propose regenerative learning using spike-timing information
to implement layer-wise weight modification that learn
representative levels for each convolutional layer. The features
from the final layer of the deep convolutional spiking network
(SpikeCNN) are then used for classification tasks.
Fully
Connected
Filter
Input
P2
P1
C2
C1
Convolution Pooling Convolution Pooling Output
Fig. 1. Standard architecture of a Deep Learning Convolutional Network
II. PRELIMINARIES
A. Convolutional Neural Networks
CNNs have proven to be very successful frameworks for
image classification tasks [13-15]. Fig. 1 shows the basic CNN
structure. They are mainly composed of three main blocks:
convolutional layer, spatial sampling/pooling layer and a fully
connected layer. The weights of a CNN are convolution
kernels. A convolution layer convolves a portion of the
previous layer with a set of weight kernels to obtain an array of
output maps. The output maps are given by
𝑥𝑥 𝑘𝑘 = 𝑓𝑓(𝛴𝛴𝑙𝑙 𝑊𝑊 𝑘𝑘 ∗ 𝑥𝑥 𝑙𝑙 )
𝑘𝑘
(1)
where f is the neuron’s activation function, 𝑥𝑥 denotes the
activation value of the neurons in the output maps k
(k=1,2…n), 𝑥𝑥 𝑙𝑙 denotes the activation of the neurons in a
previous layer’s map l, 𝑊𝑊 𝑘𝑘 are the set of weight kernels and *
denotes a 2D- valid convolution operation. The weight kernels
are replicated and moved portion-wise over the whole input
map. This sharing of weights significantly reduces the number
of parameters to be learnt during the training process and thus
enables the CNN to be scalable to high-dimensional images.
The CNNs are trained in a supervised fashion (showing the
training labels) by using standard backpropagation to train the
convolutional weight kernels along with the fully connected
weights for the final output layer as described in [16].
B. Unsupervised learning with Auto-Encoders
Unsupervised learning methods are generally used to
extract useful features from unlabeled data, to detect
oversimplified input representations and remove input
redundancies and finally to obtain only robust and
discriminative representations of the data. In fact, deep neural
network architectures have been built by stacking layers trained
in an unsupervised way. This is done to avoid local minima and
to increase the network performance [17-19].
Auto-Encoder (AE) methods are one of the most widely
used unsupervised feature extractor models for ANNs (nonX
h
WT
Input
ℎ = 𝑓𝑓𝜃𝜃 = 𝜎𝜎(𝑊𝑊𝑊𝑊)
𝑦𝑦 = 𝑓𝑓𝜃𝜃′ = 𝜎𝜎(𝑊𝑊 ′ 𝑥𝑥)
𝑊𝑊 ′ = 𝑊𝑊 𝑇𝑇
Reconstruction
Encoder-Decoder network
Fig. 2. Auto-Encoder network with input and reconstructed pattern
(2)
(3)
(4)
For a given input x, the AE first obtains the hidden
representation, h using the neuron activation function denoted
as σ. This activation value, h, is then reverse mapped to
reconstruct the input. The weights used in the reverse mapping
is generally the transposed form of the weights connecting the
input and the hidden layer. Thus, the method uses the same
weights for encoding the input and decoding the hidden values,
thereby reducing the number of parameters to be learned in the
training process. The parameters are then optimized to
minimize the error for each training pattern xi and its
reconstruction yi. Note, the training labels are nowhere used in
the weight update process. Hence, the method is completely
unsupervised. Inspired by the reconstructive model of training,
we use the auto-encoder model to update the weights for each
convolutional layer by tracking the spike information at the
input and modifying the weights such that the original input
spike pattern is reproduced. This enables the network to learn
hierarchical representative features of the input data.
III. DEEP SPIKING CONVOLUTIONAL NETWORK: LEARNING AND
IMPLEMENTATION
A. Spiking neuron model
Unlike conventional ANNs where a vector is given at the
input layer once and the corresponding output is produced after
processing through several layers of the network, SNNs require
the input to be encoded as a stream of events. At a particular
instant, each event is propagated through the layers of the
network while the neurons accumulate the events over time
causing the output neuron to fire or spike. Thus, the spike
information is used to communicate between the layers of the
network. The spiking neuron model used in this work is the
Leaky Integrate-and-Fire (LIF) model [31]. The membrane
potential 𝑣𝑣𝑚𝑚𝑚𝑚𝑚𝑚 (𝑡𝑡) of a post synaptic neuron is given by
𝜏𝜏𝑅𝑅𝑅𝑅
y
W
spiking). AE models are based on the encoder-decoder
principle [9]. The input is first transformed into a lower
dimensional space (encoder), and then expanded to reproduce
the original input (decoder) as shown in Fig. 2. This captures
the non-linear dependencies in the input. Each layer is trained
on the above principle by feeding the activation from one layer
to the next. The basic mathematical formalisms for an AE
based training of an ANN are as follows [20]:
𝑑𝑑𝑣𝑣𝑚𝑚𝑚𝑚𝑚𝑚 (𝑡𝑡)
𝑑𝑑𝑑𝑑
= −𝑣𝑣𝑚𝑚𝑚𝑚𝑚𝑚 (𝑡𝑡) + 𝐽𝐽(𝑡𝑡)
(5)
where J(t) is the input current and 𝜏𝜏𝑅𝑅𝑅𝑅 is the membrane time
constant. The neuron fires when the membrane potential
𝑣𝑣𝑚𝑚𝑚𝑚𝑚𝑚 crosses a certain user-defined threshold 𝑣𝑣𝑡𝑡ℎ . Once a
spike is generated, the membrane potential of the neuron is set
to the reset potential, 𝑣𝑣𝑟𝑟𝑟𝑟𝑟𝑟 , for a refractory period of 𝜏𝜏𝑟𝑟𝑟𝑟𝑟𝑟 . Once
the refractory period is complete, the neuron follows the
response shown in equation (5). In our simulations, we
discretize the above continuous time equation into 1 ms time
steps.
The total synaptic input current received by a neuron is
Error= Vth –Vmem(t)
𝐽𝐽(𝑡𝑡) = 𝛴𝛴𝑖𝑖 �𝛴𝛴𝑠𝑠𝜖𝜖𝑆𝑆𝑖𝑖 𝑤𝑤𝑖𝑖 𝛿𝛿(𝑡𝑡 − 𝑠𝑠)�
Input X spikes
(6)
where 𝑤𝑤𝑖𝑖 is the synaptic efficacy of the ith synapse, 𝛿𝛿 is the
delta function that contains the time of arrival of spikes at the
ith synapse denoted by Si = {ti0, ti1…}.
B. Error Backpropagation in SNNs
Weight update in Convolutional Deep Learning Networks
follow the convolutional backpropagation algorithm which is
an extension of the standard stochastic gradient descent rule for
feedforward ANNs in the convolutional context [21]. As
discussed earlier, in this work, we use the regenerative learning
method inspired from AEs to train the hierarchical
convolutional layer features of a deep spiking convolutional
network (SpikeCNN). Similar to ANNs, in the regenerative
learning for deep SNNs, the backpropagation algorithm with
gradient descent is employed to update the weights to enable
unsupervised layer wise training.
B.1. Learning theory
I1(t)
.
.
.
.
I2(t)
.
.
.
.
Input Spike trains I3(t)
w1
Observed Spikes
Sobs(t)
w2
.
.
.
.
wn
Neuron
Sdes(t)
Desired Spikes
Fig. 3. The learning problem identifies the optimal weight vector so as to
achieve the desired spike train from the given input spike pattern.
The learning problem for spiking neurons is illustrated in
Fig. 3. There are n input synapses to the spiking neuron, each
receiving an independent spike train, and the aim is to
determine the n-dimensional synaptic weight vector w = [w1 w2
. . . wn]T for the neuron such that it produces the desired spike
train 𝑆𝑆𝑑𝑑𝑑𝑑𝑑𝑑 (𝑡𝑡). Let the desired spike train be given by
𝑖𝑖
𝑆𝑆𝑑𝑑𝑑𝑑𝑑𝑑 (𝑡𝑡) = 𝛴𝛴𝑖𝑖 𝛿𝛿�𝑡𝑡 − 𝑡𝑡𝑑𝑑𝑑𝑑𝑑𝑑
�
(7)
𝑘𝑘
1
2
, 𝑡𝑡𝑑𝑑𝑑𝑑𝑑𝑑
… 𝑡𝑡𝑑𝑑𝑑𝑑𝑑𝑑
where 𝛿𝛿 is the Dirac delta function and 𝑡𝑡𝑑𝑑𝑒𝑒𝑒𝑒
are the desired spike arrival instants. We now require a learning
rule to identify the changes in synaptic weights of the neuron
so as to achieve the desired input to output transformation
under the constraint that the weight updates be spike induced.
B.2. Cost Function
The weight update in ANNs follows the backpropagation
algorithm aimed at minimizing a cost function. Similarly, in
SNNs we need to define a cost function that would drive the
learning process. Let V(t) be the membrane potential of the
neuron, with synaptic weight vector w, when the given set of
input spike trains is fed to it. Correspondingly, let the neuron
𝑘𝑘′
1
2
issue spikes at time instants 𝑡𝑡𝑜𝑜𝑜𝑜𝑜𝑜
, 𝑡𝑡𝑜𝑜𝑜𝑜𝑜𝑜
…𝑡𝑡𝑜𝑜𝑜𝑜𝑜𝑜
. Hence, in
accordance with (7) the observed spike train can be denoted as
𝑆𝑆𝑜𝑜𝑜𝑜𝑜𝑜 (𝑡𝑡) = 𝛴𝛴𝑖𝑖 𝛿𝛿�𝑡𝑡 −
𝑖𝑖
𝑡𝑡𝑜𝑜𝑜𝑜𝑜𝑜
�
Now the error function can be defined as
(8)
Error
X
X
Input Layer
Convolutional Layer
Pseudo-Visible
layer
Fig. 4. Layer-wise training of a convolutional layer using Regenerative
Learning. If a neuron X in the input layer spikes at a given time instant,
the regenerative learning model updates the weights in such a way that
the neuron X in the pseudo-visible layer also spikes. This is achieved by
propagating the error calculated at the pseudo-visible layer using gradient
descent.
𝑒𝑒(𝑡𝑡) = 𝑆𝑆𝑑𝑑𝑑𝑑𝑑𝑑 (𝑡𝑡) − 𝑆𝑆𝑜𝑜𝑜𝑜𝑜𝑜 (𝑡𝑡)
(9)
However, the desired spike train is not known in an
unsupervised learning process. In non-spiking AE models,
encoder-decoder method [9] is used to calculate the error that is
back propagated through the layers without showing any
training labels. Hence, the AE based training is completely
unsupervised. Similarly, in the regenerative layer-wise training
of SpikeCNN, we add another output layer in addition to the
input and the convolutional layer which will be interpreted as a
pseudo-visible layer that should ideally imitate the input layer
patterns as shown in Fig. 4. This will enable us to update
weights for the intermediate layers without showing any
training labels. In a non-spiking model, the weight update rule
is driven by the activation values of the output neurons. In the
spiking context, the activation values of an output neuron can
be interpreted as its membrane potential. If an input neuron
spikes at a given time instant, the corresponding neuron in the
output layer should also spike as per the regenerative approach.
This can only be achieved if the membrane potential of the
output neuron crosses the threshold.
Depending upon the input spike pattern, the error is defined
as follows:
• If an input neuron spikes at a given time instant, the
error is calculated as the difference in the threshold
value (𝑣𝑣𝑡𝑡ℎ ) and the membrane potential (𝑣𝑣𝑚𝑚𝑚𝑚𝑚𝑚 ) of the
corresponding output neuron in the pseudo-visible
layer.
• If the input neuron does not spike, the error is evaluated
as the difference in the reset value (𝑣𝑣𝑟𝑟𝑟𝑟𝑟𝑟 ) and the
membrane potential (𝑣𝑣𝑚𝑚𝑚𝑚𝑚𝑚 ) of the corresponding
output neuron in the pseudo-visible layer.
Thus, the error function can now be indicated as
𝑒𝑒(𝑡𝑡) = 𝑉𝑉𝑑𝑑𝑑𝑑𝑑𝑑 (𝑡𝑡) − 𝑉𝑉𝑜𝑜𝑜𝑜𝑜𝑜 (𝑡𝑡)
(10)
where 𝑉𝑉𝑑𝑑𝑑𝑑𝑑𝑑 is 𝑣𝑣𝑡𝑡ℎ or 𝑣𝑣𝑟𝑟𝑟𝑟𝑟𝑟 depending upon the spike event at
the input neuron. The cost function corresponding to the
synaptic weight vector w can now be defined as
1
𝑇𝑇
𝐶𝐶(𝑤𝑤) = ∫0 (𝑉𝑉𝑑𝑑𝑑𝑑𝑑𝑑 (𝑡𝑡) − 𝑉𝑉(𝑡𝑡))2 𝑑𝑑𝑑𝑑
2
(11)
𝑤𝑤𝑑𝑑𝑑𝑑𝑑𝑑 = 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑛𝑛𝑤𝑤 𝐶𝐶(𝑤𝑤)
(12)
where T denotes the duration of the training epoch. The
desired weight vector 𝑤𝑤𝑑𝑑𝑑𝑑𝑑𝑑 is
The learning process thus takes into account the spike
information and the inherent latencies. The learning rule
follows the gradient descent optimization where the weights
are adjusted depending upon the gradient of cost w.r.t the
weights. Note that the synapses can be excitatory (leading to
increase in membrane potential) or inhibitory (leading to
decrease in membrane potential) depending upon the sign of
the corresponding synaptic weight as determined by the
algorithm.
B.3. Approximate gradient descent
The cost function C(w) as given in (11) depends on V(t)
which has several discontinuities in the weight space. Hence, as
a simplification, we will try to minimize the contribution to the
cost function at each time instant independently rather than
attempting to minimize the total cost over an entire epoch.
The contribution to the cost-function at time t is obtained
by restricting the limits of integral in (11) to an infinitesimally
small interval about time t. Thus,
1
2
�𝑉𝑉 (𝑡𝑡) − 𝑉𝑉(𝑡𝑡)�
2 𝑑𝑑𝑑𝑑𝑑𝑑
Hence, its gradient with respect to w is
𝐶𝐶(𝑤𝑤, 𝑡𝑡) =
(13)
∇𝑤𝑤 𝐶𝐶(𝑤𝑤, 𝑡𝑡) = −�𝑉𝑉𝑑𝑑𝑑𝑑𝑑𝑑 (𝑡𝑡) − 𝑉𝑉(𝑡𝑡)�∇𝑤𝑤 𝑉𝑉(𝑡𝑡)
(14)
Now, the synaptic weight update corresponding to the
activity observed at time t is given as
∆𝑤𝑤(𝑡𝑡) = −𝜂𝜂∇w 𝐶𝐶(𝑤𝑤, 𝑡𝑡) = −𝜂𝜂�𝑉𝑉𝑑𝑑𝑑𝑑𝑑𝑑 (𝑡𝑡) − 𝑉𝑉(𝑡𝑡)�∇w 𝑉𝑉(𝑡𝑡) (15)
where η is the user-defined learning rate. Now, the
discontinuities in V(t) due to the LIF neuron model will render
∇𝑤𝑤 𝑉𝑉(𝑡𝑡) undefined for some particular values of w. In [22], the
authors have implemented a weight update rule and have used
certain approximations to overcome these non-linear
dependencies. In this work, we invoke similar approximations
that would allow us to replace ∇𝑤𝑤 𝑉𝑉(𝑡𝑡) with appropriate
quantities.
In non-spiking context, the weight update for a synapse
connecting neuron i and j [23] (with exponential activation
function) is computed as
∆𝑤𝑤𝑖𝑖𝑖𝑖 = −𝜂𝜂 𝜕𝜕𝜕𝜕/𝜕𝜕𝑤𝑤𝑖𝑖𝑖𝑖 ; 𝑤𝑤ℎ𝑒𝑒𝑒𝑒𝑒𝑒 𝜕𝜕𝜕𝜕/𝜕𝜕𝑤𝑤𝑖𝑖𝑖𝑖 = 𝛿𝛿𝑗𝑗 𝑜𝑜𝑖𝑖
𝛿𝛿𝑗𝑗 = 𝐸𝐸 𝑜𝑜𝑗𝑗
= �𝛴𝛴𝑙𝑙 𝛿𝛿𝑙𝑙 𝑤𝑤𝑗𝑗𝑗𝑗 � 𝑜𝑜𝑗𝑗
𝑖𝑖𝑖𝑖 𝑗𝑗 𝑖𝑖𝑖𝑖 𝑎𝑎𝑎𝑎 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛
𝑖𝑖𝑖𝑖 𝑗𝑗 𝑖𝑖𝑖𝑖 𝑎𝑎 ℎ𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛
(16)
(17)
Here, o denotes the activation value of neuron. Comparing
(15) and (16), ∇w 𝐶𝐶(𝑡𝑡) is equivalent to 𝜕𝜕𝜕𝜕/𝜕𝜕𝑤𝑤𝑖𝑖𝑖𝑖 . Interpreting
the activation value of the neuron as the membrane potential in
the spiking context, we can now replace 𝑜𝑜𝑗𝑗 in the above
equations with the membrane potential of a neuron (V(t)) at a
given time instant. Thus, the weight update equations for SNN
with approximate gradient descent can be written as
∆𝑤𝑤𝑖𝑖𝑖𝑖 (𝑡𝑡) = −𝜂𝜂 𝛿𝛿𝑗𝑗 (𝑡𝑡)𝑉𝑉𝑖𝑖 (𝑡𝑡)
(18)
𝛿𝛿𝑗𝑗 (𝑡𝑡) = (𝑉𝑉𝑑𝑑𝑑𝑑𝑑𝑑 𝑗𝑗 (𝑡𝑡) − 𝑉𝑉𝑗𝑗 (𝑡𝑡)) 𝑉𝑉𝑗𝑗 (𝑡𝑡) 𝑖𝑖𝑖𝑖 𝑗𝑗 𝑖𝑖𝑖𝑖 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛
𝑖𝑖𝑖𝑖 𝑗𝑗 𝑖𝑖𝑖𝑖 ℎ𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛
(19)
= �𝛴𝛴𝑙𝑙 𝛿𝛿𝑙𝑙 (𝑡𝑡)𝑤𝑤𝑗𝑗𝑗𝑗 �𝑉𝑉𝑗𝑗 (𝑡𝑡)
In summary, the approximate gradient descent assumes the
LIF neuron model to be an equivalent standard activation
model in non-spiking context. The gradient calculation is then
carried out by using the membrane potential of a neuron at a
given time instant as the activation value. The inherent errorresiliency of these neural networks allows us to use such
approximate models.
C. Regenerative Learning for Spike-based Convolutional
Auto-Encoder
Convolutional Auto-Encoders (CAEs) differ from
conventional AEs as their weights are shared among all
locations in the input preserving spatial locality. The CAE
architecture with weight sharing is shown in Fig. 4. The
reconstruction represented by the pseudo-visible layer is due to
a linear combination of basic image patches based on the
activations in the convolutional layer.
For a mono-channel input x, the convolutional layer
representation of the kth feature map is given by
ℎ𝑘𝑘 = 𝜎𝜎(𝑥𝑥 ∗ 𝑤𝑤 𝑘𝑘 )
(20)
𝑦𝑦 = 𝜎𝜎(𝛴𝛴𝑘𝑘 𝜀𝜀(ℎ𝑘𝑘 ) ∗ 𝑤𝑤
� 𝑘𝑘 )
(21)
𝑘𝑘
Here, ℎ denotes the membrane potential of the neuron
(𝑣𝑣𝑚𝑚𝑚𝑚𝑚𝑚 (𝑡𝑡))in the convolutional layer, σ denotes LIF model
discussed earlier to calculate 𝑣𝑣𝑚𝑚𝑚𝑚𝑚𝑚 (𝑡𝑡) at a given time instant t,
x contains the spike information from the input layer neurons
and * denotes the convolution operation. When ℎ𝑘𝑘 crosses 𝑣𝑣𝑡𝑡ℎ ,
the spikes generated at the neurons of the convolutional layers
serve as input for the pseudo-visible layer. The reconstruction
is obtained using
where 𝑤𝑤
� 𝑘𝑘 denotes the transpose (or flip in both dimensions)
operation, ε(hk) denotes the spike information corresponding to
the convolutional layer and y gives 𝑣𝑣𝑚𝑚𝑚𝑚𝑚𝑚 (𝑡𝑡) of the neurons in
the pseudo-visible layer. The convolution of a m x m matrix
with a n x n matrix may result in an (m-n+1) x (m-n+1) matrix
(valid convolution) or (m+n-1) x (m+n-1) (full convolution). In
our simulations, we perform a valid convolution from input to
the convolutional layer and a full convolution from
convolutional to pseudo-visible layer in order to get a map of
the same size as the input layer. All spike-based models and
calculations for implementing convolutional network remain
the same as described in earlier sub-section (III-B). Earlier in
(6), the input synaptic current was given by the weighed
summation of spike inputs. In this case, the only difference is
that the synaptic current is obtained by convolving the spike
information of an image patch with the weight kernel.
Please note that the hk and y are calculated at every time
instant of a given epoch and the error is calculated from the
difference in the 𝑣𝑣𝑚𝑚𝑚𝑚𝑚𝑚 (𝑡𝑡) of the neurons in the input and
pseudo-visible layer. It is clear that the regenerative learning
would activate the neurons in the pseudo-visible layer such that
they imitate the input layer spike pattern. Thus, the cost
function to minimize is the mean squared error (MSE) given by
1
2
∑𝑛𝑛 �𝑣𝑣 (𝑡𝑡) − 𝑣𝑣𝑚𝑚𝑚𝑚𝑚𝑚 (𝑡𝑡)�
𝐶𝐶(𝑤𝑤) =
(22)
2𝑛𝑛 𝑖𝑖=1 𝑑𝑑𝑑𝑑𝑑𝑑
where n is the total number of neurons in the input /pseudovisible layer. As mentioned earlier, we minimize the error at
each time instant rather than over the entire epoch duration.
(a)
(b)
(c)
Fig. 5 Reconstructed patterns observed after training the first convolutional layer of MNIST_2C with regenerative learning for different vth and Irate values
{Original pixel image (Column 1), Spike input image (Column 2), Reconstructed image (Column 3)} (a) MNIST_2C initialized with 𝑣𝑣𝑡𝑡ℎ = 1.0, Irate =100 Hz (b)
MNIST_2C with parameters (P2) 𝑣𝑣𝑡𝑡ℎ = 1.2, Irate =100 Hz (c) MNIST_2C with parameters (P1) 𝑣𝑣𝑡𝑡ℎ = 0.8, Irate =75 Hz
Now the approximate gradient descent model discussed in subsection B.3 is used to calculate the weight updates. The
gradient of the cost function now involves convolution
operations and is given by
𝜕𝜕𝜕𝜕(𝑤𝑤)
� 𝑘𝑘 ) ∗ 𝛿𝛿𝑦𝑦
= 𝑥𝑥 ∗ 𝛿𝛿ℎ𝑘𝑘 + 𝜀𝜀(ℎ
(23)
𝑘𝑘
𝛿𝛿ℎ and 𝛿𝛿𝑦𝑦 are the gradients for the convolutional and
pseudo-visible layer neurons which are evaluated using the
approximate model. 𝛿𝛿ℎ𝑘𝑘 is calculated using (19) for hidden
neuron and 𝛿𝛿𝑦𝑦 using (19) for output neuron. Several such
convolutional AEs can be assembled together to construct a
deep CNN hierarchy. Each convolutional layer in the hierarchy
is trained separately with the regenerative method described
above.
𝜕𝜕𝜕𝜕
D. Average Pooling
For CNNs, a pooling layer [24, 25] is often introduced after
the convolutional layers to obtain translational invariance. The
average pooling or subsampling layers combine the responses
from multiple neurons in the convolutional layer into one. The
representation of the averaging layer is identical to (1), except
that the kernels consist of uniform weights fixed to 1/size(wk),
where size(wk) is the size of the sampling window. After
training the convolutional layer with the regenerative method,
the membrane potentials (hk ) of the neurons across all feature
maps are obtained corresponding to a training input for a given
instant of an epoch. The membrane potentials (hk) within the
sampling window are averaged to obtain the output membrane
potential of the neuron (pk) in the pooling layer for the given
time instant. When pk crosses 𝑣𝑣𝑡𝑡ℎ , the spikes generated at the
neurons of the pooling layer serve as input for training the next
convolutional layer. It is clear that the averaging operation
down-samples the convolutional layer representation while
conserving the spike information and inherent latencies from
the previous layer.
The regenerative auto-encoder based learning, thus, trains
several convolutional layers in layer-wise fashion which can be
assembled together to form a deep hierarchy [27: SAE]. Each
layer receives its input from the previous layer as described
above. The assembled convolutional layers can be used to
initialize a CNN with the same topology prior to a fully
connected stage.
E. Supervised training with labels for classification
The Fully Connected layer (FC) at the end of the CNN
combines the inputs from the feature maps in the previous layer
to perform classification of the overall inputs at the output
layer. The training of the fully connected layer cannot follow
the encoder-decoder principle (regenerative learning) as the
aim here is to classify rather than to obtain abstract
representations of the input. At this stage, the spike information
from all end layer maps are concatenated into a vector which
serves as input from the FC layer to the output layer.
The training labels are used to fix the spike pattern of the
output layer neurons that will drive the error backpropagation
to calculate the weight updates. The weight update rules follow
the same equations as described in the approximate gradient
descent method in sub-section B.3. For a given label, a Poisson
spike pattern of a particular frequency is generated that serves
as the desired spike train for the corresponding output neuron
for all training inputs with that label. The remaining output
neurons have no spike events. The errors are calculated as per
the spike events at the output layer neurons. This method
ensures that the weights connected to the desired output neuron
are potentiated while inhibiting the activity of the remaining
connections. In our simulations, we fix the frequency of the
desired spike train for a training label at 30 Hz. After training is
done, an input is presented as a stream of events at the input
layer of the SpikeCNN. At the end of the time duration, the
spike responses from the output layer neurons are monitored.
The output neuron with the highest response is the class
predicted by the network for the given input.
IV. EXPERIMENTAL RESULTS
We evaluate our proposed learning on two datasets:
MNIST [16] and CIFAR10 [26]. MNIST is a standard dataset
of handwritten digits that contains 60,000 training and 10,000
test patterns of 28 x 28 pixel sized greyscale images of the
digits 0-9. CIFAR10 is a more challenging dataset that consists
of 60,000 colored images belonging to 10 classes. Each image
has 32x32 pixels. We used the first 50,000 images for training
and last 10,000 images for testing. Regenerative learning is
used to train spike-based convolutional AEs which are then
used to initialize a SpikeCNN with the same topology. The
1
2
3
1
2
3
1
2
3
Reconstruction for
Reconstruction for
Reconstruction for
Green channel
Red channel
Blue channel
Column1 of each figure is the original pixel image, Column 2 is corresponding
spike input image, Column 3 is the reconstructed spike pattern
Fig. 6 Reconstructed pattern observed after training the first layer of CIFAR_3C initialized with P2. CIFAR_3C has 3 maps at the input layer
corresponding to the 3 color channels (Red, Green, Blue). The figures show the reconstruction of the input pattern at the pseudo-visible layer
corresponding to the 3 separate channels.
final fully connected layer of the SpikeCNN is then trained
with a fraction of training labels from the entire dataset to
perform classification. The input is presented as Poisson
distributed spike train with firing rates proportional to the
intensity of pixels [5]. In the experiments, the epoch duration
or the time for which an input image is shown to the network is
250ms. During layer-wise training, each input is presented
multiple times (depending upon the depth of SpikeCNN) and
the weights are updated each time to minimize the
reconstruction error. Since each layer receives its input from
the previous trained layer, this process helps in maintaining an
adequate firing rate for each layer. This ensures spike
propagation as we go deeper into the network. The membrane
potentials of all neurons in the convolutional and pseudovisible layers are all set to 𝑣𝑣𝑟𝑟𝑟𝑟𝑟𝑟 before presenting a new training
pattern.
A. Network Architecture and Parameters
We implemented a SpikeCNN for MNIST (MNIST_2C)
with 2 convolutional layers: 28x28-12c5-2a-64c5-2a-10o. The
input layer is 28x28. Both convolutional layers use 5x5 kernel
size with 12 and 64 maps, respectively. A 2x2 average pooling
window is used after each convolutional layer. The final
features from the second averaging layer are then fully
connected to a 10-neuron output layer. SpikeCNN for
CIFAR10 (CIFAR_3C) consists of 3 convolutional layers:
32x32x3-32c5-2a-32c5-2a-64c4-10o. In this case, the input
layer has 3 maps corresponding to the 3 color channels RGB.
The first and second convolutional layer have 5x5 sized kernels
with 32 maps while the third layer has 4x4 kernel with 64
maps. The features from the third layer are directly fed to the
output layer without any average pooling. Please note that we
do not use any data augmentation or normalization techniques
like dropout [27] in the SpikeCNN implementation.
The network parameters like input rates (Irate) and threshold
values (𝑣𝑣𝑡𝑡ℎ ) for the networks were set by trial and error by
cross-validating a few times to get the lowest reconstruction
error from the input image layer. Fig. 5 shows the
reconstructed image patterns formed at the pseudo-visible layer
after training the first convolution layer of MNIST_2C with
regenerative learning for different values of 𝑣𝑣𝑡𝑡ℎ and Irate.
Visually, we can inspect that Fig. 5(b) with parameters 𝑣𝑣𝑡𝑡ℎ =
1.2, Irate =100 Hz (P2) and Fig. 5(c) with 𝑣𝑣𝑡𝑡ℎ = 0.8, Irate =75 Hz
(P1) give more convincing reconstruction than Fig. 5(a) 𝑣𝑣𝑡𝑡ℎ =
1.0, Irate =100 Hz. We use the parameters P2, P1 corresponding
to Fig. 5 (b), (c) to initialize MNIST_2C and evaluate the
classification accuracy in both cases. Similarly, Fig. 6 shows
the reconstructed image patterns from the first convolutional
layer of CIFAR_3C for the three color channels separately. The
parameters used are 𝑣𝑣𝑡𝑡ℎ = 1.2, Irate =100Hz to initialize
CIFAR_3C and obtain the classification accuracy. The figures
show the accumulated spike count over 250ms of simulated
time. Please note that though the deeper layers in
MNIST_2C/CIFAR_3C have significantly larger number of
maps than that of initial layers in both the configurations, the
training time still remains same due to the down-sampling of
input size with average pooling.
B. Reconstruction error across network layers
For quantitative evaluation, we plot the reconstruction
errors obtained at each layer of the network (MNIST_2C,
CIFAR_3C) as shown in Fig. 7 for the network parameters
discussed above. In order to ensure propagation of spike
information across layers, we present the input data 3/5 times
while
training
every
convolutional
layer
of
MNIST_2C/CIFAR_3C respectively. The reconstruction error
plotted in Fig. 7 is the aggregate loss over the multiple
presentation of the training data. We use the squared Euclidean
distance of the difference in the spike events as a measure of
loss. It is clearly seen that error observed for MNIST_2C with
parameters P1 is higher than that of P2 which supports the
visual reconstruction patterns shown in Fig 5(b), (c). A
noteworthy observation here is that the reconstruction error in
both networks decreases as we move towards deeper layers. In
[28], the authors have shown that in the context of deep ANNs,
blurring an image enables better reconstruction as the network
then learns more general representations. Learning the finer
details may lead to overfitting increasing the reconstruction
error. As we apply convolution and average pooling in the
Normalized
reconstruction
error
0.8
0.6
0.4
0.2
0
C1
C2
C3
Convolutional layer
CIFAR_3C with P2
MNIST_2C with P1
MNIST_2C with P2
Fig. 7 Reconstruction errors at different layers of the SpikeCNN
architecture
initial stage, certain information from the original image is lost
in this process. Thus, we can interpret that the deeper layers
work on slightly blurred details of the original image causing
lower reconstruction error.
C. Classification Accuracy
After training the convolutional layers by optimizing the
reconstruction error for the entire training dataset, the features
from the final layer are then fed to the output layer. The
weights at the output are trained in a supervised manner by
showing training labels as discussed in Section III (E).
However, in the supervised case, we use only a subset of the
training data to train the final layer. During testing, an input
test pattern is presented two times to the trained SpikeCNN.
The spikes at the output neurons are aggregated over two
simulations and the neuron with the highest response is the
predicted class for the given test input. Since the pixel intensity
values are converted to Poisson spike trains, the accuracy can
differ for different spike timings. Thus, the accuracies are
averaged over five iterations of presentation of the entire
testing dataset. Fig. 8 shows the classification error for
MNIST_2C initialized with parameters P1, P2 and CIFAR_3C
with P2 as the size of the labelled training subset is varied. For
% Classification
Error
25
20
15
10
5
0
500
MNIST_2C with P2, the lowest error achieved by showing all
the 60000 training labels at the final layer is 0.92% (99.08%
classification accuracy). It is clearly seen that the error
decreases significantly as the size of the training set is
increased from 500 to 20000. However, for subsets > 20000,
the error almost remains the same. For MNIST_2C with P1/
CIFAR_3C with P2, the minimum error obtained showing
20000 labels at the final layer is 1.81%/29.84%. In [29]/[30],
the authors have implemented a spiking deep network by
converting a deep static CNN to SNN and have achieved
22.57%/0.86% error on CIFAR-10/MNIST dataset. The fact
that our network performs favorably incorporating the inherent
latencies of a spiking neuron model in the learning process
suggests that the regenerative learning scheme can be used to
train deep spiking networks to obtain state-of-the-art results.
D. Sparsity with Regenerative Learning
The spike-based regenerative learning scheme, on account
of event-based coding, introduces sparsity over the
convolutional layer feature representations. Since the learning
is based on spiking activity of the neurons, the output at the
pseudo-visible layer is reconstructed using only the maximally
active neurons in the feature maps of the convolutional layers.
As a result, the reconstruction error at each time instant of the
training epoch is back-propagated through these active
neurons. Basically, the sparse event based computation acts as
a regularizer that prevents learning of over-complete
representations of the input. In other words, the sparsity in
features decreases the number of filters or weight kernels
required to reconstruct the input thereby forcing the filters to be
more general. Fig. 9 shows the feature maps learnt for
MNIST_2C (with P2) with accumulated spike counts over 250
ms for a particular input pattern. It is evident that the feature
maps in both convolutional layers have sparse active units.
Only a smaller section of the SpikeCNN is active in a training
epoch. Thus, we can effectively save power on the remaining
idle or inactive portions of the network. Sparsity in learning has
a key role in reducing the overall power consumption by
decreasing the spike rate in SNN architectures which is one of
the main reasons to use SNN over ANN. Fig. 9 also shows the
weight kernels learnt at the first layer from the input. Visually,
we can interpret that the weights are more diverse and global.
1000 3000 10000 20000 40000 60000
Size of the labelled training dataset
MNIST_2C with P2
MNIST_2C with P1
% Classification
Error
(a)
60
50
40
30
20
10
0
500
1000
3000
10000 20000
Size of the labelled training dataset
CIFAR_3C with P2
(b)
Fig. 8 Classification error as the size of the labelled dataset for the
supervised training of output layer is varied
Fig. 9 (a) MNIST training input to MNIST_2C initialized with P2: Original
pixel image (left), Spike input image (right) (b) The weight kernels learnt
at the input layer (c) 12 feature maps showing the sparse representations of
maximally active spiking neurons in the first convolutional layer of
MNIST_2C (d) 20 of 64 feature maps in the second convolutional layer of
MNIST_2C with sparsely active neurons
V. CONCLUSION
We introduced a spike-based learning scheme to train Spiking
Deep Networks (SpikeCNN) for object recognition problems
using leaky integrate-and-fire (LIF) neurons. The regenerative
model learns the hierarchical feature maps layer-by-layer in a
deep convolutional network in an unsupervised manner. Once
the convolutional layers are learnt, the features are then fed to
an output layer trained in a supervised manner by showing a
fraction of the labeled training dataset. The output layer
performs the overall classification of the input. While previous
work on deep spiking networks have examined the conversion
of ANNs to SNNs, we build a spiking deep CNN from scratch
with the proposed learning using spike-timing information and
inherent latencies to implement layer-wise weight
modification. Our experiments on the MNIST and CIFAR-10
dataset demonstrate comparable classification accuracy with
state-of-the-art results. Also, the sparsity in representations
introduced with regenerative learning suggests overall power
savings in the learning process which generally takes up a
significant part of the time the network is used. Finally, the
SpikeCNN system developed with our proposed learning with
the current initialized parameters is in its first generation, and
we expect its accuracy on CIFAR-10 to improve as we gain
experience with the method and tune the network for better
parameters.
REFERENCES
Alan M Turing. Computing machinery and intelligence. Mind, pages
433–460, 1950.
[2] Ben Varkey Benjamin, Peiran Gao, Emmett McQuinn, Shobhit
Choudhary, Anand R Chandrasekaran, Jean-Marie Bussat, Rodrigo
Alvarez- Icaza, John V Arthur, Paul Merolla, Kwabena Boahen, et al.
Neurogrid: A mixed-analog-digital multichip system for large-scale
neural simulations. Proceedings of the IEEE, 102(5):699–716, 2014.
[3] Jongkil Park, Sohmyung Ha, Theodore Yu, Emre Neftci, and Gert
Cauwen- berghs. A 65k-neuron 73-mevents/s 22-pj/event asynchronous
micro- pipelined integrate-and-fire array transceiver. In Biomedical
Circuits and Conference (BioCAS), IEEE, pages 675–678. IEEE, 2014.
[4] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S
Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam,
Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated
circuit with a scalable communication network and interface. Science,
345(6197):668–673, 2014.
[5] Peter O’Connor, Daniel Neil, Shih-Chii Liu, Tobi Delbruck, and
Michael Pfeiffer. Real-time classification and sensor fusion with a
spiking deep belief network. Frontiers in neuroscience, 7, 2013.
[6] Wolfgang Maass and Henry Markram. On the computational power of
circuits of spiking neurons. Journal of computer and system sciences,
69(4):593–616, 2004.
[7] J Antonio P´erez-Carrasco, Bo Zhao, Curtis Serrano, Begona Acha,
Teresa Serrano-Gotarredona, Shouchun Chen, and Bernab´e LinaresBarranco. Mapping from frame-driven to frame-free event-driven vision
systems by low-rate rate coding and coincidence processing–application
to feedforward convnets. Pattern Analysis and Machine Intelligence,
IEEE Transactions on, 35(11):2706–2719, 2013.
[8] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the
dimensionality of data with neural networks. Science, 313(5786):504–
507, 2006.
[9] Marc Aurelio Ranzato, Fu Jie Huang, Y-Lan Boureau, and Yann LeCun.
Unsupervised learning of invariant feature hierarchies with applications
to object recognition. In Computer Vision and Pattern Recognition,
2007. CVPR’07. IEEE Conference on, pages 1–8. IEEE, 2007.
[10] Honglak Lee, Roger Grosse, Rajesh Ranganath, and Andrew Y Ng.
Unsupervised learning of hierarchical representations with convolutional
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[1]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
deep belief networks. Communications of the ACM, 54(10):95–103,
2011.
Sen Song, Kenneth D Miller, and Larry F Abbott. Competitive hebbian
learning through spike-timing-dependent synaptic plasticity. Nature
neuroscience, 3(9):919–926, 2000.
Peter U Diehl, Matthew Cook, Masami Tatsuno, and Sen Song.
Unsupervised learning of digit recognition using spike-timing-dependent
plasticity. Frontiers in Computational Neuroscience, 2015.
Yoshua Bengio. Learning deep architectures for ai. Foundations and
trends in Machine Learning, 2(1):1–127, 2009.
Pierre Sermanet, David Eigen, Xiang Zhang, Micha¨el Mathieu, Rob
Fergus, and Yann LeCun. Overfeat: Integrated recognition, localization
and detection using convolutional networks. arXiv:1312.6229, 2013.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving
deep into rectifiers: Surpassing human-level performance on imagenet
classification. arXiv:1502.01852, 2015.
Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner.
Gradient-based learning applied to document recognition. Proceedings
of the IEEE, 86(11):2278–2324, 1998.
Dumitru Erhan, Yoshua Bengio, Aaron Courville, Pierre-Antoine
Manzagol, Pascal Vincent, and Samy Bengio. Why does unsupervised
pre-training help deep learning? The Journal of Machine Learning
Research, 11:625–660, 2010.
Kunihiko Fukushima. Neocognitron:
A self-organizing neural
network model for a mechanism of pattern recognition unaffected by
shift in position. Biological cybernetics, 36(4):193–202, 1980.
Geoffrey E Hinton, Simon Osindero, and Yee-Whye Teh. A fast
learning algorithm for deep belief nets. Neural computation,
18(7):1527–1554, 2006.
Yoshua Bengio, Pascal Lamblin, Dan Popovici, Hugo Larochelle, et al.
Greedy layer-wise training of deep networks. Advances in neural
information processing systems, 19:153, 2007.
Rasmus Berg Palm. Prediction as a candidate for learning deep
hierarchical models of data. Technical University of Denmark, 2012.
Navin Anwani and Bipin Rajendran. Normad-normalized approximate
descent based supervised learning rule for spiking neurons. In Neural
Networks (IJCNN), 2015 International Joint Conference on, pages 1–8.
IEEE, 2015.
David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams.
Learning representations by back-propagating errors. Cognitive
modeling, 5:3, 1988.
Dominik Scherer, Andreas Muller, and Sven Behnke. Evaluation of
pooling operations in convolutional architectures for object recognition.
In Artificial Neural Networks-ICANN, pages 92–101. Springer, 2010.
Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever,
and Ruslan R Salakhutdinov. Improving neural networks by preventing
co-adaptation of feature detectors. arXiv:1207.0580, 2012.
Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of
features from tiny images, 2009.
Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever,
and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural
networks from overfitting. The Journal of Machine Learning Research,
15(1):1929–1958, 2014.
Aravindh Mahendran and Andrea Vedaldi. Understanding deep image
representations by inverting them. arXiv:1412.0035, 2014.
Yongqiang Cao, Yang Chen, and Deepak Khosla. Spiking deep
convolutional neural networks for energy-efficient object recognition.
International Journal of Computer Vision, 113(1):54–66, 2015.
Peter U Diehl, D. Neil, J. Binas, M. Cook, S.-C. Liu, and M. Pfeiffer,
Fast-Classifying, High-Accuracy Spiking Deep Networks Through
Weight and Threshold Balancing, in IEEE International Joint
Conference on Neural Networks (IJCNN), 2015.
Andre van Schaik. Building blocks for electronic spiking neural
networks. Neural networks, 14(6):617–628, 2001.
| 9cs.NE
|
The Uncertainty Bellman Equation and Exploration
Brendan O’Donoghue, Ian Osband, Remi Munos, Volodymyr Mnih
Deepmind
{bodonoghue, iosband, munos, vmnih}@google.com
arXiv:1709.05380v1 [cs.AI] 15 Sep 2017
September 19, 2017
Abstract
We consider the exploration/exploitation problem in reinforcement learning. For exploitation, it is
well known that the Bellman equation connects the value at any time-step to the expected value at
subsequent time-steps. In this paper we consider a similar uncertainty Bellman equation (UBE), which
connects the uncertainty at any time-step to the expected uncertainties at subsequent time-steps, thereby
extending the potential exploratory benefit of a policy beyond individual time-steps. We prove that the
unique fixed point of the UBE yields an upper bound on the variance of the estimated value of any fixed
policy. This bound can be much tighter than traditional count-based bonuses that compound standard
deviation rather than variance. Importantly, and unlike several existing approaches to optimism, this
method scales naturally to large systems with complex generalization. Substituting our UBE-exploration
strategy for -greedy improves DQN performance on 51 out of 57 games in the Atari suite.
1
Introduction
We consider the reinforcement learning (RL) problem of an agent interacting with its environment to maximize cumulative rewards through time [35]. We model the environment as a Markov decision process
(MDP), but where the agent is initially uncertain of the true dynamics of the MDP [4, 5]. At each time-step,
the agent performs an action, receives a reward, and moves to the next state; from these data it can learn
which actions lead to higher payoffs. This leads to the exploration versus exploitation trade-off: Should the
agent investigate poorly understood states and actions to improve future performance or instead take actions
that maximize rewards given its current knowledge?
Separating estimation and control in RL via ‘greedy’ algorithms can lead to premature and suboptimal exploitation. To offset this, the majority of practical implementations introduce some random noise
or dithering into their action selection (such as -greedy). These algorithms will eventually explore every
reachable state and action infinitely often, but can take exponentially long to learn the optimal policy [12].
By contrast, for any set of prior beliefs the optimal exploration policy can be computed directly by dynamic
programming in the Bayesian belief space. However this approach can be computationally intractable for
even very small problems [9] while direct computational approximations can fail spectacularly badly [22].
For this reason, most provably-efficient approaches to reinforcement learning rely upon the optimism
in the face of uncertainty (OFU) heuristic [14, 13, 7]. These algorithms give a bonus to poorly-understood
states and actions and subsequently follow the policy that is optimal for this augmented optimistic MDP.
This optimism incentivises exploration but, as the agent learns more about the environment, the scale of
the bonus should decrease and the agent’s performance should approach optimality. At a high level these
approaches to OFU-RL build up confidence sets that contain the true MDP with high probability [33, 16,
1
11]. These techniques can provide performance guarantees that are ‘near-optimal’ in terms of the problem
parameters.However, apart from the simple ‘multi-armed bandit’ setting with only one state, there is still a
significant gap between the upper and lower bounds for these algorithms [15, 11, 27].
One inefficiency in these algorithms is that, although the concentration may be tight at each state and
action independently, the combination of simultaneously optimistic estimates may result in an extremely
over-optimistic estimate for the MDP as a whole [28]. Other works have suggested that a Bayesian posterior
sampling approach may not suffer from these inefficiencies and can lead to performance improvements
over OFU methods [34, 25]. In this paper we explore an alternative approach that harnesses the simple
relationship of the uncertainty Bellman equation (UBE), where we define uncertainty to be the variance
of the value estimator the agent is learning, in a sense similar to the parametric variance of Mannor et. al.
[17]. Intuitively speaking, if the agent has high uncertainty (as measured by high estimator variance) in a
region of the state-space then it should explore there, in order to get a better estimate of those Q-values. We
show that, just as the Bellman equation relates the value of a policy beyond a single time-step, so too does
the uncertainty Bellman equation propagate uncertainty values over multiple time-steps, thereby facilitating
‘deep exploration’ [26].
The benefit of our approach (which learns the solution to the UBE and uses this to guide exploration) is
that we can harness the existing machinery for deep reinforcement learning with minimal change to existing
network architectures. The resulting algorithm shares an intimate connection to the existing literature of
both OFU and intrinsic motivation [31, 30]. Recent work has further connected these approaches through
the notion of ‘pseudo-count’ [2, 29] or some generalization of the number of visits to a state and action.
Rather than pseudo-count, our work builds upon the idea that the more fundamental quantity relates to the
uncertainty of the estimated value function and that naively compounding count-based bonuses may lead to
inefficient confidence sets [28]. The key difference is that the UBE compounds the sum of the variances at
each step, rather than standard deviation.
The observation that the higher moments of a value function also satisfy a form of Bellman equation is
not new and has been observed by some of the early papers on the subject [32]. Unlike most prior work, we
focus upon the epistemic uncertainty over the mean of the value function, rather than the higher moments of
the reward-to-go [16, 1, 18]. For application to rich environments with complex generalization we will use
a deep learning architecture to learn a solution to the UBE according to our observed data, in the style of
[37].
2
Problem formulation
We consider an infinite horizon, discounted, finite state and action space MDP, with state space S, action
space A and rewards at each time period denoted by rt ∈ R. A policy π : S × A → R+ is a mapping
from state-action pair to the probability of taking that action at that state. At each time-step t the agent
receives a state st and a reward rt and selects an action at from the policy πt , and the agent moves to the
next state st+1 ∼ P (·, st , at ). Here P (s0 , s, a) is the probability of transitioning from state s to state s0 after
taking action a. TheP
goal of the agent
is maximize the expected total discounted return J under its policy
∞
t
π, where J(π) = E
t=0 γ rt | π . Here the expectation is with respect to the initial state distribution, the
state-transition probabilities, and the policy π. The discount factor γ ∈ (0, 1) controls how much the agent
prioritizes long-term versus short-term rewards.
The action-value, or Q-value, of a particular state under policy π is the expected
P∞ ttotal discounted return
from taking that action at that state and following π thereafter, Qπ (s, a) = E
t=0 γ rt | s0 = s, a0 = a, π .
π
π
The value of state s under policy π, V (s) = E [Q (s, a) | a ∼ π] is the expected total discounted return
2
of policy π from state s. The optimal action-value function Q? (s, a) = maxπ Qπ (s, a) for each (s, a). The
policy that achieves the maximum is the optimal policy π ? .
The Bellman operator T π [4] for policy π, relates the value at each time-step to the value at subsequent
time-steps via dynamic programming,
T π Qπ (s, a) = Er r(s, a) + γEs0 ,a0 Qπ (s0 , a0 ),
(1)
where this expectation is taken over next state s0 ∼ P (·, s, a), the reward r(s, a), and the action a0 from
policy πs0 . For γ ∈ (0, 1) the Bellman operator is contraction and therefore Qπ is the unique fixed point
of (1), T π Qπ = Qπ . Several reinforcement learning algorithms have been designed around minimizing
Bellman residual to propagate knowledge of immediate rewards to long term value [36, 40]. In the next
section we examine a similar relationship for propagating the uncertainties in an estimator, we call this
relationship the uncertainty Bellman equation.
3
The uncertainty Bellman equation
In this section we derive a Bellman-style relationship that propagates the uncertainty (variance) of any Qvalue estimator across multiple time-steps. Propagating the potential value of exploration over many timesteps, or deep exploration, is important for statistically efficient RL [13, 26]. Our main result, which we state
in Theorem 1, is based upon nothing more than the dynamic programming recursion (1) and some crude
upper bounds of several intermediate terms. We will show that even in very simple settings this approach
can result in well-calibrated uncertainty estimates where common count-based bonuses are inefficient [28].
3.1
Interval estimation via variance estimates
We begin by motivating why the uncertainty (or variance) of the estimator is useful for exploration [33].
Consider the problem of learning the value of a policy π from some data d. The true, but unknown, values
Qπ satisfy the Bellman equation (1) while our estimate from Q̂π = Q̂π (d) is a random variable with
potentially both bias and variance. We write (s, a) for the Bellman residual,
(s, a) := (Q̂π − T π Q̂π )(s, a)
= Q̂π (s, a) − Er r(s, a) − γEs0 ,a0 Q̂π (s0 , a0 ).
(2)
If we denote by δ(s, a) = Q̂π (s, a) − Qπ (s, a) the error at (s, a) then we have
δ(s, a) = (s, a) + γEs0 ,a0 δ(s0 , a0 ).
(3)
If we could calculate exactly everywhere, then we could solve the Bellman equation (3) and obtain the true
Q-values. However the true is intractable to calculate, since it requires an exact application of the Bellman
operator. Instead we consider the case where we can estimate the moments of the Bellman residual easily,
even though we can’t calculate it exactly.
The mean-squared error of the estimator at any state-action pair is given by
m(s, a) := Ed [(Q̂π (s, a) − Qπ (s, a))2 ]
= var Q̂π (s, a) + (bias Q̂π (s, a))2 .
Markov’s inequality [10] implies that
Prob |Q̂π (s, a) − Qπ (s, a)| > αm(s, a)1/2 ≤ 2/α2 .
3
(4)
(5)
Therefore if we know, or can bound, the variance and bias of an estimator then we can use them to construct
intervals that contain the true Q-values with high-probability. An agent can then apply the OFU-principle to
prioritize its exploration towards potentially rewarding policies [13, 22]. We argue that, for many settings of
interest, this error is dominated by the variance term and that, in this case, several simplifying relationships
emerge.
Lemma 1. For any policy π and any state-action pair (s, a), the biases satisfy a Bellman equation
bias Q̂π (s, a) = Ed (s, a) + γEs0 ,a0 bias Q̂π (s0 , a0 ).
(6)
Proof. Take the expectation of (3) with respect to d and note that Ed δ(s, a) = bias Q̂π (s, a).
For the purposes of our analysis we will assume that the Bellman residuals at any state-action pair are
uncorrelated. This property will certainly not hold in all settings, but may be a reasonable approximation in
many settings of interest.
Assumption 1. For any s, s0 ∈ S a, a0 ∈ A,
cov((s, a), (s0 , a0 )) ≤ min(var (s, a), var (s0 , a0 )),
where cov denotes the covariance.
Assumption 1 implies that the variance of the Q-value estimate satisfies a Bellman inequality.
Lemma 2. For any policy π and any estimator that satisfies assumption 1, the variance satisfies a Bellman
inequality at all (s, a),
var Q̂π (s, a) ≤ β var (s, a) + γ 2 Es0 ,a0 var Q̂π (s0 , a0 ),
(7)
∗
for some β ∈ [1, 1+γ
1−γ ] and we write β for the minimum such β that satisfies this relationship.
Proof. Let var (s, a) = Ed [((s, a) − Ed (s, a))2 ] be the variance of the Bellman residual at (s, a). We
will refer to this quantity as the local (or shallow) uncertainty from finite data. We now consider the variance
of the estimator:
h
2 i
var Q̂π (s, a) = Ed Q̂π (s, a) − Ed Q̂π (s, a)
h
2 i
= Ed Q̂π (s, a) − Qπ (s, a) − bias Q̂π (s, a)
h
2 i
= Ed Q̂π (s, a) − Er r(s, a) − γEs0 ,a0 Qπ (s0 , a0 ) − bias Q̂π (s, a)
h
2 i
= Ed (s, a) − Ed (s, a) + γEs0 ,a0 [Q̂π (s0 , a0 ) − Ed Q̂π (s0 , a0 )] ,
where in the last line we used lemma 1. Expanding the square we obtain
h
2 i
var Q̂π (s, a) = var (s, a) + γ 2 Ed Es0 ,a0 [Q̂π (s0 , a0 ) − Ed Q̂π (s0 , a0 )]
+ 2γc(s, a),
where c(s, a) is the cross-term. In the appendix, we prove that under assumption 1 this can be bounded
c(s, a) ≤ α var (s, a)),
4
(8)
1
is a constant that might depend on the MDP, the policy, and the estimator. By Jensen’s
where 0 ≤ α ≤ 1−γ
inequality we have that
Ed [(Es0 ,a0 [Q̂π (s0 , a0 ) − Ed Q̂π (s0 , a0 )])2 ] ≤ Ed,s0 ,a0 [(Q̂π (s0 , a0 ) − Ed Q̂π (s0 , a0 ))2 ]
= Es0 ,a0 var Q̂π (s0 , a0 ).
Combining this result with (8) we can then say,
var Q̂π (s, a) ≤ β var (s, a) + γ 2 Es0 ,a0 var Q̂π (s0 , a0 ),
(9)
for some 1 ≤ β ≤ (1 + γ)/(1 − γ).
With these lemmas we are ready to prove our main theorem.
Theorem 1 (Solution of the uncertainty Bellman equation). Under Assumption 1, for any policy π, let
1+γ
β ? ∈ [1, 1−γ
] be the smallest β that satisfies the conditions for Lemma 2. Then there exists a unique u? that
satisfies the uncertainty Bellman equation
u? (s, a)
= (Tuπ u? )(s, a)
:= β ? var (s, a) + γ 2 Es0 ,a0 u? (s0 , a0 )
(10)
for each (s, a), and u? ≥ var Q̂π pointwise.
Proof. To show this we use three essential properties of the Bellman operator for a fixed policy [5]. First,
the Bellman operator is a γ-contraction in the `∞ norm and so the fixed point u? exists and is unique.
Second, value iteration converges in that (Tuπ )k x → u? for any starting x. Finally, the Bellman operator is
monotonically increasing in its arguments, i.e., if x ≥ y pointwise then Tuπ x ≥ Tuπ y pointwise. As the the
variance satisfies a Bellman inequality (9), we have
var Q̂π ≤ Tuπ var Q̂π ≤ lim (Tuπ )k var Q̂π = u? .
k→∞
3.2
(11)
Comparison to traditional exploration bonus
Consider a simple decision problem with known deterministic transitions, unknown rewards and two actions.
We imagine an agent has gathered some data d and produces some unbiased value estimates. According to
these estimates, the first action leads to a single reward with expectation zero and variance σ 2 . The second
action leads to an infinite chain of independent states with expectation zero and variance σ 2 (1 − γ 2 ). These
numbers are chosen so that the variance of the estimated value of each action is mean zero and variance σ 2 .
An optimistic agent motivated by (5) has no reason to value one action over the other. Nonetheless, most
existing approaches to optimism that work via exploration bonus would lead to an inconsistent decision rule
in this setting [28].
Rather than consider the variance of the value as a whole, the majority of existing approaches to OFU
provide exploration bonuses at each state and action independently and then combine these estimates via
union bound. In this context, even a state of the art algorithm such as UCRL2 [11] would afford each state a
bonus proportional to its standard deviation of estimate. For action one this would be proportional to σ, but
for action two this would be proportional to,
r
T
X
p
1+γ
t
2
ExplorationBonus(a2 ) ∝
γ σ 1−γ =σ
.
(12)
1−γ
t=0
5
In the tabular setting, the variance of the estimated reward at each state can be approximated by the inverse
‘count’ N (s, a) of visits to each state and action [2]. Effectively, this means that action two would be visited
1+γ
a factor 1−γ
more often to obtain the same level of confidence as an estimation via the uncertainty Bellman
equation (10). The essential issue that, unlike the variance, the classic exploration bonus based upon the
standard deviation does not obey a Bellman-style relationship.
4
Local uncertainty
Section 3 outlines how the uncertainty Bellman equation can be used to propagate local estimates of the
variance of Q̂π to global estimates for the uncertainty. In this section we present some pragmatic approaches
to estimating local uncertainty that we can then use for practical learning algorithms inspired by Theorem 1.
We do not claim that these approaches are the only approaches to estimating local uncertainty, or even that
these simple approximations are in any sense the ‘best’. Investigating these choices is an important area
of future research, but outside the scope of this short paper. We present a simple progression from tabular
representations, to linear function approximation and then to non-linear neural network architectures.
Tabular value estimate. Consider the setting where each Q̂π (s, a) is learned separately. If we model the
Bellman error as IID with variance bounded by σ 2 , we can apply (5) to bound var (s, a) ≤ σ 2 /N (s, a),
where N (s, a) is the count of the number of visits to state s taking action a [11]. We can then use σ 2 /N (s, a)
as a proxy for the local uncertainty in Theorem 1.
Linear value estimate. Let us consider a linear value function estimator Q̂π (s, a) = φ(s)T wa for each
state and action with fixed basis functions φ(s) : S → RD and learned weights wa ∈ RD , one for each
action. This setting allows for some generalization between states and actions through the basis functions.
For any fixed dataset we can find the least squares solution for each action a [6],
P
T
2
minimizewa N
i=1 (φ(si ) wa − yi )2 ,
where each yi ∈ R is a regression target (e.g., a Monte Carlo return from that state-action). The solution to
this problem is wa? = (ΦTa Φa )−1 ΦTa y, where Φa is the matrix consisting of the φ(si ) vectors stacked rowwise (we use the subscript a to denote the fact that action a was taken at these states). We can consider the
variance of the estimator, if we model the targets yi as IID with variance σ 2 , then var wa? = E[wa? wa? T ] =
σ 2 (ΦTa Φa )−1 . Given a new state vector φs , the variance of the Q-value estimate at (s, a) is then var φTs wa? =
σ 2 φTs (ΦTa Φa )−1 φs . The local uncertainty is also then just
var (s, a) = Ed (φTs wa? − Er,s0 ,a0 y(s, a) − Ed (s, a))2
= σ 2 φTs (ΦTa Φa )−1 φs
(13)
where Ed (s, a) = Ed φTs wa? − Er,s0 ,a0 y(s, a) is the bias introduced by the fact that the linear approximation
might not be able to represent exactly the average of the targets everywhere. Note that in the tabular case the
quantity σ 2 φTs (ΦTa Φa )−1 φs is equal to σ 2 /N (s, a), as might be expected.
An agent using this notion of local uncertainty must maintain and update the matrix Σa = (ΦTa Φa )−1 as
it receives new data. Given new sample φ, the updated matrix Σ+
a is given by
!
T −1
Φa
Φa
+
Σa =
= (ΦTa Φa + φφT )−1
(14)
φT
φT
= Σa − (Σa φφT Σa )/(1 + φT Σa φ)
6
by the Sherman-Morrison-Woodbury formula [8], the cost of this update is one matrix multiply and one
matrix-matrix subtraction per step.
Neural networks value estimate. If we are approximating our Q-value function using a neural network
then the above analysis does not hold. However if the last layer of the network is linear, then the Q-values
are approximated as Qπ (s, a) = φ(s)T wa , where wa are the weights of the last layer associate with action a
and φ(s) is the output of the network up to the last layer for state s. In other words we can think of a neural
network as learning a useful set of basis functions such that a linear combination of them approximates the
Q-values. Then, if we ignore the uncertainty in the φ mapping, we can reuse the analysis for the purely linear
case to derive an approximate measure of local uncertainty that might be useful in practice.
This scheme has some advantages. As the agent progresses it is learning a state representation that helps
it achieve the goal of maximizing the return. The agent will learn to pay attention to small but important
details (e.g., the ball in Atari ‘breakout’) and learn to ignore large but irrelevant changes (e.g., if the background suddenly changes). This is a desirable property from the point of view of using these features to
drive exploration, because the states that differ only in irrelevant ways will be aliased to (roughly) the same
state representation, and states that differ is small but important ways will be mapped to quite different state
vectors, permitting a more task-relevant measure of local uncertainty.
5
Algorithm for Deep Reinforcement Learning
In this section we describe an exploration heuristic for deep reinforcement learning whereby we attempt to
learn both the Q-values and the uncertainties associated with them simultaneously (we assume the biases
are small enough to ignore). The goal is for the agent to explore areas where it learns that it has higher
uncertainty. This is in contrast to the commonly used -greedy [20] and Boltzmann exploration strategies
[19, 23] which simply inject noise into the agents actions. Our policy uses Thompson sampling [38], where
the action is selected as
a = argmax(Q̂π (s, b) + αζ(b)u(s, b)1/2 )
(15)
b
where ζ(b) ∼ N (0, 1) and α > 0 is a hyper-parameter. In this case the probability of selecting action a is
the probability that a has the maximum value if each action b was distributed normally with mean Q̂π (s, b)
and variance α2 u(s, b).
The technique is described in pseudo-code in algorithm 1. We refer to the technique as one-step since
the uncertainty values are updated using a one-step SARSA Bellman backup, but it is easily extendable to
the n-step case. The algorithm takes as input a neural network which has two output heads, one which is
attempting to learn the optimal Q-values as normal, the other is attempting to learn the uncertainty values
of the current policy (which is constantly changing). We do not allow the gradients from the uncertainty
head to flow into the trunk; this ensures the Q-value estimates are not perturbed by the changing uncertainty
signal. For the local uncertainty measure we use the linear basis approximation described previously. We
have dropped the constant β from the uncertainty Bellman equation (10) and the unknown σ 2 term from the
local-uncertainty in equation (13) because they are both absorbed by the hyper-parameter α in the policy.
Most of the assumptions that allowed us to bound the true Q-values in equation (5) are violated by this
scheme, in particular we have ignored the bias term, and the policy is changing as the Q-values change.
However, we might expect this strategy to provide a useful signal of novelty to the agent and therefore
perform well in practice.
7
Algorithm 1 One-step UBE exploration with linear uncertainty estimates.
// Input: Neural network outputting Q and u estimates
// Input: Q-value learning subroutine qlearn
// Input: Thompson sampling hyper-parameter α > 0
Initialize Σa = µI for each a, where µ > 0
Get initial state s, take initial action a
repeat
Retrieve feature mapping φ(s) from input to last layer of Q-head
Receive new state s0 and reward r
Calculate Q-value estimates Q(s0 , ·), uncertainty estimates u(s0 , ·)
0
0
0
1/2 ), where ζ(b) ∼ N (0, 1)
Calculate action
b (Q(s , b) + αζ(b)u(s , b)
a = argmax
T
0
φ(s) Σa φ(s), for terminal s
Calculate y =
φ(s)T Σa φ(s) + γ 2 u(s0 , a0 ), o.w.
Take gradient step in u subnetwork with respect to error (y − u(s, a))2
Update Q-values using qlearn(s, a, r, s0 , a0 )
Update Σa0 according to eq. (14)
Take action a0
until T > Tmax
5.1
Experimental results
Here we present results of algorithm (1) on the Atari suite of games [3], where the network is attempting
to learn the Q-values as in DQN [20, 21] and the uncertainties simultaneously. The only change to vanilla
DQN we made was to replace the -greedy policy with Thompson sampling over the learned uncertainty
values, where the α constant in (15) was chosen to be 0.01 for all games, by a parameter sweep. We used the
exact same network architecture, learning rate, optimizer, pre-processing and replay scheme as [21]. For the
uncertainty head we used a single fully connected hidden layer with 512 hidden units followed by the output
layer. We trained the uncertainty head using a separate RMSProp optimizer [39] with learning rate 10−3 .
The addition of the uncertainty head and the computation associated with it, only reduced the frame-rate
compared to vanilla DQN by about 10% on a GPU, so the speed cost of the approach is negligible.
We compare two versions of our approach: a 1-step method and an n-step method where we set n to
150. The n-step method accumulates the uncertainty signal over n time-steps before performing an update
which should lead to the uncertainty signal propagating to earlier encountered states faster, at the expense
of increased variance of the signal. Note that in all cases the Q-learning update is always 1-step; our n-step
implementation only affects the uncertainty update.
We compare our approaches to vanilla DQN, and also to an exploration bonus intrinsic motivation
approach, where the agent receives an augmented reward consisting of the extrinsic reward and the square
root of the linear uncertainty (13), which was scaled by a hyper-parameter chosen to be 0.1 by a sweep. In
this case a stochastic policy was still required for good performance and so we used -greedy with the DQN
annealing schedule.
In the recent work by Bellemare et al. [2], and the follow-up work [29], the authors add an intrinsic motivation signal to a DQN-style agent that has been modified to use the full Monte Carlo return of the episode
when learning the Q-values. Using Monte Carlo returns dramatically improves the performance of DQN in
a way unrelated to exploration, and due to that change we can’t compare the numerical results directly. In
order to have a point of comparison we implemented our own intrinisic motivation exploration signal, as
8
discussed above. Similarly, we can’t compare directly to the numerical results obtained by Bootstrap DQN
[24] since that agent is using Double-DQN, a variant of DQN that achieves a higher performance in a way
unrelated to exploration. However, we note that our approach achieves a higher evaluation score in 27 out of
the 48 games tested in the Bootstrap DQN paper despite using an inferior base DQN implementation, and it
runs at a significantly lower computational and memory cost.
We trained all strategies for 200M frames (about 8 days on a GPU). Each game and strategy was tested
three times per method with the same hyper-parameters but with different random seeds, and all plots and
scores correspond to an average over the seeds. All scores were normalized by subtracting the average score
achieved by an agent that takes actions uniformly at random. Every 1M frames the agents were saved and
evaluated (without learning) on 0.5M frames, where each episode is started from the random start condition
described in [21]. The final scores presented correspond to first averaging the evalution score in each period
across seeds, then taking the max average episodic score observed during any evalution period. Of the
tested strategies the n-step UBE approach was the highest performer in 32 out of 57 games, the 1-step UBE
approach in 14 games, DQN in 1 game, the exploration bonus strategy in 7 games, and there were 3 ties. In
table 1 we give the mean and median normalized scores as percentage of an expert human normalized score
across all games, and the number of games where the agent is ‘super-human’, for each tested algorithm.
Note that the mean scores are significantly affected by a single outlier with very high score (‘Atlantis’),
and therefore the median score is a better indicator of agent performance.In Figure 1 we plot the number
of games at super-human performance against frames for each method, and in Figure 2 we plot the median
performance across all games against frames for all methods. The results across all 57 games, as well as the
learning curves for all 57 games, are given in the appendix.
Of particular interest is the game ‘Montezuma’s Revenge’, a notoriously difficult exploration game
where no one-step algorithm has managed to learn anything. Our 1-step strategy learns in 200M frames
a policy that is able to consistently get about 500 points, which is the score the agent gets for picking up
the first key and moving into the second room. In Figure 3 we show the learning progress of the agents
for 500M frames where we set the Thompson sampling parameter slightly higher; 0.016 instead of 0.01
(since this game is a challenging exploration task it stands to reason that a higher exploration parameter is
required). By the end of 500M frames the n-step agent is consistently getting around 3000 points, which
is several rooms of progress. These scores are close to state-of-the-art, and are state-of-the-art for one-step
methods (like DQN) to the best of our knowledge.
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
mean
688.60
472.93
776.40
439.88
median
79.41
76.73
94.54
126.41
> human
21
24
26
35
Table 1: Scores for the Atari suite, as a percentage of human score.
6
Conclusion
In this paper we derived a Bellman equation on the uncertainties of any estimator of the Q-values of a
policy, under some conditions. This allows an agent to propagate uncertainty across many time-steps, socalled ‘deep-exploration’, in the same way that value propagates through time in the standard dynamic
9
Games at 100% human performance.
35
30
25
20
15
10
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
5
0
1 10 20
50
100
Millions of Frames
200
Figure 1: Number of games at super-human performance.
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
0.2
Median performance for all games
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
1 10 20
50
100
Millions of Frames
200
Figure 2: Median performance across all games.
montezuma_revenge
Average Episode Return
3000
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
2500
2000
1500
1000
500
0
0
100
200
300
Million Frames
400
500
Figure 3: UBE on Montezuma’s Revenge for 500M frames.
10
programming recursion. This uncertainty can be used by the agent to make decisions about which states
and actions to explore, in order to gather more data about the environment and learn a better policy. Since
the uncertainty satisfies a Bellman recursion, the agent can learn it using the same reinforcement learning
machinery that have been developed for value functions. We showed that an algorithm based on this learned
uncertainty can boost the performance of standard deep-RL techniques. Our technique was able to improve
the average performance of DQN across the Atari suite of games, when compared against DQN using greedy.
7
Acknowledgments
We thank Marc Bellemare, David Silver, Koray Kavukcuoglu, Tejas Kulkarni, Mohammad Gheshlaghi Azar,
and Bilal Piot for discussion and suggestions on the paper.
References
[1] M. G. A ZAR , R. M UNOS , AND B. K APPEN, On the sample complexity of reinforcement learning
with a generative model, in Proceedings of the 29th International Conference on Machine Learning
(ICML), 2012.
[2] M. B ELLEMARE , S. S RINIVASAN , G. O STROVSKI , T. S CHAUL , D. S AXTON , AND R. M UNOS, Unifying count-based exploration and intrinsic motivation, in Advances in Neural Information Processing
Systems, 2016, pp. 1471–1479.
[3] M. G. B ELLEMARE , Y. NADDAF, J. V ENESS , AND M. B OWLING, The arcade learning environment:
An evaluation platform for general agents, Journal of Artificial Intelligence Research, (2012).
[4] R. B ELLMAN, Dynamic programming, Princeton University Press, 1957.
[5] D. P. B ERTSEKAS, Dynamic programming and optimal control, vol. 1, Athena Scientific, 2005.
[6] J. A. B OYAN, Least-squares temporal difference learning, in ICML, 1999, pp. 49–56.
[7] R. I. B RAFMAN AND M. T ENNENHOLTZ, R-max: A general polynomial time algorithm for nearoptimal reinforcement learning, Journal of Machine Learning Research, 3 (2002), pp. 213–231.
[8] G. H. G OLUB AND C. F. VAN L OAN, Matrix computations, vol. 3, JHU Press, 2012.
[9] A. G UEZ , D. S ILVER , AND P. DAYAN, Efficient Bayes-adaptive reinforcement learning using samplebased search, in Advances in Neural Information Processing Systems, 2012, pp. 1025–1033.
[10] G. H. H ARDY, J. E. L ITTLEWOOD , AND G. P ÓLYA, Inequalities, Cambridge university press, 1952.
[11] T. JAKSCH , R. O RTNER , AND P. AUER, Near-optimal regret bounds for reinforcement learning, Journal of Machine Learning Research, 11 (2010), pp. 1563–1600.
[12] S. M. K AKADE, On the sample complexity of reinforcement learning, PhD thesis, University of London London, England, 2003.
11
[13] M. K EARNS AND S. S INGH, Near-optimal reinforcement learning in polynomial time, Machine Learning, 49 (2002), pp. 209–232.
[14] T. L. L AI AND H. ROBBINS, Asymptotically efficient adaptive allocation rules, Advances in applied
mathematics, 6 (1985), pp. 4–22.
[15] T. L ATTIMORE, Regret analysis of the anytime optimally confident ucb algorithm, arXiv preprint
arXiv:1603.08661, (2016).
[16] T. L ATTIMORE AND M. H UTTER, Pac bounds for discounted mdps, in International Conference on
Algorithmic Learning Theory, Springer, 2012, pp. 320–334.
[17] S. M ANNOR , D. S IMESTER , P. S UN , AND J. N. T SITSIKLIS, Bias and variance approximation in
value function estimates, Management Science, 53 (2007), pp. 308–322.
[18] S. M ANNOR AND J. T SITSIKLIS, Mean-variance optimization in Markov decision processes, in Proceedings of the 28th International Conference on Machine Learning (ICML), 2011, pp. 177–184.
[19] V. M NIH , A. P. BADIA , M. M IRZA , A. G RAVES , T. L ILLICRAP, T. H ARLEY, D. S ILVER , AND
K. K AVUKCUOGLU, Asynchronous methods for deep reinforcement learning, in Proceedings of the
33rd International Conference on Machine Learning (ICML), 2016, pp. 1928–1937.
[20] V. M NIH , K. K AVUKCUOGLU , D. S ILVER , A. G RAVES , I. A NTONOGLOU , D. W IERSTRA , AND
M. R IEDMILLER, Playing atari with deep reinforcement learning, in NIPS Deep Learning Workshop,
2013.
[21] V. M NIH , K. K AVUKCUOGLU , D. S ILVER , A. A. RUSU , J. V ENESS , M. G. B ELLEMARE ,
A. G RAVES , M. R IEDMILLER , A. K. F IDJELAND , G. O STROVSKI , S. P ETERSEN , C. B EATTIE ,
A. S ADIK , I. A NTONOGLOU , H. K ING , D. K UMARAN , D. W IERSTRA , S. L EGG , AND D. H ASS ABIS , Human-level control through deep reinforcement learning, Nature, 518 (2015), pp. 529–533.
[22] R. M UNOS, From bandits to monte-carlo tree search: The optimistic principle applied to optimization
and planning, Foundations and Trends R in Machine Learning, 7 (2014), pp. 1–129.
[23] B. O’D ONOGHUE , R. M UNOS , K. K AVUKCUOGLU , AND V. M NIH, Combining policy gradient and
Q-learning, in International Conference on Learning Representations (ICLR), 2017.
[24] I. O SBAND , C. B LUNDELL , A. P RITZEL , AND B. VAN ROY, Deep exploration via bootstrapped
DQN, in Advances In Neural Information Processing Systems, 2016, pp. 4026–4034.
[25] I. O SBAND , D. RUSSO , AND B. VAN ROY, (More) efficient reinforcement learning via posterior
sampling, in Advances in Neural Information Processing Systems, 2013, pp. 3003–3011.
[26] I. O SBAND , D. RUSSO , Z. W EN , AND B. VAN ROY, Deep exploration via randomized value functions, arXiv preprint arXiv:1703.07608, (2017).
[27] I. O SBAND AND B. VAN ROY, On lower bounds for regret in reinforcement learning, stat, 1050
(2016), p. 9.
[28]
, Why is posterior sampling better than optimism for reinforcement learning, arXiv preprint
arXiv:1607.00215, (2016).
12
[29] G. O STROVSKI , M. G. B ELLEMARE , A. V. D . O ORD , AND R. M UNOS, Count-based exploration
with neural density models, arXiv preprint arXiv:1703.01310, (2017).
[30] J. S CHMIDHUBER, Driven by compression progress: A simple principle explains essential aspects of
subjective beauty, novelty, surprise, interestingness, attention, curiosity, creativity, art, science, music,
jokes, in Anticipatory Behavior in Adaptive Learning Systems, Springer, 2009, pp. 48–76.
[31] S. P. S INGH , A. G. BARTO , AND N. C HENTANEZ, Intrinsically motivated reinforcement learning, in
NIPS, vol. 17, 2004, pp. 1281–1288.
[32] M. J. S OBEL, The variance of discounted Markov decision processes, Journal of Applied Probability,
19 (1982), pp. 794–802.
[33] A. S TREHL AND M. L ITTMAN, Exploration via modelbased interval estimation, 2004.
[34] M. S TRENS, A Bayesian framework for reinforcement learning, in ICML, 2000, pp. 943–950.
[35] R. S UTTON AND A. BARTO, Reinforcement Learning: an Introduction, MIT Press, 1998.
[36] R. S. S UTTON, Learning to predict by the methods of temporal differences, Machine learning, 3 (1988),
pp. 9–44.
[37] A. TAMAR , D. D I C ASTRO , AND S. M ANNOR, Learning the variance of the reward-to-go, Journal of
Machine Learning Research, 17 (2016), pp. 1–36.
[38] W. R. T HOMPSON, On the likelihood that one unknown probability exceeds another in view of the
evidence of two samples, Biometrika, 25 (1933), pp. 285–294.
[39] T. T IELEMAN AND G. H INTON, Lecture 6.5-rmsprop: Divide the gradient by a running average of its
recent magnitude, COURSERA: Neural Networks for Machine Learning, 4 (2012).
[40] C. J. C. H. WATKINS, Learning from delayed rewards, PhD thesis, University of Cambridge England,
1989.
13
8
8.1
Appendix
Bound on the cross-term
Here we prove the bound on the cross-term given in equation (8)
h
i
c(s, a) = Ed ((s, a) − Ed (s, a))Es0 ,a0 [Q̂π (s0 , a0 ) − Qπ (s0 , a0 ) − bias Q̂π (s0 , a0 )]
h
= Es0 ,a0 Ed (s, a) − Ed (s, a) (s0 , a0 ) − Ed (s0 , a0 )
i
+γEs00 ,a00 [Q̂(s00 , a00 ) − Q(s00 , a00 ) − bias Q̂π (s00 , a00 )]
h
= Es0 ,a0 ,s00 ,a00 ,... Ed (s, a) − Ed (s, a)
i
× (s0 , a0 ) − Ed (s0 , a0 ) + γ((s00 , a00 ) − Ed (s00 , a00 ) + γ 2 (. . .
h
≤ maxs0 ,a0 ,s00 ,a00 ,... Ed (s, a) − Ed (s, a)
i
× (s0 , a0 ) − Ed (s0 , a0 ) + γ((s00 , a00 ) − Ed (s00 , a00 ) + γ 2 (. . .
= var (s, a)(1 + γ + γ 2 + . . .) = var (s, a)/(1 − γ),
where we have repeatedly used the fact that Q̂π (t, b) − Qπ (t, b) = γEt0 ,b0 (Q̂π (t0 , b0 ) − Qπ (t0 , b0 )) + (t, b)
for any (t, b), the fact that the biases satisfy a Bellman equation, and assumption 1 which implies that the
max over each s0 , a0 is attained when (s0 , a0 ) = (s, a).
If we have more knowledge about the policy and the MDP then we can provide tighter bounds. For
example, if we know that under the policy the agent cannot visit the same state before T time-steps then the
cross-term can be bounded by c(s, a) ≤ γ T −1 var (s, a)/(1 − γ T ) and the multiplicative factor becomes
(1 + γ T )/(1 − γ T ). In particular a policy which never visits the same state twice (e.g., where the MDP is
acyclic) has c(s, a) = 0 and the multiplicative factor in the uncertainty Bellman equation is one.
14
Game
alien
amidar
assault
asterix
asteroids
atlantis
bank heist
battle zone
beam rider
berzerk
bowling
boxing
breakout
centipede
chopper command
crazy climber
defender
demon attack
double dunk
enduro
fishing derby
freeway
frostbite
gopher
gravitar
hero
ice hockey
jamesbond
kangaroo
krull
kung fu master
montezuma revenge
ms pacman
name this game
phoenix
pitfall
pong
private eye
qbert
riverraid
road runner
robotank
seaquest
skiing
solaris
space invaders
star gunner
surround
tennis
time pilot
tutankham
up n down
venture
video pinball
wizard of wor
yars revenge
zaxxon
DQN
40.96
58.17
479.34
67.26
1.86
28662.58
56.43
70.47
63.22
18.29
13.03
782.87
1438.45
34.54
58.06
428.59
83.94
338.43
481.10
92.01
85.03
81.08
9.86
392.09
2.38
49.27
64.17
193.96
266.39
656.32
103.20
-0.49
16.11
114.24
115.17
5.49
112.04
-0.04
79.41
43.41
524.43
770.41
7.93
11.47
3.48
87.42
309.47
16.67
145.58
92.53
148.29
92.62
4.05
1230.64
73.26
37.93
35.24
DQN Intrinsic Motivation
28.13
41.50
647.16
74.01
2.25
14382.27
54.20
77.80
58.60
21.30
20.10
784.74
1377.36
25.04
74.19
431.10
78.20
372.29
575.90
102.36
95.68
121.85
16.02
558.74
4.07
63.34
64.49
230.36
311.21
702.15
112.79
4.21
16.17
101.23
157.34
5.49
112.06
0.44
103.53
46.33
531.48
779.94
10.61
15.26
8.23
72.95
398.98
19.44
145.58
76.73
191.72
95.73
10.77
2393.54
73.96
23.24
53.14
DQN UBE 1-step
46.90
83.48
887.18
82.34
3.54
28655.71
97.72
49.63
106.79
44.92
-3.64
811.68
2045.13
22.62
69.67
496.21
94.54
765.46
750.00
9.94
97.51
0.02
11.26
656.70
2.06
47.63
38.22
198.40
202.79
1033.61
128.30
11.43
18.42
129.62
199.39
5.49
116.32
-0.44
124.96
60.73
722.15
414.29
9.01
31.55
18.56
125.29
456.86
37.42
227.93
88.36
132.62
139.35
-1.24
3354.58
187.48
46.21
62.20
DQN UBE n-step
43.61
83.14
1112.28
130.95
2.61
6889.77
162.90
101.68
149.93
2284.45
14.12
816.42
1474.66
18.94
75.90
499.34
209.70
897.89
1031.25
154.21
105.80
126.41
21.22
867.45
4.91
34.58
58.43
430.89
537.69
838.45
153.40
0.80
19.82
127.99
167.51
5.49
116.37
0.33
125.85
68.24
732.09
803.06
9.31
54.31
-6.53
138.65
547.39
60.98
145.58
121.64
138.12
142.12
8.73
1992.87
118.76
44.36
56.44
15
Table 2: Normalized scores for the Atari suite from random starts, as a percentage of human normalized
score.
Learning curves
alien
3000
2500
2000
1500
1000
500
0
0
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
Million Frames
2500
2000
1500
1000
500
0
0
50
100
Million Frames
1000
800
600
400
200
0
50
150
16
150
200
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
10000
8000
6000
4000
2000
0
200
100
Million Frames
asterix
12000
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
3000
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
1200
0
assault
3500
Average Episode Return
50
amidar
1400
Average Episode Return
Average Episode Return
3500
Average Episode Return
8.2
0
50
100
Million Frames
150
200
50
50
Average Episode Return
Average Episode Return
150
15000
10000
5000
0
50
100
Million Frames
150
bowling
50
100
Million Frames
0
50
17
200
15000
10000
5000
1600
1400
1200
1000
800
600
400
200
0
0
50
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
Million Frames
berzerk
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
0
50
100
Million Frames
150
200
boxing
80
60
40
20
0
20
200
150
battle_zone
100
150
100
Million Frames
20000
0
200
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
0
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
25000
200
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
20000
0
Average Episode Return
100
Million Frames
atlantis
30000
beam_rider
25000
60
55
50
45
40
35
30
25
20
Average Episode Return
bank_heist
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
0
8000000
7000000
6000000
5000000
4000000
3000000
2000000
1000000
0
Million Frames
Average Episode Return
900
800
700
600
500
400
300
200
100
0
0
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
Average Episode Return
Average Episode Return
Average Episode Return
asteroids
2000
1800
1600
1400
1200
1000
800
600
400
200
0
50
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
Million Frames
breakout
0
50
0
50
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
20000
15000
10000
5000
0
50
100
Million Frames
150
10
15
20
0
50
0
50
100
Million Frames
150
200
crazy_climber
80000
60000
40000
20000
0
50
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
Million Frames
demon_attack
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
20000
15000
10000
5000
Average Episode Return
0
0
50
100
Million Frames
150
200
enduro
1000
800
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
600
400
200
0
18
150
100000
25000
200
100
Million Frames
120000
200
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
5
25
1000
0
double_dunk
0
1500
Million Frames
25000
0
Average Episode Return
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
defender
30000
2000
500
Average Episode Return
Average Episode Return
3000
1000
2500
140000
4000
2000
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
3000
Million Frames
5000
0
Average Episode Return
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
150
200
chopper_command
6000
Average Episode Return
100
centipede
3500
Average Episode Return
Average Episode Return
400
350
300
250
200
150
100
50
0
0
50
100
Million Frames
150
200
fishing_derby
40
80
100
0
50
400
300
200
100
0
0
50
150
200
gravitar
350
Average Episode Return
100
Million Frames
250
200
150
100
0
50
100
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
150
200
6
8
10
12
14
16
18
0
50
20
15
10
5
0
50
18000
16000
14000
12000
10000
8000
6000
4000
2000
0
100
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
gopher
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
0
50
150
200
hero
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
5000
0
50
100
Million Frames
150
200
150
200
jamesbond
1400
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
1200
1000
800
600
400
200
0
19
100
Million Frames
10000
0
Million Frames
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
150
200
Million Frames
15000
Million Frames
ice_hockey
4
25
20000
300
50
Average Episode Return
500
Average Episode Return
Average Episode Return
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
30
0
Million Frames
frostbite
600
Average Episode Return
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
Average Episode Return
20
60
freeway
35
Average Episode Return
Average Episode Return
0
0
50
100
Million Frames
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
0
50
Average Episode Return
20000
15000
10000
5000
0
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
50
100
Million Frames
150
3000
2500
2000
1500
1000
500
0
0
50
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
8000
6000
2000
0
0
50
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
50
400
300
200
100
0
50
100
Million Frames
150
200
name_this_game
7000
6000
5000
4000
3000
2000
0
50
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
Million Frames
pitfall
100
200
300
400
500
20
Million Frames
montezuma_revenge
0
Million Frames
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
1000
Million Frames
10000
4000
0
8000
phoenix
12000
krull
500
0
200
ms_pacman
3500
9000
8000
7000
6000
5000
4000
3000
2000
1000
600
Average Episode Return
Average Episode Return
200
25000
0
Average Episode Return
150
kung_fu_master
30000
Average Episode Return
100
Million Frames
Average Episode Return
kangaroo
Average Episode Return
Average Episode Return
16000
14000
12000
10000
8000
6000
4000
2000
0
0
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
50
100
Million Frames
150
200
pong
10
0
20
0
0
50
Average Episode Return
10000
0
50
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
Million Frames
2000
500
0
0
50
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
200
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
6000
4000
2000
0
50
100
Million Frames
150
200
robotank
50
40
30
20
10
0
50
100
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
150
200
Million Frames
skiing
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
16000
18000
20000
22000
24000
26000
21
150
riverraid
14000
Million Frames
100
Million Frames
8000
0
200
2500
1000
50
60
3000
1500
0
0
seaquest
3500
0
Million Frames
30000
0
Average Episode Return
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
40000
20000
500
10000
road_runner
50000
1000
500
qbert
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
1500
Million Frames
Average Episode Return
16000
14000
12000
10000
8000
6000
4000
2000
0
50
Average Episode Return
Average Episode Return
30
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
Average Episode Return
20
10
private_eye
2000
Average Episode Return
Average Episode Return
30
0
50
100
Million Frames
150
200
solaris
2000
1500
1000
0
50
30000
20000
10000
0
15
10
5
0
5
10
15
20
25
30
0
50
100
Million Frames
150
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
50
100
Million Frames
150
50
0
50
100
Million Frames
150
6
7
22
200
8
9
0
50
100
Million Frames
150
200
150
200
time_pilot
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
7000
6000
5000
4000
3000
2000
0
50
100
Million Frames
up_n_down
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
8000
6000
4000
2000
0
200
Million Frames
surround
10000
100
0
50
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
1000
200
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
150
0
8000
tutankham
200
500
10
200
tennis
0
1000
5
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
40000
1500
0
200
Average Episode Return
Average Episode Return
150
star_gunner
50000
Average Episode Return
100
Million Frames
Average Episode Return
500
Average Episode Return
Average Episode Return
2500
space_invaders
2000
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
Average Episode Return
Average Episode Return
3000
0
50
100
Million Frames
150
200
venture
0
50
150
Million Frames
5000
4000
Average Episode Return
3000
2000
1000
0
0
50
100
Million Frames
100000
50000
0
Average Episode Return
6000
150000
150
0
50
100
Million Frames
20000
5000
0
0
50
3000
0
50
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
100
150
200
Million Frames
23
100
Million Frames
4000
0
200
10000
5000
1000
150
15000
zaxxon
2000
200
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
25000
200
6000
150
yars_revenge
30000
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
200000
200
wizard_of_wor
7000
Average Episode Return
100
video_pinball
250000
DQN
DQN Intrinsic Motivation
DQN UBE 1-step
DQN UBE n-step
Average Episode Return
Average Episode Return
90
80
70
60
50
40
30
20
10
0
| 2cs.AI
|
Fibres of Failure: Classifying errors in predictive processes
Leo Carlsson 1 Gunnar Carlsson 2 3 Mikael Vejdemo-Johansson 4
arXiv:1803.00384v1 [cs.CV] 9 Feb 2018
Abstract
We describe Fibres of Failure (F I FA), a method
to classify failure modes of predictive processes
using the M APPER algorithm from Topological
Data Analysis. Our method uses M APPER to build
a graph model of input data stratified by prediction
error. Groupings found in high-error regions of
the M APPER model then provide distinct failure
modes of the predictive process.
We demonstrate F I FA on misclassifications of
MNIST images with added noise, and demonstrate two ways to use the failure mode classification: either to produce a correction layer that adjusts predictions by similarity to the failure modes;
or to inspect members of the failure modes to illustrate and investigate what characterizes each
failure mode.
2016; Whitmore et al., 2016; Ribeiro et al., 2016b; Singh
et al., 2016; Phillips et al., 2017; Ribeiro et al., 2016a;c) to
name a few prominent directions.
In this paper, we present a method, Fibres of Failure that
draws on topological data analysis to produce model diagnostics through a classification of prediction failure modes
in feature space. Our method relates to both the input data
segmentation and the model diagnostics directions of research by finding and classifying input regions that behave
unexpectedly or erroneously as compared to what the model
is designed to predict.
Noisy input as well as adversarial learning has been used
to motivate and to generate examples and insights for interpretability (Kindermans et al., 2016). We will use the same
basic idea to illustrate our method – by studying prediction
failures on MNIST images with added noise.
2. Related work
1. Introduction
In recent years the interest in transparent, interpretable and
explainable models in machine learning has grown dramatically, with dedicated workshops at NIPS 2016(Wilson et al.,
2016), NIPS 2017 (Tosi et al., 2017; Wilson et al., 2017)
and ICML 2017 (Varshney et al., 2017) as well as attention
from grant agencies (Gunning, 2016).
The approaches to interpretable models go in several distinct directions – producing sparse models (Hara & Maehara,
2016; Wisdom et al., 2016; Hayete et al., 2016; Tansey et al.,
2017), visualization techniques (Smilkov et al., 2016; Selvaraju et al., 2016; Thiagarajan et al., 2016; Gallego-Ortiz
& Martel, 2016; Krause et al., 2016; Zrihem et al., 2016;
Handler et al., 2016), hybrid models (Krakovna & DoshiVelez, 2016; Reing et al., 2016), input data segmentation
(Samek et al., 2016; Hechtlinger, 2016; Thiagarajan et al.,
2016), and model diagnostics with or without blackbox interpretation layers (Lundberg & Lee, 2016; Vidovic et al.,
1
KTH Royal Institute of Technology 2 Ayasdi Inc. 3 Stanford
University 4 CUNY College of Staten Island. Correspondence
to: Leo Carlsson <[email protected]>, Mikael Vejdemo-Johansson
<[email protected]>.
Preliminary work. Under review. Copyright 2018 by the author(s).
One interpretability method with a large impact on the field,
LIME (Ribeiro et al., 2016c), inspects single instances by
perturbing the input and tracing how predictions change with
the perturbation. Other interpretability methods focus closer
on aggregates of inputs, such as TreeView (Thiagarajan
et al., 2016), which visualizes deep neural networks by first
clustering neurons by activation patterns, then clusters these
groups by prediction labels, and finally trains a predictor to
predict the meta-clusters from the input data directly.
The F I FA method builds on M APPER, an algorithm from
Topological Data Analysis that constructs a graph (or simplicial complex) model of arbitrary data. M APPER has had
success in a wide range of application areas, from medical
research studying cancer, diabetes, asthma and many more
topics (Nicolau et al., 2011; Li et al., 2015; Hinks et al.,
2016; Schneider et al., 2016), genetics and phenotype studies (Romano et al., 2014; Carlsson, 2017; Cámara, 2017;
Savir et al., 2017; Bowman et al., 2008), to hyperspectral
imaging, material science, sports and politics (Duponchel,
2018a;b; Lee et al., 2017; Lum et al., 2013). Of note for
our approach are in particular the contributions on cancer,
diabetes and fragile X syndrom (Nicolau et al., 2011; Romano et al., 2014; Li et al., 2015) where M APPER was used
to extract new subgroups from a segmentation of the input
Fibres of Failure
space.
Our results build on two fundamental concepts: viewing
predictive models as functions and therefore usable as input to M APPER, and the M APPER technique for producing
intrinsic graph models of arbitrary data sets.
As a running illustration in this paper we will be looking
at how a CNN trained on the MNIST dataset fails when
encountering noisy images derived from MNIST. The influence of noise on learning algorithm performance has
been studied. (Zhou et al., 2017; Dodge & Karam, 2016)
found a dramatic increase in error rates with increased image distortion, confirming our choice of illustrative test case.
Adversarial learning is another method that has been proven
successful at deteriorating performance for trained networks
(Cisse et al., 2017a; Yuan et al., 2018; Moosavi-Dezfooli
et al., 2016; Chen et al., 2018).
A lot of work has been done on making deep networks more
robust against perturbations: both against noise deterioration
and against adversarial manipulation (Fawzi et al., 2016;
Hein & Andriushchenko, 2017; Noh et al., 2017; Cisse et al.,
2017b; Huang et al., 2015; Tramèr et al., 2018).
3. Proposed method
The proposed method, F IBRES OF FAILURE (F I FA ), takes a
different approach from the related work. We do not intend
to modify deep neural network models, rather we create
classifiers on top of the model that recognizes specific types
of faulty predictions (failure modes) from a deep learning
model trained to recognize MNIST images.
3.1. M APPER
M APPER (Singh et al., 2007) is an algorithm that constructs
a graph (more generally a simplicial complex) model for a
point cloud data set. The graph is constructed systematically
from some well defined input data. It was defined in (Singh
et al., 2007), and has been shown to have great utility in the
study of various kinds of data sets (as described in Section
2). It can be viewed as a method of unsupervised analysis
of data, in the same way as principal component analysis,
multidimensional scaling, and projection pursuit can, but it
is more flexible than any of these methods. Comparisons
of the method with standard methods in the context of hyperspectral imaging have been documented in (Duponchel,
2018a;b).
In topological language, M APPER starts with the choice of
a collection of continuous filter functions and an open cover
over their range. The fibres, or preimages of this open cover
produces an open cover on the data space, which can be
refined using connected components. Doing this with a fine
enough cover and non-degenerate filter functions produces a
good cover in the sense of the nerve lemma (Hatcher, 2002),
so the nerve complex is homotopy equivalent with the data
source.
An open cover here is almost, but not quite the same thing
as a partition. In order to track connectivity information,
the partition cannot be allowed to become disconnected –
that would miss parts of the space, and introduce artificial
disconnects. The open cover most cleanly translates into
a “fattened” partition, or a partition with overlaps between
adjacent parts.
In more detail, and using a more data-focused and less
topological description, M APPER proceeds by the following
steps. We let X (the dataset) be a finite metric space.
1. Select arbitrary functions f1 , f2 , . . . , fk : X → R. We
call these filter functions and they encode a separation
of datapoints. In practice, the number k is usually 1, 2,
or 3. Common filter functions are statistically meaningful quantities such as the values of a density estimator
or centrality measure, or outputs from a machine learning algorithm such as PCA or MDS, or a variable used
in defining the data set.
2. For each of the functions, pick parameters to produce
an overlapping partition of R: a number Ni of partitions and a proportion of overlap 0 < p < 1.
3. For each function fi , let ai and bi denote the minimum
and maximum values taken by fi , and construct an
open cover of the interval J i = [ai , bi ] by introducing
Ni subintervals
Jsi = [ai + (s − 1 − p/2)∆i , ai + (s + p/2)∆i ] ⊆ J i
where ∆i = (bi − ai )/Ni and 1 ≤ s ≤ Ni .
4. Construct a (likely overlapping) partition of X by letting each
Us1 ,...,sk = f1−1 (Js11 ) ∩ · · · ∩ fk−1 (Jskk )
with 1 ≤ si ≤ Ni be a part in in the partition.
5. Apply some clustering algorithm to each Us1 ,...,sk to
decompose it into disjoint sets Us1 ,...,sk ;j . For our
experiments, we use single linkage clustering with a
heuristic for cutoff based on the histogram of distances
in U , described in detail in (Singh et al., 2007).
6. Construct a graph by setting the vertices to σ =
(s1 , . . . , sk ; j) and connecting σ, . . . , to σ 0 with
anedge precisely when
Uσ ∩ Uσ0 6= ∅
For the simplicial complex version, vertices
[σ0 , . . . , σd ] are connected when their joint intersection is non-empty.
Fibres of Failure
See (Singh et al., 2007; Carlsson, 2009) for more details.
M APPER has several implementations available: Python
Mapper (Müllner & Babu, 2013), Kepler Mapper (Saul
& van Veen, 2017) and TDAmapper1 are all open source,
while Ayasdi Inc.2 provides a commercial implementation
of the algorithm. For our work we are using the Ayasdi
implementation of M APPER.
3.1.1. M APPER ON PREDICTION FAILURE
The filters in the M APPER function have the effect of ensuring separation of features in the data that are separated
by the filter functions themselves. Step one of F I FA specifically uses a M APPER analysis with prediction error as one
of the filter functions. By including prediction error this
way, the F I FA algorithm guarantees that any groups that are
extracted are homogenous with respect to prediction failure,
and thus useable as a failure mode designation.
We name a M APPER model with prediction failure as a filter
a F I FA model.
3.2. Extract subgroups
Subgroups of the F I FA model with tight connectivity in the
graph structure and with homogenous and large average prediction failure per component cluster provide a classification
of failure modes. These can be selected either manually, or
using a community detection algorithm.
When selecting failure modes manually, a visualization such
as in Figure 2 is most helpful. Here, flares (tightly connected subgraphs emanating from a core, such as Group
40) or tightly connected components, loosely connected to
surrounding parts of the graph, are the most compelling
characterizations of a good failure mode subgroup.
3.3. Quantitative: model correction layer
Once failure modes have been identified, one way to use the
identification is to add a correction layer to the predictive
process. Use a classifier to recognize input data similar to a
known failure mode, and adjust the predictive process output
according to the behavior of the failure mode in available
training data.
3.3.1. T RAIN CLASSIFIERS
For our illustrative examples, we demonstrate several “one
vs rest” binary classifier ensembles where each classifier
is trained to recognize one of the failure modes (extracted
subgroups) from the Mapper graph. We demonstrate performance of F I FA for model correction using Linear SVM,
1
http://cran.r-project.org/web/packages/
TDAmapper
2
http://ayasdi.com
Logistic Regression, and Naı̈ve Bayes classifiers.
3.3.2. E VALUATE BIAS
A classifier trained on a failure mode may well capture
larger parts of test data than expected. As long as the space
identified as a failure mode has consistent bias, it remains
useful for model correction: by evaluating the bias in data
captured by a failure mode classifier we can calibrate the
correction layer.
3.3.3. A DJUST MODEL
The actual correction on new data is a type of ensemble
model, and has flexibility on how to reconcile the bias prediction with the original model prediction – or even how to
reconcile several bias predictions with each other. For our
example in this paper we choose to override the CNN prediction with the observed ground truth in the failure mode
from the training data used to create the classifier. For regression tasks we have also used the average of the failure
mode training group as an offset to subtract from the model
prediction.
3.4. Qualitative: model inspection
Identifying distinct failure modes and giving examples of
these is valuable for model inspection and debugging. Statistical methods, such as Kolmogorov-Smirnov testing, can
provide measures of how influential any one feature is in
distinguishing one group from another and can give notions
of what characterizes any one failure mode from other parts
of input space. With examples and distinguishing features
in hand, we can go back to the original model design and
evaluate how to adapt the model to handle the failure modes
better.
Much of the work in interpretability for machine learning
provides tools to inspect examples, and for providing a
model explanation for a specific example. These work well
in conjunction with F I FA to find explanations for the identified failure modes.
4. Experiments
In order to evaluate the F I FA method we have trained a
CNN classifier on the MNIST data set, created prediction
failures by adding noise to the data, and gone through the
F I FA pipeline for the resulting erroneous predictions. With
distinct failure modes extracted, we then illustrate both a
quantitative and a qualitative approach to handling the output from F I FA: on the one hand we adjust predictions using
classifiers trained on recognizing each failure mode and
measure the improvement in classification on the resulting
ensemble approach, on the other hand we compare several
failure modes that misclassify versions of the same digit
Fibres of Failure
9216
26x26x32
24x24x64
classifiers.
12x12x64 12x12x64
128
28x28x1
128
10
C(k)
Conv2D
Conv2D
Max
pooling
Dropout
25%
Flatten
Dense
Dense
Dropout
50%
Softmax
Figure 1. The topology for the CNN model. The numbers display
the dimension of each layer in the model. The abbreviations,
such as Conv2D, describes the specific transformations performed
between layers in the model. The activation functions for the
classification layer was ’Softmax’ and for the other layers ’ReLU’.
The optimizer used was ’Adadelta’.
The use of probabilities for predicted and ground truth digit
as filters guarantees that M APPER separates regions of correct predictions from those of wrong predictions. After all,
these probabilities are measures of error for the CNN model.
We purposely omitted the activations from the Dense-10
layer as input variables because of the direct reference to the
probabilities for both the ground truth digit and the predicted
digit.
The following variables were included in the analysis but
were not used to create the F I FA model:
10 activations from the Dense-10 layer, which consists of
the probabilities for each digit, 0-9.
(the digit 5) in different ways.
784 pixel values representing the flattened MNIST image
of size 28x28x1.
4.1. Protocol
6 variables: prediction by the CNN model, ground truth
digit, corrupt or original data (binary), correct or incorrect
prediction(binary), probability of the Predicted digit (highest value of the Dense-10 layer), and probability of ground
truth digit.
We created a CNN model with a topology shown in Figure 1.
The network topology and parameters was chosen arbitrarily
with the only condition that it performs well on the original
MNIST data set. The activation functions was ’Softmax’
for the classification layer and ’ReLU’ for all other layers.
The optimizer was Adadelta with learningrate = 1.0,
ρ = 0.95, and = 1e − 7. We trained the model on 60,000
clean MNIST training images and tested it on 10,000 clean
MNIST images through 12 epochs. The accuracy on the
test-set of 10,000 clean MNIST images was 99.05%. We
created 10,000 corrupt MNIST images using 25% random
binary flips on the clean test images[source for code]. The
accuracy on the corrupt MNIST images was 40.45%.
To create the M APPER graph we used the following:
Filters: Principal Component 1, probability of Predicted
digit, probability of Ground truth digit, and Ground truth
digit. Our measure of predictive error is the probability of
Ground truth digit. By including the Ground truth digit
itself we separate the model on ground truth, guaranteeing
that any one one failure mode has a consistent ground truth
that can be used for corrections.
Metric: Variance Normalized Euclidean
Variables: 9472 network activations: all activations after
the Dropout layer that finishes the convolutional part in the
network and before the softmax layer that provides the final
predictions. These are the layers with 9216, 128 and 128
nodes displayed in Figure 1.
Instances: We randomly shuffled the data from the 10,000
clean and 10,000 corrupt images that were used to test the
CNN model, and split the 20,000 instances into 5 training
sets of size 16,000 each and 5 test sets of size 4,000 each.
The training sets was used to create 5 M APPER graphs.
This is in order to perform 5-fold cross validation on the
Hence, the total number of variables in our analysis were
10272.
To extract failure modes from the F I FA model we used a
supervised community detection method to find groups of
approximately constant prediction error. In the M APPER
implementation we are using a grouping method based on
Agglomerative Hierarchical Clustering (AHCL) (Edwards
& Cavalli-Sforza, 1965; Murtagh & Contreras, 2012) and
Louvain Modularity(Blondel et al., 2008) is included. As
supervision, a function on the data is chosen – for F I FA,
choose the measure of prediction error. The difference in
means of the supervision function produces a graph edge
weighting: edges are weighted as “strong” if they have similar supervision function values, and “weak” if the supervision function values are different. With the graph weighting
in place, hierarchical clustering produces a clustering tree
using the weighted edges to generate a graph metric to cluster over. Finally, Louvain modularity identifies an optimal
graph partition from the clustering tree.
From partitioned groups, we retain as failure modes those
groups that have at least 15 data points and have less than
99.05% correct predictions, which is the accuracy of the
CNN model on the original MNIST test data.
We trained classifiers in a one vs. rest scheme on each
group in the 5 folds of data that were used to create the 5
M APPER graphs. We used the following types of classifiers
with varying parameters shown in square brackets:
Linear-SVM Loss function: squared hinge, Penalty
function:
`2 .
Regularization parameter, C =
Fibres of Failure
[0.001, 0.01, 0.1, 1, 10, 100, 1000].
Logistic Regression Penalty function: `2 . Regularization
parameter, C = [0.001, 0.01, 0.1, 1, 100, 10, 1000].
Naı̈ve Bayes Gaussian Naı̈ve Bayes using class priors for
each group in the training data set.
We used the parameters from each best performing classifier
to train new models. This time, we evaluated each model
on second test data set, called ’Corrupt’, which consisted of
10,000 new corrupt images using 25% binary flips on the
original MNIST test dataset. Hence, we used the same noise
setup as the corrupt images used for testing the CNN model.
For the test data sets, we evaluated to what extent each
classifier predicted member points with the same ground
truth digit as that of the group the model was trained on. As
we trained the classifiers on groups containing a lot of wrong
predictions, it is expected that the classifiers will classify
member points with wrong predictions on the test data sets.
Hence, we offset the predicted digits with the ground truth
digit of the group it was trained on. We attempt to exploit
the consistent bias of the classifiers to improve the accuracy
of the now combined CNN and classifier ensemble.
4.2. Quantitative Results
The following parameters were chosen for the three classifiers we evaluated as model correction layers:
Linear-SVM C = 1. (chosen as highest accuracy in a
5-fold crossvalidation)
Logistic Regression C = 1000 (chosen as highest accuracy
in a 5-fold crossvalidation)
Naı̈ve Bayes Gaussian Naı̈ve Bayes using priors induced
from data.
The average number of data points in all failure mode groups
in the 5 folds were 4937 of the total 16,000. The average
number of clean data points in all groups in the 5 folds were
10.4, accounting for a fraction of 0.21% of the 4937 data
points. This also means that the failure mode groups encompasses roughly 62% of all corrupt data points in the training
set. The number of failure modes (extracted subgroups) in
each fold were 41, 41, 41, 41, and 37, respectively.
Table 1 shows the accuracy on the two test data sets using CNN with and without F I FA. The linear-SVM classifiers performed best on both data sets with an improvement by 6.43%pt on the 5-fold cross validation test sets and
19.33%pt on the ’Corrupt data’.
4.3. Qualitative Results
For the qualitative analysis, we chose to focus on four
groups with digit 5 as the ground truth digit. Group 50,
G40
G47
G30
G50
Figure 2. The Mapper graph for the CNN on MNIST dataset colored with probability of predicting the ground truth digit. The
colorbar is for interpreting the values of the coloring. The circled
nodes and edges are the groups Group30, Group40, Group47, and
Group50. The 5-fold Mapper graphs are shown in the Supplement.
which is not one of the failure mode groups and Groups 30,
40, and 47, all part of the total 39 failure mode groups. The
locations of each group are shown in Figure 2. The distribution of predicted probabilities for each label is shown in
Figure 4: group 30 is the group with highest probability to
the digit 5, while 40 and 47 are more focused on 8, 2, and 3.
All three groups favor digit 8 as their mean probabilities are
between 0.5-0.9.
We compared these three failure modes with the non-failure
Group 50 and extracted the 5 activations with the highest KSvalues from the Dense-128 layer. See Figure 1. To illustrate
the differences between the three failure modes regarding
the activations, we have provided a selection of saliency
maps (Simonyan et al., 2013) for all images considered
as true members of each of the three failure mode groups.
These were all produced using the keras-vis Python
package.
Figure 3 shows a selection of noisy images and their saliency
maps for some of the activations highest KS-values within
the Dense-128 layer. The two leftmost image pairs were
selected based on visual clear saliency maps with respect
to digits. The two rightmost were selected based on most
unclear/noisy saliency maps. The full collection of saliency
Fibres of Failure
5-fold
(4 000)
% clean
49.7%
Corrupt
(10 000)
CNN
69.40%
48.53%
41.14%
CNN+LR
CNN+SVM
CNN+NB
75.45%
75.83%
73.33%
0.21%
0.32%
6.85%
59.57%
60.47%
48.34%
Table 1. Performance of the CNN as compared to CNN with F I FA
driven improvements both on the average of the 5 folds of test
data and on entirely corrupted test data. The improvements by
each classifier ensemble are for the best performing parameters. In
addition to prediction accuracy, we also report the average proportion of uncorrupted (clean) data points in the 5-fold test data set
as well as in the predicted data points by each classifier ensemble.
Bold face marks the best performances (highest accuracy; lowest
percentage of clean digits caught). Noticeably linear classifiers
perform well, producing an almost 20%pt increase in accuracy
on corrupted data while imposing corrections on almost no clean
images.
Activation 24 – one row each for groups 30, 40 and 47
Activation 81 – one row each for groups 30, 40 and 47
Group 30, activation 89
maps for these groups can be found in our supplemental
material.
The activations 24 and 81, present in all three groups, display activity that is consistent with an activation detecting
features of the digit 5, while the activations 89 and 99 correspond closer to an activation for the digit 3 and 119, 122 and
124 correspond to activations for the digit 8. In particular in
the last three groups, noise that closes loops in a written 5
tend to have high saliency.
In Table 2 we show the percentage of blank saliency maps,
indicating that an activation is missing completely for a
particular input.
Group 30, activation 124
Group 40, activation 89
Group 40, activation 99
Group 40, activation 119
5. Discussion and Conclusion
For the quantitative approach to handling failure modes we
could see significant improvement even using quite simplistic classifiers for constructing a correction layer: an increase
by almost 20%pt, avoiding corrections on almost all uncorrupted images was seen in both the linear separation
methods: both with logistic regression and SVM.
On the qualitative side, an inspection of the saliency maps
– see Figure 3 for a selection of particularly illustrative
maps, and the supplementary material for a full collection –
showed us that the groups were distinguished from group 50
containing correctly predicted digits 5 differed in network
activity either by an activation tuned to detecting 5s, or in
an activation that often looked for closing loops and found
them in the added noise. Blank saliency maps were common for the 5-detecting neurons, as can be seen in Table 2,
overwhelmingly so for the groups 40 and 47 where correct
Group 47, activation 89
Group 47, activation 122
Figure 3. Example noisy images and saliency maps for activations
in the penultimate dense layer for the three main failure modes
identified for noisy 5s. The two leftmost images were chosen as the
most clear saliency maps with respect to digits. The two rightmost
were selected based on unclear/noisy saliency maps. All saliency
maps are from images classified as members of the respective
failure mode group.
Fibres of Failure
Group 30
Neuron
%Blank
Group 40
24
36.1%
33
26.2%
81
60.7%
89
0%
124
8.2%
Neuron
%Blank
Group 47
24
82.2%
81
91.8%
89
0%
99
4.1%
119
17.9%
Neuron
%Blank
24
70.6%
49
54.3%
81
84.3%
89
0.5%
122
3.6%
Table 2. The percentage of blank (all zero) saliency maps for each
of the 5 neurons with the highest absolute KS-values (compared
to group 50) in the Dense-128 layer. The bold neuron numbers
are the neurons qualitatively identified as encoding digit 5. We
observe that the neurons encoding digit 5 have predominantly
larger percentages of blank saliency maps.
1.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
5
6
●
●
●
●
●
●
0.5
0.0
1.0
●
●
●
●
●
●
●
●
40
Softmax
0.0
1.0
●
●
●
●
●
●
●
●
●
●
●
●
3
4
47
●
●
0.5
0.0
30
0.5
●
●
●
●
●
0
1
2
●
●
●
●
●
●
7
8
9
Ground Truth
Figure 4. The failure modes for a ground truth of 5. We see the
distributions of predictions for the three failure modes: only group
30 attaches any significant likelihood to the digit 5 at all, while all
three favor 8. For group 40, the digits 2 and 3 are also commonly
suggested, while this happens somewhat more rarely in groups 30
and 47.
predictions were rare, and much less commonly in group 30
where as can be seen in Figure 4 a correct prediction still
came with significant strength in the softmax layer.
Using F I FA on a CNN-based MNIST digit classifier that
had to cope with severely corrupted MNIST images we were
able to find 39 distinct failure modes based on activations
in the antepenultimate and penultimate layers of our CNN
model. When inspecting the digit 5 in particular, we found
that the three identified failure modes could be distinguished
from the wellbehaved parts of input space by specific activations that seemed to code for features corresponding closely
to the kinds of misclassifications that were observed.
In addition to inspecting examples, we explored the addition
of a correction layer to the CNN model. The failure modes
act as seeds for training a classifier. The classifier can assign
new data to a known failure mode, so that the correction
layer can adjust for known behaviour of that failure mode.
For regression models, our suggestion would be to treat the
prediction error as bias, and subtract the mean prediction
error for the identified failure mode from the model prediction. In the CNN on corrupted MNIST example we use
to illustrate the methodology, we impose the ground truth
digit from which the identified failure mode emerged as a
replacement prediction. By doing this, we could observe
up to a 19.33%pt improvement in prediction accuracy on
corrupted data while accidentially including only 0.32% of
uncorrupted observations in the correction groups. The percentage of clean data is in well accordance with that in the
failure mode groups; 0.21%.
F I FA is generically applicable. While developing the
method we have used it to analyze an energy based regression model used to predict temperatures in electric arc steel
furnaces. In that application, we found failure modes that
consistently over-predicted and under-predicted by close to
100o C . Adjusting the regression by the mean prediction
error of the failure group provided significant improvement
in the energy model and a qualitative analysis of the failure
modes uncovered metallurgically important observations
about material composition related to high prediction error.
The F I FA method picks out high prediction error regions
from input space of an arbitrary predictive process, and classifies failure modes that are internally similar but that have
significant separation either in the predictive behaviour of
the process or in the distance measure of input space. Having identified failure modes we can view them as witnesses
for misbehaviour in different ways, and produce correspondingly different developments of the predictive process. On
the one hand, a failure mode witnesses a region of input
space with local bias to the predictive process, and we can
correct specifically for that bias by classifying new data as
belonging to that failure mode (or not) and correct predictions for the failure mode members. On the other hand the
failure mode is a witness for some coherent collection of
predictive failures. By inspecting features of input space
that distinguish these from other parts of input space we can
gain insights about types of failure that could be handled by
adjusting the design of the predictive process itself.
References
Blondel, Vincent D, Guillaume, Jean-Loup, Lambiotte, Renaud, and Lefebvre, Etienne. Fast unfolding of communities in large networks. Journal of Statistical Mechanics:
Theory and Experiment, 2008(10):P10008, 2008.
Bowman, G.R., Huang, X., Yao, Y., Sun, J., Carlsson, G.,
Guibas, L.J., and Pande, V.S. Structural insight into rna
hairpin folding intermediates. JACS Communications, pp.
Fibres of Failure
9676–9678, 2008.
Carlsson, Gunnar. Topology and data. American Mathematical Society, 46(2):255–308, 2009.
Carlsson, Gunnar. The shape of biomedical data. Current
Opinion in Systems Biology, 1, 2017.
Chen, Sen, Xue, Minhui, Fan, Lingling, Hao, Shuang, Xu,
Lihua, Zhu, Haojin, and Li, Bo. Automated poisoning
attacks and defenses in malware detection systems: An
adversarial machine learning approach. Computers and
Security, 73:326––344, 2018.
Cisse, Moustapha, Adi, Yossi, Neverova, Natalia, and
Keshet, Joseph. Houdini: Fooling deep structured visual
and speech recognition models with adversarial examples.
In Advances in Neural Information Processing Systems
30, 2017a.
Cisse, Moustapha, Bojanowski, Piotr, Grave, Edouard,
Dauphin, Yann, and Usunier, Nicolas. Parseval networks:
Improving robustness to adversarial examples. In Proceedings of the 34th International Conference on Machine Learning, 2017b.
Cámara, Pablo G. Topological methods for genomics:
Present and future direction. Current Opinion in Systems Biology, 1:95–101, 2017.
Dodge, Samuel and Karam, Lina. Understanding how image quality affects deep neural networks. In 2016 Eighth
International Conference on Quality of Multimedia Experience (QoMEX). IEEE, June 2016.
Duponchel, Ludovic. Exploring hyperspectral imaging data
sets with topological data analysis. Analytica Chimica
Acta, 1000:123–131, 2018a.
Duponchel, Ludovic. When remote sensing meets topological data analysis. Journal of Spectral Imaging, 2018b.
Edwards, Anthony WF and Cavalli-Sforza, L Luka. A
method for cluster analysis. Biometrics, pp. 362–375,
1965.
Fawzi, Alhussein, Moosavi-Dezfooli, Seyed-Mohsen, and
Frossard, Pascal. Robustness of classifiers: from adversarial to random noise. In Advances in Neural Information
Processing Systems 29, 2016.
Gallego-Ortiz, Cristina and Martel, Anne L.
Interpreting extracted rules from ensemble of trees: Application to computer-aided diagnosis of breast MRI.
arXiv:1606.08288 [cs, stat], June 2016. URL http://
arxiv.org/abs/1606.08288. WHI 2016 (ICML
Workshop).
Gunning, David. Explainable artificial intelligence (XAI).
DARPA Broad Agency Announcement DARPA-BAA16-53, 2016.
Handler, Abram, Blodgett, Su Lin, and O’Connor, Brendan. Visualizing textual models with in-text and wordas-pixel highlighting. arXiv:1606.06352 [cs, stat],
June 2016. URL http://arxiv.org/abs/1606.
06352. WHI 2016 (ICML Workshop).
Hara, Satoshi and Maehara, Takanori. Finding Alternate
Features in Lasso. arXiv:1611.05940 [stat], November 2016. URL http://arxiv.org/abs/1611.
05940. NIPS 2016 InterpretML.
Hatcher, Allen. Algebraic Topology. Cambridge University
Press, 2002.
Hayete, Boris, Valko, Matthew, Greenfield, Alex, and Yan,
Raymond. MDL-motivated compression of GLM ensembles increases interpretability and retains predictive power.
arXiv:1611.06800 [stat], November 2016. URL http:
//arxiv.org/abs/1611.06800. NIPS 2016 InterpretML.
Hechtlinger, Yotam. Interpretation of Prediction Models
Using the Input Gradient. arXiv:1611.07634 [cs, stat],
November 2016. URL http://arxiv.org/abs/
1611.07634. NIPS 2016 InterpretML.
Hein, Matthias and Andriushchenko, Maksym. Formal guarantees on the robustness of a classifier against adversarial
manipulation. In Advances in Neural Information Processing Systems 30, 2017.
Hinks, T.S., T, Brown, LC, Lau, H, Rupani, C, Barber,
S, Elliott, JA, Ward, J, Ono, S, Ohta, K, Izuhara, R,
Djukanović, RJ, Kurukulaaratchy, A, Chauhan, and P.,
Howarth. Multidimensional endotyping in patients with
severe asthma reveals inflammatory heterogeneity in matrix metalloproteinases and chitinase 3-like protein 1.
J.Allergy Clin Immunol, 138(1), 2016.
Huang, Ruitong, Xu, Bing, Schuurmans, Dale, and
Szepesvári, Csaba. Learning with a strong adversary.
11 2015.
Kindermans, Pieter-Jan, Schütt, Kristof, Müller, KlausRobert, and Dähne, Sven. Investigating the influence
of noise and distractors on the interpretation of neural networks. arXiv:1611.07270 [cs, stat], November 2016. URL http://arxiv.org/abs/1611.
07270. NIPS 2016 InterpretML.
Krakovna, Viktoriya and Doshi-Velez, Finale. Increasing
the Interpretability of Recurrent Neural Networks Using
Hidden Markov Models. arXiv:1611.05934 [cs, stat],
November 2016. URL http://arxiv.org/abs/
1611.05934. NIPS 2016 InterpretML.
Fibres of Failure
Krause, Josua, Perer, Adam, and Bertini, Enrico. Using Visual Analytics to Interpret Predictive Machine Learning
Models. arXiv:1606.05685 [cs, stat], June 2016. URL
http://arxiv.org/abs/1606.05685.
WHI
2016 (ICML Workshop).
Phillips, Richard L., Chang, Kyu Hyun, and
Friedler, Sorelle A.
Interpretable Active Learning. arXiv:1708.00049 [cs, stat], July 2017. URL
http://arxiv.org/abs/1708.00049.
WHI
2017 (ICML Workshop).
Lee, Yongjin, arthel, Senja D. B, Dlotko, Pawel, Moosavi,
S. Mohamad, Hess, Kathryn, and Smit, Berend. Quantifying similarity of pore-geometry in nanoporous materials.
Nature Communications, 2017.
Reing, Kyle, Kale, David C., Steeg, Greg Ver, and Galstyan,
Aram. Toward Interpretable Topic Discovery via Anchored Correlation Explanation. arXiv:1606.07043 [cs,
stat], June 2016. URL http://arxiv.org/abs/
1606.07043. ICML 2016 Workshop #Data4Good.
Li, Li, Cheng, Wei-Yi, Glicksberg, Benjamin S., Gottesman,
Omri, Tamler, Ronald, Chen, Rong, Bottinger, Erwin P.,
and Dudley, Joel T. Identification of type 2 diabetes subgroups through topological analysis of patient similarity.
Science Translational Medicine, 7(311), 2015.
Lum, Pek Y, Singh, Gurjeet, Lehman, Alan, Ishkanov,
Tigran, Vejdemo-Johansson, Mikael, Alagappan,
Muthu, Carlsson, John, and Carlsson, Gunnar. Extracting insights from the shape of complex data
using topology.
Scientific Reports, 3, February
2013. ISSN 2045-2322. doi: 10.1038/srep01236.
URL http://www.nature.com/srep/2013/
130207/srep01236/full/srep01236.html.
Lundberg, Scott and Lee, Su-In. An unexpected unity
among methods for interpreting model predictions.
arXiv:1611.07478 [cs], November 2016. URL http:
//arxiv.org/abs/1611.07478. NIPS 2016 InterpretML.
Moosavi-Dezfooli, Seyed-Mohsen, Alhussein, Fawzi, and
Frossard, Pascal. Deepfool: a simple and accurate method
to fool deep neural networks. 11 2016.
Murtagh, Fionn and Contreras, Pedro. Algorithms for hierarchical clustering: an overview. Wiley Interdisciplinary
Reviews: Data Mining and Knowledge Discovery, 2(1):
86–97, 2012.
Müllner, Daniel and Babu, Aravindakshan. Python mapper:
An open-source toolchain for data exploration, analysis
and visualization, 2013. URL http://danifold.
net/mapper.
Ribeiro, Marco Tulio, Singh, Sameer, and Guestrin, Carlos.
Model-Agnostic Interpretability of Machine Learning.
arXiv:1606.05386 [cs, stat], June 2016a. URL http://
arxiv.org/abs/1606.05386. WHI 2016 (ICML
Workshop).
Ribeiro, Marco Tulio, Singh, Sameer, and Guestrin, Carlos.
Nothing Else Matters: Model-Agnostic Explanations By
Identifying Prediction Invariance. arXiv:1611.05817 [cs,
stat], November 2016b. URL http://arxiv.org/
abs/1611.05817. NIPS 2016 InterpretML.
Ribeiro, Marco Tulio, Singh, Sameer, and Guestrin, Carlos.
”Why Should I Trust You?”: Explaining the Predictions
of Any Classifier. arXiv:1602.04938 [cs, stat], February 2016c. URL http://arxiv.org/abs/1602.
04938. arXiv: 1602.04938.
Romano, David, Nicolau, Monica, Quintin, Eve-Marie,
Mazaika, Paul K., Lightbody, Amy A., Hazlett,
Heather Cody, Piven, Joseph, Carlsson, Gunnar, and
Reiss, Allan L. Topological methods reveal high and
low functioning neuro-phenotypes within fragile x syndrome. Human Brain Mapping, 35:4904—-4915, 2014.
Samek, Wojciech, Montavon, Grégoire, Binder, Alexander,
Lapuschkin, Sebastian, and Müller, Klaus-Robert. Interpreting the Predictions of Complex ML Models by Layerwise Relevance Propagation. arXiv:1611.08191 [cs, stat],
November 2016. URL http://arxiv.org/abs/
1611.08191. NIPS 2016 InterpretML.
Saul, Nathaniel and van Veen, Hendrik Jacob.
Mlwave/kepler-mapper: 186f (version 1.0.1), 2017. URL
http://doi.org/10.5281/zenodo.1054444.
Nicolau, Monica, Levine, Arnold J., and Carlsson, Gunnar. Topology based data analysis identifies a subgroup
of breast cancers with a unique mutational profile and
excellent survival. PNAS, 108:7265–7270, 2011.
Savir, Aleksandar, Toth, Gergely, and Duponchel, Ludovic.
Topological data analysis (tda) applied to reveal pedogenetic principles of european topsoil system. Science of
the Total Environment, 586(2):1091–1100, 2017.
Noh, Hyeonwoo, You, Tackgeun, Mun, Jonghwan, and Han,
Bohyung. Regularizing deep neural networks by noise:
Its interpretation and optimization. In Advances in Neural
Information Processing Systems 30, 2017.
Schneider, David S., Torres, Brenda Y., Oliveira, Jose Henrique M., Tate, Ann Thomas, Rath, Poonam, and Cumnock, Katherine. Tracking resilience to infections by
mapping disease space. PLOS Biology, 14(6), 2016.
Fibres of Failure
Selvaraju, Ramprasaath R., Das, Abhishek, Vedantam,
Ramakrishna, Cogswell, Michael, Parikh, Devi, and
Batra, Dhruv. Grad-CAM: Why did you say that?
arXiv:1611.07450 [cs, stat], November 2016. URL
http://arxiv.org/abs/1611.07450. NIPS
2016 InterpretML.
Simonyan, Karen, Vedaldi, Andrea, and Zisserman, Andrew.
Deep inside convolutional networks: Visualising image
classification models and saliency maps. arXiv preprint
arXiv:1312.6034, 2013.
Singh, Gurjeet, Mémoli, Facundo, and Carlsson, Gunnar.
Topological Methods for the Analysis of High Dimensional Data Sets and 3d Object Recognition. In SPBG, pp.
91–100, 2007. URL http://comptop.stanford.
edu/preprints/mapperPBG.pdf.
Singh, Sameer, Ribeiro, Marco Tulio, and Guestrin, Carlos.
Programs as Black-Box Explanations. arXiv:1611.07579
[cs, stat], November 2016. URL http://arxiv.
org/abs/1611.07579. NIPS 2016 InterpretML.
Smilkov, Daniel, Thorat, Nikhil, Nicholson, Charles, Reif,
Emily, Viégas, Fernanda B., and Wattenberg, Martin. Embedding Projector: Interactive Visualization and Interpretation of Embeddings. arXiv:1611.05469 [cs, stat],
November 2016. URL http://arxiv.org/abs/
1611.05469. NIPS 2016 InterpretML.
Tansey, Wesley, Thomason, Jesse, and Scott, James G.
Interpretable Low-Dimensional Regression via DataAdaptive Smoothing. arXiv:1708.01947 [stat], August 2017. URL http://arxiv.org/abs/1708.
01947. WHI 2017 (ICML Workshop).
Thiagarajan, Jayaraman J., Kailkhura, Bhavya, Sattigeri,
Prasanna, and Ramamurthy, Karthikeyan Natesan. TreeView: Peeking into Deep Neural Networks Via FeatureSpace Partitioning. arXiv:1611.07429 [cs, stat], November 2016. URL http://arxiv.org/abs/1611.
07429. NIPS 2016 InterpretML.
Tosi, Alessandra, Vellido, Alfredo, and Alvarez, Mauricio
(eds.). Transparent and Interpretable Machine Learning
in Safety Critical Environments, NIPS 2017 Workshop,
2017.
Tramèr, Florian, Kurakin, Alexey, Papernot, Nicolas, Goodfellow, Ian, Boneh, Dan, and McDaniel, Patrick. Ensemble adversarial training: Attacks and defenses. International Conference on Learning Representations,
2018. URL https://openreview.net/forum?
id=rkZvSe-RZ. accepted as poster.
Varshney, Kush, Weller, Adrian, Kim, Been, and Malioutov,
Dmitry (eds.). Human Interpretability in Machine Learning, ICML 2017 Workshop, 2017.
Vidovic, Marina M.-C., Görnitz, Nico, Müller, KlausRobert, and Kloft, Marius. Feature Importance Measure
for Non-linear Learning Algorithms. arXiv:1611.07567
[cs, stat], November 2016. URL http://arxiv.
org/abs/1611.07567. NIPS 2016 InterpretML.
Whitmore, Leanne S., George, Anthe, and Hudson,
Corey M. Mapping chemical performance on molecular structures using locally interpretable explanations.
arXiv:1611.07443 [physics, stat], November 2016. URL
http://arxiv.org/abs/1611.07443. NIPS
2016 InterpretML.
Wilson, Andrew Gordon, Kim, Been, and Herlands, William
(eds.). Interpretable Machine Learning for Complex Systems, NIPS 2016 Workshop, 2016.
Wilson, Andrew Gordon, Yosinski, Jason, Simard, Patrice,
Caruana, Rich, and Herlands, William (eds.). Interpretable ML Symposium, NIPS 2017 Workshop, 2017.
Wisdom, Scott, Powers, Thomas, Pitton, James, and Atlas,
Les. Interpretable Recurrent Neural Networks Using
Sequential Sparse Recovery. arXiv:1611.07252 [cs, stat],
November 2016. URL http://arxiv.org/abs/
1611.07252. NIPS 2016 InterpretML.
Yuan, Xiaoyong, He, Pan, Zhu, Qile, Bhat, Rajendra Rana,
and Li, Xiaolin. Adversarial examples: Attacks and
defenses for deep learning. 1 2018.
Zhou, Yiren, Song, Sibo, and Cheung, Ngai-Man. On classification of distorted images with deep convolutional
neural networks. arXiV, 2017.
Zrihem, Nir Ben, Zahavy, Tom, and Mannor, Shie.
Visualizing Dynamics: from t-SNE to SEMI-MDPs.
arXiv:1606.07112 [cs, stat], June 2016. URL http://
arxiv.org/abs/1606.07112. WHI 2016 (ICML
Workshop).
| 1cs.CV
|
ON THE CLASSIFICATION OF TOGLIATTI SYSTEMS
arXiv:1710.03579v1 [math.AG] 7 Oct 2017
ROSA M. MIRÓ-ROIG AND MARTÍ SALAT
Abstract. In [5], Mezzetti and Miró-Roig proved that the minimal number of generators µ(I) of
a minimal (smooth) monomial Togliatti system I ⊂ k[x0 , . . . , xn ] satisfies 2n + 1 ≤ µ(I) ≤ n+d−1
n−1
and they classify all smooth minimal monomial Togliatti systems I ⊂ k[x0 , . . . , xn ] with 2n + 1 ≤
µ(I) ≤ 2n + 2.
In this paper, we address the first open case. We classify all smooth monomial Togliatti systems
I ⊂ k[x0 , . . . , xn ] of forms of degree d ≥ 4 with µ(I) = 2n + 3 and n ≥ 2 and all monomial Togliatti
systems I ⊂ k[x0 , x1 , x2 ] of forms of degree d ≥ 6 with µ(I) = 7.
Contents
1. Introduction
2. Preliminaries
3. The classification of Togliatti systems with 2n + 3 generators
References
1
3
5
21
October 11, 2017
1. Introduction
The study and classification of smooth projective varieties satisfying at least one Laplace
equation is a long standing problem in algebraic geometry. In [4], shedding new light on
this subject, it was related to another long standing problem in commutative algebra: the
study and classification of homogeneous artinian ideals failing the Weak Lefschetz Property
(WLP). We contribute to these two problems resolving the first question that was left open
in [5].
To be more precise. Let k be an algebraically closed field of characteristic 0, R =
k[x0 , . . . , xn ] and I = (F1 , . . . , Fs ) ⊂ R a homogeneous artinian ideal generated by forms
of the same degree d. Set A = R/I. We say that A fails the WLP from degree d − 1 to
degree d if the homomorphism ×ℓ : [A]d−1 → [A]d induced by a general linear form ℓ has not
Acknowledgments: The first author was partially supported by MTM2013-45075-P.
Key words and phrases. Osculating space, weak Lefschetz property, Laplace equations, monomial ideals, toric varieties.
2010 Mathematic Subject Classification. 13E10, 14M25, 14N05, 14N15, 53A20.
1
2
R. M. Miró-Roig, M. Salat
maximal rank. As shown in [4], if s ≤ n+d−1
then this assertion is equivalent to saying that
d
I
the projection Xn,d of the dth Veronese variety V (n, d) ⊂ Pn from hF1 , . . . , Fs i satisfies at
least one Laplace equation of order d −1. We call Togliatti systems the ideals satisfying these
two equivalent statements (see Definition 2.3). The name is in honour of E. Togliatti who
gave a complete classification of rational surfaces parameterized by cubics and satisfying at
least one Laplace equation of order 2 (see for instance [15],[16]). Narrowing the field of study
I
we deal only with monomial ideals I, so Xn,d
turns out to be a toric variety. In this sense,
one can apply pure combinatoric tools due to Perkinson in [12] to see whether I is a minimal monomial (smooth) Togliatti system (see Definition 2.3 and Propositions 3.4 and 3.6).
In [4], using these tools, Mezzetti, Miró-Roig and Ottaviani classified all smooth minimal
monomial Togliatti systems of cubics in four variables and conjectured a further classification for n ≥ 3. By means of graph theory, this conjecture was proved by Miró-Roig and
Michalek in [7] where a classification of smooth minimal Togliatti systems I ⊂ k[x0 , . . . , xn ]
of quadrics and cubics is achieved. When d ≥ 4 the picture becomes much more involved
and a complete classification seems out of reach for now. Therefore, in [5] it was introduced
another strategy: First to establish lower and upper bounds, depending on n and d ≥ 2,
for the minimal number of generators of a monomial Togliatti system and then to study the
monomial Togliatti systems with fixed number of generators.
In fact, in [5], Mezzetti and Miró-Roig bounded the number of generators of monomial
Togliatti systems and classified all minimal monomial (smooth) Togliatti systems reaching
the lower bound or close to reach it; namely those generated by 2n + 1 and 2n + 2 forms of
degree d ≥ 4. In this paper, we use again combinatoric tools and we classify the first open
case, i.e., all minimal monomial Togliatti systems generated by 2n + 3 forms of degree d ≥ 4
and n ≥ 2.
Next we outline the structure of this note. In Section 2 we fix the notation and we collect
the basic results needed in the sequel. Then, in Section 3 we expose the main results of this
note. Firstly, we give a complete classification of all minimal monomial Togliatti systems
generated by 7 forms of degree d ≥ 10 in three variables (see Theorem 3.8). In order
to achieve this classification we have had to consider all possible configurations of these 7
monomials regarded in the integer standard simplex d∆2 ⊂ Z3 and then apply combinatorial
criteria to each configuration. Separating the problem in two basic cases which we have
also had to separate into a few more subcases has helped so as to reduce the number of
configurations to study. Once seen this classification we compute all minimal monomial
Togliatti systems generated by 7 forms of degree 6 ≤ d ≤ 9 getting a complete scene of what
occurs in three variables. From this result we can look apart all minimal monomial smooth
On the classification of Togliatti systems
3
Togliatti systems in three variables generated by 7 forms of degree d ≥ 6 and close the open
question we were dealing with.
Acknowledgement. The first author of this paper warmly thanks Emilia Mezzetti for
interesting conversations and many ideas developed in this paper.
2. Preliminaries
We fix k an algebraically closed field of characteristic zero, R = k[x0 , . . . , xn ] and Pn =
Proj(k[x0 , . . . , xn ]). Given a homogeneous artinian ideal I ⊂ k[x0 , . . . , xn ], we denote by I −1
the ideal generated by the Macaulay inverse system of I (see [4, §3] for details). In this
section we fix the notations and the main results that we use throughout this paper. In
particular, we quickly recall the relationship between the existence of homogeneous artinian
ideals I ⊂ k[x0 , . . . , xn ] failing the weak Lefschetz property; and the existence of (smooth)
projective varieties X ⊂ PN satisfying at least one Laplace equation of order s ≥ 2. For
more details, see [4] and [7].
Definition 2.1. Let I ⊂ R be a homogeneous artinian ideal. We say that R/I has the weak
Lefschetz property (WLP, for short) if there is a linear form L ∈ (R/I)1 such that, for all
integers j, the multiplication map
×L : (R/I)j → (R/I)j+1
has maximal rank, i.e. it is injective or surjective. We often abuse notation and say that the
ideal I has the WLP. If for the general form L ∈ (R/I)1 and for an integer number j the
map ×L has not maximal rank we say that the ideal I fails the WLP in degree j.
Though many algebras are expected to have the WLP, establishing this property is often
rather difficult. For example, it was shown by R. Stanley [14] and J. Watanabe [18] that
a monomial artinian complete intersection ideal I ⊂ R has the WLP. By semicontinuity, it
follows that a general artinian complete intersection ideal I ⊂ R has the WLP but it is open
whether every artinian complete intersection of height ≥ 4 over a field of characteristic zero
has the WLP. It is worthwhile to point out that the WLP of an artinian ideal I strongly
depends on the characteristic of the ground field k and, in positive characteristic, there are
examples of artinian complete intersection ideals I ⊂ k[x0 , x1 , x2 ] failing the WLP (see, e.g.
[10, Remark 7.10]).
In [4], Mezzetti, Miró-Roig, and Ottaviani showed that the failure of the WLP can be used
to construct (smooth) varieties satisfying at least one Laplace equation of order s ≥ 2 (see
also [7], [5] and [6]). We have:
4
R. M. Miró-Roig, M. Salat
Theorem 2.2. Let I ⊂ R be an artinian ideal generated by r forms F1 , ..., Fr of degree d
with r ≤ n+d−1
. The following conditions are equivalent:
n−1
(1) the ideal I fails the WLP in degree d − 1;
(2) the homogeneous forms F1 , ..., Fr become k-linearly dependent on a general hyperplane
H of Pn ;
n+d
(3) the closure X := Im(ϕ(I −1 )d ) ⊂ P( d )−r−1 of the image of the rational map
ϕ(I −1 )d : Pn 99K P(
n+d
d
)−r−1
associated to (I −1 )d satisfies at least one Laplace equation of order d − 1.
Proof. See [4, Theorem 3.2].
The above result motivates the following definition:
Definition 2.3. Let I ⊂ R be an artinian ideal generated by r forms F1 , ..., Fr of degree d,
r ≤ n+d−1
. We say:
n−1
(i) I is a Togliatti system if it satisfies the three equivalent conditions in Theorem 2.2.
(ii) I is a monomial Togliatti system if, in addition, I (and hence I −1 ) can be generated
by monomials.
(iii) I is a smooth Togliatti system if, in addition, the n-dimensional variety X is smooth.
(iv) A monomial Togliatti system I is minimal if I is generated by monomials m1 , . . . , mr
and there is no proper subset mi1 , . . . , mir−1 defining a monomial Togliatti system.
The names are in honor of Eugenio Togliatti who proved that for n = 2 the only smooth
Togliatti system of cubics is I = (x30 , x31 , x32 , x0 x1 x2 ) ⊂ k[x0 , x1 , x2 ] (see [15], [16]). This
result has been reproved recently by Brenner and Kaid [1] in the context of weak Lefschetz
property. Indeed, Togliatti gave a classification of rational surfaces parameterized by cubics
and satisfying at least one Laplace equation of order 2: There is only one rational surface in
P5 parameterized by cubics and satisfying a Laplace equation of order 2; it is obtained from
the 3rd Veronese embedding V (2, 3) of P2 by a suitable projection from four points on it.
In [4], the first author together with Mezzetti and Ottaviani classified all smooth rational
3-folds parameterized by cubics and satisfying a Laplace equation of order 2, and gave a
conjecture to extend this result to varieties of higher dimension. This conjecture has been
recently proved in [7]. Indeed, the first author together with Michalek classified all smooth
minimal Togliatti systems of quadrics and cubics. For d ≥ 4, the picture becomes soon
much more involved than in the case of quadrics and cubics, and for the moment a complete
classification appears out of reach unless we introduce other invariants as, for example, the
number of generators of I.
On the classification of Togliatti systems
5
3. The classification of Togliatti systems with 2n + 3 generators
From now on, we restrict our attention to monomial artinian ideals I ⊂ k[x0 , . . . , xn ],
n ≥ 2, generated by forms of degree d ≥ 4. It is worthwhile to recall that for monomial
artinian ideals to test the WLP there is no need to consider a general linear form. In fact,
we have
Proposition 3.1. Let I ⊂ R := k[x0 , . . . , xn ] be an artinian monomial ideal. Then R/I has
the WLP if and only if x0 + x1 + · · · + xn is a Lefschetz element for R/I.
Proof. See [10, Proposition 2.2].
Let I ⊂ k[x0 , . . . , xn ] be a minimal monomial Togliatti systems of forms of degree d and
denote by µ(I) the minimal number of generators of I. In [8], the first author and Mezzetti
proved:
n+d−1
.
2n + 1 ≤ µ(I) ≤
n−1
In addition, the Togliatti systems with number of generators reaching the lower bound or
close to the lower bound were classified. Indeed, we have
Theorem 3.2. Let I ⊂ k[x0 , . . . , xn ] be a minimal monomial Togliatti system of forms of
degree d ≥ 4. Assume that µ(I) = 2n + 1. Then, up to a permutation of the coordinates,
one of the following cases holds:
(i) n ≥ 2 and I = (xd1 , . . . , xdn ) + xd−1
0 (x0 , . . . , xn ), or
5
5
(ii) (n, d) = (2, 5) and I = (x0 , x1 , x52 , x30 x1 x2 , x0 x21 x22 ), or
(iii) (n, d) = (2, 4) and I = (x40 , x41 , x42 , x0 x1 x22 , x20 x21 ).
Furthermore, (i) and (ii) are smooth while (iii) is not smooth.
Proof. See [5, Theorem 3.7].
Theorem 3.3. Let I ⊂ k[x0 , . . . , xn ] be a smooth minimal monomial Togliatti system of
forms of degree d ≥ 4. Assume that µ(I) = 2n + 2. Then, up to a permutation of the
coordinates, one of the following cases holds:
(i) n ≥ 2 and I = (xd0 , . . . , xdn ) + m(x0 , . . . , xn ) with m = xi00 xi11 · · · xinn where i0 ≥ i1 ≥
· · · ≥ in ≥ 0, i2 > 0 and i0 + i1 + · · · + in = d − 1.
(ii) (n, d) = (2, 5) and I = (x50 , x51 , x52 , x30 x1 x2 , x20 x21 x2 , x0 x31 x2 ) or (x50 , x51 , x52 , x30 x1 x2 ,
x0 x31 x2 , x0 x1 x32 ) or (x50 , x51 , x52 , x20 x21 x2 , x20 x1 x22 , x0 x21 x22 ).
(iii) (n, d) = (2, 7) and d = 7 and I = (x70 , x71 , x72 , x30 x31 x2 , x30 x1 x32 , x0 x31 x32 ) or (x70 , x71 , x72 ,
x50 x1 x2 , x0 x51 x2 , x0 x1 x52 ), (x70 , x71 , x72 , x0 x1 x52 , x30 x31 x2 , x20 x21 x32 ) or (x70 , x71 , x72 , x40 x1 x22 ,
x20 x41 x2 , x0 x21 x42 ).
6
R. M. Miró-Roig, M. Salat
Proof. See [5, Theorem 3.17 and Proposition 3.19].
In this paper, we address the first open case and we classify all smooth minimal monomial
Togliatti systems I ⊂ k[x0 , . . . , xn ] of forms of degree d ≥ 4 with µ(I) = 2n + 3 (see Theorem
3.9) as well as all minimal monomial Togliatti systems I ⊂ k[x0 , x1 , x2 ] of forms of degree
d ≥ 6 with µ(I) = 7 (see Theorem 3.8).
In order to achieve this classification, we associate to any artinian monomial ideal a polytope and we tackle our problem with tools coming from combinatorics. In fact, the failure
of the WLP of an artinian monomial ideal I ⊂ k[x0 , . . . , xn ] can be established by purely
combinatoric properties of the associated polytope PI . To state this result we need to fix
some extra notation.
Given an artinian monomial ideal I ⊂ k[x0 , . . . , xn ] generated by monomials of degree
d and its inverse system I −1 , we denote by ∆n the standard n-dimensional simplex in the
lattice Zn+1 , we consider d∆n and we define the polytope PI as the convex hull of the finite
subset AI ⊂ Zn+1 corresponding to monomials of degree d in I −1 . As usual we define:
X
X
nx = 1}
nx · x | nx ∈ Z,
AffZ (AI ) := {
x∈AI
x∈AI
the sublattice AffZ (AI ) in Zn+1 generated by AI . We have the following criterion which will
play an important role in the proof of our main result.
Proposition 3.4. Let I ⊂ k[x0 , . . . , xn ] be an artinian monomial ideal generated by mono
mials of degree d. Assume r ≤ n+d−1
. Then, I is a Togliatti system if and only if there
n−1
exists a hypersurface of degree d − 1 containing AI ⊂ Zn+1 . In addition, I is a minimal
Togliatti system if and only if any such hypersurface F does not contain any integral point
of d∆n \ AI except possibly some of the vertices of d∆n .
Proof. This follows from Theorem 2.2 and [12, Proposition 1.1].
Example 3.5. The artinian monomial ideal
I = (x0 , x1 )3 + (x2 , x3 )3 + (x34 , x0 x2 x4 , x0 x3 x4 , x1 x2 x4 , x1 x3 x4 ) ⊂ k[x0 , x1 , x2 , x3 , x4 ]
defines a minimal monomial Togliatti system of cubics. In fact, the set AI ⊂ Z5 is:
AI = {(2, 0, 1, 0, 0), (1, 0, 2, 0, 0), (2, 0, 0, 1, 0), (1, 0, 0, 2, 0), (2, 0, 0, 0, 1), (1, 0, 0, 0, 2), (0, 2, 1, 0, 0),
(0, 1, 2, 0, 0), (0, 2, 0, 1, 0), (0, 1, 0, 2, 0), (0, 2, 0, 0, 1), (0, 1, 0, 0, 2), (0, 0, 2, 0, 1), (0, 0, 1, 0, 2), (0, 0, 0, 2, 1),
(0, 0, 0, 1, 2), (1, 1, 1, 0, 0), (1, 1, 0, 1, 0), (1, 1, 0, 0, 1), (1, 0, 1, 1, 0), (0, 1, 1, 1, 0), (0, 0, 1, 1, 1)}.
On the classification of Togliatti systems
7
There is a quadric, and only one, containing all points of AI and no integral point of 4∆4 \AI ,
namely,
4
X
X
Q(x0 , x1 , x2 , x3 , x4 ) = 2
x2i + 9(x0 x1 + x2 x3 ) − 5
xi xj .
i=0
0≤i<j≤4
The following criterion allows us to check if a subset A of points in the lattice Zn+1 defines
a smooth toric variety XA or not.
Proposition 3.6. Let I ⊂ k[x0 , . . . , xn ] be an artinian monomial ideal generated by r monomials of degree d. Let AI ⊂ Zn+1 be the set of integral points corresponding to monomials
in (I −1 )d , SI the semigroup generated by AI and 0, PI the convex hull of AI and XAI the
projective toric variety associated to the polytope PI . Then XAI is smooth if and only if for
any non-empty face Γ of PI the following conditions hold:
(i) The semigroup SI /Γ is isomorphic to Zm
+ with m = dim(PI ) − dim Γ + 1.
(ii) The lattices Zn+1 ∩ AffR (Γ) and AffZ (AI ∩ Γ) coincide.
Proof. See [2, Chapter 5, Corollary 3.2]. Note that in this case XAI = X where X is the
n+d
closure of the image of the rational map ϕ[I −1]d : Pn −→ P( d )−r−1 .
As a direct application of this criterion we get:
Example 3.7. Let
I = (x0 , x1 )3 + (x2 , x3 )3 + (x34 , x0 x2 x4 , x0 x3 x4 , x1 x2 x4 , x1 x3 x4 ) ⊂ k[x0 , x1 , x2 , x3 , x4 ]
be the minimal monomial Togliatti system of cubics described in Example 3.5. Applying the
above smoothness criterion we get that the toric variety XAI is smooth.
For any integer d ≥ 3, we define M(d) := {xa0 xb1 xc2 | a + b + c = d and a, b, c ≤ d − 1} and
we consider the sets of ideals:
A = {(x20 x2 , x0 x21 , x31 , x1 x22 ), (x20 x2 , x0 x1 x2 , x31 , x21 x2 ), (x20 x2 , x0 x1 x2 , x31 , x1 x22 ), (x20 x1 , x0 x22 , x31 , x21 x2 ),
(x20 x1 , x0 x22 , x31 , x1 x22 ), (x20 x2 , x0 x22 , x31 , x1 x22 ), (x0 x21 , x0 x22 , x31 , x1 x22 ), (x20 x2 , x21 x2 , x31 , x32 ), (x0 x22 , x21 x2 , x31 , x32 ),
(x20 x2 , x0 x21 , x21 x2 , x1 x22 ),(x0 x21 , x0 x22 , x21 x2 , x1 x22 ),(x20 x1 , x0 x21 , x31 , x32 ),(x20 x1 , x21 x2 , x31 , x32 ),(x21 x2 , x1 x22 , x31 , x32 ),
(x0 x22 , x21 x2 , x1 x22 , x31 ),(x0 x1 x2 , x0 x22 , x31 , x1 x22 ),(x20 x2 , x0 x22 , x31 , x21 x2 ),(x0 x21 , x0 x22 , x31 , x32 ), (x20 x2 , x0 x21 , x31 , x32 ),
(x0 x1 x2 , x0 x22 , x31 , x21 x2 ), (x20 x2 , x21 x2 , x1 x22 , x31 )},
B = {(x20 x1 x2 , x0 x32 , x41 , x31 x2 ), (x30 x2 , x0 x21 , x2 , x41 , x1 x32 ), (x20 x22 , x0 x21 x2 , x41 , x42 ), (x20 x1 x2 , x21 x22 , x41 , x42 )}
and
C = {(x20 x1 x22 , x0 x31 x2 , x51 , x52 ), (x30 x1 x2 , x0 x21 x22 , x51 , x52 )}.
8
R. M. Miró-Roig, M. Salat
Theorem 3.8. Let I ⊂ k[x0 , x1 , x2 ] be a minimal monomial Togliatti system of forms of
degree d ≥ 10. Assume that µ(I) = 7. Then, up to a permutation of the coordinates, one of
the following cases holds:
(1)
(2)
(3)
(4)
(5)
(6)
I
I
I
I
I
I
= (xd0 , xd1 , xd2 ) + m(x20 , x21 , x0 x2 , x1 x2 ) with m ∈ M(d − 2), or
= (xd0 , xd1 , xd2 ) + m(x20 , x21 , x0 x1 , x22 ) with m ∈ M(d − 2), or
= (xd0 , xd1 , xd2 ) + m(x30 , x31 , x32 , x0 x1 x2 ) with m ∈ M(d − 3), or
= (xd0 , xd1 , xd2 ) + xd−3
0 J with J ∈ A, or
d
d
d
= (x0 , x1 , x2 ) + xd−4
0 J with J ∈ B, or
d
d
d
= (x0 , x1 , x2 ) + xd−5
0 J with J ∈ C.
Proof. It is easy to check that all of these ideals are minimal Togliatti systems. Vice versa,
let us write I = (xd0 , xd1 , xd2 , m1 , m2 , m3 , m4 ) where for 1 ≤ i ≤ 4, mi = xa0i xb1i xc2i with
ai + bi + ci = d. We consider AI ⊂ d∆2 ∩ Z3 and we slice AI with planes in three possible
manners:
(i,j)
For 0 ≤ j ≤ 2 and 0 ≤ i ≤ d, we define Hij := {(t0 , t1 , t2 )|tj = i} and AI := AI ∩ Hij .
We divide the proof in two cases:
Case 1: There exist 1 ≤ ia , ib , ic ≤ 4 such that aia , bib , cic ≤ 1.
Case 2: There exists one variable whose square divides all monomials mi .
Case 1: None of the squares of the variables divide the four monomials m1 , m2 , m3 and m4 .
Up to permutation of the variables, we have two possibilities:
1
3 e3
) with 0 ≤
Case 1A: I = (xd0 , xd1 , xd2 , xe01 xa1 xd−a−e
, xb0 xe12 x2d−b−e2 , xc0 xd−c−e
x2 , xα0 xβ1 xd−α−β
2
2
1
e1 , e2 , e3 ≤ 1, d − 2 − e1 ≥ a ≥ 2, d − 2 − e2 ≥ b ≥ 2, d − 2 − e3 ≥ c ≥ 2 and only one of the
exponents α, β, d − α − β is ≤ 1.
1 −e2
3 e3
Case 1B: I = (xd0 , xd1 , xd2 , xe01 xe12 xd−e
, xa0 xd−a−e
x2 , xα0 xβ1 xd−α−β
, xγ0 xδ1 xd−γ−δ
) with 0 ≤
2
2
2
1
e1 , e2 , e3 ≤ 1.
In both cases, a straightforward computation using the hypothesis d ≥ 10 shows that when
we restrict to x0 + x1 + x2 the 7 monomials remain k−linearly independents. Therefore, I is
not a Togliatti system.
Case 2: Without loss of generality we can suppose that x20 divides each monomial mi . We
can also assume that a1 ≥ a2 ≥ a3 ≥ a4 = s ≥ 2.
Let Fd−1 be a plane curve of degree d − 1 containing all integral points of AI . Since s ≥ 2,
it factorizes as Fd−1 = L00 L01 · · · L0s−1 Fd−s−1 . Indeed, Fd−1 has degree d − 1 and contains the d
(1,0)
(0,0)
points in AI . So, Fd−1 = L01 Fd−2 . Since Fd−2 contains all d − 1 points of AI it factorizes
as Fd−1 = L00 L01 Fd−3 . Repeating the process we arrive to Fd−1 = L00 L01 · · · L0s−1 Fd−s−1 . We
On the classification of Togliatti systems
9
(s,0)
claim that a3 = a4 = s ≥ 2. If a3 > a4 = s then AI
has d − s points and Fd−s−1 contains
0
them. Hence, Fd−s−1 = Ls Fd−s−2 contradicting the minimality of I (Proposition 3.4).
So far we have a3 = a4 = s ≥ 2 and Fd−1 = L00 · · · L0s−1 Fd−s−1 where Fd−s−1 is aplane
(k,0)
s−1
curve of degree d − s − 1 which contains all integer points of ÃI := AI \ ∪k=0
AI
(i,j)
ÃI
. Set
= ÃI ∩ Hij . We distinguish four subcases:
Case
Case
Case
Case
2A: a1 = a2 = a3 = a4 =: s ≥ 2.
2B: u := a1 > a2 = a3 = a4 =: s ≥ 2.
2C: u := a1 = a2 > a3 = a4 =: s ≥ 2.
2D: u := a1 > v := a2 > a3 = a4 =: s ≥ 2.
Case 2A: We assume a1 = a2 = a3 = a4 =: s ≥ 2. In this case, Fd−1 = L00 · · · L0s−1 L0s+1 · · ·
3
2
2
3
L0d−1 . Therefore, s = d − 3 and I = (xd0 , xd1 , xd2 ) + xd−3
0 (x1 , x1 x2 , x1 x2 , x2 ), which is of type
(4).
Case 2B: We assume u := a1 > a2 = a3 = a4 =: s ≥ 2. In this case, u ≤ s + 2. Indeed, if
u > s + 2 we have, Fd−s−1 = L0s+1 · · · L0u−1 Fd−u with Fd−u a plane curve of degree d − u which
(s,0)
(s,0)
contains in particular AI . By minimality, #(Fd−u ∩AI ) = d −s −2 > d −u (Proposition
3.4) and we have Fd−u = L0s Fd−u−1 , which is a contradiction. Then, up to permutation of
variables, I is as one of the following cases:
, xe1 x2d−e−s ).
, xc1 xd−c−s
, xb1 xd−b−s
Case b1: u = s + 1 and I = (xd0 , xd1 , xd2 ) + xs0 (x0 xa1 xd−a−s−1
2
2
2
Case b2: u = s + 2 and I = (xd0 , xd1 , xd2 ) + xs0 (x20 xa1 xd−a−s−2
, xb1 xd−b−s
, xc1 xd−c−s
, xe1 x2d−e−s ).
2
2
2
0
Case b1: In this case we are removing three points from Hs0 and one from Hs+1
. Up
d−s−1
to permutation of the variables y and z, we can assume d − s − 1 ≥ a ≥ ⌊ 2 ⌋ and
⌋ > e ≥ 0. In this case
d − s ≥ b > c > e ≥ 0. Let us first suppose that ⌊ d−s−1
2
(
d−s
e≥1
(0,1)
#(Fd−s−1 ∩ ÃI ) =
d − s − 1 e = 0.
(0,1)
If e ≥ 1, then #ÃI
= d − s, Fd−s−1 = L10 Fd−s−2 = L10 · · · L1e−1 Fd−s−e−1 and Fd−s−e−1
(e,1)
(e,1)
contains the integer points of ÃI . Since a > e and b > c > e, we have #ÃI = d − s − e
and Fd−s−e−1 = L1e Fd−s−e−2 contradicting the minimality of I. Therefore it must be e = 0,
and m4 = xs0 xd−s
with d − s − 1 ≥ c ≥ 1. Let us consider
2
d−s
a, c ≥ 2
d − s − 1 a = 1, c ≥ 2
(1,1)
#(Fd−s−1 ∩ ÃI ) =
d − s − 1 a ≥ 2, c = 1
d−s−2 a= c =1
10
R. M. Miró-Roig, M. Salat
and we study the four possibilities.
If a, c ≥ 2 then we have the factorization Fd−s−1 = L11 Fd−s−2 . In particular, Fd−s−2 is a
(0,1)
plane curve of degree d−s−2 containing the d−s−1 points of ÃI . So, Fd−s−2 = L10 Fd−s−3
which contradicts the minimality of I. Therefore, if a ≥ 2, then c = 1.
If a = 1, we have d − 2 ≥ s ≥ d − 3. If s = d − 2, then c = 1 and I = (xd0 , xd1 , xd2 , xd−1
0 x1 ,
d−2
d−2 2
d−2 2
x0 x1 , x0 x1 x2 , x0 x2 ) which is not a Togliatti system. Otherwise, s = d − 3, then we
have several possibilities:
d−3 2
3
3
(i) c ≥ 2 and I = (xd0 , xd1 , xd2 ) + xd−3
0 (x1 , x2 , x0 x1 x2 ) + (x0 x1 x2 ) which is not minimal.
d−3
3
3
2
d
d
d
(ii) c = 1 and I = (xd0 , xd1 , xd2 ) + xd−3
0 (x0 x1 x2 , x1 , x2 ) + (x0 x1 x2 ) or I = (x0 , x1 , x2 ) +
d−3 3
xd−3
0 x1 x2 (x0 , x1 , x2 ) + (x0 x2 ). Both of them are not minimal.
If d − s − 1 ≥ a ≥ 2 and s ≤ d − 3. We have e = 0, c = 1 and (m1 , m2 , m3 , m4 ) =
s+1 a d−a−s−1
(x0 x1 x2
, xs0 xb1 xd−b−s
, xs0 x1 x2d−s−1 , xs0 xd−s
2
2 ) with d − s ≥ b ≥ 2. Let us consider
d−s
d−s−1
(0,2)
#(Fd−s−1 ∩ ÃI ) =
d−s−1
d−s−2
d − s − 1 ≥ b, d − s − 2 ≥ a
a = d − s − 1, d − s − 1 ≥ b
b = d − s, d − s − 2 ≥ a
b = d − s, a = d − s − 1.
In the first case, we have the factorization Fd−s−1 = L20 Fd−s−2 = L20 L01 Fd−s−3 and it
contradicts the minimality of I.
If a = d − s − 1 and d − s − 1 ≥ b, then b = d − s − 1. Otherwise, we would have Fd−s−1 =
2
L1 Fd−s−2 and it would contradict the minimality of I. Therefore we have I = (xd0 , xd1 , xd2 ) +
xs0 (x0 x1d−s−1 , x1d−s−1 x2 , x1 x2d−s−1 , xd−s
2 ) with s ≤ d − 3. For s = d − 3 it corresponds to a
Togliatti system of type (4), while for s ≤ d − 4 is not Togliatti because when we restrict to
x0 + x1 + x2 = 0 the generators, they remain k−linearly independent.
If d − s − 2 ≥ a and b = d − s, then a = d − s − 2. Hence we have s ≤ d − 4 and
d−s−1
I = (xd0 , xd1 , xd2 ) + xs0 (x0 x1d−s−2 x2 , xd−s
, xd−s
1 , x1 x2
2 ) which is never a Togliatti system.
Finally, if b = d − s and a = d − s − 1, then s ≤ d − 3 and I = (xd0 , xd1 , xd2 ) +
d−s−1
xs0 (x0 x1d−s−1 , xd−s
, xd−s
1 , x1 x2
2 ) which is a Togliatti system of type (4) for s = d − 3 while
for s ≤ d − 4 it is not Togliatti.
⌋. In
To finish with the case b1, we have to see what happens when d − s − 2 ≥ e ≥ ⌊ d−s−1
2
this case s ≤ d − 3. Let us see that a = e. Otherwise, we can suppose a > e (the other case
is symmetric) and we have the factorization Fd−s−1 = L10 · · · L1e−1 Fd−s−e−1 . Since a > e and
(e,1)
b > c > e, ÃI
has d − s − e points and we have the factorization Fd−s−e−1 = L1e Fd−s−e−2
which contradicts the minimality of I. Hence a = e and in particular d − s − 1 > a and
d − s ≥ b > c > a.
On the classification of Togliatti systems
11
˜(0,2)
Let us consider ØI := ÃI \ ∪a−1
k=0 in the same spirit as AI and ÃI . If b = d − s, then ÃI
(0,2)
consists in d − s − e different points. Otherwise, d − s − 1 ≥ b and #Ø
= d − s + 1 − e.
I
(0,2)
In both cases ØI
have more points than the degree of the curve Fd−s−e−1 which passes
through them. Therefore, Fd−s−e−1 = L20 Fd−s−e−2 and d − s − 1 ≥ b. Since m2 cannot be
aligned vertically with any other monomial mi , we can repeat this argument until we get
that b = c + 1 and Fd−s−e−1 = L20 · · · L2b−1 Fd−s−e−b−1 . Now Fd−s−e−b−1 is a plane curve of
(b,2)
degree d − s − e − b − 1 containing all d − s − e − b points of Ø . Hence, we can factorize
I
Fd−s−e−b−1 = L2b Fd−s−e−b−2 contradicting the minimality assumption.
0
Case b2: We are removing from d∆2 to get ÃI : three points of Hs0 and one from Hs+2
. Up
d−s−2
to permutation of the variables y and z, we can suppose that d − s − 2 ≥ a ≥ ⌊ 2 ⌋ and
d − s ≥ b > c > e ≥ 0.
Let us suppose first that ⌊ d−s−2
⌋ > e ≥ 0. We argue as in the case u = s + 1 to prove that
2
(1,1)
e = 0. Let us consider #(Fd−s−1 ∩ ÃI ). Using the same argumentation we prove that if
a, c ≥ 2 we get a contradiction. If a = 1, then either s = d − 3 or s = d − 4 and we have the
following cases:
d−3 2
2
3
2
3
3
2
3
(i) If s = d − 3, then (m1 , m2 , m3 , m4 ) is xd−3
0 (x0 x1 , x1 , x1 x2 , x2 ), x0 (x0 x1 , x1 , x1 x2 , x2 ) or
2
2
2
3
xd−3
0 (x0 x1 , x1 x2 , x1 x2 , x2 ). All of them are Togliatti systems of type (4).
d−4 2
2
4
3
4
4
2 2
4
(ii) If s = d−4, then (m1 , m2 , m3 , m4 ) is xd−4
0 (x0 x1 x2 , x1 , x1 x2 , x2 ), x0 (x0 x1 x2 , x1 , x1 x2 , x2 ),
d−4 2
d−4 2
d−4
2
4
3
4
3
2 2
4
3
3
4
xd−4
0 (x0 x1 x2 , x1 , x1 x2 , x2 ), x0 (x0 x1 x2 , x1 x2 , x1 x2 , x2 ), x0 (x0 x1 x2 , x1 x2 , x1 x2 , x2 ) or x0 (
x20 x1 x2 , x21 x22 , x1 x32 , x42 ). The only one which is a minimal Togliatti system is the second one
and it is of type (5).
Now, we assume e = 0, c = 1 and d − s − 2 ≥ a ≥ 2. In particular, s ≤ d − 4. We consider
(0,2)
#(Fd−s−1 ∩ ÃI ), and see that if d − s − 1 ≥ b ≥ 2 and d − s − 3 ≥ a ≥ 2, there is a
contradiction with the minimality of I. If b = d − s and a ≤ d − s − 3 (resp. a = d − s − 2
and b ≤ d − s − 1) then a = d − s − 3 (resp. b = d − s − 1). Otherwise we would incur again
to a contradiction with the minimality of I. So, we have three possibilities.
(i) a = d − s − 3 ≥ 2, b = d − s, s ≤ d − 5 and I = (xd0 , xd1 , xd2 ) + xs0 (x20 x1d−s−3 x2 , xd−s
1 ,
d−s−1
d−s
x1 x2
, x2 ).
(ii) a = d − s − 2, b = d − s − 1, s ≤ d − 4 and I = (xd0 , xd1 , xd2 ) + xs0 (x20 x1d−s−2 , x1d−s−1 x2 ,
x1 x2d−s−1 , xd−s
2 ).
d−s−1
(iii) a = d − s − 2, b = d − s, s ≤ d − 4 and I = (xd0 , xd1 , xd2 ) + xs0 (x20 x1d−s−2 , xd−s
, xd−s
1 , x1 x2
2 ).
After restricting to x0 + x1 + x2 = 0, we see that none of them corresponds to a Togliatti
system.
⌋. With
To finish with the case b2, we see what happens when d − s − 2 ≥ e ≥ ⌊ d−s−2
2
the same argument that we use before, we can see that a = e. The difference with the case
12
R. M. Miró-Roig, M. Salat
u = s+1 is that in this case we can have m1 and m2 aligned vertically. This condition can be
translated as the case when d − b − s = d − a − s − 2. If this does not happen (i.e. b > a + 2),
then it will contradict the minimality of I. Indeed: let us suppose that 0 ≤ k := d − b − s <
d − a − s − 2. Inductively we have the factorization Fd−s−e−1 = L20 · · · L2k−1 Fd−s−e−k−1.
Fd−s−e−k−1 is a plane curve of degree d − s − e − k − 1 which passes through all d − s − e − k
points of ØkI . Hence, we have the factorization Fd−s−e−k−1 = L2k Fd−s−e−k−2 , contradicting
the minimality assumption.
Therefore it must be b = a + 2 and, since b > c > a we have c = a + 1. Finally we get:
I = (xd0 , xd1 , xd2 ) + xs0 xa1 xd−a−s−2
(x20 , x21 , x1 x2 , x22 ) which is of type (1).
2
Case 2C: We assume that u := a1 = a2 > a3 = a4 =: s ≥ 2. Arguing as in case 2B we get
u = s + 1 and I = (xd0 , xd1 , xd2 ) + xs0 (x0 xa1 xd−a−s−1
, x0 xb1 x2d−b−s−1 , xc1 xd−c−s
, xe1 x2d−e−s ). We can
2
2
⌋, a > b and d − s ≥ c > e ≥ 0. Let us suppose first that
assume d − s − 1 ≥ a ≥ ⌊ d−s−1
2
d−s−1
⌊ 2 ⌋ > e ≥ 0. We consider
e ≥ 1, b ≥ 1 c1
d−s
d − s − 1 e = 0, b ≥ 1 c2
(0,1)
#(Fd−s−1 ∩ ÃI ) =
d − s − 1 e ≥ 1, b = 0 c3
d−s−2 e=b =0
c4
Case c1: Since Fd−s−1 is a plane curve of degree d − s − 1 which contains all d − s points
(0,1)
(1,1)
of ÃI
we have the factorization Fd−s−1 = L10 Fd−s−2 . Now, intersecting Fd−s−2 with ÃI
and using the minimality assumption, we see that the only two possibilities are either e ≥ 2
and b ≥ 2 or e = b = 1. In the first case Fd−s−2 factorizes as Fd−s−2 = L11 Fd−s−3 . Repeating
the same argument we get that it must be e = b in any case. Now, we consider ØI as before
and we take
d − s − e + 1 d − s − 2 ≥ a, d − s − 1 ≥ c
d−s−e
a = d − s − 1, d − s − 1 ≥ c
(0,2)
#(Fd−s−e−1 ∩ ØI ) =
d−s−e
d − s − 2 ≥ a, c = d − s
d − s − e − 1 a = d − s − 1, c = d − s
In the second and third cases we obtain directly a contradiction with the minimality of I. In
(1,2)
the fourth case, ØI consists in d−s−e different points and we have Fd−s−e−1 = L21 Fd−s−e−2 .
(0,2)
Since Ø
has d − s − e − 1 different points, we get a contradiction with the minimality
I
of I. Finally, in the first case we obtain a factorization Fd−s−e−1 = L20 Fd−s−e−2 and we
repeat the same argument until we get that m1 and m3 are always vertically aligned. Then,
c = a + 1 and we have the factorization Fd−s−e−1 = L20 · · · L2d−s−a−2 Fa−e . If a ≥ e + 2 we have
the factorization Fa−e = L1e+1 · · · L1a−1 F1 , which contradicts the minimality of I. Therefore,
On the classification of Togliatti systems
13
d−s−e−2
d−s−e−2
a = e + 1 and I = (xd0 , xd1 , xd2 ) + xs0 (x0 xe+1
, x0 xe1 xd−s−e−1
, xe+2
, xe1 x2d−s−e ) =
1 x2
2
1 x2
(xd0 , xd1 , xd2 ) + xs0 xe1 xd−s−e−2
(x0 x1 , x0 x2 , x21 , x22 ) which is of type (1).
2
Case c2: We assume e = 0 and b ≥ 1. Let us consider
d−s
d−s−1
(1,1)
#(Fd−s−1 ∩ ÃI ) =
d−s−1
d−s−2
b ≥ 2, c ≥ 2
b = 1, c ≥ 2
b ≥ 2, c = 1
b=c=1
(i)
(ii)
(iii)
(iv)
In Case (i) we factorize Fd−s−1 = L11 Fd−s−2 . Since Fd−s−2 is a plane curve of degree d−s−2
(0,1)
containing all d − s − 1 different points of ÃI
it factorizes as Fd−s−2 = L10 Fd−s−3 . This
contradicts the minimality of I.
(0,2)
Case (ii): assume e = 0, b = 1 and c ≥ 2. We consider #(Fd−s−1 ∩ ÃI ) and arguing as
in the previous cases we get three possibilities:
a = d − s − 1, c = d − s − 1 and I = (xd0 , xd1 , xd2 ) + xs0 (x0 x1d−s−1 , x0 x1 x2d−s−2 , x1d−s−1 x2 , xd−s
2 ).
d−s−2
d−s−2
d−s
d−s
d
d
d
s
a = d − s − 2, c = d − s and I = (x0 , x1 , x2 ) + x0 (x0 x1
x2 , x0 x1 x2
, x1 , x2 ).
d−s−1
d−s−2
d−s
a = d − s − 1, c = d − s and I = (xd0 , xd1 , xd2 ) + xs0 (x0 x1
, x0 x1 x2
, xd−s
1 , x2 ).
Restricting the generators to the hyperplane x0 + x1 + x2 , we see that each of them is a
Togliatti system if, and only if s = d − 3.
Case (iii): assume e = 0, c = 1 and b ≥ 2. In particular a ≥ 3 and s ≤ d − 4. Arguing as
before, we see that the only viable possibility is a = d − s − 1 and b = d − s − 2. Therefore
I = (xd0 , xd1 , xd2 ) + xs0 (x0 x1d−s−1 , x0 x1d−s−3 x22 , x1 x2d−s−1 , xd−s
2 ), which is never a Togliatti system.
Case (iv): assume e = 0 and b = c = 1. Now it only remains to determinate a. If
(d−s−a−1,2)
consists in a + 1
d − s − 2 ≥ a, we have Fd−s−1 = L20 · · · L2d−s−a−2 Fa . Since ÃI
different points, we get a contradiction with the minimality of I. Therefore, a = d − s − 1.
Using the same argumentation we see that a = b + 1 = 2. Thus s = d − 4 and I =
3
2
3
4
(xd0 , xd1 , xd2 ) + xd−4
0 (x0 x1 , x0 x1 x2 , x1 x2 , x2 ) which is not a Togliatti system.
Case c3: Now, assume that b = 0 and e ≥ 1. Since a > e by hypothesis, considering
(1,1)
(0,2)
#(Fd−s−1 ∩ ÃI ) we see that e = 1 and s ≤ d − 3. If we consider #(Fd−s−1 ∩ ÃI ) we get
that the only viable possibilities are:
a = d−s−1, c = d−s−1 and I = (xd0 , xd1 , xd2 )+xs0 (x0 x1d−s−1 , x0 x2d−s−1 , x1d−s−1 x2 , x1 x2d−s−1 ).
d−s−1
a = d − s − 2, c = d − s and I = (xd0 , xd1 , xd2 ) + xs0 (x0 x1d−s−2 x2 , x0 x2d−s−1 , xd−s
).
1 , x1 x2
d−s−1
d−s
d−s−1
d−s−1
d
d
d
s
).
, x1 , x1 x2
, x0 x2
a = d − s − 1, c = d − s and I = (x0 , x1 , x2 ) + x0 (x0 x1
Each of them are Togliatti systems if, and only if s = d − 3.
Case c4: in this case we assume that e = b = 0. If a, c ≥ 3, we have the factorization
(0,1)
Fd−s−1 = L11 L12 Fd−s−3 . Since Fd−s−3 is a plane curve containing all d − s − 2 points of ÃI ,
14
R. M. Miró-Roig, M. Salat
we get Fd−s−3 = L00 Fd−s−4 which contradicts the minimality of I. Therefore we can consider
three subcases:
Case (i). Assume that a = 1, then it has to be either s = d − 2 or s = d − 3. Hence
2
2
d
d
d
I is one of the following possibilities: (xd0 , xd1 , xd2 ) + xd−2
0 (x0 x1 , x0 x2 , x1 , x2 ), (x0 , x1 , x2 ) +
d−3
d−3
d−1
d
d
d
3
3
2
d
d
d
xd−2
0 x2 (x0 , x1 , x2 ) + (x0 x1 ), (x0 , x1 , x2 ) + x0 (x0 x1 x2 , x0 x2 , x1 , x2 ), (x0 , x1 , x2 ) + x0 x2 (
d−2
2
x0 x1 , x0 x2 , x21 , x22 ) and (xd0 , xd1 , xd2 ) + xd−3
0 x2 (x0 , x1 , x2 ) + (x0 x1 x2 ). And only the first and
the third possibilities give to minimal Togliatti systems of type (1) and type (4) respectively.
Case (ii). Assume that a = 2, then it can be either s = d − 3, s = d − 4 or s = d − 5.
2
2
3
3
d
d
d
Therefore, I is one of the next ideals: (xd0 , xd1 , xd2 ) + xd−3
0 (x0 x1 , x0 x2 , x1 , x2 ), (x0 , x1 , x2 ) +
d−3
d−2 2
d−4
2
2
2
3
d
d
d
d
d
d
2
xd−3
0 (x0 x1 , x0 x2 , x1 x2 , x2 ), (x0 , x1 , x2 ) + x0 (x0 , x1 , x2 ) + (x0 x1 ), (x0 , x1 , x2 ) + x0 (x0 x1 x2 ,
d−4
2
3
3
4
d
d
d
2
3
2 2
x0 x32 , x41 , x42 ), (xd0 , xd1 , xd2 ) + xd−4
0 (x0 x1 x2 , x0 x2 , x1 x2 , x2 ), (x0 , x1 , x2 ) + x0 (x0 x1 x2 , x0 x2 , x1 x2 ,
d−3 2
d−5
3
d
d
d
2 2
4
5
5
x42 ), (xd0 , xd1 , xd2 ) + xd−4
0 x2 (x0 , x1 , x2 ) + (x0 x1 x2 ), (x0 , x1 , x2 ) + x0 (x0 x1 x2 , x0 x2 , x1 , x2 ),
d−5
d
d
5
3 2
4
2 2
d
d
d
5
4
4
2 2
(xd0 , xd1 , xd2 ) + xd−5
0 (x0 x1 x2 , x0 x2 , x1 x2 , x2 ), (x0 , x1 , x2 ) + x0 (x0 x1 x2 , x0 x2 , x1 x2 , x2 ), (x0 , x1 ,
d−5 4
d−4 2 2
d
d
5
d
2 3
2 2
4
xd2 ) + xd−5
0 (x0 x1 x2 , x0 x2 , x1 x2 , x2 ) and (x0 , x1 , x2 ) + x0 x2 (x0 , x1 , x2 ) + (x0 x1 x2 ).
Only the first and the second ones correspond to minimal Togliatii systems.
Case (iii). Assume that a ≥ 3 which implies that either c = 1 or c = 2 and we have
s ≤ d − 4. In both cases, since a ≥ 3 and c ≤ 2, m1 cannot be aligned vertically with any
mi . Therefore, in both cases, we get a contradiction with the minimality of I.
To finish case 2C, let us assume e ≥ ⌊ d−s−1
⌋ > 0. We will separate two cases: when b = 0
2
and when b ≥ 1.
(1,1)
Case (i). We assume b = 0, then considering #(Fd−s−1 ∩ ÃI ) and using the bound for e,
we obtain that a = 1 and therefore it is either s = d − 2 or d − 3. So, I is one of the following
d−1
d−3
d
d
d
2
3
2
ideals: (xd0 , xd1 , xd2 ) + xd−2
0 x1 (x0 , x1 , x2 ) + (x0 x2 ), (x0 , x1 , x2 ) + x0 (x0 x1 x2 , x0 x2 , x1 , x1 x2 ),
d−3
2
3
2
d
d
d
(xd0 , xd1 , xd2 ) + xd−3
0 (x0 x1 x2 , x0 x2 , x1 , x1 x2 ) and (x0 , x1 , x2 ) + x0 x1
2
x2 (x0 , x1 , x2 ) + (xd−2
0 x2 ). And any of them are minimal Togliatti systems.
Case (ii). We assume b ≥ 1. In this case, we can assume e ≥ b (the other case is symmetric)
and we obtain the factorization Fd−s−1 = L10 · · · L1b−1 Fd−s−b−1 . If e > b we get a contradiction
with the minimality of I. Hence, e = b. Now we consider as before Ø and we have
I
d−s−b+1
d−s−b
(0,2)
#(Fd−s−b−1 ∩ ØI ) =
d−s−b
d−s−b−1
d − s − 2 ≥ a, d − s − 1 ≥ c
a = d − s − 1, d − s − 1 ≥ c
d − s − 2 ≥ a, c = d − s
a = d − s − 1, c = d − s
In the second and third cases we get immediately a contradiction with the minimality. In
the first case, we can repeat the same argument and we get contradiction unless m1 and m3
are aligned vertically. Hence, we always obtain that c = a + 1, and we have the factorization
On the classification of Togliatti systems
15
(d−s−a,2)
Fd−s−b−1 = L20 · · · L2d−s−a−2 Fa−b . If a ≥ b + 2, then ØI
consists in a − b + 1 different
points, so we have the factorization Fa−b = L2d−s−a Fa−b−1 . Now Fa−b−1 is a plane curve of
(d−s−a−1,2)
degree a − b − 1 which contains all a − b points of Ø
and then it factorizes as
I
Fa−b−1 = L2d−s−a−1 Fa−b−2 which contradicts the minimality of I. Therefore, a = b + 1 and
d−s−b−2
d−s−b−2
I = (xd0 , xd1 , xd2 ) + xs0 (x0 xb+1
, x0 xb1 x2d−s−b−1 , xb+2
, xb1 x2d−s−b ) = (xd0 , xd1 , xd2 ) +
1 x2
1 x2
xs0 xb1 x2d−s−b−2 (x0 x1 , x0 x2 , x21 , x22 ) which is of type (1).
Case 2D: We assume that u := a1 > v := a2 > a3 = a4 =: s ≥ 2. Recall that we have the
factorization Fd−1 = L00 L01 · · · L0s−1 Fd−s−1 and we easily check that the minimality of I forces
, xe1 x2d−e−s )
, x0 xb1 x2d−b−s−1 , xc1 xd−c−s
v = s+1. So we can write I = (xd0 , xd1 , xd2 )+xs0 (xr0 xa1 xd−a−s−r
2
2
with u = s + r and d − s − 1 ≥ r ≥ 2. We can assume d − s − 1 ≥ b ≥ ⌊ d−s−1
⌋ and
2
d − s ≥ c > e ≥ 0, and we have d − s − r ≥ a ≥ 0 and s ≤ d − 3. Let us suppose first that
⌋ > e ≥ 0. We consider
⌊ d−s−1
2
d−s
e ≥ 1, a ≥ 1 (d1)
d − s − 1 e = 0, a ≥ 1 (d2)
(0,1)
#(Fd−s−1 ∩ ÃI ) =
d − s − 1 e ≥ 1, a = 0 (d3)
d−s−2 e= a=0
(d4).
Case d1: In this case a = e. Indeed, if a > e ≥ 1 (the other case is symmetric) we have the
factorization Fd−s−1 = L10 . . . L1e−1 Fd−s−e−1 = L10 . . . L1e−1 L1e Fd−s−e−2 which contradicts the
minimality of I. Let us now consider
d−s−e
b ≥ e + 2, c ≥ e + 2 (i)
d − s − e − 1 b = e + 1, c ≥ e + 2 (ii)
(e+1,1)
#(Fd−s−e−1 ∩ ÃI
)=
d − s − e − 1 b ≥ e + 2, c = e + 1 (iii)
d−s−e−2 b= c=e+1
(iv).
Case (i). We have Fd−s−e−1 = L1e+1 Fd−s−e−2 , and since Fd−s−e−2 passes through all d −
(e,1)
s − e − 1 points of ÃI
we contradicts the minimality of I.
Case (ii). We assume b = e + 1. Let us consider ØI as we did before and we examine
d − s − e + 1 d − s − 1 ≥ c, 1 ≤ d − s − e − r
d−s−e
c = d − s, 1 ≤ d − s − e − r
(0,2)
#(Fd−s−e−1 ∩ ØI ) =
d−s−e
d − s − 1 ≥ c, d − s − e − r = 0
d − s − e − 1 c = d − s, d − s − e − r = 0
In the second and third possibilities we obtain directly a contradiction with the minimality
of I. In the last possibility we also obtain a contradiction. In fact, if c = d−s and s+r = d−e,
(1,2)
we do not remove any point of H12 and we have #(Fd−s−e−1 ∩ ØI ) = d − s − e. Then
Fd−s−e−1 = L21 Fd−s−e−2 = L21 L20 Fd−s−e−3 , which contradicts the minimality of I.
16
R. M. Miró-Roig, M. Salat
Therefore if b = e + 1, it must be d − s − 1 ≥ c ≥ e + 2 and 1 ≤ d − s − e − r. Iterating
this argument we conclude that either c = e + 2 and r = 2 or c = e + 3 and r = 3.
Therefore, either I = (xd0 , xd1 , xd2 ) + xs0 xe1 xd−s−e−2
(x20 , x0 x1 , x21 , x22 ) which is of type (2); or
2
I = (xd0 , xd1 , xd2 ) + xs0 xe1 xd−s−e−3
(x30 , x0 x1 x2 , x31 , x32 ) which is of type (3).
2
Case (iii). Arguing as in case (ii) we get b = e + 2 and r = 2. Therefore, I = (xd0 , xd1 , xd2 ) +
(x20 , x1 x, x21 , x22 ) which is of type (2).
xs0 xe1 xd−s−e−2
2
Case (iv). Arguing as in case (ii) we get that r = 2 and I is of type (1).
Case d2: In this case we assume e = 0 and a ≥ 1. We will separate the case b = 1 from the
case b ≥ 2.
3
3
If b = 1 ≥ ⌊ d−s−1
⌋ then s = d−3 and r = 2. Hence, I = (xd0 , xd1 , xd2 )+xd−3
0 (x0 x1 x2 , x1 , x2 )+
2
d−3 2
d−3 2
2
d
d
d
2
3
d
d
d
2
3
(xd−3
0 x0 x1 ),(x0 , x1 , x2 )+x0 (x0 x1 , x0 x1 x2 , x1 x2 , x2 ) or (x0 , x1 , x2 )+x0 (x0 x1 , x0 x1 x2 , x1 x2 , x2 ).
The first one is not minimal and the remaining two are of type (4).
Assume b ≥ 2. Let us first suppose d − s − r − 1 ≥ 0 (i.e. m1 ∈
/ H02 ) and we consider
d−s
a ≥ 2, c ≥ 2 (i)
d − s − 1 a = 1, c ≥ 2 (ii)
(1,1)
#(Fd−s−1 ∩ ÃI ) =
d − s − 1 a ≥ 2, c = 1 (iii)
d−s−2 a= c =1
(iv).
Case (i). We get Fd−s−1 = L11 Fd−s−2 = L10 L11 Fd−s−3 which contradicts the minimality of I.
Case (ii). Assume that a = 1 and c ≥ 2. Suppose that d − s − r − 1 > 0 and let us consider
d−s
d − s − 2 ≥ b, d − s − 1 ≥ c
d − s − 1 b = d − s − 1, d − s − 1 ≥ c
(0,2)
#(Fd−s−1 ∩ ÃI ) =
d − s − 1 d − s − 2 ≥ b, c = d − s
d − s − 2 b = d − s − 1, c = d − s.
The first possibility contradicts the minimality of I.
Now let us suppose that d − s − r − 1 > 1. In this case, the second (resp. third) possibility
can occur if, and only if b = c = d − s − 1 (resp. b = d − s − 2 and c = d − s). Therefore I
is one of the next types:
I = (xd0 , xd1 , xd2 ) + xs0 (xr0 x1 x2d−s−r−1 , x0 x1d−s−1 , x1d−s−1 x2 , xd−s
2 ) which does not correspond to
a Togliatti system.
d−s
I = (xd0 , xd1 , xd2 ) + xs0 (xr0 x1 x2d−s−r−1 , x0 x1d−s−2 x2 , xd−s
1 , x2 ) which is a Togliatti system if,
and only if r = 2 and s = d − 5 (of type (6)).
If d − s − r − 1 = 1, then there are no special restrictions for the second and third case.
Therefore I is one of the next types:
I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−2 x1 x2 , x0 x1d−s−1 , xc1 x2d−s−c , xd−s
2 ) which is a Togliatti system if,
and only if s = d − 4 and c = 3 (of type (5)), or
On the classification of Togliatti systems
17
d−s
I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−2 x1 x2 , x0 xb1 x2d−s−b−1 , xd−s
1 , x2 ) which is a Togliatti system if,
and only if s = d − 5 and b = 2 (of type (6)), or
d−s
I = (xd0 , xd1 , xd2 ) + xs0 (xr0 x1 x2d−s−r−1 , x0 x1d−s−1 , xd−s
1 , x2 ) which is a Togliatti system if, and
only if r = 2 and s = d − 3 (of type (3)).
Now, let us suppose that d − s − r − 1 = 0. Arguing as usual, we see that it cannot be
d − s − 2 ≥ c and d − s − 3 ≥ b. Therefore d − s ≥ c ≥ d − s − 1 or d − s − 1 ≥ b ≥ d − s − 2,
and we have the following possibilities:
b = d − s − 2, d − s − 2 ≥ c and I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−1 x1 , x0 x1d−s−2 x2 , xc1 x2d−s−c , xd−s
2 )
which is a Togliatti system if, and only if s = d − 3 and c = d − s − 2 (of type (4)).
d − s − 3 ≥ b, c = d − s − 1 and I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−1 x1 , x0 xb1 x2d−s−b−1 , x1d−s−1 x2 , xd−s
2 )
which is a Togliatti system if, and only if d − 3 ≥ s ≥ d − 4 and b = d − s − 3 (resp. of type
(4) and (5)).
b = d − s − 2, c = d − s − 1 and I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−1 x1 , x0 x1d−s−2 x2 , x1d−s−1 x2 , xd−s
2 )
which is a Togliatti system if, and only if s = d − 3 (of type (4)).
b = d − s − 1, d − s − 1 ≥ c and I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−1 x1 , x0 x1d−s−1 , xc1 x2d−s−c , xd−s
2 ) which
is a Togliatti system if, and only if s = d − 3 and d − s − 1 ≥ c ≥ d − s − 2 (of type (4)).
d−s
d − s − 2 ≥ b, c = d − s and I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−1 x1 , x0 xb1 x2d−s−b−1 , xd−s
1 , x2 ) which
is a Togliatti system if, and only if s = d − 3 and d − s − 2 ≥ b ≥ d − s − 3 (of type (4)).
d−s
b = d − s − 1, c = d − s and I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−1 x1 , x0 x1d−s−1 , xd−s
1 , x2 ) which is a
Togliatti system if, and only if s = d − 3 (of type (4)).
(0,2)
Case (iii). Assume c = 1 and a ≥ 2. We consider #(Fd−s−1 ∩ ÃI ) and we obtain that I
is one of the next types:
I = (xd0 , xd1 , xd2 ) + xs0 (xr0 x1d−s−r−1 x2 , x0 x1d−s−1 , x1 x2d−s−1 , xd−s
2 ) which is a Togliatti system if,
and only if r = 2 and d − 3 ≥ s ≥ d − 4 (of type (4) and (5)).
I = (xd0 , xd1 , xd2 ) + xs0 (xr0 x1d−s−r , x0 x1d−s−2 x2 , x1 x2d−s−1 , xd−s
2 ) which is a Togliatti system if,
and only if r = 2 and s = d − 3 (of type (4)).
I = (xd0 , xd1 , xd2 ) + xs0 (xr0 x1d−s−r , x0 x1d−s−1 , x1 x2d−s−1 , xd−s
2 ). In this case, let us consider
(1,2)
#(Fd−s−1 ∩ ÃI ) = d − s and inductively we obtain Fd−s−1 = L21 · · · L1d−s−2 F1 . Therefore it must be r = 2 and s = d − 3, and I is of type (4).
Case (iv). Assume a = c = 1. Let us first suppose that d − s − r − 1 > 0. If d − s − 2 ≥ b
we factorize Fd−s−1 as Fd−s−1 = L20 Fd−s−2 which contradicts the minimality of I. Therefore,
(0,2)
b = d − s − 1 and we factorize Fd−s−1 = L21 · · · Ld−s−r−2 Fr+1 . Since #(Fr ∩ ÃI ) = d − s − 1
we have r + 1 ≥ d − s − 1 and then d − s − r − 1 ≤ 1. Therefore, d − s − r − 1 = 1 and
I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−2 x1 x2 , x0 x1d−s−1 , x1 x2d−s−1 , xd−s
2 ). It is a Togliatti system if, and
only if s = d − 4 (of type (5)).
18
R. M. Miró-Roig, M. Salat
If d − s − r − 1 = 0, we use the same argumentation to prove that d − s − 1 ≥ b ≥ d − s − 2
and then we have two possibilities:
b = d − s − 1 and I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−1 x1 , x0 x1d−s−1 , x1 x2d−s−1 , xd−s
2 )
d−s−1
d−s−2
d−s−1
d
d
d
s
b = d − s − 2 and I = (x0 , x1 , x2 ) + x0 (x0
x1 , x0 x1
x2 , x1 x2
, xd−s
2 )
They are Togliatti systems if, and only if s = d − 3 (of type (4)).
Case d3: Let us assume e ≥ 1 and a = 0. Actually, it must be e = 1. Otherwise, e > 1 and
(1,1)
#(Fd−s−1 ∩ ÃI ) = d − s, and we have seen that this cannot happen.
Now, let us suppose d−s−r > 1. Arguing as before we see that there are three possibilities:
b = d − s − 1, c = d − s − 1 and I = (xd0 , xd1 , xd2 ) + xs0 (xr0 x2d−s−r , x0 x1d−s−1 , x1d−s−1 x2 , x1 x2d−s−1 ).
d−s−1
b = d − s − 2, c = d − s and I = (xd0 , xd1 , xd2 ) + xs0 (xr0 x2d−s−r , x0 x1d−s−2 x2 , xd−s
).
1 , x1 x2
They do not correspond to a Togliatti system.
b = d − s − 1, c = d − s. If d − 2 > s + r, then we have the factorization Fd−s−1 =
(0,2)
L21 · · · Ld−s−r−1 Fr and #(Fr ∩ ÃI ) = d − s − 2 > r, which contradicts the minimality of
d−s−1
I. Hence we have s + r = d − 2 and I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−2 x22 , x0 x1d−s−1 , xd−s
)
1 , x1 x2
which is never a Togliatti system since s ≤ d − 4.
To finish, assume d − s − r = 1. Arguing in the same manner, we see that it cannot occur
d − s − 3 ≥ b and d − s − 2 ≥ c ≥ 1. Therefore we have the following possibilities:
d−s−3 ≥ b, c = d−s−1 and I = (xd0 , xd1 , xd2 )+xs0 (x0d−s−1 x2 , x0 xb1 x2d−s−b−1 , x1d−s−1 x2 , x1 x2d−s−1 ),
it is a Togliatti system if, and only if s = d − 3 and b = d − s − 3 (of type (4)).
b = d−s−2, c = d−s−1 and I = (xd0 , xd1 , xd2 )+xs0 (x0d−s−1 x2 , x0 x1d−s−2 x2 , x1d−s−1 x2 , x1 x2d−s−1 ),
it is a Togliatti system if, and only if s = d − 3 but it is not minimal.
b = d − s − 1, d − s − 1 ≥ c and I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−1 x2 , x0 x1d−s−1 , xc1 x2d−s−c , x1 x2d−s−1 ),
it is a Togliatti system if, and only if s = d − 3 and c = d − s − 1 (of type (4)).
d−s−1
d − s − 2 ≥ b, c = d − s and I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−1 x2 , x0 xb1 x2d−s−b−1 , xd−s
). It
1 , x1 x2
is a Togliatti system if, and only if s = d − 3 and d − s − 2 ≥ b ≥ d − s − 3 (of type (4)), or
s = d − 4 and b = d − s − 2 (of type (5)).
d−s−1
), it is
b = d − s − 1, c = d − s and I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−1 x2 , x0 x1d−s−2 x2 , xd−s
1 , x1 x2
a Togliatti system if, and only if s = d − 3 (of type (4)).
Case d4: Let us assume e = a = 0. If b ≥ 3 and c ≥ 3, we have the factorization
(0,1)
Fd−s−1 = L11 L12 Fd−s−3 and #(Fd−s−3 ∩ ÃI ) = d − s − 2 and we contradict the minimality
of I. Hence we distinguish three cases: b = 1, b = 2 and b ≥ 3.
Case (i). We assume b = 1. Since b ≥ ⌊ d−s−1
⌋ it must be s = d − 3. Therefore
2
d−3
d−1
d
d
d
3
3
d
2
2
3
I = x0 , x1 , x2 ) + x0 (x0 x1 x2 , x1 , x2 ) + (x0 x2 ), (x0 , xd1 , xd2 ) + xd−3
0 (x0 x2 , x0 x1 x2 , x1 x2 , x2 ) or
2
2
3
(xd0 , xd1 , xd2 ) + xd−3
0 (x0 x2 , x0 x1 x2 , x1 x2 , x2 ). The first one is not minimal while the remaining
two are of type (1).
⌋ it must be d − 3 ≥ s ≥ d − 5.
Case (ii). We assume b = 2. Since b ≥ ⌊ d−s−1
2
On the classification of Togliatti systems
19
d−3 2
2
2
3
3
d
d
d
2
2
3
If s = d−3, I = (xd0 , xd1 , xd2 )+xd−3
0 (x0 x2 , x0 x1 , x1 , x2 ), (x0 , x1 , x2 )+x0 (x0 x2 , x0 x1 , x1 x2 , x2 )
2
2
2
3
or (xd0 , xd1 , xd2 ) + xd−3
0 (x0 x2 , x0 x1 , x1 x2 , x2 ). All of them are of type (4).
d−4 3
3
2
4
4
d
d
d
2
If s = d − 4, I = (xd0 , xd1 , xd2 ) + xd−4
0 (x0 x2 , x0 x1 x2 , x1 , x2 ), (x0 , x1 , x2 ) + x0 (x0 x2 , x0 x1 x2 ,
d−4 3
4
3
2
d
d
d
4
2 2
2
3
x31 x2 , x42 ), (xd0 , xd1 , xd2 )+xd−4
0 (x0 x2 , x0 x1 x2 , x1 x2 , x2 ), (x0 , x1 , x2 )+x0 (x0 x2 , x0 x1 x2 , x1 x2 , x2 ),
d−4 2 2
d
d
d
4
3
2
d
d
d
4
4
2
2 2
(xd0 , xd1 , xd2 )+xd−4
0 (x0 x2 , x0 x1 x2 , x1 , x2 ), (x0 , x1 , x2 )+x0 (x0 x2 , x0 x1 x2 , x1 x2 , x2 ), (x0 , x1 , x2 )+
d−4 2 2
2 2
2
2 2
4
d
d
d
2
3
4
xd−4
0 (x0 x2 , x0 x1 x2 , x1 x2 , x2 ), (x0 , x1 , x2 ) + x0 (x0 x2 , x0 x1 x2 , x1 x2 , x2 ). Only the fifth one is
a minimal Togliatti system, and it is of type (5).
Finally, if s = d − 5, I has 15 possibilities, but any of them is a minimal Togliatti system.
Case (iii). We assume b ≥ 3. Then, either c = 1 or c = 2.
Case c = 1. We will see that b = d − s − 1. Suppose d − s − 2 ≥ b ≥ 3, then #(Fd−s−1 ∩
(0,2)
ÃI ) = d − s and Fd−s−1 = L20 Fd−s−2 . First we will see that this implies that m1 and
m2 are aligned vertically (i.e. d − s − b − 1 = d − s − r). We suppose that d − s − b −
1 ≤ d − s − r, and then b ≥ r − 1 (the other case is symmetric). Inductively we obtain
(d−s−b−1,2)
Fd−s−1 = L20 L21 · · · L2d−s−b−2 Fb . If b > r − 1, then #ÃI
= b + 1 and it would mean
to a contradiction with the minimality of I. Hence, b = r − 1 and we get the factorization
(d−s−b−1,2)
Fd−s−1 = L20 L21 · · · L2d−s−b−2 L2d−s−b · · · Ld−s−2 F1 . Since #ÃI
= b ≥ 3 we have again
a contradiction with the minimality.
Once we have seen that b = d−s−1, using the usual argumentation we see that d−s−r = 1.
Therefore I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−1 x2 , x0 x1d−s−1 , x1 x2d−s−1 , xd−s
2 ) with s ≤ d − 3, which is
Togliatti if, and only if s = d − 3 and it is of type (4).
(1,1)
Case c = 2. Since #ÃI = d − s we have Fd−s−1 = L11 Fd−s−2 . If d − s − 2 ≥ b ≥ 3, then
(0,2)
#ØI
= d − s − 1 and Fd−s−1 would factorize as Fd−s−1 = L11 L20 Fd−s−3 . This contradicts
(0,1)
the minimality of I because #Ø
= d − s − 2 which would force the factorization F
=
I
d−s−1
L11 L20 L10 Fd−s−4 . Therefore b = d − s − 1 and again by minimality we see that d − s − r = 1.
Hence, I = (xd0 , xd1 , xd2 ) + xs0 (x0d−s−1 x2 , x0 x1d−s−1 , x21 x2d−s−2 , xd−s
2 ), which is Togliatti if, and
only if s = d − 3 and in this case it is of type (4)
To finish case 2D we see what happens when d − s ≥ c > e ≥ ⌊ d−s−1
⌋. We see using the
2
minimality that either a ≥ b = e, b ≥ a = e or e ≥ a = b.
Arguing as before we see that in the first possibility m1 and m3 must be vertically
aligned and in particular c = e + 2, a = e and r = 2. Therefore I = (xd0 , xd1 , xd2 ) +
xs0 xe1 xd−s−e−2
(x20 , x0 x1 , x21 , x22 ) which is of type (1).
2
Now we assume b ≥ a = e. If b, c ≥ e + 1, then we have the factorization Fd−s−e−1 =
(e,1)
L1e+1 Fd−s−e−2 and, since #ØI
= d − s − e − 1 we get Fd−s−e−1 = L1e L1e+1 Fd−s−e−3 which
contradicts the minimality. Now, suppose b = e + 1 and c ≥ e + 2 (resp. b ≥ e + 2 and
c = e + 1). As we have seen earlier, m1 and m3 (resp. m2 ) must be aligned. Therefore,
20
R. M. Miró-Roig, M. Salat
we can see using the minimality assumption that r = 2 and c = e + 2 (resp. r = 2 and
b = e + 2). In both cases I is of type (1).
Finally, let us assume e ≥ a = b. If e ≥ a + 2, we get a contradiction with the minimality
of I. Hence either e = a or e = a + 1. If e = a we see that c = a + 2 and r = 2. Therefore I
is of type (1). Otherwise e = a + 1 and we get c = a + 2 and r = 2 and I is of type (1).
For any integer d ≥ 3, set M 0 (d) = {xa0 xb1 xc2 | a + b + c = d and a, b, c ≥ 1}.
Theorem 3.9. Let I ⊂ k[x0 , . . . , xn ] be a smooth minimal monomial Togliatti system of
forms of degree d ≥ 10. Assume that µ(I) = 2n + 3. Then, n = 2 and, up to a permutation
of the coordinates, one of the following cases holds:
(i) I = (xd0 , xd1 , xd2 ) + m(x20 , x21 , x0 x2 , x1 x2 ) with m ∈ M 0 (d − 2), or
(ii) I = (xd0 , xd1 , xd2 ) + m(x20 , x21 , x0 x1 , x22 ) with m ∈ M 0 (d − 2), or
(iii) I = (xd0 , xd1 , xd2 ) + m(x30 , x31 , x32 , x0 x1 x2 ) with m ∈ M 0 (d − 3).
Proof. By [5, Proposition 4.1], for n ≥ 3 and d ≥ 4 there are no smooth minimal monomial
Togliatti systems I ⊂ k[x0 , . . . , xn ] of forms of degree d with µ(I) = 2n + 3. So, n = 2.
For n = 2, the result follows from Theorem 3.8 together with the smoothness criterion
Proposition 3.6.
The following remarks shows that in the above Theorem the hypothesis of being smooth
cannot be deleted.
Remark 3.10. If n = 3 and d ≥ 10 one can easily check that I = (xd0 , xd1 , xd2 , xd3 ) +
2
2
2
xd−2
0 (x0 x1 , x2 x3 , x1 , x2 , x3 ) is a minimal monomial Togliatti systems of forms of degree d
with µ(I) = 2n + 3 = 9 and it is non-smooth.
Remark 3.11. For n = 2 and 6 ≤ d ≤ 9 one can check with the help of Macaulay2 [3] that
there exist other examples of minimal monomial Togliatti systems I = (xd0 , xd1 , xd2 ) + J ⊂
k[x0 , x1 , x2 ] with µ(I) = 7. For seek of completeness we give the full list of possible J’s not
included in Theorem 3.8:
d = 6 : (x50 x1 , x30 x32 , x20 x31 x2 , x51 x2 ), (x50 x2 , x30 x31 , x20 x21 x22 , x51 x2 ), (x30 x32 , x20 x41 , x20 x21 x22 , x51 x2 ), (x50 x2 , x30 x31 ,
x0 x1 x42 , x51 x2 ), (x40 x22 , x30 x31 , x20 x21 x22 , x41 x22 ), (x30 x32 , x20 x41 , x20 x21 x22 , x41 x22 ), (x40 x22 , x30 x31 , x0 x1 x42 , x41 x22 ), (x30 x31 , x30 x32 ,
x20 x21 x22 , x31 x32 ), x0 x1 (x40 , x20 x21 , x0 x1 x22 , x41 ), x0 x1 (x30 x2 , x20 x21 , x0 x1 x22 , x31 x2 ), x0 x1 (x20 x21 , x20 x22 , x0 x1 x22 , x21 x22 ),
x0 x1 (x20 x21 , x20 x22 , x0 x32 , x21 x22 ), x0 x1 (x40 , x0 x32 , x41 , x21 x22 ), x0 x1 (x40 , x20 x21 , x41 , x42 ), x0 x1 (x40 , x0 x1 x22 , x41 , x42 ),
x0 x1 (x30 x2 , x20 x21 , x31 x2 , x42 ), x0 x2 (x30 x2 , x20 x22 , x0 x1 x22 , x41 ), x0 x2 (x20 x1 x2 , x20 x22 , x0 x1 x22 , x41 ), x0 x2 (x30 x2 , x0 x21 x2 ,
x0 x1 x22 , x41 ), x0 x2 (x30 x2 , x20 x1 x2 , x0 x32 , x41 ), x0 x2 (x30 x2 , x20 x22 , x0 x32 , x41 ), x0 x2 (x30 x2 , x0 x21 x2 , x0 x32 , x41 ), x0 x2 (x20 x21 ,
x20 x22 , x0 x31 , x31 x2 ), x0 x2 (x20 x21 , x20 x22 , x0 x21 x2 , x31 x2 ), x0 x2 (x20 x22 , x0 x31 , x0 x21 x2 , x31 x2 ), x0 x2 (x20 x21 , x20 x22 , x0 x31 , x41 ),
x0 x2 (x20 x21 , x20 x22 , x41 , x31 x2 ), x0 x2 (x20 x21 , x20 x22 , x41 , x21 x22 ), x0 x2 (x30 x2 , x20 x1 x2 , x0 x21 x2 , x41 ), x0 x2 (x30 x2 , x20 x1 x2 ,
x0 x1 x22 , x41 ), x0 (x0 x41 , x0 x1 x32 , x0 x42 , x31 x22 ), x0 (x40 x2 , x20 x31 , x0 x21 x22 , x51 ), x0 (x40 x2 , x0 x1 x32 , x51 , x31 x22 ), x0 (x20 x32 ,
x0 x41 , x0 x21 x22 , x51 ), x0 (x40 x2 , x20 x1 x22 , x51 , x21 x32 ), x0 (x20 x32 , x0 x41 , x0 x1 x32 , x31 x22 ), x0 (x40 x2 , x20 x32 , x0 x31 x2 , x51 ),
On the classification of Togliatti systems
21
x0 (x40 x2 , x20 x31 , x51 , x1 x42 ), x0 (x30 x22 , x20 x31 , x0 x42 , x31 x22 ), x0 (x40 x2 , x0 x21 x22 , x51 , x1 x42 ), x0 (x20 x32 , x0 x41 , x0 x42 , x31 x22 ),
x0 (x20 x31 , x20 x32 , x41 x2 , x1 x42 ), x0 (x40 x1 , x20 x32 , x0 x31 x2 , x51 ), x0 (x20 x1 x22 , x0 x31 x2 , x51 , x52 )
d = 7 : x0 x1 (x20 x32 , x0 x41 , x0 x21 x22 , x51 ), x0 x1 (x50 , x20 x21 x2 , x0 x1 x32 , x51 ), x0 x1 (x40 x1 , x30 x21 , x0 x42 , x31 x22 ), x0 x1 (x30 x21 ,
x20 x31 , x20 x32 , x21 x32 ), x0 x1 (x50 , x20 x21 x2 , x51 , x52 ), x0 x1 (x40 x2 , x0 x41 , x51 , x52 ), x0 x1 (x50 , x0 x1 x32 , x51 , x52 ), x0 x2 (x30 x22 ,
x20 x32 , x0 x31 x2 , x51 ), x0 x2 (x40 x2 , x20 x1 x22 , x0 x42 , x51 ), x0 x2 (x40 x2 , x0 x31 x2 , x0 x42 , x51 ), x0 (x50 x2 , x20 x31 x2 , x0 x21 x32 , x61 ),
x0 (x0 x51 , x0 x21 x32 , x0 x52 , x41 x22 ), x0 (x50 x2 , x40 x21 , x20 x21 x22 , x31 x32 ), x0 (x40 x21 , x40 x22 , x20 x21 x22 , x31 x32 ), x0 (x30 x31 , x30 x32 ,
x20 x21 x22 , x31 x32 ), x0 (x40 x1 x2 , x20 x41 , x20 x42 , x31 x32 ), x0 (x20 x41 , x20 x21 x22 , x20 x42 , x31 x32 ), x0 (x40 x1 x2 , x0 x51 , x0 x52 , x31 x32 ),
x0 (x20 x21 x22 , x0 x51 , x0 x52 , x31 x32 ), x0 (x50 x2 , x20 x31 x2 , x61 , x1 x52 ), x0 (x50 x2 , x0 x21 x32 , x61 , x1 x52 ), x0 (x50 x2 , x40 x21 , x51 x2 ,
x1 x52 ), x0 (x40 x21 , x40 x22 , x51 x2 , x1 x52 ), x0 (x30 x31 , x30 x32 , x51 x2 , x1 x52 ), x0 (x40 x1 x2 , x20 x21 x22 , x61 , x62 ), x0 (x40 x1 x2 , x61 , x31 x32 ,
x62 ), x0 (x20 x21 x22 , x61 , x31 x32 , x62 ), x0 x1 x2 (x20 x21 , x20 x22 , x0 x31 , x41 ), x0 x1 x2 (x30 x2 , x20 x1 x2 , x0 x21 x2 , x41 ), x0 x1 x2 (x40 , x20 x21 ,
x0 x1 x22 , x41 ), x0 x1 x2 (x30 x2 , x20 x1 x2 , x0 x1 x22 , x41 ), x0 x1 x2 (x30 x2 , x20 x22 , x0 x1 x22 , x41 ), x0 x1 x2 (x20 x1 x2 , x20 x22 , x0 x1 x22 ,
x41 ), x0 x1 x2 (x30 x2 , x0 x21 x2 , x0 x1 x22 , x41 ), x0 x1 x2 (x30 x2 , x20 x1 x2 , x0 x32 , x41 ), x0 x1 x2 (x30 x2 , x20 x22 , x0 x32 , x41 ),
x0 x1 x2 (x30 x1 , x0 x31 , x0 x32 , x41 ), x0 x1 x2 (x30 x2 , x0 x21 x2 , x0 x32 , x41 ), x0 x1 x2 (x20 x21 , x20 x22 , x0 x31 , x31 x2 ), x0 x1 x2 (x20 x21 ,
x20 x22 , x0 x21 x2 , x31 x2 ), x0 x1 x2 (x20 x22 , x0 x31 , x0 x21 x2 , x31 x2 ), x0 x1 x2 (x30 x2 , x20 x21 , x0 x1 x22 , x31 x2 ), x0 x1 x2 (x30 x1 , x20 x22 ,
x0 x1 x22 , x31 x2 ), x0 x1 x2 (x20 x21 , x20 x22 , x41 , x31 x2 ), x0 x1 x2 (x40 , x0 x31 , x0 x32 , x21 x22 ), x0 x1 x2 (x40 , x0 x32 , x41 , x1 x32 ).
d = 8 : x0 x1 (x40 x22 , x30 x31 , x0 x1 x42 , x41 x22 ), x0 x2 (x30 x32 , x20 x21 x22 , x0 x41 x2 , x61 )
d = 9 : x0 x1 x2 (x30 x32 , x20 x21 x22 , x0 x41 x2 , x61 ), x0 x1 x2 (x30 x31 , x30 x32 , x20 x21 x22 , x31 x32 ), x0 x1 x2 (x60 , x20 x21 x22 , x61 , x62 ).
References
[1] H. Brenner and A. Kaid, Syzygy bundles on P2 and the Weak Lefschetz Property, Illinois J. Math. 51 (2007),
1299–1308.
[2] I.M. Gelfand, M.M. Kapranov and A.V. Zelevinsky, Discriminants, Resultants and Multidimensional Determinants, Birkhäuser, Boston 1994.
[3] D.R. Grayson and M.E. Stillman, Macaulay2, a software system for research in algebraic geometry, Available at
http://www.math.uiuc.edu/Macaulay2/
[4] E. Mezzetti, R.M. Miró-Roig and G. Ottaviani: Laplace Equations and the Weak Lefschetz Property, Canad. J.
Math. 65 (2013), 634–654.
[5] E. Mezzetti and R.M. Miró-Roig: The minimal number of generators of a Togliatti system, Annali di Matematica
Pura ed Applicata 195 (2016), 2077-2098. DOI .10.1007/s10231-016-0554-y.
[6] E. Mezzetti and R.M. Miró-Roig: Togliatti systems and Galois coverings, Preprint 2016.
[7] M. Michalek and R.M. Miró-Roig, Smooth monomial Togliatti systems of cubics, Journal of Combinatorial Theory,
Ser. A 143 (2016), 66-87. http://dx.doi.org /10.1016/j.jcta.2016.05.004.
[8] J. Migliore and R.M. Miró-Roig: Ideals of general forms and the ubiquity of the Weak Lefschetz property, J. Pure
Appl. Algebra 182 (2003), 79–107.
[9] J. Migliore, R.M. Miró-Roig and U. Nagel: On the Weak Lefschetz Property for Powers of Linear Forms. Algebra
and Number Theory, 6 (2012), No. 3, 487-526.
[10] J. Migliore, R.M. Miró-Roig and U. Nagel: Monomial ideals, almost complete intersections, and the Weak
Lefschetz property, Trans. Amer. Math. Soc. 363 (2011), no. 1, 229–257.
[11] R.M. Miró-Roig: Ordinary curves, webs and the ubiquity of the Weak Lefschetz Property, Algebras and Representation Theory 17 (2014), 1587 – 1596.
[12] D. Perkinson, Inflections of toric varieties, Michigan Math. J. 48 (2000), 483–515.
[13] R. Stanley, The number of faces of a simplicial convex polytope, Adv. Math. 35 (1980), 236–238.
22
R. M. Miró-Roig, M. Salat
[14] R. Stanley, Weyl groups, the hard Lefschetz theorem, and the Sperner property, SIAM J. Algebraic Discrete
Methods 1 (1980), 168–184.
[15] E. Togliatti, Alcuni esempi di superfici algebriche degli iperspazi che rappresentano un’equazione di Laplace,
Comm. Math. Helvetici 1 (1929), 255-272.
[16] E. Togliatti, Alcune osservazioni sulle superfici razionali che rappresentano equazioni di Laplace, Ann. Mat. Pura
Appl. (4) 25 (1946) 325-339.
[17] J. Vallès, Variétés de type Togliatti, C. R. Acad. Sci. Paris, Ser. I 343 (2006), 411-414.
[18] J. Watanabe, The Dilworth number of Artinian rings and finite posets with rank function, Commutative Algebra
and Combinatorics, Advanced Studies in Pure Math. Vol. 11, Kinokuniya Co. North Holland, Amsterdam (1987),
303–312.
Department de matemàtiques i Informàtica, Universitat de Barcelona, Gran Via de les Corts
Catalanes 585, 08007 Barcelona, Spain
E-mail address: [email protected]
Department de matemàtiques i Informàtica, Universitat de Barcelona, Gran Via de les Corts
Catalanes 585, 08007 Barcelona, Spain
E-mail address: [email protected]
| 0math.AC
|
ON THE DENSITY OF IMAGES OF THE POWER MAPS IN LIE
GROUPS
arXiv:1701.00331v1 [math.GR] 2 Jan 2017
SAURAV BHAUMIK AND ARUNAVA MANDAL
Abstract. Let G be a connected Lie group. In this paper, we study the density of the
images of individual power maps Pk : G → G : g 7→ g k . We give criteria for the density
of Pk (G) in terms of regular elements, as well as Cartan subgroups. In fact, we prove
that if Reg(G) is the set of regular elements of G, then Pk (G) ∩ Reg(G) is closed in
Reg(G). On the other hand, the weak exponentiality of G turns out to be equivalent to
the density of all the power maps Pk . In linear Lie groups, weak exponentiality reduces
to the density of P2 (G). We also prove that the density of the image of Pk for G implies
the same for any connected full rank subgroup.
1. Introduction
Let G be a connected Lie group and let g be the Lie algebra associated to G. The
question of whether the exponential map exp : g → G is surjective or has dense image
has been addressed by many authors (see [H], [HM], etc). A connected Lie group G is
called weakly exponential (resp. exponential) if the image of the exponential map exp is
dense (resp. surjective). McCrudden [Mc] showed that exp : g → G is surjective if and
only if the power map Pk : G → G, defined by g → gk is surjective for all k ≥ 2. The
surjectivity of the power maps has been studied by P. Chatterjee ([C1], [C2], [C3], and see
there references), R. Steinberg [St], Dani and Mandal [DM]. If a connected Lie group G
is weakly exponential, then Pk (G) is dense for all k ≥ 1.
For an algebraic group over R or C, the power maps can have dense images without
being surjective. For a connected complex algebraic group which is not exponential (e.g.,
SL(2, C)), there is some k > 1 for which Pk is not surjective, but have dense images.
One can take Weil restriction to produce such an example over R. For an algebraic group
G over Qp , the question of dense images of the power map is equivalent to the study of
surjectivity (see Remark 3.1). The surjectivity of Pk in G(Qp ) (Qp -points of the algebraic
group G defined over Qp ) has been settled by P. Chatterjee in [C2]. However there does
not seem to be any study in the literature on the density of the individual power maps.
This motivated us to consider the following natural question. Given k > 1, when is the
image Pk (G) dense in G?
On the other hand, weak exponentiality of a Lie group is closely related to the regular elements and Cartan subgroups. A. Borel showed that a connected semisimple Lie
group is weakly exponential if and only if all of its Cartan subgroups are connected (see
[HM, Theorem 2.10]). Later, K.H. Hofmann showed that for a connected Lie group G
(not necessarily semisimple), the set of its regular elements Reg(G) has a property that,
Reg(G) ∩ exp(g) is closed in Reg(G) (see [H, Theorem 17]). Also, he deduced criteria
for weak exponentiality of G in terms of regular elements and Cartan subgroups (see [H,
Corollary 18]). Following this, K. H. Neeb proved that, a connected Lie group is weakly
exponential if and only if all of its Cartan subgroups are connected (see [N, Theorem I.2]).
It is therefore natural to ask whether Pk (G) ∩ Reg(G) is closed in Reg(G) for all k ≥ 1.
Key words and phrases. Power maps of Lie groups, regular elements, Cartan subgroups, weak exponentiality, full rank subgroups.
1
2
S. BHAUMIK AND A. MANDAL
In the following we answer these questions.
Theorem 1.1. Let G be a connected Lie group. Let k > 1 be an integer. The following
are equivalent:
(a) Pk (G) is dense in G.
(b) Reg(G) ⊂ Pk (G).
(c) If C is a Cartan subgroup of G, then Pk (C) = C.
Theorem 1.2. Let G be a connected Lie group. Let k > 1 be an integer. Then Pk (G) ∩
Reg(G) is closed in Reg(G).
The proof of Theorem 1.1 relies crucially on Theorem 1.2. A result of Wüstner, that
every Cartan subgroup is compatible with some Levi decomposition, also turned out to
be very useful in many of our proofs.
Analogous results due to Hofmann ([H, Theorem 17, Corollary 18]), and K. H. Neeb
([N, Theorem I.2]) in the context of exponential maps, can be deduced from Theorem 1.2
and Theorem 1.1 (see Corollary 3.2).
In analogy to McCrudden’s criterion for exponential images for a connected Lie group,
in terms of divisibility, we have the following corollary.
Corollary 1.3. Let G be a connected Lie group. Then Pk : G → G is dense for all k if
and only if G is weakly exponential.
The next result also gives a characterization of the density of Pk .
Corollary 1.4. Let G be a connected Lie group and let C be a Cartan subgroup of G. Let
C ∗ denote the connected component of C and k > 1 be an integer. Then Pk : G → G is
dense if and only if Pk : C/C ∗ → C/C ∗ is surjective. In particular, G is weakly exponential
if and only if C/C ∗ is divisible for any Cartan subgroup C.
The density of k-th power map Pk only depends on its Levi part of the group G. More
precisely, if R is the solvable radical (i.e., maximal solvable connected normal subgroup)
of G, then Pk (G) is dense in G if and only if Pk (G/R) is dense in G/R (Proposition 3.3).
We list some conditions for the density of the images of power maps Pk for simple Lie
groups in §4.
The following provides a criterion for weak exponentiality in terms of P2 only.
Corollary 1.5. Let G be a connected linear Lie group. Then G is weakly exponential if
and only if P2 : G → G is dense.
This means that (compare with Theorem 1.1), in a connected linear Lie group G, if
every element that belongs to a Cartan subgroup admits a square root then G is weakly
exponential. For linear groups, this extends Neeb’s result, which says that G is weakly
exponential if and only if all of its Cartan subgroups are connected. On the other hand it
strengthens one of P. Chatterjee’s results ([C3, Theorem 1.6]). However, the corollary is
no longer valid for non-linearizable groups, as we observe in Example 5.1.
The following can be thought of as a power map analogue of Neeb’s result (see [N,
Proposition I.6]).
Theorem 1.6. Let G be a connected Lie group and let A be a connected full rank subgroup
of G. Let k > 1 be an integer. Then Pk : G → G is dense implies Pk : A → A is dense.
Let G be a connected linear reductive Lie group and g be its associated Lie algebra. Let
θ be the Cartan involution on g. Then there is a natural Cartan decomposition g = k + p,
where k and p are the eigenspaces corresponding to the eigenvalues 1 and −1 respectively.
ON THE DENSITY OF IMAGES OF THE POWER MAPS IN LIE GROUPS
3
It follows that [k, k] ⊆ k, [k, p] ⊆ p, [p, p] ⊆ k and z(g) ⊆ k. Let a be a maximal abelian
subspace of p and take a Cartan subalgebra ta of zk (a). Define c := ta + a. It is noted
in Neeb [N] that c is a Cartan subalgebra of g and every Cartan subalgebra c′ of g is
conjugate to a Cartan subalgebra is of the form t′a + a′ , where a′ ⊆ a and t′a ⊇ ta .
We give below another criterion for the density of images of power maps in case of
reductive groups, which is mostly important for non-split case.
Corollary 1.7. Let G be a connected reductive Lie group. Let k > 1 be an integer. Then
Pk : G → G is dense if and only if Pk : ZG (ta ) → ZG (ta ) is dense.
The paper is organized as follows. We recall some definitions and prove some results on
regular elements, Cartan subgroups, including Theorem 1.2, in §2. Theorem 1.1, Corollaries 1.3, 1.4 are proved in §3. We discuss the density of power maps in simple Lie groups
in §4. Corollary 1.5 is proved in §5. Finally, in §6 we prove Theorem 1.6 and Corollary
1.7.
2. Regular elements, Cartan subgroups and power maps
In this section we recall some definitions, introduce notations, prove a few lemmas and
Theorem 1.2.
Definition 1. [H] An element g in a Lie group G is called regular if the nilspace N (Adg −
1) ⊂ g is of minimal possible dimension. The set of regular elements of G is denoted by
Reg(G).
Definition 2. [C1] Let G be a Lie group. An element g in G is called Pk -regular if dPk
is nonsingular at g.
Lemma 2.1. Let G be a real Lie group. Suppose g, h ∈ G, and hk = g. Then g is regular
if and only if h is both regular and Pk -regular.
Proof. Let g be regular. Writing T = Adg − 1, U = Adh − 1 and
V = Adh(k−1) + Adh(k−2) + . . . + Adh + 1,
we get T = U V . Then N (U ) ⊂ N (T ), and by minimality of dim(N (T )), we conclude that
h is regular.
Now, the complexification gC of the real Lie algebra g, splits into direct sum of generalized eigenspaces Vα under Adh , i.e., gC = ⊕α∈∆ Vα (where ∆ denotes the set of eigenvalues
of Adh ). We note that V1 = N (U ) ⊗ C. Let αk = 1 but α 6= 1. Then in the splitting of gc
under Adkh = Adg , the generalized eigenspace for 1 will contain V1 ⊕ Vα . By minimality of
dimC (V1 ) = dimR (N (U )), Vα = 0. By [C1, Lemma 2.1], it follows that h is Pk -regular.
Conversely, let h be Pk -regular and regular. By [C1, Lemma 2.1], V is invertible on gC
and N (U ) ⊗ C = N (T ) ⊗ C and hence N (T ) is of minimal dimension.
We now recall the definition of Cartan subgroup of a Lie group. Let G be a connected
Lie group with its Lie algebra g. Let c be the Cartan subalgebra of g and let ∆ be the
set of roots of gC belonging to cC . Thus gC = cC + Σα∈∆ gαC . The normalizer NG (c) of c is
defined by NG (c) := {g ∈ G|Adg (c) = c}. Note that NG (c) is a closed subgroup of G with
Lie algebra ng (c) = {X ∈ g|[X, c] ⊆ c}. Let C(c) := {g ∈ NG (c)|α ◦ Adg |cC = α, ∀α ∈ ∆}.
A closed subgroup C of a Lie group G is called a Cartan subgroup if its Lie algebra c is
a Cartan subalgebra of g and C(c) = C. This agrees with the usual definition of a Cartan
subgroup (see Appendix of [N]). In particular, a Cartan subgroup is nilpotent.
Let us recall a theorem due to M. Wüstner ([W, Theorem 1.9 (ii)]). Let G be a connected
Lie group and let R be its solvable radical. Then for any Cartan subgroup C of G, there is
4
S. BHAUMIK AND A. MANDAL
a Levi subgroup S of G such that C can be decomposed as C = (C ∩ S)(C ∩ R). Moreover,
C ∩ S is a Cartan subgroup of S, C ∩ R is connected and C ∩ R ⊆ ZG (S).
Now we can immediately observe that, if G is connected linear Lie group then any
Cartan subgroup can be written as C = T N (almost direct product), where T is abelian
and N is connected nilpotent subgroup of C. Indeed, if G is linear, so is S. This means,
S is isomorphic to the identity component (in Euclidean topology) of a geometrically
connected semisimple algebraic group defined over R. In this case, T = C ∩ S is abelian,
and N = C ∩ R is connected nilpotent. This will be used later.
Lemma 2.2. Let G be a Lie group, and let Z ⊂ Z(G) be a closed central subgroup. Let
G1 = G/Z, and let π : G → G1 be the projection.
(a) If C is a Cartan subgroup of G, then C1 = π(C) is a Cartan subgroup of G1 , and
C = π −1 (C1 ).
(b) Suppose Z is connected. If Pk (C1 ) is both closed and open in C1 , then Pk (C) is both
closed and open in C.
Proof. (a) It follows from the description of Cartan subgroup in terms of its adjoint action
on the Lie algebra. Indeed, if c is a Cartan subalgebra of g, then c1 = c/z is a Cartan
subalgebra of g1 = g/z, while z ⊂ z(g) ⊂ zg (c). For any x ∈ G, Adx preserves z(g), hence
z. Therefore Adπ(x) normalizes c1 if and only if Adx normalizes c. The roots α of g1 with
respect to c1 are induced from roots α of g with respect to c, and gα = gα1 . Hence C1 is a
Cartan subgroup of G1 .
Again, C ⊂ π −1 (C1 ). Let x ∈ π −1 (C1 ). Since C1 = π(C), there is some y ∈ C such
that π(x) = π(y). This means, x = yz, for some z ∈ Z ⊂ Z(G) ⊂ C, which proves
C = π −1 (C1 ).
π
(b) By (a), we get a short exact sequence 1 → Z → C → C1 → 1. Now, since Z
is a connected abelian group, for any x ∈ C, xZ ∩ Pk (C) is nonempty if and only if
xZ ⊂ Pk (C). Therefore Pk (C) = π −1 (Pk (C1 )).
Lemma 2.3. Let C be a Lie group and let C̃ be a cover of C. If Pk (C̃) is both open and
closed then so is Pk (C).
Proof. Let π : C̃ → C be the covering map. Then π(Pk (C̃)) = Pk (C). By hypothesis,
Pk (C̃) is disjoint union of connected components of C̃. Let Pk (C̃) = ⊔i∈I X̃i , where X̃i ’s
are connected component of C̃ and I is the indexing set. For any i, π(X̃i ) is connected,
hence there is a connected component Xi of C such that π(X̃i ) ⊆ Xi . We claim that for
i ∈ I, π(X̃i ) is a connected component of C. Suppose the claim holds, then π(X̃i ) = Xi
and π(Pk (C̃)) = ⊔i∈I π(X̃) = ⊔i∈I Xi . As Xi ’s are both open and closed in C, Pk (C) is
both open and closed. This will prove the lemma.
To prove our claim, we note that, as π is a covering map, π(X̃i ) is open in Xi , therefore
it is enough to prove that π(X̃i ) is closed in Xi . Now, let x be an element in the closure of
π(X̃i ). As π is covering of a Lie group C, there is a connected open set U containing x such
that π −1 (U ) = ⊔r Vr , where Vr ’s are open in X̃ and π restricted to Vr is a diffeomorphism
onto U. Then, there exists r0 , such that X̃i ∩ Vr0 6= ∅. Since Vr0 is connected, it must lie
entirely in the connected component X̃i . Therefore x ∈ U = π(Vr0 ) ⊂ π(X̃i ).
Proposition 2.4. Let G be any connected real Lie group and let C be a Cartan subgroup
of G. For any k ≥ 0, Pk (C) is both open and closed in C.
Proof. First note that it is enough to consider the case when Z(G) is discrete. Indeed, let
G1 = G/Z(G)∗ , where Z(G)∗ is the identity component of Z(G), and let π1 : G → G1 be
the projection. Then by Lemma 2.2, it is enough to prove that Pk (C1 ) is closed and open
in C1 . Inductively, Gi+1 = Gi /Z(Gi )∗ , and let Ci+1 be the image of Ci in Gi+1 . Since the
ON THE DENSITY OF IMAGES OF THE POWER MAPS IN LIE GROUPS
5
dimension of G is finite, there is some i such that Z(Gi ) is discrete. Henceforth, we will
assume that Z = Z(G) is discrete.
Note that G is a covering group of the linear group G′ = G/Z. Let π : G → G′ denote
the covering projection. Therefore, by Lemma 2.2, π(C) is a Cartan subgroup of G′ . Then
by Wüstner, there is a Levi subgroup S of G′ such that π(C) = (π(C) ∩ S)(π(C) ∩ R)
(almost direct product), where R is the solvable radical of G′ .
Now, write G′ = S · R (almost semidirect product). As S acts on R, we consider an
external semidirect product of S and R, denoted by S ⋉ R. We observe that there is a
surjective group homomorphism P : S ⋉ R → S · R, whose kernel is isomorphic to D =
{(z, z −1 )|z ∈ S ∩R}. Since D is discrete normal, it is central in S ⋉R. Hence by Lemma 2.2,
P −1 (π(C)) is a Cartan subgroup of S ⋉R and in fact P −1 (π(C)) = (π(C)∩S)×(π(C)∩R).
As π(C) ∩ R is connected linear nilpotent, it can be written as T1 × N for some torus T1
and simply connected subgroup N of π(C) ∩ R. Therefore P −1 (π(C)) is of the form T × N ,
where T = (π(C) ∩ S) × T1 . Hence Pk (P −1 (π(C)) is both open and closed in P −1 (π(C))
implies Pk (π(C)) is so in π(C), as S ⋉ R is a finite covering of G′ . Therefore, to prove
Pk (π(C)) is both open and closed, it is enough to show the same for Pk (P −1 (π(C))).
Hence, by Lemma 2.3, we may assume that π(C) = T × N , where T is an abelian group
and N is a simply connected nilpotent group. Now, π : C → T ×N is a covering projection.
Therefore C has to be of the form A × N , where A is a covering space of T . Indeed, let
H1 = π −1 (T × {1}) and H2 = π −1 ({1} × N ), where {1} denotes the identity element in
G. Then H1 , H2 are normal in C. Again, N is simply connected, so the covering space
H2 → N has a section, while the kernel Z is central. Hence H2 = N × Z. Now, let x ∈ H1 .
Then ιx induces an action on H2 which preserves the central Z. Let (y, z) ∈ N × Z. Then
ιx (y, z) = (y ′ , z ′ ). Now, taking projection to N , since x ∈ H1 , we see that y = y ′ . Hence
the action of ιx is given by a matrix
idN
{1}
φ
idZ
where φ is a homomorphism N → Z. Now, since N is connected and Z is discrete, φ is
zero, hence H1 and H2 commute. Therefore we have a surjective homomorphism of groups
H1 × H2 → C given by inclusions and multiplication in C. The kernel is {(z, z −1 )|z ∈ Z}.
Since H2 = N ×Z, we see that the induced homomorphism H1 ×N → C is an isomorphism.
We set, A = H1 .
This means, it is enough to prove that Pk (A) is closed in A. So we can assume that C
is a covering group of T with central kernel Z.
Since G′ is linear, T is of the form F × T2 , where F is a finite Z/2Z-vector space, and
T2 is a connected abelian group. Now, in order to prove that Pk (C) is both open and
closed in C, by Lemma 2.3, it is enough to do so after replacing C by a covering group.
In particular, we can replace T2 by its simply connected covering, which will be a vector
space, say V , and replace C by the fiber product of V and C over T2 .
So we can assume that C is an extension of F × V by the central Z.
Pr
π
Let p : C → F be the composite C → F × V →1 F. Let us consider the normal subgroup
B = π −1 (0 × V ). This is a covering group of the simply connected V by Z, hence is of
the form Z × V . Since V is connected abelian and Z is discrete, B is central by similar
argument as above.
Therefore, we can express C as an extension of F by the central B = Z × V as follows.
p
0 → B → C → F → 0.
Let the number of elements in F be r, and let F = {p(x1 ), . . . , p(xr )}. Now, C is the
disjoint union of the cosets xi B. Therefore, Pk (C) is the union of Pk (xi B). Note that,
6
S. BHAUMIK AND A. MANDAL
since B is central in C, Pk (xB) = xk Pk (B), and hence is closed and open in xk B. This
implies that Pk (C) = ∪ri=1 Pk (xi B) is closed and open in C.
Proof of Theorem 1.2: By [B, Ch. VII, §4 n◦ 2, Proposition 5], for each regular g ∈ G and
for each identity neighbourhood W there is a neighbourhood V of g such that for each
x ∈ V there is a w ∈ W with C(x) = ιw C(g) (equivalently, c(x) = Adw c(g) at the level of
Lie subalgebras).
Let g ∈ Reg(G) and let gn ∈ Reg(G) be a sequence such that gn = hkn and g is
the limit of the sequence i.e. gn → g in G. Note that Pk : G → G is nonsingular at
identity, because dPk at identity is multiplication by k. Hence there are sequences of
identity neighbourhoods Wi , Wi′ such that both the diameters of Wi , Wi′ tend to zero,
and Wi = Pk (Wi′ ). Now, let Vi be neighbourhoods of g corresponding to Wi′ . Since
gn → g, for each i there is some gni ∈ Vi . We can choose ni+1 > ni . Hence we can
assume without loss of generality that gi ∈ Vi . We can therefore find ηi ∈ Wi′ such that
ηi −1 C(gi )ηi = C(g). Since C(gi ) = C(hi ), ηi −1 hi ηi ∈ C(g) for all i, and hence writing
ξi = ηik ∈ Wi , ξi −1 gi ξi ∈ Pk (C(g)). Since ξi → 1 and gi → g, ξi −1 gi ξi → g. Hence, by
closedness of Pk (C(g)) (Proposition 2.4), we get g ∈ Pk (C(g)).
3. Characterization of the density of the power map Pk
In this section we prove Theorem 1.1 by using Theorem 1.2. Also, we deduce many
characterization about the dense images of the power map, which can be used in the rest
of the paper.
Proof of Theorem 1.1: (a) ⇔ (b) If Pk (G) is dense in G then Pk (G) ∩ Reg(G) is dense
in Reg(G), hence equal to Reg(G), by Theorem 1.2. As Reg(G) is dense in G and it is
contained in Pk (G), the converse follows.
(b) ⇔ (c) Suppose Reg(G) ⊂ Pk (G). It is enough to show that C ∩ Reg(G) ⊂ Pk (C),
because C ∩ Reg(G) is dense in C (see [H1, Proposition 1.6]), while Pk (C) is closed in C
(Proposition 2.4). If x ∈ C ∩ Reg(G) then by assumption, there is y ∈ G such that x = y k .
This implies y ∈ Reg(G), by Lemma 2.1. Then y lies in a unique Cartan subgroup, which
has to be C because y k ∈ C. Conversely, any regular element g can be found in a unique
Cartan subgroup C.
Remark 3.1. For algebraic groups we can give a more straightforward argument.
(1) Let G be an algebraic group over a field F. Let C be a Cartan subgroup of G. Then
for F = R or Qp , the regular elements of G(F) (group of F-rational point of the algebraic
group G) which can be conjugated to C(F) form an open subset of G(F) in the Euclidean
or the ultrametric topology.
To see this, let Φ : G × C → G : (h, y) 7→ hyh−1 and fixing y ∈ C, let Φy : G → G is
given by Φy (h) = hyh−1 . Note that the differential of Φ at (e, y), where y ∈ Reg(G) ∩ C,
is surjective. Write Ty C = dLy Te C,
dΦ(ξ, dLy v) = dΦ(ξ, 0) + dLy v = dΦy (ξ) + dLy v = dLy (df (ξ) + v),
where f : G → G : h 7→ y −1 hyh−1 . Note that, f is the composition of G → G × G : h 7→
µ
(y −1 hy, h−1 ) and G × G → G : (x, y) 7→ xy. Then,
df (ξ) = dµ(Ady −1 ξ, −ξ) = (Ady −1 − I)ξ.
Now, since y ∈ C is regular, Te C = N (Ady −1 − I). Therefore we get the following.
dΦ(g × Ty C) = dLy (im(Ady −1 − I) + N (Ady −1 − I)) = g.
(2) Now, for F = Qp , if the group is not anisotropic, there is a nonempty open subset
of regular elements which can be conjugated to a Cartan subgroup C which contains a
×
copy of Q×
p = Z × Zp . In this case, for any k > 0, Pk (C) is proper, open and closed
ON THE DENSITY OF IMAGES OF THE POWER MAPS IN LIE GROUPS
7
in C. Therefore none of (a), (b), (c) of Theorem 1.1 will hold. On the other hand, if G
is anisotropic then it is weakly exponential. In this case, density of Pk is equivalent to
surjectivity.
(3) For F = R, if the group contains any disconnected Cartan subgroup C, then
P2k (C) = C ∗ , the identity component. Therefore, P2k is never dense. On the other
hand, if P2k+1 (C) = C, so that P2k+1 (G) contains all regular elements, hence is dense. In
case all Cartan subgroups are connected G is weakly exponential.
Proof of Corollary 1.3: Suppose that Pk (G) is dense in G. Then by Theorem 1.1, Pk (G)
contains a dense open set (namely Reg(G)). Hence by Baire category theorem, ∩k≥1 Pk (G)
contains a non empty open set and therefore G is weakly exponential. Now the converse
follows from McCrudden criterion.
Theorem 1.1 can be used to obtain the following criteria of weak exponentiality, which
is well known ([N, Theorem I.2]).
Corollary 3.2. Let G be a connected Lie group and let g be the Lie algebra associated to
G. The following are equivalent.
(a) G is weakly exponential.
(b) Reg(G) ⊆ exp(g).
(c) All Cartan subgroups of G are connected.
Proof. (a) ⇔ (b) By Corollary 1.1, G is weakly exponential if and only if Pk (G) is dense
for all k, which is equivalent to saying that Reg(G) ⊆ ∩k≥1 Pk (G) ⊆ exp(g), by Theorem
1.1.
(a) ⇔ (c) By using Theorem 1.1, G is weakly exponential if and only if all Cartan
subgroups of G are divisible (i.e., Pk (C) = C for all k ≥ 1). By [HL, Proposition 1], this
is equivalent to saying that C is connected.
Proposition 3.3. Let G be a connected Lie group, and let R be the radical of G. Let k > 0
be an integer. Then Pk : G → G is dense if and only if Pk : G/R → G/R is dense.
Proof. The right implication is obvious. For the converse, let H be a Cartan subgroup of
G. Then there exists a Levi subgroup S of G such that H = (H ∩ S)(H ∩ R). We write,
G = S · R. Let Z = S ∩ R and ∆(Z) := {(z, z −1 )|z ∈ Z}. Now, consider S ⋉ R as an
external semidirect product of S and R. The map π : S ⋉ R → S · R sending (s, r) → s · r,
for s ∈ S, r ∈ R, is a covering map and G can be thought of as a quotient of S ⋉R by ∆(Z).
Then Cartan subgroup π −1 (H) of S ⋉ R is of the form C × N where C = H ∩ S is a Cartan
subgroup of S, N = H ∩ R is connected nilpotent subgroup of R, and H = (C × N )/∆(Z).
Also, note that as S → S/Z is covering, then C/Z is a Cartan subgroup of G/R. Now,
there is a short exact sequence of groups
1 → N → (C × N )/∆(Z) → C/Z → 1.
As Pk (G/R) is dense in G/R, by Theorem 1.1, Pk (C/Z) = C/Z. Now, let x = x1 x2 ∈
(C × N )/∆(Z) for some x1 ∈ C and x2 ∈ N. As Pk (C/Z) = C/Z, there exists y1 ∈ C
and z ∈ Z such that x1 = y1k z. Thus x1 x2 = y1k zx2 = y1k y2k in H, for some y2 ∈ N , as z is
contained in N and N is connected nilpotent Lie group. Now we notice that x1 x2 = y1 y2 k ,
as C and N commute. Hence the power map Pk : H → H is surjective. Therefore, by
applying Theorem 1.1, the proposition follows.
To deduce Corollary 1.4, we need next two lemmas.
Lemma 3.4. Let π : G → H be a covering of connected Lie groups, with discrete kernel
D. Let C ⊂ H be a connected Cartan subgroup which is abelian. Then the corresponding
Cartan subgroup M = π −1 (C) is also abelian.
8
S. BHAUMIK AND A. MANDAL
Proof. Note that the identity component M ∗ is a covering of C. Since C is connected
abelian, M ∗ is abelian. Any element x ∈ M is of the form x = mz for some m ∈ M ∗ and
z ∈ D. Now we note that, the group homomorphism M ∗ × D → M , given by (m′ , z ′ ) →
m′ z ′ , is surjective. Hence M is abelian.
Lemma 3.5. Let G̃ be a connected Lie group with discrete center. Let C̃ be any Cartan
subgroup of G̃. Then the connected component C̃ ∗ of C̃ is central.
Proof. Let G = G̃/Z(G̃) and C := π(C̃) be the corresponding Cartan subgroup of G.
We can write C = A × F , where A is connected and F is a Z/2Z vector space. If F is
trivial, by Lemma 3.4, C̃ is abelian and hence the conclusion holds. Suppose that F is
not trivial. We observe that π −1 (F ) is a normal subgroup of C̃, where π : C̃ → C is
the natural projection map. Therefore, C̃ ∗ acts trivially on π −1 (F ), as C̃ ∗ is connected
and π −1 (F ) is discrete. Thus elements of C̃ ∗ and π −1 (F ) commute each other. Note that
C̃ is a quotient of C̃ ∗ and π −1 (F ). Indeed, for x ∈ C̃, let π(x) = (a, b) ∈ A × F. We
−1
can choose x1 ∈ C̃ ∗ such that x1 = π −1 (a). Then it follows that xx−1
1 ∈ π (F ). Thus
−1
∗
∗
x = (xx−1
1 )x1 ∈ π (F ) × C̃ and hence C̃ is central in C̃.
Proof of Corollary 1.4: (:⇒) This follows from Theorem 1.1.
(⇐:) Let R be the solvable radical of G. We recall that, there is a decomposition of
the Cartan subgroup C ([W, Theorem 1.9 (ii)]). For C, there is a Levi subgroup S of G
such that C = Cs · N (almost direct product), where Cs = C ∩ S is a Cartan subgroup
of S and N = C ∩ R is connected nilpotent contained in ZG (S). Therefore we observe
that C/C ∗ ≃ Cs /Cs∗ Z, where Z = Cs ∩ N , a central subgroup. Also, we note that if
C1 is a Cartan subgroup of G/R, then C1 /C1∗ ≃ Cs /Cs∗ Z. Hence Pk : C/C ∗ → C/C ∗ is
surjective if and only if Pk : C1 /C1∗ → C1 /C1∗ is surjective. Applying Lemma 3.5 for the
semisimple group G/R, we get Pk : C1 → C1 is surjective and hence by Theorem 1.1,
Pk : G/R → G/R is dense. Now by Proposition 3.3, we conclude that Pk (G) is dense in
G.
4. Density of the power maps on simple Lie groups
In this section we study the dense images of the power maps in simple Lie groups. We
use results in §3, to determine the integer k, for which Pk has dense images.
We recall from Remark 3.1 that, for a connected real algebraic group G with disconnected Cartan subgroup, Pk is dense if and only if k is odd. However, in general this is
not true.
^R) be the simply connected cover of SL(2, R). Then
Example 4.1. Let G̃ = SL(2,
^R) → SL(2,
^R) is not dense for all k > 1. Indeed, the center of G̃ is infiPk : SL(2,
nite cyclic, and the adjoint group PSL(2, R) has a Cartan subgroup isomorphic to R. The
corresponding Cartan subgroup in G̃ is therefore isomorphic to R × Z, where the component Z is contributed by the center. Therefore for k > 1, Pk is never surjective on this
Cartan subgroup, so the assertion follows by Theorem 1.1.
There is a list of weakly exponential non-compact non-complex simple Lie algebras in
Neeb ([N]). Neeb also mentioned exactly which in the list are completely weakly exponential (i.e., whose corresponding simply connected group is weakly exponential). Thus
recalling Borel’s criterion about weak exponentiality, one can understand the connectedness of the Cartan subgroup.
For a connected Lie group G, if Cartan subgroups are connected, then Pk is dense in
G for all k, as G is weakly exponential. Therefore we concentrate the case when Cartan
subgroups are disconnected. As we noted before (Proposition 3.3), the study of density
ON THE DENSITY OF IMAGES OF THE POWER MAPS IN LIE GROUPS
9
boils down to semisimple Lie group, at this moment we now investigate this phenomenon
for simple Lie groups.
Let us fix some notations for this section. Let G̃ be the universal cover of the connected
simple adjoint Lie group G. Let π denote the covering map and C̃ be a Cartan subgroup
of G̃. Then let C = π(C̃) be the corresponding Cartan subgroup of G. Let us denote the
fundamental group of G by π1 (G). So we get a natural short exact sequence
1 → π1 (G) → C̃ → C → 1.
For a connected simple adjoint Lie group G, we notice that its fundamental group is
finitely generated, whose possible factors are isomorphic to Z, Z/2Z, Z/4Z, Z/8Z, or Z/3Z.
We discuss some cases. In the following cases we assume that some Cartan subgroup of
G̃, say C̃ is disconnected.
Case 1. Let π1 (G) ≃ Z2 or non cyclic group of order 4.
This happens for An I (n ≥ 2), Bn I, E66 I, E88 V III, F44 I, G22 I, Cn II, E7−5 V I, E8−24 IX,
−22
F4 II. Note that none of these is weakly exponential.
Lemma 4.1. In Case 1, Pk : G̃ → G̃ is dense if and only if k is odd.
Proof. First assume that the Cartan subgroup C = π(C̃) of G is disconnected. By hypothesis, it is enough to check for odd k, as for even k, Pk (G) is not dense in G. Note that
π1 (G) is central in C̃. Since for k odd, Pk is surjective on C and π1 (G) as well, Pk : C̃ → C̃
is surjective and hence by Theorem 1.1, the lemma follows.
Now, suppose that C is connected. Using the same argument as before, for odd k, Pk
is dense in G̃. Now Suppose that P2 is dense in G̃. Then Pk is dense in G̃ for all k, which
further implies that Pk (C̃) = C̃ for all k, by Theorem 1.1. Thus by [HL, Proposition 1], C̃
is connected, which is a contradiction. This implies that, for even k, Pk (G̃) is not dense
in G̃, as desired.
Case 2(a). C is connected, and π1 (G) ≃ Z. This occurs for E7−25 V II, A1 I.
As C is connected, we get the following short exact sequence
1 → C̃ ∗ ∩ Z → C̃ ∗ → C → 1.
Suppose that C̃ ∗ ∩ Z is the subgroup nZ of Z. Thus we have C̃/C̃ ∗ ≃ Z/nZ.
In this case, Pk : G̃ → G̃ is dense if and only if k is coprime to n.
Indeed, as Pk : Z/nZ → Z/nZ is surjective if and only if k is coprime to n, the result
follows from Theorem 1.4.
Case 2(b). C is disconnected, and π1 (G) ≃ Z. Then there is a surjective group homomorphism C̃/C̃ ∗ → C/C ∗ . Let D = Ker(C̃/C̃ ∗ → C/C ∗ ). As π1 (G) is infinite cyclic, D is
isomorphic to Z/nZ for some n.
This happens for A1 I, Cn I (n ≥ 3), Dn III (n ≥ 4 even).
Lemma 4.2. In case 2(b), Pk : G̃ → G̃ is dense if and only if k is odd and coprime to n.
Proof. First observe that, as C is disconnected, for even k, Pk is not dense.
In this case we get the following short exact sequence.
1 → Z/nZ → C̃/C̃ ∗ → C/C ∗ → 1.
Let σ be any element in Z/nZ. If π : C̃/C̃ ∗ → C/C ∗ , then π(σ) = 0. Suppose that σ = tk
for some t ∈ C̃/C̃ ∗ . Then we have 0 = π(σ) = π(tk ) = π(t)k . As for odd integer, Pk is
identity on C/C ∗ , we conclude that t ∈ Z/nZ. Therefore, if k is not coprime to n, Pk
is not surjective on C̃/C̃ ∗ . On the other hand, if k is odd and coprime to n, then both
10
S. BHAUMIK AND A. MANDAL
Pk : Z/nZ → Z/nZ and Pk : C/C ∗ → C/C ∗ are surjective and the result follows from the
short exact sequence.
Case 3. C is not connected and π1 (G) ≃ Z/6Z. This happens for E62 II.
Lemma 4.3. Pk : G̃ → G̃ is dense if k is odd and coprime to 3 and for even k, Pk : G̃ → G̃
is not dense.
Proof. We first note that, if k is odd and coprime to 3 (hence to 6), then Pk is surjective
both on Z/6Z and C. Hence Pk is surjective on C̃. Then the result follows by applying
Theorem 1.1.
Since C is disconnected, P2 can not be dense in G̃.
Case 4. Let π1 (G) ≃ Z/4Z or Z/8Z and C is disconnected. This occurs for Dn I(n ≥
4 and n odd) and Dn I(n ≥ 4 and n even) respectively. The following lemma is very
straightforward and hence we omit the details.
Lemma 4.4. In case 4, Pk : G̃ → G̃ is dense if and only if k is odd. Moreover, if G′ be
any cover of G, then Pk : G′ → G′ is dense if and only if k is odd.
Case 5. G is split.
As usual, let G̃ be its universal cover. Let Z denote the discrete group π1 (G). Suppose
that there is a Cartan subgroup C of the form A × F , where A is simply connected and
F is a Z/2Z vector space. Then we get the following short exact sequence,
1 → Z → C̃ → A × F → 1,
where C̃ is the corresponding Cartan subgroup of C. Let K be the kernel of the map
C̃ → A. Then we have the following two short exact sequences
1 → Z → K → F → 1.
and
1 → K → C̃ → A → 1.
Now, as A is simply connected, we get a section of the above short exact sequence, so that
C̃ = A ⋉ K. It turns out that C̃ is a direct product of A and K, as A is connected and
K is discrete. Hence Pk (C̃) = C̃ if and only if Pk (K) = K. Furthermore, let σ ∈ Z. Let
σ = tk for some t ∈ K. We denote π is the map from K → F in the above short exact
sequence. Note that for even k, Pk (G̃) is not dense. For odd k, as Pk : C → C is identity,
0 = π(σ) = π(tk ) = Pk (π(t)) = π(t),
we get t ∈ Z. Thus we have Pk (G̃) is dense in G̃ if and only if Pk (Z) = Z.
Let D be a subgroup of Z and G′ = G̃/D. Then by above procedure we can conclude
that Pk (G′ ) is dense in G′ if and only if Pk (Z/D) = Z/D.
Remark 4.5. In Case 3, the authors have not addressed the case when k is odd and
divisible by 3, as they are not able to calculate the number of connected components of
the Cartan subgroups involved. They have also left unaddressed the case when G is a
non-split, non-compact simple Lie group and π1 (G) contains an infinite cyclic subgroup,
for the same reason.
ON THE DENSITY OF IMAGES OF THE POWER MAPS IN LIE GROUPS
11
5. Relation between square map and weak exponentiality
In this section we discuss the relation between the density of P2 and weak exponentiality
of Lie groups. Also, we prove Corollary 1.5.
Proof of Corollary 1.5: Let R be the solvable radical of G. Then G/R is a connected
linear semisimple Lie group. As any connected semisimple Linear Lie group is precisely
the Hausdorff connected component of some Zariski connected semisimple real algebraic
group, Cartan subgroups of a connected linear Lie group are either connected or it has 2m
(for some positive integer m) number of disconnected components. Let C be any Cartan
subgroup and C ∗ be its connected component. As P2 (G) is dense in G, by Corollary 1.4,
we get P2 : C/C ∗ → C/C ∗ is surjective. This implies that C = C ∗ , and hence G is weakly
exponential.
The converse follows from Corollary 1.3.
Example 5.1. We give an example of a non-linearizable Lie group G which is not weakly
^R). We
exponential, but P2 (G) is dense. Let us consider G = PSL(2, R) and G̃ = SL(2,
know that Z(G̃) is infinite cyclic. Fix a generator σ. Let Z to be the subgroup of Z(G̃)
generated by 3σ and define G′ = G̃/Z. There is a Cartan subgroup of G isomorphic to R,
and the corresponding Cartan subgroup for G̃ is isomorphic to R × Z, where the factor Z
is contributed by the center. Therefore, the corresponding Cartan subgroup of G′ would
be isomorphic to R × Z/3Z, which is not connected. By Theorem 1.1, P2 (G′ ) is dense in
G′ , but it has a disconnected Cartan subgroup, and hence G′ is not weakly exponential.
Remark 5.1. Let G be a connected Lie group, not necessarily linear, and let R be its
radical. Suppose the adjoint semisimple group Ad(G/R) does not contain any simple factor
whose fundamental group contains an infinite cyclic subgroup. It follows from the cases
discussed in the previous section (§4) that if P2 (G) is dense then G is weakly exponential.
6. Density on full rank subgroup
We recall that a Lie subgroup A of G is of full rank if its Lie algebra a contains a
Cartan subalgebra of g = Lie(G). In this case, every Cartan subalgebra of a is a Cartan
subalgebra of g ([B, Proposition 3, §3, Ch. VII]).
Proposition 6.1. Let G be a connected Lie group and let A be a connected subgroup of
full rank. Let k ≥ 1 be an odd integer. If Pk (G) is dense in G, then Pk (A) is dense in A.
Proof. Note that since A is of full rank, Z(G)∗ ⊂ A. By killing Z(G)∗ , it is easy to see that
A1 = A/Z(G)∗ is a full rank subgroup of G1 = G/Z(G)∗ . Inductively, set Zi = Z(Gi )∗
and Ai = Ai−1 /Zi−1 . Then Ai is of full rank in Gi . Since the dimension of G is finite,
Zi is discrete for some i. It is enough to show that Pk (Ai ) is dense in Ai . For, if H is a
connected Lie group and Z ⊂ H is central connected, then Pk (H) is dense in H if and
only if Pk (H/Z) is dense in H/Z. This is because, the projection π : H → H/Z is a
submersion, hence an open map, and π −1 (Pk (H/Z)) = Pk (H).
Thus we can assume that Z(G) is discrete. Since Z(G) is the fundamental group of
linear Lie group Ad(G), it is a finitely generated abelian group.
We know that Reg(G) ∩ A is dense in A (see proof of [N, I.6]). It is enough to prove
that elements in Reg(G) ∩ A admit k-th roots in A. Let g ∈ Reg(G) ∩ A. There is a unique
Cartan subgroup C of G such that g ∈ C. The subgroup C ′ := C ∩ A contains the identity
component C ∗ , which is also the identity component of the Cartan subgroup of A that
contains g, with Lie algebra given by Na (Adg − I) = Ng (Adg − I). Therefore, C ′ is the
disjoint union of cosets of C ∗ , hence open in C ′ . It is enough to show that Pk (C ′ ) = C ′ .
12
S. BHAUMIK AND A. MANDAL
Writing D = Z(G), let π : G → G/D be the projection. Now, by Theorem 1.1,
Pk (C) = C. We get an exact sequence as follows, where E is the corresponding Cartan
subgroup of G/D:
π
1 → D → C → E → 1.
This gives another short exact sequence
π
1 → D/(D ∩ C ∗ ) → C/C ∗ → E/E ∗ → 1.
However, E/E ∗ is an Z/2Z-vector space, so for odd k ≥ 0, Pk induces identity on E/E ∗ .
Therefore Pk is surjective on D/(D ∩ C ∗ ). Indeed, let x ∈ D/(D ∩ C ∗ ). Since Pk (C) = C,
there is some y ∈ C/C ∗ such that x = y k . But 0 = π(x) = π(y)k = π(y), so y ∈
D/(D ∩ C ∗ ). This means, D/(D ∩ C ∗ ) is a torsion abelian group with torsion coprime to
k. However, C ′ fits into a short exact sequence
1 → D ′ → C ′ → F → 1,
where F = π(C ′ ) is an open subgroup of E, and D′ = D ∩ C ′ . This gives rise to the short
exact sequence as follows.
1 → D ′ /(D′ ∩ C ∗ ) → C ′ /C ∗ → F/F ∗ → 1.
Here F/F ∗ is the group of connected component of F ⊂ E, which must be an F2 -vector
space, and D′ /(D ′ ∩ C ∗ ) is a subgroup of D/D ∩ C ∗ , hence a torsion abelian group with
torsion coprime to k. Therefore Pk is surjective on C ′ /C ∗ .
Lemma 6.2. Let G be a connected Lie group and let A be a connected full rank subgroup.
Let C be any Cartan subgroup and C ∗ be its connected component. Suppose that C/C ∗
is finitely generated abelian. Let k > 0 be an integer. If Pk : G → G is dense then
Pk : A → A is dense.
Proof. In view of Proposition 6.1, it is enough to prove for integer k = 2. As P2 (G) is dense
in G, by Corollary 1.4, P2 : C/C ∗ → C/C ∗ is surjective. By hypothesis C/C ∗ is a torsion
abelian group and torsion is coprime to 2. We recall from the proof of Proposition 6.1
that, to prove Pk : A → A is dense, it only needs to show that P2 : A ∩ C/C ∗ → A ∩ C/C ∗
is surjective. Note that A ∩ C/C ∗ is also torsion subgroup of C/C ∗ and torsion is coprime
to 2 as well. This proves the lemma.
Proposition 6.3. Let G be a connected Lie group and let A be a connected full rank
subgroup of G. Then P2 : G → G is dense implies P2 : A → A is dense.
Proof. Arguing as before (see the proof of Proposition 6.1), we may assume that Z(G)
is discrete. Now, let π : G → Ad(G) be the covering map and C be a Cartan subgroup
of G. Let CAd = π(C) be the corresponding Cartan subgroup of Ad(G). Since P2 (G) is
dense in G, P2 : Ad(G) → Ad(G) is dense and hence CAd is connected. Let C ∗ denote the
identity component of C. We observe that the restriction map π : C ∗ → CAd is surjective.
Indeed, π(C ∗ ) is non empty open and closed in CAd (for details, see the proof of Lemma
2.3). Therefore we get a surjective homomorphism from Z(G) × C ∗ → CAd . Hence C/C ∗
is finitely generated abelian group (as Z(G) is finitely generated abelian group). Thus by
Lemma 6.2, it follows that P2 : A → A is dense.
Proof of Theorem 1.6: The theorem follows from Proposition 6.1 and Proposition 6.3.
Proof of Corollary 1.7: It is noted in Neeb ([N]) that, ZG (ta ) is a connected, full rank,
reductive subgroup of G. Therefore forward implication of the statement follows from Theorem 1.6. Now for the converse, let C ′ be any Cartan subgroup of G. Let c′ be the Lie
algebra associated to C ′ . Since g is reductive, c′ is reductive in g and its associated Cartan
ON THE DENSITY OF IMAGES OF THE POWER MAPS IN LIE GROUPS
13
subgroup is C ′ = ZG (c′ ). We may assume that c′ = t′a + a′ , where a′ ⊆ a and t′a ⊇ ta .
Then it is immediate that C ′ ⊆ ZG (ta ) and by [N, Lemma I.5], C ′ is a Cartan subgroup
of ZG (ta ). Now, by applying Theorem 1.1 on ZG (ta ), we conclude that Pk (C ′ ) = C ′ , as
required.
Acknowledgement. We thank Professor S. G. Dani for his useful comments, particularly Remark 3.1.
References
[B] Bourbaki, N. Lie Groups and Lie Algebras, Elements of Mathematics, Springer. Chapters 7-9.
[C1] Chatterjee, P. On surjectivity of the power maps of solvable Lie groups, J. Algebra 248 (2002), no.
2, 669 - 687.
[C2] Chatterjee, P. On the power maps, orders and exponentiality of p-adic algebraic groups, J. Reine
Angew. Math., (2009), no. 629, 201 - 220.
[C3] Chatterjee, P. Automorphism invariant Cartan subgroups and power maps of disconnected groups,
Math. Z. 269 (2011), no. 4, 221 - 233.
[DM] Dani, S.G. and Mandal, A. On the surjectivity of the power maps of a class of solvable groups,
Preprint.
[F] Fell, J. A Hausdorff topology for the closed subsets of a locally compact non Hausdorff space,
[H] Hofmann, K.H. A memo on the exponential function and regular points, Arch. Math., (1992), no. 59,
24 - 37.
[H1] Hofmann, K.H. Near-Cartan algebras and groups, Seminar Sophus Lie, (1992), no. 2, 135 - 151.
[HL] K.H. Hofmann, J.D. Lawson, Divisible subsemigroups of Lie groups, J. London Math. Soc. 27 (1983),
427 - 437.
[HM] Hofmann, K.H., and A. Mukherjea, On the density of the image of the exponential function, Math.
Ann., (1978), 263 -273.
[Mc] McCrudden, M. On nth roots and infinitely divisible element, Arch. Math., (1992), no. 59, 24 - 37.
[N] Neeb, K-H. Weakly exponential Lie groups, J. Algebra. 179 (1996), no. 15, 331-361.
[St] Steinberg, R. On the power maps in algebraic groups, Math. Res. Letters, (2003), no. 10, 621 - 624.
[W] Wüstner, M. A generalization of the Jordan decomposition. Forum Math. volume 15 (2003) 395-408.
Saurav Bhaumik
Department of Mathematics,
Indian Institute of Technology Bombay,
Powai, Mumbai 400076, India
[email protected]
Arunava Mandal
Department of Mathematics,
Indian Institute of Technology Bombay,
Powai, Mumbai 400076, India
[email protected]
| 4math.GR
|
LEMPEL-ZIV COMPLEXITY REFERENCE
arXiv:1707.09848v1 [cs.IT] 28 Jul 2017
G. RUFFINI - JAN 2017 — STARLAB TECHNICAL NOTE, TN00344 (V1.0)
Contents
Abstract
1. LZW compression: the main concept
2. The process of digitization
3. LZW and entropy rate for stochastic processes
4. Metrics
5. Fundamental Normalization of LZW
6. Other normalizations or measures
7. LZW and Kolmogorov complexity
References
Annex: StarLZW.py
1
2
2
3
4
4
4
5
5
6
Abstract
The aim of this note is to provide some reference facts for LZW—mostly from Thomas
and Cover [1]—adapted to the needs of the Luminous project. LZW is an algorithm to
compute a Kolmogorov Complexity estimate derived from a limited programming language
that only allows copy and insertion in strings (not Turing complete set).
Despite its delightful simplicity, it is rather powerful and fast. We then focus on definitions of LZW derived complexity metrics consistent with the notion of descriptive length,
and discuss different normalizations, which result in a set of metrics we call ρ0 , ρ1 and ρ2 ,
in addition to the Description Length lLZW and the Entropy Rate.
1
2
G. RUFFINI - JAN 2017 — STARLAB TECHNICAL NOTE, TN00344 (V1.0)
1. LZW compression: the main concept
The main idea in LZW is to look for repeating patterns in the data, and instead of
rewriting repeating sequences, refer to the last one seen [3]. As Kaspar clearly states, LZW
is the Kolmogorov Complexity computed with a limited set of programs that only allow
copy and insertion in strings [2, 5].
“We do not profess to offer a new absolute measure for complexity which, as mentioned
already, we believe to be nonexistent. Rather, we propose to evaluate the complexity of a
finite sequence from the point of view of a simple self- delimiting learning machine which,
as it scans a given n- digit sequence S = s1 · s1 · ...sn , from left to right, adds a new
word to its memory every time it discovers a substring of consecutive digits not previously
encountered. The size of the vocabulary, and the rate at which new words are encountered
along S, serve as the basic ingredients in the proposed evaluation of the complexity of S.”
We consider a string of characters in and alphabet with A symbols (typically binary) of
length n. From wikipedia: A high level view of the encoding algorithm is shown here:
2.
3.
4.
5.
Find t h e l o n g e s t s t r i n g W i n t h e d i c t i o n a r y t h a t matches t h e c u r r e n t i n p u t .
Emit t h e d i c t i o n a r y i n d e x f o r W t o output and remove W from t h e i n p u t .
Add W f o l l o w e d by t h e nex t symbol i n t h e i n p u t t o t h e d i c t i o n a r y .
Go t o Step 2 .
After applying LZW, we will end up with a set of words (or phrases, as they are sometimes called) c(n) that go into a dictionary. The length of the compressed string will be
lLZW ≤ n (the analog of Kolmogorov or algorithmic complexity).
The description length of the sequence encoded by LZW would have length less
or equal to the number of phrases times the number of bits needed to identify a seen phrase
plus the bits to specify a new symbol (to form a new phrase), hence1
(1)
lLZW ≤ c(n) log2 [c(n) + log2 A] ≈ c(n) log2 [c(n)]
2. The process of digitization
When we digitize (e.g., binarize) a signal prior LZW, we are creating a new string from
the data, and we make an explicit choice on what aspects of the data we wish to compress.
In this process we destroy information—we are going to do lossy compression. Thus, the
choice of digitization results in us having access to a subset of the features of the original
string.
1 Actually, we can do a bit better than this. In practice, not all dictionary entries are used. We can use
the max dictionary key ID and state that “n bits are needed to describe any key entry, and there are m of
them (and here they are)”, leading to lLZW ≤ log2 (log2 max(output)) + length(output) ∗ log2 [max(output)],
since we need log2 (log2 max(output)) bits to describe n. This is how it is implemented in the appended
code.
LEMPEL-ZIV COMPLEXITY REFERENCE
3
A reasonable strategy is to preserve as much information as possible in the resulting
transformed string. In this sense, using methods that maximize the entropy of the resulting series are recommended, such as using the median for thresholding (this is guaranteed
to result in H0 = 1)2.
On the other hand, other methods that destroy more information may tap and highlight
other, also relevant features of the data. At this stage, then, how to binarize or preprocess
(e.g., filter) the original string is an empirical question. The same applies to the choice of
compression method, of course, as LZW is just one framework for compression.
3. LZW and entropy rate for stochastic processes
The main fact from Thomas and Cover [1] refers to stochastic random processes {Xi }.
A key concept is the entropy rate of the stochastic process, given by
1
H(X1 , ..., Xn ),
n
when this limit exists, with H denoting the usual multivariate entropy of X, H(X) =
−EX [log(P (X)]. It is an important theorem that for stationary processes,
(2)
(3)
H(X) = lim
n→∞
H(X) = lim H(Xn |Xn−1 , Xn−2 ..., X1 ).
n→∞
Let also
H0 (p) = −p log p − (1 − p) log(1 − p)
denote the univariate entropy, with p the probability of a Bernoulli (binary) process
(Markov chain3 of order zero).
We note that entropy rate of a stochastic processes is non-increasing as a function of
order, that is, 0 ≤ H ≤ .. ≤ Hq ≤ ... ≤ H0 ≤ 1.
The fundamental relation is that description length is closely related to entropy rate,
(4)
lLZW = c(n) log2 [c(n) + log2 A] ≈ c(n) log 2 [c(n)] −→ nH
Another important result in what follows is that with probability 1 (Thomas and Cover
Theorem 13.5.3)
lim sup lLZW ≤ nH
n→∞
2Can we generalize this idea? Can we, e.g., binarize the data so that it has maximal H and H ?
0
1
3We denote a Markov chain of order m to be one where the future state depends on the past m
states (time-translation invariantly), P (Xn |Xn−1 , ..., X1 ) = P (Xn |Xn−1 , ..., Xn−m ) for n > m.
4
G. RUFFINI - JAN 2017 — STARLAB TECHNICAL NOTE, TN00344 (V1.0)
which can rewrite as
lim sup c(n) log2 c(n) ≤ nH ≤ nH0
n→∞
and use to rewrite (in the limit above)
nH
nH0
nH
nH
≤
≤
∼
(5)
c(n) ≤
nH
log2 c(n)
log2 n
log2 n
log2 log c(n)
2
which we use below for normalization purposes.
4. Metrics
Two metrics are used in the field, one is c(n) and the other lLZW . Of the two the latter is
more closely related to Kolmogorov complexity or description length. Both contain similar
information (in fact one is a monotonic function of the other).
5. Fundamental Normalization of LZW
The purest way to normalize this metric is to normalize by the original string length n
c(n) log 2 [c(n) + A]
→H
n
with units of bits per character. This is the LZW compression ratio.
ρ0 = lLZW /n =
6. Other normalizations or measures
A typical normalization adopted by the literature is to “divide by entropy”. By this we
mean ρ1 = lLZW /H0 . In the literature this is usually defined through c(n),
ρ1 =
c(n)
nH0
log2 n
∼
H
lLZW
H
∼
→
H0
nH0
H0
(with units of bits per character). Essentially the same can be computed from the randomly
reshuffled data series, which with high probability forces lLZW ∼ nH0 by destroying 2nd
order interactions. Hence,
lLZW
lLZW
≈ shuf
ρ1 ≈
H0
l
LZW
This ratio tells us how much information density is hidden in 2nd and higher order entropy
rate as compared to first order one.
We can think of this a being the comparison of “first order apparent complexity” (entropy) and an estimate of the entropy rate (which provides and upper bound to algorithmic
complexity). This is an important comparison, as the proposal in [4] is that conscious level
may be associated to systems that exhibit high apparent entropy with low algorithmic
complexity.
Alternatively, we could define
ρ2 = H0 − ρ0 > 0
LEMPEL-ZIV COMPLEXITY REFERENCE
5
which can be interpreted as the extra apparent extra entropy (bits/char) incurred by using
first order methods instead estimating the true entropy rate.
At any rate, from a machine learning point of view it is probably best to compute ρ0 ,
H0 , ..., Hq as separate measures.
Also, it is known that LZW or entropy estimates are sensitive to string length. When
comparing metrics across datasets make sure you keep string length constant and as long
as possible.
7. LZW and Kolmogorov complexity
As mentioned above, LZW description length is only an estimate of algorithmic complexity. We can easily provide examples of sequences that have low algorithmic complexity,
but which are rather hard to compress using LZW. For example, consider the firs n digits of π. Or consider “the digits of the smallest prime with 109 digits”. The algorithmic
complexity of these numbers is very low, but LZW won’t compress them much if at all.
Does this mean that LZW is useless? No, but we should keep in mind that it provides only
an upper bound on algorithmic complexity. In order to get better bounds, we may
consider a family of compressors and use the lowest complexity that any of them can find
as a better upper bound.
References
[1] Thomas M. Cover and Joy A. Thomas. Elements of information theory. John Wiley & sons, 2 edition,
2006.
[2] F. Kaspar and H. G. Schuster. Easily calculable measure for the complexity of spatiotemporal patterns.
Phys. Rev. A, 36(2):842–848, 1987.
[3] A Lempel and J Ziv. On the complexity of finite sequences. IEEE Transactions on Information Theory,
IT-22(1):75–81, January 1976.
[4] G. Ruffini. An algorithmic information theory of consciousness. Submitted to the Neuroscience of Consciousness, August 2016.
[5] Giulio Ruffini, David Iba nez, Marta Castellano, Stephen Dunne, and Aureli Soria-Frisch. Eeg-driven
rnn classification for prognosis of neurodegeneration in at-risk patients. ICANN 2016, 2016.
6
G. RUFFINI - JAN 2017 — STARLAB TECHNICAL NOTE, TN00344 (V1.0)
Annex: StarLZW.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
#
#
#
#
#
LZW Update May 2017 , Feb 12 2017 to match TN00344 .
G . Ruffini with Eleni Kroupi / Starlab + N e u r o e l e c t r i c s (2017)
Updated May 2017 with Entropy rate c a l c u l a t o r s
Part of this code from sources as indicated , but modified to
be used with binary data .
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
P e r m i s s i o n is hereby granted , free of charge , to any person o b t a i n i n g a copy of
this software and a s s o c i a t e d d o c u m e n t a t i o n files ( the " Software ") , to deal in
the Software without restriction , i n c l u d i n g without l i m i t a t i o n the rights to
use , copy , modify , merge , publish , distribute , sublicense , and / or sell copies of
the Software , and to permit persons to whom the Software is f u r n i s h e d to do so ,
subject to the f o l l o w i n g c o n d i t i o n s:
The above c o p y r i g h t notice and this p e r m i s s i o n notice shall be included in all
copies or s u b s t a n t i a l portions of the Software .
THE SOFTWARE IS PROVIDED " AS IS " , WITHOUT WARRANTY OF ANY KIND , EXPRESS OR
IMPLIED , I N C L U D I N G BUT NOT LIMITED TO THE W A R R A N T I E S OF MERCHANTABILITY , FITNESS
FOR A P A R T I C U L A R PURPOSE AND N O N I N F R I N G E M E N T. IN NO EVENT SHALL THE AUTHORS OR
C O P Y R I G H T HOLDERS BE LIABLE FOR ANY CLAIM , DAMAGES OR OTHER LIABILITY , WHETHER
IN AN ACTION OF CONTRACT , TORT OR OTHERWISE , ARISING FROM , OUT OF OR IN
# C O N N E C T I O N WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE .
import numpy as np
# #########################################################################
def b i n a r i z e b y( data , method = " median " ):
""" Binarize 1 D numpy data array using d i f f e r e n t methods : mean or median """
d = data *0
if method == " median " :
thr = np . median ( data )
d [ np . where ( data >= thr )]=1
print " Using median as t h r e s h o l d"
print " The median is : " , str ( thr )
else : # use mean
thr = np . mean ( data )
d [ np . where ( data >= thr )]=1
print " Using mean as t h r e s h o l d"
print " The mean is : " , str ( thr )
print " B i n a r i z e d data average " , str ( np . mean ( d . astype ( np . int8 )))
return d . astype ( np . int8 )
# adapted from https :// r o s e t t a c o d e. org / wiki / L Z W _ c o m p r e s s i o n# Python
def compress ( theString , mode = ’ binary ’ , verbose = True ):
""" Compress a * string * to a list of output symbols . Starts from two sympols , 0 and 1.
Returns the c o m p r e s s e d string and the length of the d i c t i o n a r y
If you need to , convert first arrays to a string ,
e . g . , entry ="". join ([ np . str ( np . int ( x )) for x in theArray ]) """
if mode == ’ binary ’:
d i c t _ s i z e=2
d i c t i o n a r y={ ’0 ’ :0 , ’1 ’ :1}
elif mode == ’ ascii ’:
# Build the d i c t i o n a r y for generic ascii .
LEMPEL-ZIV COMPLEXITY REFERENCE
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
d i c t _ s i z e = 256
d i c t i o n a r y = dict (( chr ( i ) , i ) for i in xrange ( d i c t _ s i z e))
else :
print " u n r e c o g n i z e d mode , please use binary or ascii "
w = ""
result = []
for c in t h e S t r i n g:
wc = w + c
if wc in d i c t i o n a r y:
w = wc
else :
result . append ( d i c t i o n a r y[ w ])
# Add wc to the d i c t i o n a r y.
d i c t i o n a r y[ wc ] = d i c t _ s i z e
d i c t _ s i z e += 1
w = c
# Output the code for w .
if w :
result . append ( d i c t i o n a r y[ w ])
if verbose :
print " length of input string : " , len ( t h e S t r i n g)
print " length of d i c t i o n a r y: " , len ( d i c t i o n a r y)
print " length of result : " , len ( result )
return result , len ( d i c t i o n a r y)
def d e c o m p r e s s( compressed , mode = ’ binary ’ ):
""" D e c o m p r e s s a list of output ks to a string . """
from c S t r i n g I O import StringIO
if mode == ’ binary ’:
d i c t _ s i z e=2
d i c t i o n a r y ={0: ’0 ’ , 1: ’1 ’}
w = str ( c o m p r e s s e d. pop (0))
elif mode == ’ ascii ’:
# Build the d i c t i o n a r y for generic ascii .
d i c t _ s i z e = 256
d i c t i o n a r y = dict (( i , chr ( i )) for i in xrange ( d i c t _ s i z e))
w = chr ( c o m p r e s s e d. pop (0))
else :
print " u n r e c o g n i z e d mode , please use binary or ascii "
# use StringIO , o t h e r w i s e this becomes O ( N ^2)
# due to string c o n c a t e n a t i o n in a loop
result = StringIO ()
result . write ( w )
for k in c o m p r e s s e d:
if k in d i c t i o n a r y:
entry = d i c t i o n a r y[ k ]
elif k == d i c t _ s i z e:
entry = w + w [0]
else :
raise V a l u e E r r o r( ’ Bad c o m p r e s s e d k : % s ’ % k )
result . write ( entry )
7
8
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
G. RUFFINI - JAN 2017 — STARLAB TECHNICAL NOTE, TN00344 (V1.0)
# Add w + entry [0] to the d i c t i o n a r y.
d i c t i o n a r y[ d i c t _ s i z e] = w + entry [0]
d i c t _ s i z e += 1
w = entry
return result . getvalue ()
# #############################################################################
def C o m p u t e _ r h o 0( theArray ):
""" Computes rho0 metric ( bits / Sample ) as d e s c r i b e d in TN000344 Starlab / Luminous """
return
C o m p u t e D e s c r i p t i o n L e n g t h( theArray ) / len ( theArray )*1.0
def C o m p u t e _ r h o 1( theArray ):
""" Computes rho1 metric ( bits / Sample ) as d e s c r i b e d in TN000344 / Starlab Luminous """
return
C o m p u t e D e s c r i p t i o n L e n g t h( theArray ) / S h a n n o n E n t r o p y( theArray )
/ len ( theArray )
def C o m p u t e _ r h o 2( theArray ):
""" Computes rho2 metric ( bits / Sample ) as d e s c r i b e d in TN00044 """
return
C o m p u t e _ r h o 0( theArray ) - S h a n n o n E n t r o p y( theArray )
def C o m p u t e D e s c r i p t i o n L e n g t h( theArray , classic = False ):
""" Computes d e s c r i p t i o n lenght l_ { LZW } as d e s c r i b e d in TN000344 """
entry = " " . join ([ np . str ( np . int ( x )) for x in theArray ])
compressedstring , len_dict = compress ( entry )
ndigits = len ( np . unique ( theArray ))
print " distinct digits : " , ndigits
# returns
# you need n bits log ( max ( comp ))) to describe this sequence . And there are this many : len ( a )...
# DL : " you need n bits . There are these ( m ) of them ".
DL = np . log2 ( np . log2 ( max ( c o m p r e s s e d s t r i n g )))+ np . log2 ( max ( c o m p r e s s e d s t r i n g))* len ( c o m p r e s s e d s t r i n g)
if classic :
# old way ... more for LZ than LZW :
DL = len_dict * np . log2 ( len_dict )
return DL
def S h a n n o n E n t r o p y( labels ):
""" Computes entropy of label d i s t r i b u t i o n. numpy array int in """
n_labels = len ( labels )
if n_labels <= 1:
return 0
counts = np . bincount ( labels )
probs = counts *1.0 / n_labels *1.0
n _ c l a s s e s = np . c o u n t _ n o n z e r o( probs )
if n _ c l a s s e s <= 1:
LEMPEL-ZIV COMPLEXITY REFERENCE
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
return 0
# Compute standard entropy .
ent = 0.
for i in probs :
ent -= i * np . log2 ( i )
return ent
# #########################################################################
# Entropy rate of English ...
import re , codecs , random , math , textwrap
from c o l l e c t i o n s import defaultdict , deque , Counter
def C o m p u t e E R( theArray , m a r k o v _ o r d e r =4):
""" Compute entropy rate of array with some memomory / m a r k o v _ o r d e r """
# compute entropy rate ...
# np . savetxt (" rule . txt " , theArray . astype ( int ) , fmt = ’% i ’, d e l i m i t e r ="")
# model , stats = m a r k o v _ m o d e l( chars (" rule . txt ") , m a r k o v _ o r d e r)
d a t a s t r e a m= s t r e a m A r r a y( theArray . astype ( int ))
model , stats = m a r k o v _ m o d e l( datastream , m a r k o v _ o r d e r)
del d a t a s t r e a m
return e n t r o p y _ r a t e( model , stats )
def s t r e a m A r r a y( theArray ):
""" G e n e r a t o r for array elements ( c o n v e n i e n t) """
for element in theArray :
yield element
# The f o l l o w i n g 3 are adapted from :
# Clement Pit - - Claudel ( http :// pit - claudel . fr / clement / blog )
def m a r k o v _ m o d e l( stream , m o d e l _ o r d e r):
""" model d e s c r i b e s prob of a digit a p p e a r i n g f o l l o w i n g n prior ones """
""" Stats provide f r e q u e n c i e s of groups of n symbols """
model , stats = d e f a u l t d i c t( Counter ) , Counter ()
c i r c u l a r _ b u f f e r = deque ( maxlen = m o d e l _ o r d e r)
for token in stream :
prefix = tuple ( c i r c u l a r _ b u f f e r)
c i r c u l a r _ b u f f e r. append ( token )
if len ( prefix ) == m o d e l _ o r d e r:
stats [ prefix ] += 1.0
model [ prefix ][ token ] += 1.0
return model , stats
def entropy ( stats , n o r m a l i z a t i o n _ f a c t o r):
return - sum ( proba / n o r m a l i z a t i o n _ f a c t o r * math . log ( proba
/ n o r m a l i z a t i o n_ fa cto r , 2) for proba in stats . values ())
def e n t r o p y _ r a t e( model , stats ):
return sum ( stats [ prefix ] * entropy ( model [ prefix ] , stats [ prefix ])
for prefix in stats ) / sum ( stats . values ())
9
| 7cs.IT
|
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN
SUBGROUP OF INDEX p: III
arXiv:1708.08710v3 [math.GR] 14 Dec 2017
BOB OLIVER AND ALBERT RUIZ
Abstract. We finish the classification, begun in two earlier papers, of all simple fusion
systems over finite nonabelian p-groups with an abelian subgroup of index p. In particular,
this gives many new examples illustrating the enormous variety of exotic examples that can
arise. In addition, we classify all simple fusion systems over infinite nonabelian discrete
p-toral groups with an abelian subgroup of index p. In all of these cases (finite or infinite),
we reduce the problem to one of listing all Fp G-modules (for G finite) satisfying certain
conditions: a problem which was solved in the earlier paper [CrOS] using the classification
of finite simple groups.
A saturated fusion system over a finite p-group S is a category whose objects are the
subgroups of S, and whose morphisms are injective homomorphisms between the subgroups,
and which satisfy some additional conditions first formulated by Puig (who called them
“Frobenius S-categories” in [Pg]) and motivated in part by the Sylow theorems for finite
groups. For example, if G is a finite group and S ∈ Sylp (G), then the category FS (G),
whose objects are the subgroups of S and whose morphisms are the homomorphisms between
subgroups defined via conjugation in G, is a saturated fusion system over S. We refer to [Pg],
[AKO, Part I], or [Cr] for the basic definitions and properties of saturated fusion systems.
A saturated fusion system is realizable if it is isomorphic to FS (G) for some finite group
G and some S ∈ Sylp (G); it is exotic otherwise. Here, by an isomorphism of fusion systems
we mean an isomorphism of categories that is induced by an isomorphism between the
underlying p-groups. Exotic fusion systems over finite p-groups seem to be quite rare for
p = 2 (the only known examples are those constructed in [LO] and others easily derived
from them), but many examples of them are known for odd primes p.
A discrete p-torus is a group of the form (Z/p∞ )r for some r ≥ 0, where Z/p∞ is the union
of the cyclic groups Z/pk via the obvious inclusions Z/pk < Z/pk+1 . A discrete p-toral group
is a group containing a discrete p-torus as a normal subgroup of p-power index. Saturated
fusion systems over discrete p-toral groups were defined and studied in [BLO3], motivated
by the special case of fusion systems for compact Lie groups and p-compact groups.
A fusion system is simple if it is saturated and contains no proper nontrivial normal
fusion subsystems (see Definition 1.4). As a special case, very rich in exotic examples,
we have been looking at simple fusion systems F over finite nonabelian p-groups S with
an abelian subgroup A of index p. By [AOV2, Proposition 5.2(a)], if p = 2, then S is
dihedral, semidihedral, or a wreath product of the form C2k ≀ C2 , and hence F is isomorphic
to the fusion system of PSL2 (q) or PSL3 (q) for some odd q. Fusion systems over extraspecial
groups of order p3 and exponent p were listed in [RV], and by [Ol, Theorem 2.1], these include
the only simple fusion systems over nonabelian p-groups containing more than one abelian
2000 Mathematics Subject Classification. Primary 20D20. Secondary 20C20, 20D05, 20E45.
Key words and phrases. finite groups, fusion, finite simple groups, modular representations, Sylow
subgroups.
B. Oliver is partially supported by UMR 7539 of the CNRS.
A. Ruiz is partially supported by MICINN-FEDER project number MTM2016-80439-P.
1
2
BOB OLIVER AND ALBERT RUIZ
subgroup of index p. The other cases where p is odd and A is not essential (equivalently,
not radical) in F were handled in [Ol, Theorem 2.8], while those where A is essential and of
exponent p was handled in [CrOS]. So it remains to describe those cases where A is essential
and not elementary abelian (and the unique abelian subgroup of index p). This, together
with analogous results about simple fusion systems over infinite discrete p-toral groups with
abelian subgroup of index p, are the main results of this paper.
To simplify the following summary of our results, we use the term “index-p-triple” to
denote a triple (F , S, A), where S is a nonabelian discrete p-toral group (finite or infinite)
with abelian subgroup A of index p, and F is a simple fusion system over S. Our main results
are shown in Sections 4 and 5, where we handle separately the finite and infinite cases. In
each of these sections, we first list, in Theorems 4.5 and 5.11, all index-p-triples (F , S, A),
for S finite or infinite, in terms of the pair (G, A) where G = AutF (A) and A is regarded as a
Zp G-module. Theorem 4.5 is taken directly from [CrOS, Theorem 2.8], while Theorem 5.11
is new. For completeness in the infinite case, we also show that each index-2-triple (F , S, A)
with |S| = ∞ is isomorphic to that of SO(3) or PSU(3) (Theorem 5.6), and that for each p
there is (up to isomorphism) a unique index-p-triple (F , S, A) where |S| = ∞ and A is not
essential (Theorem 5.12).
The main theorems, Theorems A and B, appear at the ends of Sections 4 and 5, respectively. In Theorem A, for p odd, we prove that each index-p-triple (F , S, A), where A is finite,
essential in F , and not elementary abelian, is determined by G = AutF (A), V = Ω1 (A) regarded as an Fp G-module, the exponent of A, and some additional information needed when
A is not homocyclic. In all cases, rk(A) ≥ p − 1, and A is homocyclic whenever rk(A) ≥ p.
Also, A is always isomorphic to some quotient of a Zp G-lattice.
Theorem B can be thought of as a “limiting case” of the classification in Theorem A. It
says that each index-p-triple (F , S, A) such that A is infinite and essential in F is determined
by the pair (G, V ), where G = AutF (A), and V = Ω1 (A) is regarded as an Fp G-module. In
all such cases, A is a discrete p-torus of rank at least p − 1. We also determine which of
the fusion systems we list are realized as fusion systems of compact Lie groups or p-compact
groups.
Theorems A and B reduce our classification problems to questions about Fp G-modules with
certain properties. These questions were already studied in [CrOS], using the classification of
finite simple groups, and the results in that paper that are relevant in this one are summarized
in Section 6. Theorems A and B together with Proposition 6.1 and Table 6.1 allow us to
completely list all simple fusion systems over nonabelian discrete p-toral groups (finite or
infinite) with abelian subgroup of index p that is not elementary abelian. In particular, as
in the earlier papers [Ol] and [CrOS], we find a very large, very rich variety of exotic fusion
systems over finite p-groups (at least for p ≥ 5).
This work was motivated in part by the following questions and problems, all of which are
familiar to people working in this field.
Q1: For a fixed odd prime p, a complete classification of all simple fusion systems over finite
p-groups, or even a conjecture as to how they could be classified, seems way out of reach
for now. But based on the many examples already known, is there any meaningful
way in which one could begin to systematize them; for example, by splitting up the
problem into simpler cases? Alternatively, is there a class of simple fusion systems
over finite p-groups, much less restrictive than the one we look at here, for which there
might be some chance of classifying its members?
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
3
Q2: Find some criterion which can be used to prove that some (or at least one!) of the
examples constructed here or earlier (over finite p-groups for odd primes p) are exotic,
without invoking the classification of finite simple groups.
Q3: A torsion linear group in defining characteristic q is a subgroup Γ ≤ GLn (K), for some
n ≥ 1 and some field K of characteristic q, such that all elements of Γ have finite order.
If p is a prime and Γ is a torsion linear group in defining characteristic different from
p, then by [BLO3, § 8], there is a maximal discrete p-toral subgroup S ≤ Γ, unique
up to conjugation, and FS (Γ) is a saturated fusion system. Are there any saturated
fusion systems over discrete p-toral groups (for any prime p) which we can prove are
not fusion systems of torsion linear groups?
The notation used in this paper is mostly standard. We let A ◦ B denote a central product
of A and B. When g and h are in a group G, we set g h = ghg −1 and hg = g −1 hg. When
A is an abelian group and β ∈ Aut(A), we write [β, A] = hβ(x)x−1 | x ∈ Ai. When P is a
p-group, we let Fr(P ) denote its Frattini subgroup, and for k ≥ 1 set
k
Ωk (P ) = hg ∈ P | g p = 1i
and
k
℧k (P ) = hg p | g ∈ P i.
We would like to thank the Centre for Symmetry and Deformation at Copenhagen University, and the Universitat Autònoma de Barcelona, for their hospitality in allowing us to
get together on several different occasions.
1. Background
We first recall some of the definitions and standard terminology used when working with
fusion systems. Recall that a discrete p-toral group is a group that contains a normal
subgroup of p-power index isomorphic to (Z/p∞ )r for some r ≥ 0. A fusion system over a
discrete p-toral group S is a category F whose objects are the subgroups of S, and where for
each P, Q ≤ S, the set HomF (P, Q) is a set of injective homomorphisms from P to Q that
includes all those induced by conjugation in S, and such that for each ϕ ∈ HomF (P, Q), we
have ϕ ∈ HomF (P, ϕ(P )) and ϕ−1 ∈ HomF (ϕ(P ), P ).
Define the rank rk(S) of a discrete p-torus S by setting rk(S) = r if S ∼
= (Z/p∞ )r . If
S is a discrete p-toral group with normal discrete p-torus S0 E S of p-power
index, then
we refer to S0 as the identity component of S, and set |S| = rk(S0 ), |S/S0 | , where such
pairs are ordered lexicographically. Thus if T is another discrete p-toral group with identity
component T0 , then |S| ≤ |T | if rk(S0 ) < rk(T0 ), or if rk(S0 ) = rk(T0 ) and |S/S0 | ≤ |T /T0 |.
Note that the identity component of S, and hence |S|, are uniquely determined since a
discrete p-torus has no proper subgroups of finite index.
Definition 1.1. Fix a prime p, a discrete p-toral group S, and a fusion system F over S.
• For each P ≤ S and each g ∈ S, P F denotes the set of subgroups of S which are F conjugate (isomorphic in F ) to P , and g F denotes the F -conjugacy class of g (the
set of images of g under morphisms in F ).
• A subgroup P ≤ S is fully normalized in F ( fully centralized in F ) if |NS (P )| ≥ |NS (Q)|
(|CS (P )| ≤ |CS (Q)|) for each Q ∈ P F .
def
• A subgroup P ≤ S is fully automized in F if OutF (P ) = AutF (P )/Inn(P ) is finite and
OutS (P ) ∈ Sylp (OutF (P )). The subgroup P is receptive in F if for each Q ∈ P F
and each ϕ ∈ IsoF (Q, P ), there is ϕ ∈ HomF (Nϕ , S) such that ϕ|P = ϕ, where
Nϕ = g ∈ NS (Q) ϕcg ϕ−1 ∈ AutS (P ) .
4
BOB OLIVER AND ALBERT RUIZ
• The fusion system F is saturated if
– (Sylow axiom) each fully normalized subgroup of S is fully automized and fully centralized;
– (extension axiom) each fully centralized subgroup of S is receptive; and
– (continuity axiom, when |S| = ∞)Sif P1 ≤ P2 ≤ P3 ≤ · · · is an increasing sequence
of subgroups of S with P = ∞
i=1 Pi , and ϕ ∈ Hom(P, S) is such that ϕ|Pi ∈
HomF (Pi , S) for each i ≥ 1, then ϕ ∈ HomF (P, S).
The above definition of a saturated fusion system is the one given in [BLO2] and [BLO3,
Definition 2.2]. It will not be used directly in this paper (saturation of the fusion systems we
construct will be shown using later theorems), but we will frequently refer to the extension
axiom as a property of saturated fusion systems.
We now need some additional definitions, to describe certain subgroups in a saturated
fusion system.
Definition 1.2. Fix a prime p, a discrete p-toral group S, and a saturated fusion system F
over S. Let P ≤ S be any subgroup. Note that by Definition 1.1, OutF (P ) is finite whether
or not P is fully normalized (see also [BLO3, Proposition 2.3]).
• P is F -centric if CS (Q) = Z(Q) for each Q ∈ P F , and is F -radical if Op (OutF (P )) = 1.
• P is F -essential if P < S, P is F -centric and fully normalized in F , and OutF (P )
contains a strongly p-embedded subgroup. Here, a proper subgroup H < G of a finite
group G is strongly p-embedded if p |H|, and p∤|H ∩ gHg −1| for each g ∈ GrH. Let
EF denote the set of all F -essential subgroups of S.
• P is normal in F (P E F ) if each morphism ϕ ∈ HomF (Q, R) in F extends to a morphism ϕ ∈ HomF (P Q, P R) such that ϕ(P ) = P . The maximal normal p-subgroup of
a saturated fusion system F is denoted Op (F ).
• P is strongly closed in F if for each g ∈ P , g F ⊆ P .
Proposition 1.3. Let F be a saturated fusion system over a discrete p-toral group S.
(a) Each morphism in F is a composite of restrictions of elements in AutF (P ) for P ≤ S
that is fully normalized in F , F -centric and F -radical.
(b) Each morphism in F is a composite of restrictions of elements in AutF (P ) for P ∈
EF ∪ {S}.
(c) For each Q E S, Q E F if and only if for each P ∈ EF ∪ {S}, Q ≤ P and Q is
AutF (P )-invariant.
Proof. Point (a) is shown in [BLO3, Theorem 3.6].
By [BLO3, Proposition 2.3], OutF (P ) = AutF (P )/Inn(P ) is always finite. For each such
P < S that is not F -essential, AutF (P ) is generated by automorphisms that can be extended
to strictly larger subgroups: this is shown in [AKO, Proposition I.3.3] in the finite case, and
the same argument applies when S is infinite. Point (b) now follows from (a) and induction,
and (in the infinite case) since there are only finitely many S-conjugacy classes of subgroups
of S that are F -centric and F -radical [BLO3, Corollary 3.5].
Point (c) follows easily from (b), just as in the finite case [AKO, Proposition I.4.5].
Definition 1.4. Let F be a saturated fusion system over a discrete p-toral group S. A
saturated fusion subsystem E over T ≤ S is normal in F (E E F ) if
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
5
• T is strongly closed in F (in particular, T E S);
• (invariance condition) each α ∈ AutF (T ) is fusion preserving in the sense that it extends
to an automorphism of E;
• (Frattini condition) for each P ≤ T and each ϕ ∈ HomF (P, T ), there are α ∈ AutF (T )
and ϕ0 ∈ HomE (P, T ) such that ϕ = α ◦ ϕ0 ; and
• (extension condition) each α ∈ AutE (T ) extends to some α ∈ AutF (T CS (T )) such that
[α, CS (T )] ≤ Z(T ).
The fusion system F is simple if it contains no proper nontrivial normal subsystems.
For further discussion of the definition and properties of normal fusion subsystems, we refer
to [AKO, § I.6] or [Cr, §§ 5.4 & 8.1] (when S and T are finite) and to [Gon, Definition 2.8]
(in the general case). Note in particular the different definition used in [Cr] and in [Gon]: a
saturated fusion subsystem E ≤ F over a subgroup T that is strongly closed in F is normal if
the extension condition holds, and also the strong invariance condition: for each P ≤ Q ≤ T ,
and each ϕ ∈ HomE (P, Q) and ψ ∈ HomF (Q, T ), ψ ◦ ϕ ◦ (ψ|P )−1 ∈ HomE (ψ(P ), T ). When S
is finite, this is equivalent to the above definition by [AKO, Proposition I.6.4], and a similar
argument (made more complicated because there can be infinitely many subgroups) applies
when S and T are p-toral.
Since the Frattini condition will be important in Section 5, we work here with the above
definition. However, none of the examples over infinite discrete p-toral groups considered
here contains a nontrivial proper strongly closed subgroup (see Lemma 5.8), so these details
make no difference as to which of them are simple or not.
Proposition 1.5. Fix a prime p, and let F be a saturated fusion system over an infinite
discrete p-toral group S. Let S0 be the identity component of S, and assume that each element
of S is F -conjugate to an element of S0 .
(a) If F is realized by a compact Lie group G with identity connected component G0 , then
G/G0 has order prime to p. If in addition, F is simple, then F is realized by the
connected, simple group G0 /Z(G0 ), where Z(G0 ) is finite of order prime to p.
(b) If F is realized by a p-compact group X, then X is connected. If in addition, F is
simple, then so is X.
In either case, if F is simple, then the action of the Weyl group AutF (S0 ) on the Qp -vector
space Q ⊗Z Hom(S0 , Qp /Zp ) is irreducible and generated by pseudoreflections.
Proof. If F = FS (G) where G is a compact Lie group with identity connected component G0
and maximal discrete p-toral subgroup S, then S ∩ G0 is strongly closed in F since G0 E G,
S0 ≤ S ∩ G0 , and so S ≤ G0 . Hence G/G0 has order prime to p. Also, FS (G0 ) E FS (G):
the invariance and extension conditions are easily checked, and the Frattini condition holds
since G = G0 NG (S) by the Frattini argument.
If in addition, F is simple, then FS (G0 ) = FS (G) = F , and Z(G0 ) is finite of order prime
to p since Z(F ) = 1. Hence F is also realized by G0 /Z(G0 ), which is simple.
If F ∼
= FS (X) for some p-compact group X with S ∈ Sylp (X), then X is connected by
[GLR, Proposition 4.9(a)]. Then X is a central product of connected, simple p-compact
groups, and hence is simple if F is simple.
Whenever F is realized by a connected p-compact group X (possibly a compact connected
Lie group), then by [DW, Theorem 9.7(ii)], the action of the Weyl group AutF (S0 ) on
6
BOB OLIVER AND ALBERT RUIZ
Q ⊗Z H 2 (BS0 ; Zp ) is generated by pseudoreflections, where
H 2 (BS0 ; Zp ) ∼
= Hom(S0 , Qp /Zp ).
= H 1 (S0 ; Qp /Zp ) ∼
= H 2 (S0 ; Zp ) ∼
If this is not irreducible as a group generated by pseudoreflections, then by the classification
of connected p-compact groups in [AGMV, Theorem 1.2] (for p odd) and in [AG, Theorem
1.1] or [Mø, Corollary 1.2] (for p = 2), X must be a nontrivial central product of simple
factors, and hence F is not simple.
We also recall the definition of a reduced fusion system, but only for fusion systems over
′
finite p-groups. Recall [AKO, § I.7] that in this setting, O p (F ) and O p (F ) are the smallest
(normal) fusion systems in F of p-power index and of index prime to p, respectively.
Definition 1.6. A saturated fusion system F over a finite p-group S is reduced if Op (F ) = 1,
′
and O p (F ) = F = O p (F ).
For each saturated fusion system F over a finite p-group S, FOp (F ) (Op (F )), O p (F ), and
′
O p (F ) are all normal fusion subsystems. Hence F is reduced if it is simple. Conversely, if
E E F is any normal subsystem over the subgroup T E S, then by definition of normality, T
is strongly closed in F . Since each normal fusion subsystem over S itself has index prime to
p, a reduced fusion system is simple if it has no proper nontrivial strongly closed subgroups.
When F is a saturated fusion system over an infinite discrete p-toral group S, there are
′
well defined normal subsystems O p (F ) (see [Gon, Appendix B]), and O p (F ) (see [GLR,
A.14–A.16]), with the same properties as in the finite case. So we could define reduced
fusion systems in this context just as in the finite case. However, to simplify the discussion,
and because we don’t know whether or not infinite reduced fusion systems have the same
properties that motivated the definition in the finite case (see [AOV1, Theorems A & B]),
we restrict attention to simple fusion systems in the infinite setting.
2. Reduced or simple fusion systems over nonabelian discrete p-toral
groups with index p abelian subgroup
In this section, p is an arbitrary prime. We want to study simple fusion systems over
nonabelian discrete p-toral groups (possibly finite) which contain an abelian subgroup of
index p. Most of the results here were shown in [CrOS], but only in the case where |A| < ∞
and p is odd.
We first fix some notation which will be used throughout the rest of the paper. As
usual, for a group S, we define Zm (S) for all m ≥ 1 by setting Z1 (S) = Z(S), and setting
Zm (S)/Zm−1 (S) = Z(S/Zm−1 (S)) for m ≥ 2.
Notation 2.1. Fix a nonabelian discrete p-toral group S with a unique abelian subgroup A
of index p, and a saturated fusion system F over S. Define
S ′ = [S, S] = [S, A] ,
Z = Z(S) = CA (S) ,
Z0 = Z ∩ S ′ ,
Z2 = Z2 (S) .
′
Thus Z0 ≤ Z ≤ Z2 and Z0 ≤ S ≤ A. Also, set
H = Zhxi x ∈ SrA
B = Z2 hxi x ∈ SrA
G = AutF (A)
U = AutS (A) ∈ Sylp (G) .
Recall that by [Ol, Theorem 2.1], if p is odd, and S is finite and nonabelian and has
more than one abelian subgroup of index p, then either S is extraspecial of order p3 (and
the reduced fusion systems over S were described in [RV]), or there are no reduced fusion
systems over S. So in the finite case, the restriction about the uniqueness of A is just a
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
7
convenient way to remove certain cases that have already been handled. We will show later
(Corollary 5.2) that in the infinite case (also when p = 2), A is unique whenever Op (F ) = 1.
Lemma 2.2. Assume Notation 2.1. Then EF ⊆ {A} ∪ H ∪ B, and |NS (P )/P | = p for each
P ∈ EF . If EF 6⊆ {A}, then Z2 ≤ A and |Z2 /Z| = p.
Proof. Fix some P ∈ EF r {A}. Then P A since P is F -centric. Set P0 = P ∩ A, and fix
some element x ∈ P rP0 . Since OutF (P ) is finite (Definition 1.1) and contains a strongly
p-embedded subgroup, we have that Op (OutF (P )) = 1 (cf. [AKO, Proposition A.7(c)]).
We must show that P ∈ H ∪ B, |NS (P )/P | = p, Z2 ≤ A, and |Z2 /Z| = p. (Clearly,
|NS (P )/P | = p if P = A.)
Case 1: Assume P is nonabelian. Since Z ≤ P (P is F -centric), Z(P ) = CP0 (x) = Z. For
each g ∈ NA (P )rP , cg is the identity on P0 and on P/P0 . If P0 is characteristic in P , then
cg ∈ Op (AutF (P )) by Lemma A.1, which is impossible since Op (OutF (P )) = 1. Thus P0 is
not characteristic in P , and hence is not the unique abelian subgroup of index p in P . So by
Lemma A.3, |P0 /Z| = p and |[P, P ]| = |[x, P0 ]| = p. Also, P/Z is abelian since |P/Z| = p2 ,
so [x, P0 ] ≤ Z, and hence P0 ≤ Z2 . Note that P0 > Z, since P is nonabelian.
If P0 < Z2 , then for y ∈ Z2 r P , [y, P ] ≤ Z = Z(P ), so y ∈ NS (P ) r P and cy ∈
Op (Aut(P )), contradicting the assumption that P is F -essential. Thus P0 = Z2 , so P ∈ B,
Z2 ≤ A, and |Z2 /Z| = p. Finally, |NS (P )/P | = p by Lemma A.6, applied with OutF (P ) in
the role of G, OutS (P ) ∼
= NS (P )/P in the role of S, and P/Z0 (if P ∈ B) in the role of A.
Note that [P, P ] = [x, Z2 ] ≤ Z0 and CP/Z0 (NS (P )) = Z2 /Z0.
Case 2: If P ∈ EF is abelian, then P0 = Z: it contains Z since P is centric, and cannot
be larger since then P would be nonabelian. Hence P = Zhxi ∈ H. Also, AutA (P ) =
AutS (P ) ∈ Sylp (AutF (P )) centralizes P0 . The conditions of Lemma A.6 thus hold (with
P and AutF (P ) in the roles of A and G), so |NS (P )/P | = |AutS (P )| = p. Since [S:P ] =
|A/Z| > p by Lemma A.3 and since A is the unique abelian subgroup of index p, this implies
that S/Z is nonabelian, so [x, A] Z, and Z2 ≤ A.
For each g ∈ A, g ∈ NS (P ) if and only if [g, x] ∈ P0 = Z, if and only if gZ ∈ CA/Z (x) =
Z(S/Z) = Z2 /Z. Thus NA (P ) = Z2 , NS (P ) = Z2 hxi = Z2 P , and |Z2 /Z| = |NS (P )/P | =
p.
Lemma 2.3. Let S, F , etc. be as in Notation 2.1. If, for some x ∈ S r A, Z2 hxi ∈ EF ,
then Zhxi is not F -centric, and hence Zhxi ∈
/ EF .
Proof. Assume x ∈ S r A and Z2 hxi ∈ EF , and set P = Z2 hxi ∈ B. In particular, Z2 < A
since Z2 hxi < S. Also, |Z2 /Z| = p by Lemma 2.2, so |P/Z| = p2 , and Z2 is not normalized
by AutF (P ) since P is essential.
Let P be the set of all subgroups of index p in P which contain Z. Then P % {Z2 },
so P/Z ∼
= Cp2 (i.e., is not cyclic), and AutS (P ) permutes transitively the p members of
P r {Z2 }. So AutF (P ) must act transitively on P, hence Zhxi is F -conjugate to Z2 , and is
not F -centric (recall Z2 < A). So Zhxi ∈
/ EF in this case.
Lemma 2.4. Assume Notation 2.1, and also A 5 F ( ⇐⇒ EF 6⊆ {A}). Then |Z0 | = p,
and |Zi (S)/Zi−1(S)| = p for all i > 1 such that Zi (S) < S. If |A| < ∞, then |A/ZS ′| = p.
Proof. Fix x ∈ SrA. Let ψ ∈ End(A) be the homomorphism ψ(g) = [g, x]. Thus Ker(ψ) =
Z and Im(ψ) = S ′ .
By Lemma 2.2 and since EF 6⊆ {A}, (H ∪ B) ∩ EF 6= ∅, Z2 ≤ A, and |Z2 /Z| = p. Since
Z2 /Z = CA/Z (x), Z2 = ψ −1 (Z), and so ψ sends Z2 onto Z0 = Z ∩ S ′ with kernel Z. Thus
|Z0 | = |Z2 /Z| = p.
8
BOB OLIVER AND ALBERT RUIZ
Set Zi = Zi (S) for each i ≥ 0, and let k > 2 be the smallest index such that Zk = S. Thus
S/Zk−2 is nonabelian, so Zk−2 ≤ A, and Zk−1 /Zk−2 = Z(S/Zk−2) ≤ A/Zk−2 . Hence Zi ≤ A,
and Zi = ψ −1 (Zi−1 ), for all i < k. In particular, ψ induces a monomorphism from Zi /Zi−1
into Zi−1 /Zi−2 for each 3 ≤ i < k, so |Zi /Zi−1 | ≤ p, with equality since Zi (S) > Zi−1 (S)
whenever Zi−1 (S) < S.
If |A| < ∞, then |ZS ′ | = |Z|·|S ′| |Z0| = |A|/|Z0|, and hence ZS ′ has index p in A.
Lemma 2.5. Let A E S, F , H, B, etc., be as in Notation 2.1. Assume P ∈ EF where
P ∈ H ∪ B, and set X = 1 if P ∈ H and X = Z0 if P ∈ B. Define P1 , P2 ≤ P by setting
′
′
P1 /X = CP/X (O p (AutF (P )))
and
P2 = [O p (AutF (P )), P ].
′
Then O p (OutF (P )) ∼
= SL2 (p), and the following hold.
(a) If P ∈ H, then P1 < Z, Z = P1 × Z0 , Z0 < P2 ∼
= Cp2 , and P = P1 × P2 . If p is odd,
then P1 is the unique AutF (Z)-invariant subgroup of Z such that Z = P1 × Z0 .
(b) If P ∈ B, then P1 = Z, P2 is extraspecial of order p3 , P2 ∼
= Q8 if p = 2 while P2 has
exponent p if p is odd, and P1 ∩ P2 = Z(P2 ) = Z0 = [P, P ]. Thus P = P1 ×Z0 P2 .
′
Proof. To simplify notation, set H = OutF (P ), H0 = O p (H), and T = OutS (P ) ∈ Sylp (H).
If P ∈ B, then [P, P ] ≤ Z(P ) ∩ S ′ = Z ∩ S ′ = Z0 , with equality since |Z0 | = p by Lemma
2.4. Thus P/X is abelian in both cases. Also, [NS (P ), P/X] = Z0 (if P ∈ H) or Z2 /Z0 (if
P ∈ B), and thus has order p in both cases. So by Proposition A.7, applied to the H-action
on P/X, we have H0 ∼
= Cp2 .
= SL2 (p), P/X = (P1 /X) × (P2 /X), and P2 /X ∼
If P ∈ H (so X = 1), then P1 = CP (H0 ) ≤ CP (T ) = Z, and [Z:P1 ] = p since [P :P1] = p2 .
Also, P2 ≥ [T, P ] = Z0 , so P1 ∩ Z0 = 1, and Z = P1 × Z0 . If p is odd, then NH0 (T ) is
a semidirect product of the form Cp ⋊ Cp−1 . Fix α ∈ NH0 (T ) of order p − 1; then α acts
on Z0 = [T, P ] with order p − 1 and acts trivially on P1 . Thus α|Z ∈ AutF (Z), and P1 is
the only subgroup which is a complement to Z0 in Z and could be normalized by AutF (Z).
Since AutF (Z) has order prime to p, there is at least one such subgroup, and hence P1 is
AutF (Z)-invariant.
If P ∈ B, then X = Z0 , and P2 /Z0 ∼
= SL2 (p) acts faithfully on P2 , and
= Cp2 . Also, H0 ∼
this is possible only if Z(P2 ) = Z0 , and P2 ∼
= Q8 (if p = 2) or P2 is extraspecial of exponent p
(if p is odd). Also, P1 has index p2 in P since P1 ∩ P2 = Z0 , P1 ≤ Z(P ) since P = P1 ×Z0 P2
is a central product, and hence P1 = Z(P ) = Z.
Corollary 2.6. In the situation of Notation 2.1, if A 5 F (i.e., if EF 6⊆ {A}) and p is odd,
then S splits over A: there is x ∈ S r A of order p.
Proof. Fix P ∈ EF r {A}. By Lemma 2.2, P ∈ H ∪ B. In either case, by Lemma 2.5, there
is x ∈ P r A of order p.
We now restrict to the case where p is odd. Recall that G = AutF (A) by Notation 2.1.
Lemma 2.7. Assume Notation 2.1, and also that p is odd and A 5 F . Then Op (F ) = 1 if
and only if either there are no nontrivial G-invariant subgroups of Z, or EF ∩ H =
6 ∅ and
Z0 is the only G-invariant subgroup of Z.
Proof. The following proof is essentially the same as the proof in [CrOS, Lemma 2.7(a)] in
the finite case.
def
Assume first that Q = Op (F ) 6= 1. Since A 5 F , there is P ∈ EF r {A} ⊆ B ∪ H.
If P ∈ H, then Q ≤ Z: the intersection of the subgroups S-conjugate to P . If P ∈ B,
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
9
then Q ≤ Z2 by a similar argument, and then Q ≤ Z since that is the intersection of the
subgroups in the AutF (P )-orbit of Z2 . Thus Q is a non-trivial G-invariant subgroup of Z.
If Q = Z0 , then EF ∩ H = ∅, since for P ∈ EF ∩ H, Z0 is not normalized by AutF (P ). This
proves one implication.
Conversely, assume that 1 6= R ≤ Z is G-invariant. For each α ∈ AutF (S), α(A) = A
since A is the unique abelian subgroup of index p, so α|A ∈ G, and thus α(R) = R. Since
each element of AutF (Z) extends to S by the extension axiom, R is also normalized by
AutF (Z). Also, for each P ∈ EF ∩ B, Z = Z(P ) is characteristic in P and so R is also
normalized by AutF (P ). In particular, if EF ∩ H = ∅, then R E F , and so Op (F ) 6= 1.
Now assume that EF ∩ H =
6 ∅, and also that R 6= Z0 . By [CrOS, Lemma 2.3(b)] (the
argument easily extends to the infinite case), there is a unique AutF (Z)-invariant factorizae Set R
e = R ∩ Z.
e If R ≥ Z0 , then R = R
e × Z0 . Otherwise, R ∩ Z0 = 1
tion Z = Z0 × Z.
(recall |Z0 | = p), and since R is AutF (Z)-invariant, the uniqueness of the splitting implies
e Since R 6= Z0 , we have R
e 6= 1 in either case.
that R ≤ Ze and hence R = R.
e ≤ R ≤ Z, so by the extension axiom, ϕ| e extends
For each ϕ ∈ AutF (A) = G, ϕ(R)
R
e = ϕ(R)
e = R
e since R
e is AutF (Z)-invariant. So by the
to some ϕ ∈ AutF (S), and ϕ(R)
e is normalized by AutF (P ) for each P ∈
same arguments as those applied above to R, R
e = Ze by [CrOS, Lemma
({S} ∪ EF ) r H. If P ∈ EF ∩ H, then for each α ∈ AutF (P ), α(Z)
2.3(b)], so α|Ze extends to an element of AutF (S) and hence of AutF (Z), and in particular,
e = R.
e Thus 1 6= R
e E F , and hence Op (F ) 6= 1.
α(R)
Without the assumption that p be odd in Lemma 2.7, the 2-fusion system F of PΣL2 (q 2 )
is a counterexample for each prime power q ≡ ±1 (mod 8). Here, PΣL2 (q 2 ) = PSL2 (q 2 )hθi
where θ acts on PSL2 (q 2 ) as a field automorphism of order 2 (and θ2 = 1). Then O2 (F ) = 1,
S∼
= C2m−1 × C2 , Z = Z(S) = Ω1 (A), and
= D2m × C2 for some m ≥ 4 depending on q, A ∼
G = AutS (A) acts trivially on Z (so that all subgroups of Z are G-invariant).
Lemma 2.8. Assume Notation 2.1, and also that p is odd and Op (F ) = 1. Let A2 ≤ A1 ≤ A
be G-invariant subgroups such that A1 ≤ ZA2 . Then either A1 = A2 , or A1 = Z0 × A2 and
Z0 is G-invariant.
Proof. Fix a class xA2 ∈ A1 /A2 . By assumption, we can assume x ∈ Z. Since G acts
′
on A1 /A2 and U acts trivially on this quotient, G0 = O p (G) also acts trivially. Hence
α(x) ∈ xA2 for each α ∈ G0 . Let α1 , . . . , αk ∈ G0 be left coset representatives for U (so
1/k
Q
k
α
(x)
p ∤ k = [G0 :U]), and set y =
. Then y ∈ xA2 since αi (x) ∈ xA2 for each i,
i
i=1
and y ∈ CA (G0 ). This shows that A1 ≤ CA (G0 )A2 .
Now, CA (G0 ) is a subgroup of Z = CA (U) normalized by G. So by Lemma 2.7 and since
Op (F ) = 1, CA (G0 ) ≤ Z0 . Thus A1 ≤ Z0 A2 . If A1 > A2 , then A1 = A2 × Z0 since |Z0 | = p,
and Z0 = CA (G0 ) is G-invariant.
The following notation, taken from [CrOS, Notation 2.4], will be used throughout the rest
of the paper.
Notation 2.9. Assume Notation 2.1, and also that |Z0| = p. Set
∆ = (Z/p)× × (Z/p)× ,
and
∆i = {(r, r i ) | r ∈ (Z/p)× } ≤ ∆
(for i ∈ Z).
10
BOB OLIVER AND ALBERT RUIZ
Set
Aut∨ (S) = α ∈ Aut(S) [α, Z] ≤ Z0 , Aut∨ (A) = α|A α ∈ Aut∨ (S)
Aut∨F (S) = Aut∨ (S) ∩ AutF (S)
Define
µ : Aut∨ (S) −−−−−−→ ∆
by setting, for α ∈ Aut∨ (S),
µ(α) = (r, s) if
Aut∨F (A) = Aut∨ (A) ∩ AutF (A)
= β ∈ NAutF (A) (AutS (A)) [β, Z] ≤ Z0 .
and
µA : Aut∨ (A) −−−−−−→ ∆
(
α(x) ∈ xr A for x ∈ S r A
α(g) = g s
for g ∈ Z0
and µA (α|A ) = µ(α) if α ∈ Aut∨F (S).
3. Minimally active modules
In the earlier paper [CrOS], the concept of “minimally active” modules played a central
role when identifying the pairs (A, AutF (A)) that can occur in a simple fusion system F over
a p-group S that contains an elementary abelian group A with index p. Before continuing
to study the structure of such F , we need to recall some of the notation and results in that
paper, beginning with [CrOS, Definitions 3.1 & 3.3], and describe how they relate to the
more general situation here.
Definition 3.1. For each prime p,
• Gp is the class of finite groups Γ with U ∈ Sylp (Γ ) such that |U| = p and U 5 Γ ; and
• Gp∧ is the class of those Γ ∈ Gp such that |OutΓ (U)| = p − 1 for U ∈ Sylp (Γ ).
For Γ ∈ Gp , an Fp Γ -module is minimally active if its restriction to U ∈ Sylp (Γ ) has exactly
one Jordan block with nontrivial action.
The next lemma explains the importance of minimally active modules here. In particular,
it means that many of the tables and results in [CrOS, § 4–5] can be applied to get information
about AutF (A) and Ω1 (A).
Lemma 3.2. Assume Notation 2.1 and 2.9, and also that p is odd, A ∈ EF , and Op (F ) = 1.
Set V = Ω1 (A) and
Aut∨F (V ) = β ∈ NAutF (V ) (AutS (V )) [β, Ω1 (Z)] ≤ Z0
= α|V α ∈ AutF (S), [α, Ω1 (Z)] ≤ Z0 ,
and define µV : Aut∨F (V ) −→ ∆ by setting µV (α|V ) = µ(α). Then
(a) G = AutF (A) ∈ Gp∧ ;
(b) V , and A/Fr(A) if |A| < ∞, are both faithful, minimally active, and indecomposable as
Fp G-modules; and
(
if Z0 Fr(A)
µV (Aut∨F (V ))
(c) µA (Aut∨F (A)) =
∨
µV (AutF (V )) ∩ ∆0 if Z0 ≤ Fr(A).
Proof. (a) By assumption, U = AutS (A) ∈ Sylp (G) has order p. Since A ∈ EF , U is not
normal in G = AutF (A), and hence G ∈ Gp .
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
11
′
Since Op (F ) = 1, there is P ∈ EF ∩(H∪B). By Lemma 2.5(a,b), O p (OutF (P )) ∼
= SL2 (p).
p′
Choose α ∈ O (AutF (P )) of order p − 1 whose class in OutF (P ) normalizes OutS (P ) ∼
= Cp ;
then α extends to an element of AutF (NS (P )) and hence (since P is maximal among F essential subgroups) to some α ∈ AutF (S). Then α|A normalizes U and its class in AutG (U)
has order p − 1, so G ∈ Gp∧ .
(b) Set A = A/Fr(A). If |A| < ∞, then since G acts faithfully on A, [G, Theorems 5.2.4
& 5.3.5] imply that CG (V ) and CG (A) are both normal p-subgroups of G. Since U is not
normal by (a), G acts faithfully on V and on A in this case. If |A| = ∞, then G acts faithfully
on Ωm (A) for m large enough, and hence acts faithfully on V by the above argument.
Since |Z0 | = p by Lemma 2.4, where Z0 = S ′ ∩ Z = [U, V ] ∩ CV (U), the Fp U-module
V |U has exactly one Jordan block with nontrivial action of U. So V is minimally active. If
|A| < ∞, then ZS ′ = CA (U)[U, A] has index p in A by Lemma 2.4, so CA (U)[U, A] has
index at most p in A, and hence A is minimally active.
If V = V1 × V2 , where each Vi is a nontrivial Fp G-submodule, then by [CrOS, Lemma
3.4(a)], we can assume (after exchanging indices if needed) that V1 ≤ Z and (since it is a
summand) V1 ∩ Z0 = 1. But this contradicts Lemma 2.7.
Assume |A| < ∞. If A = X × Y where X, Y ≤ A are Fp G-submodules, then by the
Krull-Schmidt theorem, one of the factors, say X, contains a nontrivial Jordan block, while
U acts trivially on the other factor. Thus X ≥ [U, A] and X ZS ′ /Fr(A). Let X ≤ A
be such that Fr(A) ≤ X and X/Fr(A) = X. Then X ≥ S ′ and X ZS ′ , so XZ = A.
By Lemma 2.8, either X = A (and X = A) or A = X × Z0 (which is impossible since
Z0 ≤ S ′ ≤ X). Thus X = A, and A is indecomposable.
(c) By definition, the restriction to V of each element in Aut∨F (A) lies in Aut∨F (V ), and hence
µA (Aut∨F (A)) ≤ µV (Aut∨F (V )). If Z0 ≤ Fr(Z), then choose z ∈ Z such that 1 6= z p ∈ Z0 .
For each β ∈ Aut∨F (A), β(z) = z pk+1 for some k since [β, Z] ≤ Z0 , and hence β|Z0 = Id and
µA (β) ∈ ∆0 . Thus µA (Aut∨F (A)) is contained in the right hand side in (c).
Now assume that β ∈ Aut∨F (V ), where β = α|V for α ∈ AutF (S). If Z0 ≤ Fr(A), then
k
assume also that µV (β) ≤ ∆0 ; i.e., that β|Z0 = Id. Upon replacing β by β p and α by
k
αp for appropriate k, we can also assume that α has order prime to p without changing
µ(α). Then Z = CZ (α) × [α, Z] by [G, Theorem 5.2.3] and since Z is the union of the
finite abelian p-groups Ωi (Z), and [α, Ω1 (Z)] = [β, Ω1 (Z)] ≤ Z0 since β ∈ Aut∨F (V ). Also,
Ω1 ([α, Z]) = [α, Ω1 (Z)] ≤ Z0 (since it can’t be any larger). If Z0 Fr(Z), then this implies
that [α, Z] ≤ Z0 , hence that α ∈ Aut∨F (S). If Z0 ≤ Fr(Z), then Ω1 ([α, Z]) ≤ Z0 ≤ CZ (α)
implies that Ω1 ([α, Z]) = 1 and hence [α, Z] = 1, so again α ∈ Aut∨F (S). Thus µV (β) =
µA (α|A ) ∈ µA (Aut∨F (A)), and the right hand side in (c) is contained in µA (Aut∨F (A)).
The following basic properties of minimally active indecomposable modules, taken from
[CrOS], play an important role in the rest of the paper.
Lemma 3.3 ([CrOS, Proposition 3.7]). Fix an odd prime p, a finite group Γ ∈ Gp , and
U ∈ Sylp (Γ ). Let V be a faithful, minimally active, indecomposable Fp Γ -module. Then
(a) dim(V ) ≤ p implies that V |U is indecomposable and thus contains a unique Jordan block;
(b) dim(V ) ≥ p + 1 implies that V |U is the direct sum of a Jordan block of dimension p and
a module with trivial action of U; and
(c) dim(CV (U)) = 1 if dim(V ) ≤ p, while dim(CV (U)) = dim(V ) − p + 1 if dim(V ) ≥ p.
12
BOB OLIVER AND ALBERT RUIZ
The next lemma is closely related to [CrOS, Lemma 1.11].
Lemma 3.4. Fix an odd prime p, let Γ be a finite group such that U ∈ Sylp (Γ ) has order
p, and set N = NΓ (U). Let V be a faithful, minimally active, indecomposable Fp Γ -module
such that dim(V ) ≤ p. Then CV (U) and V /[U, V ] are both 1-dimensional, and the following
hold.
(a) If dim(V ) = p, then V /[U, V ] and CV (U) are 1-dimensional, and isomorphic as Fp [N/U]modules.
(b) The projective cover and the injective envelope of V |N are both p-dimensional.
(c) If dim(V ) = p − 1, and there is an Fp Γ -submodule V0 < V with dim(V0 ) = 1, then
there is a projective Fp Γ -module W such that dim(W ) = p and W has a submodule
isomorphic to V .
(d) Let W be another Fp Γ -module such that dim(W ) = dim(V ), and assume that CW (U) ∼
=
V
as
F
[N
(U)]-modules,
and
as
F
Γ
CV (U) as Fp [NΓ (U)/U]-modules. Then W ∼
=
p
Γ
p modules if dim(V ) < p.
Proof. (a) By Lemma 3.3, U acts on V with only one Jordan block, so dim(CV (U)) = 1
and dim(V /[U, V ]) = 1. By [CrOS, Lemma 1.11(b)], if g ∈ NΓ (U) and t ∈ (Z/p)× are
such that g acts on V /[U, V ] via multiplication by t, then for some r ∈ (Z/p)× , g acts on
CV (U) via multiplication by tr m−1 = tr p−1 = t. Thus V /[U, V ] and CV (U) are isomorphic
as Fp [NΓ (U)/U]-modules.
(b) By the Schur-Zassenhaus theorem, there is H < N of index p such that N = HU. Set
V0 = CV (U), regarded as an Fp [N/U]-module, and also as an Fp H-module via the natural
isomorphism H ∼
= N/U. Set Vb = IndN
H (V0 ): a projective and injective p-dimensional Fp Nmodule. Then
Vb /[U, Vb ] ∼
= V0 ,
= Fp [N/U] ⊗F H V0 ∼
= Fp [N/U] ⊗F N Fp N ⊗F H V0 ∼
p
p
p
and so CVb (U) ∼
= V0 by (a). Thus Vb is the injective envelope of V0 when regarded as an
Fp N-module. In particular, an isomorphism CV (U) ∼
= CVb (U) extends to an Fp N-linear
homomorphism V −→ Vb which is injective since it sends the socle CV (U) injectively. Thus
Vb is an injective envelope of V |N . The statement about projective covers is shown in a
similar way (or by dualizing).
(c) Assume that dim(V ) = p − 1, and that there is an Fp Γ -submodule V0 < V with
dim(V0 ) = 1. Then V0 = CV (U), since this is the unique 1-dimensional submodule of V as
c containing
an Fp U-module. By (b), there is an injective (hence projective) Fp N-module W
c /V ∼
V |N as a submodule. By (a), W
= V0 |N as Fp N-modules.
Consider the homomorphisms
Φ
Φ
Ext1Fp Γ (V0 , V ) −−−−1−→ Ext1Fp N (V0 , V ) −−−−2−→ Ext1Fp U (V0 , V )
induced by restrictions of rings. Since U has index prime to p in Γ , Φ1 and Φ2 are injective,
and the images of Φ2 and Φ2 Φ1 are certain subgroups of stable elements (see [Ben, Proposition
3.8.2] for this version of the stable elements theorem). Since Ext1Fp (V0 , V ) = 0, we need only
consider stability of elements with respect to automorphisms of U, and hence Im(Φ2 ) =
Im(Φ2 Φ1 ).
Thus Φ1 is an isomorphism. Interpreted in terms of extensions, this implies that there
c. In
is an extension 0 −→ V −→ W −→ V0 −→ 0 of Fp Γ -modules such that W |N ∼
= W
c is projective as an Fp N-module.
particular, W is projective since W
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
13
(d) By (b), the injective envelope Vb of V |N is p-dimensional. Hence CVb (U) ∼
=
= CV (U) ∼
b
CW (U), so V is also the injective envelope of CW (U), and hence of W |N since CW (U) is its
socle. Since dim(V ) = dim(W ), and Vb contains a unique Fp N-submodule of each dimension
m ≤ p, we conclude that V ∼
= W are isomorphic as Fp N-modules.
If dim(V ) < p, then U is a vertex of V and of W and they are the Green correspondents
of V |N and W |N , respectively (see [Ben, § 3.12]). So V ∼
= W as Fp Γ -modules.
Point (d) need not hold if dim(V ) = p. As an example, fix p ≥ 5, set Γ = SL2 (p)
and choose U ∈ Sylp (Γ ), let V be the simple p-dimensional Fp Γ -module, and let W be
the projective cover of the trivial 1-dimensional Fp Γ -module. Using the fact that V is the
(p − 1)-st symmetric power of the natural 2-dimensional Fp Γ -module, it is not hard to
see that CV (U) is 1-dimensional with trivial NΓ (U)/U-action. The same holds for W by
construction, where dim(W ) = p. We refer to [Al, pp. 48–52] or the discussion in [CrOS,
§ 6] for more detail.
A minimally active indecomposable Fp Γ -module of dimension at least p + 2 is simple by
[CrOS, Proposition 3.7(c)]. This is not true for modules of dimension p + 1, but the following
lemma gives some information about such modules.
Lemma 3.5. Fix a finite group Γ ∈ Gp with U ∈ Sylp (Γ ). Let V be a finite, minimally
active, indecomposable Fp Γ -module of rank p + 1. If 0 6= V0 < V is a proper nontrivial
submodule, then V0 |U and (V /V0 )|U are both indecomposable Fp U-modules with nontrivial
action. In particular, 2 ≤ dim(V0 ) ≤ p − 1.
Proof. Recall |U| = p since Γ ∈ Gp . By [CrOS, Proposition 3.7(a)], V |U ∼
= Fp U ⊕ Fp ; i.e.,
V |U has Jordan blocks of dimension p and 1.
Fix a proper nontrivial submodule 0 6= V0 < V , and assume that V0 |U is decomposable or has trivial action as an Fp U-module. We will first show that V0 always has a
1-dimensional Fp Γ -submodule, and then show that this is impossible. In particular, this
shows that dim(V0 ) ≥ 2. The corresponding results for V /V0 then follow by dualizing.
If V0 |U is decomposable with nontrivial action, then V0 |U is the sum of a 1-dimensional
module with trivial action and an indecomposalble module of dimension at most p − 1.
By [CrOS, Proposition 3.7(a)], V0 is decomposable as an Fp Γ -module, and thus has a 1dimensional summand.
′
If U acts trivially on V0 , then dim(V0 ) ≤ 2 and O p (Γ ) (the normal closure of U in Γ ) acts
trivially on V0 . If dim(V0 ) = 2, then V0 = CV (U), and [U, V ] ∩ V0 is a 1-dimensional subspace
′
normalized by NΓ (U). Since Γ = O p (Γ )NΓ (U) by the Frattini argument, [U, V ] ∩ V0 is a
1-dimensional Fp Γ -submodule.
We are thus reduced to the case where dim(V0 ) = 1. If V0 6= [U, V ], then (V /V0 )|U is
indecomposable (consists of one Jordan block), and hence is Fp U-free. So V /V0 is projective,
contradicting the assumption that V is indecomposable.
Thus V0 = [U, V ] is an Fp Γ -submodule, and V /V0 has Jordan blocks of length 1 and p − 1.
So by [CrOS, Proposition 3.7(a)], it is decomposable: there are submodules W1 , W2 < V such
that V /V0 = (W1 /V0 )⊕(W2 /V0 ), where dim(W1 ) = p, and (W1 /V0 )|U is an (indecomposable)
Jordan block. If [U, W1 ] = 0, then [U, W2 ] = V0 , and V |U contains Jordan blocks of dimension
p−1 and 2, which we saw is impossible. Thus [U, W1 ] = V0 , so W1 |U is indecomposable, W1 is
projective and injective, and this again contradicts the assumption that V is indecomposable.
14
BOB OLIVER AND ALBERT RUIZ
This proves that U acts nontrivially on V0 , and in particular, dim(V0 ) ≥ 2. A similar
argument applied to the dual V ∗ shows that U acts nontrivially on V /V0, and that dim(V0 ) ≤
p − 1.
The following definitions will be useful.
Definition 3.6. For a finite group Γ , a Zp Γ -lattice is a finitely generated Zp Γ -module
that is free as a Zp -module (hence a lattice in a finitely generated Qp Γ -module). A discrete
Γ -p-torus is a discrete p-torus equipped with an action of Γ by automorphisms.
Let Qp (ζ) ⊇ Zp [ζ] denote the extensions of Qp ⊇ Zp by a primitive p-th root of unity ζ.
When U is a group of order p, we regard Qp (ζ) and Zp [ζ] as Zp U-modules under some choice
of identification U ∼
= hζi.
Lemma 3.7. Fix an odd prime p, a group Γ ∈ Gp , and U ∈ Sylp (Γ ).
(a) Let Λ be a Zp Γ -lattice such that Λ/pΛ is faithful and minimally active as an Fp Γ module. Then Λ/CΛ (U) ∼
= Zp [ζ] as Zp U-modules, and [U, Λ] + CΛ (U) has index p in
Λ.
(b) Let A be a discrete Γ -p-torus such that Ω1 (A) is faithful and minimally active as an
Fp Γ -module. Then A/CA (U) ∼
= Qp (ζ)/Zp [ζ] as Zp U-modules, and |[U, A] ∩ CA (U)| = p.
(c) Let X be a finite, faithful Zp Γ -module. Then Γ acts faithfully on Ω1 (X) and on X/pX.
Among the following conditions:
(1) Ω1 (X) is minimally active as an Fp Γ -module.
(2) X/pX is minimally active as an Fp Γ -module.
(3) |[U, X] ∩ CX (U)| = p.
(4) [U, X] + CX (U) has index p in X.
we have (1) ⇐= (3) ⇐⇒ (4) =⇒ (2). If X ∼
= Λ/Λ0 for some Zp Γ -lattice Λ and
some submodule Λ0 ≤ pΛ, then all four conditions are equivalent.
(d) If X is a finite, faithful Zp Γ -module such that condition (c.4) holds, then for each
x ∈ X r ([U, X] + CX (U)), X = CX (U) + Zp U·x.
Proof. Fix a generator u ∈ U.
(a) Set M = Qp ⊗Zp Λ. By Lemma A.5(a), M/CM (U) is isomorphic, as a Qp U-module,
to a sum of copies of Qp (ζ), where ζ is a primitive p-th root of unity. In particular, each
Jordan block for the action of U on Λ/(CΛ (U) + pΛ) has length at most p − 1. Since Λ/pΛ is
minimally active, it follows that M/CM (U) ∼
= Qp (ζ), since otherwise Λ/(CΛ (U) + pΛ) would
have rank at least 2(p − 1) and hence U would be fixed by a submodule of rank at least
p ≥ 3. Hence Λ/CΛ (U) ∼
= Zp [ζ] by Lemma A.5(c).
Consider the short exact sequence
incl
ϕ
0 −−−→ CΛ (U) −−−−−→ Λ −−−−−→ [U, Λ] −−−→ 0,
where ϕ(x) = u(x) − x for all x ∈ Λ. We just saw that [U, Λ] ∼
= Zp [ζ]. Under this identification, ϕ|[U,Λ] is multiplication by 1 − ζ, and so its image has index p in [U, Λ]. Thus
CΛ (U) + [U, Λ] has index p in Λ.
(b) Let A be a discrete Γ -p-torus such that Ω1 (A) is faithful and minimally active as
an Fp Γ -module. Set Λ = HomZp (A, Qp /Zp ), regarded as a Zp Γ -lattice. Then Ω1 (A) ∼
=
Λ/pΛ by Proposition A.4, so Λ/pΛ is minimally active. We just saw, in the proof of (a),
that this implies that Λ/CΛ (U) ∼
= Zp [ζ]. So after taking tensor products with Qp /Zp and
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
15
applying Proposition A.4 again, we get that A/(Qp /Zp ⊗Zp CΛ (U)) ∼
= Qp (ζ)/Zp [ζ]. Hence
∼
A/CA (U) = Qp (ζ)/R for some Zp U-lattice R < Qp (ζ) that contains Zp [ζ] with finite index.
Then R ∼
= Qp (ζ)/Zp [ζ].
= Zp [ζ] by Lemma A.5(c), Qp (ζ) = Qp ·R, and so A/CA (U) ∼
Consider the short exact sequence
incl
ϕ
0 −−−→ CA (U) −−−−−→ A −−−−−→ [U, A] −−−→ 0,
where ϕ(x) = u(x) − x for all x ∈ A. We just saw that [U, A] ∼
= Qp (ζ)/Zp [ζ]. Under
this identification, ϕ|[U,A] is multiplication by 1 − ζ, and so its kernel has order p. Thus
|CA (U) ∩ [U, A]| = p.
(c) We have |X| = |CX (U)|·|[U, X]|: X is finite, and [U, X] is the image of the homomor1−u
phism X −−−→ X while CX (U) is its kernel. Hence (3) and (4) are equivalent.
If (3) holds, then [U, Ω1 (X)] ∩ CΩ1 (X) (U) also has order p (since it cannot be trivial).
Since the rank of this intersection is the number of Jordan blocks in Ω1 (X) with nontrivial
U-action, we see that Ω1 (X) is minimally active in this case. So (3) implies (1); and a similar
argument shows that (4) implies (2).
Assume that Λ0 < Λ are Zp Γ -lattices such that Λ0 ≤ pΛ and Λ/Λ0 ∼
= X. In particular,
∼
Λ/pΛ = X/pX. So if X/pX is minimally active, then [U, Λ] + CΛ (U) has index p in Λ by (a),
and hence [U, X] + CX (U) has index p in X (since it cannot be all of X). Thus (2) implies
(4) in this case.
We continue to assume that X ∼
= Λ/Λ0, and set Rp = Qp /Zp for short. We have an exact
sequence
0 −−−→ TorZp (Rp , X) −−−−→ Rp ⊗Zp Λ0 −−−−→ Rp ⊗Zp Λ −−−−→ Rp ⊗Zp X −−−→ 0.
Also, by tensoring the short exact sequence 0 → Zp → Qp → Rp → 0 by X and using the
fact that Qp is flat over Zp , we see that Rp ⊗Zp X = 0, and TorZp (Rp , X) ∼
= X as Zp Γ modules. Thus X is isomorphic to a subgroup of the discrete Γ -p-torus A = Rp ⊗Zp Λ0 (see
Proposition A.4), where Ω1 (X) ∼
= Ω1 (A). With the help of (b), we now see that (1) implies
(3).
(d) Set X = X/pX for short. By (c), X is minimally active as an Fp Γ -module. Hence
there is y ∈ X such that X = CX (U) + Fp U·y. Thus [U, X] = (1 − u)Fp U·y. Choose y ∈ X
whose class modulo p is y; then [U, X] ≤ (1 − u)Zp U·y + pX.
By assumption, X = CX (U) + [U, X] + Zp ·x. So there are ξ ∈ Zp U and r ∈ Zp such
that y ∈ rx + (1 − u)ξ·y + CX (U) + pX. Then (1 − (1 − u)ξ)y ∈ rx + CX (U) + pX, where
1 − (1 − u)ξ is invertible in Zp U since (1 − u)p ∈ pZp U. Thus y ∈ Zp U·x + CX (U) + pX, and
hence X = Zp U·x + CX (U) + pX. Since pX is the Frattini subgroup of X, it now follows
that X = Zp U·x + CX (U).
In the rest of the section, we look at questions of existence and uniqueness of finite Zp Γ modules or discrete Γ -p-tori A for which Ω1 (A) is isomorphic to a given minimally active,
indecomposable Fp Γ -module.
Proposition 3.8. Fix an odd prime p and a group Γ ∈ Gp , and let V be a faithful, minimally
active, indecomposable Fp Γ -module.
(a) If dim(V ) ≥ p − 1, then there are a Qp Γ -module M and a Zp Γ -lattice Λ ≤ M such that
V ∼
= Λ/pΛ.
(b) If dim(V ) = p, and there is an Fp Γ -submodule V1 < V of dimension 1, then M and Λ
can be chosen as in (a) such that M contains a 1-dimensional Qp Γ -submodule.
16
BOB OLIVER AND ALBERT RUIZ
Proof. Fix U ∈ Sylp (Γ ), and let u ∈ U be a generator. Let ζ be a p-th root of unity, and
regard Qp (ζ) as a Qp U-module where u acts by multiplication by ζ. Thus Zp [ζ] is a Zp Ulattice in Qp (ζ). We also write Fp [ζ] = Zp [ζ]/pZp [ζ]. Thus Fp [ζ] ∼
= Fp [u]/h(1 − u)p−1i, and
V |U ∼
= Fp [ζ] since V is minimally active and indecomposable [CrOS, Proposition 3.7(a)].
(a) If dim(V ) ≥ p, then V is a trivial source module by [CrOS, Proposition 3.7(b)], and
hence V is the mod p reduction of some Zp Γ -lattice (see [Ben, Corollary 3.11.4.i]). So for
the rest of the proof, we assume that dim(V ) = p − 1.
b = IndΓ (Zp [ζ]) and Vb = IndΓ (V |U ). Then Vb ∼
b Λ.
b Since induction of repSet Λ
= Λ/p
U
U
resentations is adjoint to restriction, the identity on V extends to a surjective Fp Γ -linear
homomorphism α : Vb −→ V . This is split by an Fp U-linear map, and hence (by averaging
over cosets of U) by an Fp Γ -linear homomorphism β : V −→ Vb . Set e0 = βα ∈ EndFp Γ (Vb ).
Thus e0 is an idempotent in this endomorphism ring, and e0 Vb ∼
= V.
b By the Mackey double coset formula,
We want to lift e0 to an idempotent in EndZp Γ (Λ).
m
n
∼
b U = IndΓU (Zp [ζ])
Λ|
Z
[ζ]
Z
U
×
=
p
p
U
as Zp U-modules for some m, n ≥ 0. Hence as Zp -modules,
n
m
b ∼
b ∼
EndZp Γ (Λ)
= HomZp U (Zp [ζ], Λ)
= EndZp U (Zp [ζ]) × HomZp U (Zp [ζ], Zp U)
∼
= (Zp [ζ])m × ((1 − u)Zp U)n
where the last isomorphism follows upon sending a homomorphism ϕ to ϕ(1). Since
b ∼
EndFp Γ (Λ)
= (Fp [ζ])m × ((1 − u)Fp U)n
b to EndFp Γ (Vb ) is surby a similar argument, the natural homomorphism from EndZp Γ (Λ)
b (see [Ben,
jective (and reduction mod p). So e0 lifts to an idempotent e ∈ EndZp U (Λ)
Proposition 1.9.4]).
b Then Λ/pΛ ∼
Now set Λ = eΛ.
= V , and Λ is a Zp Γ -lattice in the Qp Γ -module M =
Qp ⊗Zp Λ.
(b) We repeat the proof of (a), but keeping control of the submodule as well as V . Set
V2 = V and V3 = V2 /V1 , and set Vbi = IndΓU (Vi |U ) for i = 1, 2, 3. We thus have short exact
sequences
f
fb
g
gb
0 −−−→ V1 −−−→ V2 −−−→ V3 −−−→ 0 and 0 −−−→ Vb1 −−−→ Vb2 −−−→ Vb3 −−−→ 0.
Let αi : Vbi −→ Vi be the natural map, and let βi : Vi −→ Vbi be the Fp Γ -linear splitting of αi
obtained by taking the natural Fp U-linear inclusion and then averaging over cosets of U in
def
Γ . Thus αi ◦ βi = IdVi (upon composing from right to left), while ei = βi ◦ αi is an idempotent
in EndFp Γ (Vbi ). All of these commute with the natural homomorphisms f , fb, g, and b
g , and
so we get a commutative diagram with exact rows:
0
/
Vb1
fb
e1
0
/
Vb1
Vb2
/
gb
e2
fb
/
Vb2
Vb3
/
0
/
(1)
e3
gb
/
Vb3
/
0.
Now set Λ01 = Zp , Λ02 = Zp U, Λ03 = Zp [ζ], and let ϕ and ψ be such that
ϕ
ψ
0 −−−→ Λ01 −−−→ Λ02 −−−→ Λ03 −−−→ 0
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
17
is a short exact sequence of Zp U-modules. We identify Vi |U = Λ0i /pΛ0i in such a way that
b i = IndΓU (Λ0i ), so that
f and g are the reductions modulo p of ϕ and ψ, respectively. Set Λ
b i /pΛ
b i as Fp Γ -modules, and let ϕ
Vbi = Λ
b and ψb be the homomorphisms induced by ϕ and
b i ) that make the following
ψ. We claim that the ei can be lifted to elements εi ∈ EndZp Γ (Λ
diagram commute:
0
/
b1
Λ
ϕ
b
ε1
0
To see this, we identify
/
b1
Λ
b2
Λ
/
b
ψ
ε2
ϕ
b
/
b2
Λ
b3
Λ
/
0
/
(2)
ε3
b
ψ
/
b3
Λ
/
0.
b i) ∼
EndZp Γ (Λ
= HomZp U Λ0i , (IndΓU (Λ0i ))|U ,
where by the Mackey double coset formula, for some indexing sets J and K independent of
i ∈ {1, 2, 3},
M
M
(Zp U ⊗Zp Λ0i ).
Λ0i ⊕
(IndΓU (Λ0i ))|U ∼
=
j∈J
k∈K
b 2 ), and set ε2 (1) = (uj )j∈J , (vk )k∈K with respect to the above decomFix ε2 ∈ EndZp Γ (Λ
position (and where 1 is the identity in Λ02 = Zp U). Then ε2 induces endomorphisms ε1 and
P
i
b
ε3 such that (2) commutes if and only if p−1
i=0 u (vk ) ∈ Ker(ψ) for each k ∈ K. Note that
b after restriction to the summands for some k ∈ K, is a surjection of one free Zp U-module
ψ,
onto another. Hence ε2 can always be chosen (as a lifting of e2 ) to induce ε1 and ε3 since
the above condition holds modulo p by the commutativity of (1).
k
k
Since εi is a lifting of the idempotent ei , we have ε2i ≡ εi (mod p). Hence (εi )2p ≡ (εi )p
k
(mod pk+1 ) for each k ≥ 1. Upon replacing εi by the limit of the (εi )p , we can arrange that
b i ) (and that the above diagram still commutes). Set
each εi is an idempotent in EndZp Γ (Λ
b 2 and M = Qp ⊗Zp Λ. Thus M is a Qp Γ -module with a 1-dimensional submodule,
Λ = ε2 Λ
and has a Zp Γ -lattice Λ such that Λ/pΛ ∼
= V.
We now turn to questions of uniqueness, looking first at the finite case. When R is a ring
and M is an R-module, we let AnnR (x) denote the annihilator of an element x ∈ M.
Proposition 3.9. Fix an odd prime p, a finite group Γ ∈ Gp , and U ∈ Sylp (Γ ). Let A1
and P
A2 be finite Zp Γ -modules, and assume that CAi (U) ∩ [U, Ai ] = p for i = 1, 2. Set
σ = u∈U u ∈ Zp U.
Assume that there is a Zp U-linear isomorphism ϕ : A1 −→ A2 whose reduction modulo p is
Fp Γ -linear. Then A1 ∼
= A2 /pA2
= A2 as Zp Γ -modules. In particular, this happens if A1 /pA1 ∼
and either
(a) A1 and A2 are homocyclic of the same exponent and σ·Ai pAi for i = 1, 2; or
(b) there are elements a1 ∈ A1 and a2 ∈ A2 such that A1 = Zp U·a1 , A2 = Zp U·a2 , and
AnnZp U (a1 ) = AnnZp U (a2 ).
Proof. Set Ai = Ai /pAi for i = 1, 2. For each X ≤ Ai and g ∈ Ai , set X = (X + pAi )/pAi
and g = g + pAi . Set Zi = CAi (U) and Si′ = [U, Ai ]. By assumption, |Zi ∩ Si′ | = p, and hence
|Ai /(Zi + Si′ )| = p by Lemma 3.7(c).
Assume ϕ : A1 −→ A2 is a Zp U-linear isomorphism whose reduction ϕ : A1 −→ A2 modulo
p is Fp Γ -linear. Let g1 , . . . , gk be a set of representatives for the left cosets gU in Γ (where
18
BOB OLIVER AND ALBERT RUIZ
Pk
−1
k = |Γ/U| is prime to p), and define ψ : A1 −−−→ A2 by setting ψ(λ) = k1
j=1 gj ϕ(gj λ) .
Then ψ is Zp Γ -linear, its reduction modulo p is equal to ϕ since ϕ is Fp Γ -linear, it is surjective
since the reduction mod p is surjective, and is an isomorphism since |A1 | = |A2 |.
It remains to prove that each of (a) and (b) implies the existence of the homomorphism
ϕ. Fix an Fp Γ -linear isomorphism ϕ : A1 −→ A2 .
(a) Assume A1 and A2 are homocyclic of the same exponent pk , and for i = 1, 2, σ·Ai pAi .
Choose a1 ∈ A1 such that σ·a1 ∈
/ pA1 , and let a2 ∈ A2 be such that ϕ(a1 ) = a2 . Thus for
i = 1, 2, σ·ai 6= 0, and so {u(ai ) | u ∈ U} is a basis for Fp U·ai , and AnnFp U (ai ) = 0. Note
that ai ∈
/ Zi + Si′ , since σ·Zi ≤ pZi and σ·Si′ = 0.
We claim that each element of CAi (U) lifts to an element of CAi (U); i.e., that
CAi (U) = Z i .
(1)
To see this, let g ∈ Ai be such that g ∈ CAi (U). By Lemma 3.7(d), g = ξ·ai + z for some
ξ ∈ Zp U and z ∈ Zi . Then ξ·ai is fixed by U since g is, and CZp U ·ai (U) = hσ·ai i since
AnnFp U (ai ) = 0. Hence there is k ∈ Z such that kσ·ai = ξ·ai ; and g = kσ·ai + z where
kσ·ai + z ∈ Zi . This proves (1).
Set m = rk(A1 ) − p = rk(A2 ) − p (possibly m = 0). By (1), we can choose elements
x1 , . . . , xm ∈ Z1 such that A1 = Fp U·a1 ⊕ hx1 , . . . , xm i. By (1) again, there are elements
y1 , . . . , ym ∈ Z2 such that ϕ(xi ) = y i for each i. Then
{u(a1 ) | u ∈ U} ∪ {x1 , . . . , xm }
and
{u(a2 ) | u ∈ U} ∪ {y 1 , . . . , y m }
are bases for A1 and A2 , respectively, and ϕ sends the first basis to the second. Since
A1 and A2 are both homocyclic of exponent pk , the sets {u(a1 ), x1 , . . . , xm | u ∈ U} and
{u(a2 ), y1 , . . . , ym | u ∈ U} are bases for A1 and A2 , respectively, as Z/pk -modules. Thus ϕ
lifts to a Zp U-linear isomorphism ϕ : A1 −→ A2 , defined by setting ϕ(u(a1 )) = u(a2 ) for
u ∈ U and ϕ(xi ) = yi for each i.
(b) Let ai ∈ Ai (for i = 1, 2) be such that Zp U·ai = Ai and AnnZp U (a1 ) = AnnZp U (a2 ). Let
ξ ∈ Zp U be such that ϕ(a1 ) = ξ·a2 . Thus ξ·a2 generates A2 as an Fp U-module, and since
(1 − u)A2 ≤ S2′ for 1 6= u ∈ U, ξ is not in the ideal (1 − u)Zp U + pZp U of index p in Zp U.
Since this is the unique maximal ideal in Zp U, ξ is invertible, and we can replace a2 by ξ·a2
without changing AnnZp U (a2 ).
Let ϕ : A1 −→ A2 be the unique Zp U-linear homomorphism such that ϕ(a1 ) = a2 . Its
reduction modulo p is ϕ, since ϕ(a1 ) = a2 and ϕ is Fp U-linear.
It remains to prove the analogous uniqueness result for discrete p-tori.
Lemma 3.10. Fix an odd prime p, a finite group Γ ∈ Gp , and U ∈ Sylp (Γ ). Let A1 and A2
be discrete, Γ -p-tori, and assume that
(i) Ω1 (A1 ) and Ω1 (A2 ) are faithful, minimally active, and indecomposable as Fp Γ -modules;
and
(ii) Ω1 (A1 ) ∼
= Ω1 (A2 ) as Fp Γ -modules.
Then A1 ∼
= A2 as Zp Γ -modules.
Proof. By Lemma 3.7(b), we also have that
(a) [U, Ai ] ∩ CAi (U) has order p for i = 1, 2.
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
19
Assume, for each k ≥ 1, that Ωk (A1 ) ∼
= Ωk (A2 ) as Zp Γ -modules, and let Xk be the set of
∼
=
Zp Γ -linear isomorphisms Ωk (A1 ) −−−→ Ωk (A2 ). Then Xk is finite since the Ωk (Ai ) are finite,
Xk 6= ∅ by assumption, and if k ≥ 2, restriction to Ωk−1 (Ai ) defines a map Xk −→ Xk−1 .
So the inverse limit of the Xk is nonempty, and each element in the inverse limit determines
a Zp Γ -linear isomorphism A1 ∼
= A2 .
It remains to show that Ωk (A1 ) ∼
= Ω1 (Ai ) as
= Ωk (A2 ) for each k. Since Ωk (Ai )/pΩk (Ai ) ∼
k−1
Fp Γ -modules (multiplication by p
defines an isomorphism), we have Ωk (A1 )/pΩk (A1 ) ∼
=
Ωk (A2 )/pΩk (A2 ), and both are faithful, minimally active, and indecomposable.
P
Set σ = u∈U u ∈ Zp U, as usual. If rk(Ai ) ≥ p, then by Lemma 3.3(a,b), σ·Ωk (Ai )
pΩk (Ai ). Since Ωk (A1 ) and Ωk (A2 ) are both homocyclic of exponent pk , they are isomorphic
as Zp Γ -modules by Proposition 3.9(a).
If rk(Ai ) = p − 1 for i = 1, 2, then by Proposition A.4, Ai ∼
= (Qp /Zp ) ⊗Zp Λi for some
(p − 1)-dimensional Zp Γ -lattice Λi . Since Γ acts faithfully on the lattices, Λ1 ∼
= Λ2 ∼
= Zp [ζ]
as Zp U-modules by Lemma A.5(a,c) (where ζ is a primitive p-th root of unity). Hence
for i = 1, 2, Ai ∼
= Qp (ζ)/Zp [ζ], and Ωk (Ai ) ∼
= Zp [ζ]/pk Zp [ζ], as Zp U-modules. So there is
ai ∈ Ωk (Ai ) such that Zp U·ai = Ωk (Ai ) and AnnZp U (ai ) is the ideal generated by pk and σ.
Proposition 3.9(b) now applies to conclude that Ωk (A1 ) ∼
= Ωk (A2 ) as Zp Γ -modules.
4. Reduced fusion systems over finite nonabelian p-groups with index p
abelian subgroup (p odd)
Throughout this section, p is an odd prime, and A is finite. As noted in the introduction,
the corresponding question for finite 2-groups was answered in [AOV2, Proposition 5.2(a)].
Lemma 4.1 ([CrOS, Lemma 2.2(d,e,f)]). Assume the notation and hypotheses of 2.1, and
also that p is odd and |A| < ∞. Set A0 = ZS ′ . Then the following hold.
(a) If A 5 F , then there are elements x ∈ SrA and a ∈ ArA0 such that A0 hxi and S ′ hai
are normalized by AutF (S). If some element of SrA has order p, then we can choose
x to have order p.
(b) For each P ∈ EF and each α ∈ NAutF (P ) (AutS (P )), α extends to some α ∈ AutF (S).
(c) For each x ∈ SrA and each g ∈ A0 , Zhxi is S-conjugate to Zhgxi, and Z2 hxi is
S-conjugate to Z2 hgxi.
We now fix some more notation, based on Lemma 4.1.
Notation 4.2. Assume Notation 2.1. Assume also that p is odd, S is finite, and A 5 F ,
and hence that |Z0 | = |A/ZS ′| = p by Lemma 2.4. Fix a ∈ A r ZS ′ and x ∈ S r A, chosen
such that ZS ′ hxi and S ′ hai are each normalized by AutF (S), and such that xp = 1 if any
element of S r A has order p (Lemma 4.1(a)). For each i = 0, 1, . . . , p − 1, define
Hi = Zhxai i ∈ H
and
Bi = Z2 hxai i ∈ B .
Let Hi and Bi denote the S-conjugacy classes of Hi and Bi , respectively, and set
H∗ = H1 ∪ · · · ∪ Hp−1
and B∗ = B1 ∪ · · · ∪ Bp−1 .
For each P ≤ S, set
′
(P )
AutF (S) = α ∈ AutF (S) α(P ) = P, α|P ∈ O p (AutF (P )) .
When |Z0 | = p, then by Lemma 4.1(c), H = H0 ∪ H∗ and B = B0 ∪ B∗ . Note that
for x, x′ ∈ S r A, Zhxi is S-conjugate to Zhx′ i or Z2 hxi is S-conjugate to Z2 hx′ i only if
20
BOB OLIVER AND ALBERT RUIZ
x′ x−1 ∈ ZS ′ . So in fact, each of the sets H and B is a union of p distinct S-conjugacy
classes: the classes Hi and Bi for 0 ≤ i ≤ p − 1.
Lemma 4.3 ([CrOS, Lemma 2.5(a,b)]). Let p be an odd prime, let S be a finite nonabelian
p-group with a unique abelian subgroup A E S of index p, and let F be a saturated fusion
system over S such that A 5 F . We use the conventions of Notation 2.1 and 2.9, set
A0 = ZS ′ , and let m ≥ 3 be such that |A/Z| = pm−1 . Then the following hold.
(a) µ
b|Out∨F (S) is injective.
(b) Fix α ∈ Aut(S), set (r, s) = µ(α), and let t be such that α(g) ∈ g t A0 for each g ∈ ArA0 .
Then s ≡ tr m−1 (mod p).
Lemma 4.4 ([CrOS, Lemma 2.6(a)]). Let p be an odd prime, let S be a finite nonabelian pgroup with a unique abelian subgroup A E S of index p, and let F be a saturated fusion system
over S. We use the notation of Notation 2.1 and 2.9. Let m be such that |A/Z| = pm−1 . Fix
P ∈ H ∪ B, and set
(
−1 if P ∈ H
b P = α|P α ∈ HP ,
HP = NAutF (S) (P ) ,
H
and
t=
0
if P ∈ B.
(P )
(P )
If P ∈ EF , then AutF (S) ≤ Aut∨F (S) and µ(AutF (S)) = ∆t . If P ∈ H∗ or P ∈ B∗ , then
m ≡ t (mod p − 1).
Theorem 4.5 ([CrOS, Theorem 2.8]). Fix an odd prime p, and a finite nonabelian p-group
S which contains a unique abelian subgroup A E S of index p. Let F be a reduced fusion
system over S for which A is F -essential. We use the notation of 2.1, 2.9, and 4.2, and
also set A0 = ZS ′ , E0 = EF r{A}, and G = AutF (A). Thus U = AutS (A) ∈ Sylp (G). Let
m ≥ 3 be such that |A/Z| = pm−1 . Then the following hold:
(a) Z0 = CA (U) ∩ [U, A] has order p, and hence A0 = CA (U)[U, A] has index p in A.
(b) There are no nontrivial G-invariant subgroups of Z = CA (U), aside (possibly) from Z0 .
(c) [G, A] = A.
(d) One of the conditions (i)–(iv) holds, described in Table 4.1, where σ =
′
µA (Aut∨F (A)) G = O p (G)X where m (mod p − 1)
P
u∈U
σ·A
u ∈ Zp U.
E0
(i)
∆
X = Aut∨F (A)
≡0
≤ Fr(Z) H0 ∪ B∗
(ii)
∆
X = Aut∨F (A)
≡ −1
≥ ∆−1
X = µ−1
A (∆−1 )
≤ Fr(Z) B0 ∪ H∗
S
≤ Fr(Z)
i∈I Hi
(iii′ )
(iii′′ )
(iv′ )
(iv′′ )
≥ ∆0
≡ −1
−
−
X = µ−1
A (∆0 )
≡0
≤ Fr(Z)
Z0 not G-invariant
−
−
S
H0
i∈I
Bi
B0
Table 4.1.
Conversely, for each G, A, U ∈ Sylp (G), and E0 ⊆ H∪B which satisfy conditions (a)–(d),
where |U| = p and U 5 G, there is a simple fusion system F over A ⋊ U with AutF (A) = G
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
21
and EF = E0 ∪ {A}, unique up to isomorphism. When A is not elementary abelian, all such
fusion systems are exotic, except for the fusion systems of the simple groups listed in Table
4.2.
Such a fusion system F has a proper strongly closed subgroup if and only if A0 = CA (U)[U, A]
is G-invariant, and E0 = Hi or Bi for some i, in which case A0 Hi = A0 Bi is strongly closed.
conditions
rk(A)
e
m
G = AutΓ (A)
E0
p2 |(q−1), p > 3
p−1
vp (q−1)
e(p−1) − 1
Σp
H0 ∪ H∗
PSLn (q) p p2 |(q−1), p<n<2p
n−1
vp (q−1)
e(p−1) + 1
Σn
B0
2
P Ω+
2n (q) p p |(q−1), p≤n<2p
n
vp (q−1)
e(p−1) + 1
C2n−1 ⋊ Σn
B0
2
v3 (q+1)
2e
GL2 (3)
B0 ∪ B∗
Γ
p
PSLp (q) p
2F (q)
4
3
En (q)
5
n = 6, 7, p2 |(q−1)
n
v5 (q−1)
4e + 1
W (En )
B0
En (q)
7
n = 7, 8, p2 |(q−1)
n
v7 (q−1)
6e + 1
W (En )
B0
E8 (q)
5
v5 (q 2 + 1) ≥ 2
4
v5 (q 4 −1)
4e
(4 ◦ 21+4 ).Σ6
B0 ∪ B∗
Table 4.2. In this table, e is such that pe is the exponent of A, and we restrict
to the cases where e ≥ 2. In all cases except when Γ ∼
= PSLp (q), A is homocyclic.
We now look for a more precise description of the group A when it is finite but not
elementary abelian. The following notation will be useful when describing more precisely
elements and subgroups of A.
Notation 4.6. Assume Notation 2.1 and 4.2, and set u = cx ∈ U = AutS (A). Set σ =
1 + u + u2 + . . . + up−1 ∈ Zp U. Regard A as a Zp U-module, and define
Ψ : Zp U −−−−−→ A
Qp−1 i n
Pp−1
i
by setting Ψ(ξ) = ξ·a. Thus Ψ
n
u
= i=0 u (a) i for ni ∈ Zp .
i
i=0
Set ζ = e2πi/p , R = Zp [ζ], and p = (1−ζ)R. Thus p is the unique maximal ideal in R. We
identify R = Zp U/σZp U, by sending ζ ∈ R to the class of u modulo hσi.
The basic properties of Ψ, and the role of Ψ(σ), are described in the following lemma.
k
Recall that ℧k (P ) = hg p | g ∈ P i, when P is a p-group and k ≥ 1.
Lemma 4.7. Assume Notation 2.1 and 4.6, where A ∈ EF , and A 5 F is finite and not
elementary abelian. Let m be such that |A/Z| = pm−1 . Then
(a) Im(Ψ) ∩ Z = Z0 hΨ(σ)i;
(b) Ψ induces an isomorphism A/Z ∼
= R/pm−1 via the identification R = Zp U/hσi; and
(c) Ψ((1 − u)m ) = 1 and Z0 = hΨ((u − 1)m−1 )i.
Furthermore, the following all hold.
(d) The homomorphism Ψ is surjective if and only if rk(A) ≤ p, if and only if Z is cyclic.
If Ψ(σ) ∈ Fr(A), then rk(A) < p and Ψ is surjective.
(e) Either
22
BOB OLIVER AND ALBERT RUIZ
• Ψ(σ) = 1, in which case rk(A) = p − 1, Z = Z0 , and Ψ induces an isomorphism
A∼
= R/pm via the identification R = Zp U/hσi; or
• Ψ(σ) ∈
/ Fr(Z), in which case EF ⊆ {A} ∪ H0 or EF ⊆ {A} ∪ B0 .
(f) If Ψ(σ) 6= 1 and Ψ(σ) ∈ Z0 , then m ≡ 1 (mod p − 1). If Ψ(σ) ∈
/ Z0 , then µ(Aut∨F (S)) =
∆m−1 ; and either m ≡ 1 (mod p − 1) and EF r {A} = B0 , or m ≡ 0 (mod p − 1) and
EF r {A} = H0 .
(g) If Ψ is not surjective, then A is homocyclic.
Proof. Set A = A/Fr(A) for short. For B ≤ A or g ∈ A, let B ≤ A or g ∈ A denote
their images in A under projection. Let Ψ : Zp U −−−→ A be the composite of Ψ followed by
projection to A.
(a) Since (1 − u)Zp U + σZ has index p in Zp U,
Im(Ψ) = Ψ (1 − u)Zp U hΨ(σ)ihΨ(1)i = S ′ hΨ(σ)ihai,
where ap ∈ S ′ hΨ(σ)i. Since a ∈
/ Z and Ψ(σ) ∈ Z, we have
Im(Ψ) ∩ Z = CIm(Ψ) (U) = CS ′ hΨ(σ)i (U) = (S ′ ∩ Z)·hΨ(σ)i = Z0 hΨ(σ)i.
(b,c) Since Ψ(σ) ∈ Z, Ψ induces a homomorphism from Zp U/hσi ∼
= R to A/Z, which
m−1
is onto since A = Z·Im(Ψ) by Lemma 3.7(d). Since |A/Z| = p
by assumption (and
since p is the unique maximal ideal in R that contains p), we have A/Z ∼
= R/pm−1 . Hence
Ψ((u − 1)m−1 ) ∈ Z and Ψ((u − 1)m−2 ) ∈
/ Z, and the latter implies that Ψ((u − 1)m−1 ) 6= 1.
Thus in all cases (and since |Z0| = p), Ψ((u − 1)m ) = 1 and hΨ((u − 1)m−1 )i = Z0 .
(d) If rk(A) ≤ p, then rk(A) ≤ p, and by [CrOS, Proposition 3.7(a)], A|U is indecomposable.
Hence Ψ is onto in this case, and so Ψ is also onto. Conversely, if rk(A) > p = rk(Zp U),
then Ψ is clearly not surjective.
By Lemmas 3.2(b) and 3.3(c), Ω1 (Z) = CΩ1 (A) (U) has rank 1 if and only if rk(Ω1 (A)) ≤ p.
Hence Z is cyclic if and only if rk(A) ≤ p.
If Ψ(σ) ∈ Fr(A), then rk(Im(Ψ)) ≤ p − 1. Hence A has no nontrivial Jordan block of
rank p, and by [CrOS, Proposition 3.7(a)] again, A is indecomposable as an Fp U-module.
So rk(A) = rk(A) < p, and Ψ is onto.
(e) If Ψ(σ) ∈ Fr(Z) ≤ Fr(A), then Ψ is surjective by (a), so Z = Z0 hΨ(σ)i ≤ Z0 ·Fr(Z), and
hence Z = Z0 and Ψ(σ) ∈ Fr(Z0 ) = 1. Thus Ψ factors through a surjection Ψ∗ : Zp U/hσi ∼
=
∼
R −−−→ A, and induces an isomorphism A = R/I for some ideal I in R. Since p is the only
prime ideal in R of p-power index (and |R/p| = p), and since |A| = pm−1 |Z| = pm (recall
Z = Z0 by (a)), we have I = pm .
Since EF 6⊆ {A} (Notation 4.6), xp = 1 by Notation 4.6 and Lemma 4.1. For each b ∈ A,
Q
2
p−1
i
(bx)p = (bx)p x−p = b·xb·x b · · · x b = p−1
i=0 u (b).
Q
Q
Sp−1 i ′
i
i
′
If Ψ(σ) = p−1
/ Fr(Z), then p−1
i=0 u (a) ∈
i=0 u (b) 6= 1 for each b ∈ A r ZS =
i=1 a ZS . So
by Lemma 2.5, no member of H∗ ∪ B∗ can be essential, and EF ⊆ {A} ∪ H0 ∪ B0 .
(P )
(f) Assume A 5 F , and thus EF 6⊆ {A}. Fix P ∈ EF ∩ (H ∪ B) and α ∈ AutF (S) ≤
Aut∨F (S) (Lemma 4.4).
Set µ(α) = (r, s), and let t be as in Lemma 4.3(b). Thus s ≡ tr m−1 (mod p) and
α(a) ≡ at (mod ZS ′ ), so α(a) = Ψ(ξ) for some ξ ≡ t (mod h1 − u, pi). Also, α(x) ∈ xr A,
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
23
so α(ui (g)) = uri (α(g)) for all i and g ∈ A. Thus
p−1
α(Ψ(σ)) =
Y
i
α(u (a)) =
p−1
Y
ri
u (α(a)) = Ψ
i=0
i=0
p−1
X
ξuri
i=0
= Ψ(ξσ) ≡ Ψ(tσ) .
(mod Ψ(pσ)) (1)
In other words, α(Ψ(σ)) ≡ Ψ(σ)t (mod hΨ(σ)p i).
If Ψ(σ) 6= 1 and Ψ(σ) ∈ Z0 , then t ≡ s (mod p) by (1) (and by definition of µ), and hence
m−1
r
≡ 1 (mod p). Since this holds for arbitary α and hence for arbitary r prime to p by
[CrOS, Lemma 2.6(a)] and since P ∈ EF ∩ (H ∪ B), it follows that m ≡ 1 (mod p − 1).
Now assume Ψ(σ) ∈
/ Z0 . By (1) and since [α, Z] ≤ Z0 , we have t ≡ 1 and s ≡ r m−1 . Since
(P )
this holds for arbitrary α ∈ AutF (S) (in particular, for arbitrary r prime to p), it follows
(P )
that µ(AutF (S)) ≤ µ(Aut∨F (S)) ≤ ∆m−1 , with equality by Lemma 4.4. So by Lemma 4.4,
P ∈
/ H∗ ∪ B∗ , and either P ∈ H0 and ∆m−1 = ∆−1 (so m ≡ 0 (mod p − 1)); or P ∈ B0 and
∆m−1 = ∆0 (so m ≡ 1 (mod p − 1)).
(g) Assume that Ψ is not onto, and hence by (d) that rk(A) ≥ p + 1 and Ψ(σ) ∈
/ Fr(A).
Let k ≥ 2 be such that A has exponent pk . If A/Z has strictly smaller exponent, then
1 6= ℧k−1 (A) ≤ Z, and thus ℧k−1 (A) is an Fp G-submodule of the minimally active, indecomposable module Ω1 (A) upon which U acts trivially. If rk(A) = p + 1, this contradicts
Lemma 3.5, while if rk(A) ≥ p + 2, this is impossible since Ω1 (A) is simple by [CrOS, Proposition 3.7(c)]. Thus A/Z ∼
= R/pm−1 also has exponent pk ≥ p2 , and hence m − 1 ≥ p. So by
(c), and since (u − 1)p ∈ pZp U, we have Z0 = hΨ((u − 1)m−1 )i ≤ Fr(A).
Now, rk(A/Z) = rk(R/pm−1 ) = p − 1 since m ≥ p, and rk(Z) = rk(CΩ1 (A) (U)) =
rk(A) − (p − 1) by Lemma 3.3(c). If Z0 is a direct factor in Z, then rk(Z/Z0) = rk(Z) − 1,
so rk(A/Z0 ) ≤ rk(A/Z) + rk(Z/Z0) = rk(A) − 1. Thus no minimal generating set for A/Z0
lifts to a generating set for A, so Z0 Fr(A), which contradicts what we just showed. Thus
Z0 is not a direct factor in Z, and so EF ∩ H = ∅ by Lemma 2.5(a).
In particular, m ≡ 1 (mod p − 1) by (f), and hence A/Z ∼
= R/pm−1 is homocyclic of
rank p − 1 and exponent pk . Thus A and A/Z are both Z/pk -modules and A/Z is free, so
A∼
= Z × (A/Z) as abelian groups. Since ℧k−1 (A) ∩ Z = C℧k−1 (A) (U) 6= 1, this shows that
rk(℧k−1 (A)) ≥ p.
If A is not homocyclic, then ℧k−1 (A) < Ω1 (A) is a nontrivial proper Fp G-submodule,
where Ω1 (A) is faithful, minimally active, and indecomposable by Lemma 3.2(b). Hence
rk(A) = dim(Ω1 (A)) = p + 1, since Ω1 (A) is simple if dim(Ω1 (A)) ≥ p + 2 by [CrOS,
Proposition 3.7(c)]. So dim(℧k−1 (A)) ≤ p − 1 by Lemma 3.5. This contradicts what was
shown in the last paragraph, and we conclude that A is homocyclic.
P
Lemma 4.8. Let p be an odd prime, let U be a group of order p, and set σ = u∈U u ∈ ZU.
Then for each 1 6= u ∈ U and each k ≥ 1,
(u − 1)k(p−1) ≡ (−1)k−1 (pk−1 σ − pk )
Proof. Since
pZU). Hence
p−1
k
(mod pk (u − 1)ZU).
≡ (−1)k (mod p) for each 0 ≤ k ≤ p − 1, we have (u − 1)p−1 ≡ σ (mod
(u − 1)p−1 ≡ σ − p
(mod p(u − 1)ZU)
(1)
since they are congruent modulo p and modulo u − 1. This proves the lemma when k = 1.
24
BOB OLIVER AND ALBERT RUIZ
When k > 1, (1) together with the congruence for (u − 1)(k−1)(p−1) give
(u − 1)k(p−1) = (u − 1)p−1 ·(u − 1)(k−1)(p−1)
≡ (u − 1)p−1·(−1)k−2 (pk−2 σ − pk−1)
(mod (u − 1)p−1 ·pk−1(u − 1))
≡ (σ − p)·(−1)k−2 (pk−2 σ − pk−1 )
(mod p(u − 1)(pk−2σ − pk−1 ))
= (−1)k−1 (pk−1 σ − pk );
and the congruences hold modulo pk (u−1) since p(u−1) divides (u−1)p and (u−1)σ = 0.
Proposition 4.9. Assume the notation of 2.1, 2.9, 4.2, and 4.6. Assume also that A is
finite and not elementary abelian, that A ∈ EF , and that Op (F ) = 1. Let m ≥ 3 be such
that |A/Z| = pm−1 , and let k ≥ 2 be such that A has exponent pk . Then one of the following
holds, as summarized in Table 4.3, where G = AutF (A).
(a) If Ψ(σ) = 1, then Ψ is onto, Ker(Ψ) = hσ, (u − 1)m i, rk(A) = p − 1, Z = Z0 =
hΨ((u − 1)m−1 )i, and A ∼
= R/pm as Zp U-modules.
(b) If Ψ(σ) ∈
/ Fr(Z) and A is homocyclic, then rk(A) ≥ rk(Im(Ψ)) = p, rk(Z) = rk(A) −
p + 1, and Im(Ψ) and Z are both direct factors in A and homocyclic of exponent pk .
Also, EF = {A} ∪ B0 . Either Ψ is onto and rk(A) = p, or Ψ is not onto and rk(A) > p.
If rk(A) ≥ p + 2, then A/Fr(A) ∼
= Ω1 (A) are irreducible Fp [AutF (A)]-modules.
(c) If Ψ(σ) ∈
/ Fr(Z) and A is not homocyclic, then Ψ is onto, rk(A) = p − 1, m ≡ 1 (mod
p − 1), Ker(Ψ) = hpk , pk−1 − ℓσi for some ℓ prime to p, and EF = {A} ∪ H0 . Also,
A∼
= (Cpk−1 )p−2 × Cpk , where ℧k−1(A) = Z = Z0 = hΨ(σ)i. If k = 2, then ℓ 6≡ 1 (mod
p).
Case
(a)
(b)
Ψ(σ)
Ψ(σ) = 1
A homocyclic?
yes if (p − 1) | m
no if (p − 1) ∤ m
yes
no
Ψ onto?
yes
yes if rk(A) = p
no if rk(A) > p
yes
rk(A)
p−1
r≥p
p−1
Ker(Ψ)
h(u − 1)m , σi
pk Zp U
hpk , pk−1 − ℓσi (p ∤ ℓ)
A
∼
= R/pm
∼
= (Cpk )r
∼
= (Cpk−1 )p−2 × Cpk
Z
= Z0
∼
= (Cpk )r−p+1
= Z0
Z0
hΨ((u − 1)m−1 )i
hΨ(pk−1 σ)i
hΨ(pk−1)i = hΨ(σ)i
k(p − 1) + 1
(k − 1)(p − 1) + 1
B0
H0
Ψ(σ) ∈
/ Fr(Z)
m
EF r{A}
(c)
(see Table 4.1)
Table 4.3.
Proof. If EF = {A}, then A E F by Proposition 1.3(c), contradicting the assumption that
Op (F ) = 1. Thus EF % {A}.
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
25
Case 1: Ψ(σ) ∈ Fr(Z). In this case, Ψ is surjective by Lemma 4.7(d) and since Fr(Z) ≤
Fr(A). By Lemma 4.7(e), Ψ(σ) = 1, rk(A) = p − 1, Z = Z0 , and A ∼
= R/pm . In particular,
m
m−1
Ker(Ψ) = hσ, (u − 1) i, while Z0 = hΨ((u − 1)
)i by Lemma 4.7(c). We are thus in the
situation of (a).
Case 2: Ψ(σ) ∈
/ Fr(Z) and A is homocyclic. Recall that k ≥ 2 is such that A has
k
exponent p .
If rk(A) < p, then Ψ is onto by Lemma 4.7(d), and Ω1 (Z) = CΩ1 (A) (U) has rank 1 by
Lemma 3.3(c) and since Ω1 (A) is minimally active and indecomposable by Lemma 3.2(b).
Thus Z is cyclic, and since A is homocyclic of rank at least 2,A/Z ∼
= R/pm−1 also has
exponent pk ≥ p2 . Hence rk(A) = rk(A/Z) = p − 1. Also, (A/Z) ℧k−1 (A/Z) ∼
= (Cpk−1 )p−1 ,
m
so |Z| = p, and Z = Z0 . Thus |A| = |A/Z|·|Z| = p , and m ≡ 0 (mod p − 1) since A is
homocyclic of rank p − 1. But then Ψ(σ) ∈
/ Z0 by Lemma 4.7(f), a contradiction.
Thus rk(A) ≥ p, and Ψ(σ) ∈
/ Fr(A) by Lemma 4.7(d). So the homomorphism Ψ : Fp U −→
A/Fr(A) is injective, and A/Fr(A) contains a Jordan block Im(Ψ) of rank p. Since A is
homocyclic of exponent pk , |Im(Ψ)| ≥ ppk , and thus Ker(Ψ) = pk Zp U. So Im(Ψ) ∼
= Z/pk U.
Also,
pm−1 = |A/Z| = |Im(Ψ)/hΨ(σ)i| = pk(p−1) ,
and so m = k(p − 1) + 1 ≡ 1 (mod p − 1).
Now, Z0 = hΨ(pk−1 σ)i, and Ψ(σ) ∈
/ Z0 since k ≥ 2. Hence EF r {A} = B0 by Lemma
4.7(f). Also, Ψ is surjective if and only if rk(A) = p (Lemma 4.7(d)), and we are in the
situation of case (b).
Case 3: Ψ(σ) ∈
/ Fr(Z) and A is not homocyclic. Set k ′ = [m/(p − 1)]. Since m ≡ 0, 1
′
′
(mod p − 1) by Lemma 4.7(f), A/Z ∼
= R/pm−1 has exponent pk , and hence ℧k (A) ≤ Z. So
′
pk (u − 1) ∈ Ker(Ψ).
Now, Ψ is onto by Lemma 4.7(g), and hence rk(A) ≤ p and Z is cyclic by Lemma 4.7(d).
If Ψ(σ) ∈
/ Z0 , then Z0 < Z and is not a direct factor, so EF ∩ H = ∅ by Lemma 2.5(a).
Thus
Ψ onto
′
=⇒
′
Ψ(σ) ∈ Z0 or EF ∩ H = ∅.
(1)
Case 3.1: ℧k (A) 6= 1. Since ℧k (A) ≤ Z is invariant under the action of G = AutF (A),
′
′
′
6 ∅. In particular, ℧k (A) = hΨ(pk )i has
Lemma 2.7 implies that ℧k (A) = Z0 and EF ∩ H =
′
′
order p, and Ψ(σ) ∈ Z0 by (1). Thus hΨ(σ)i = Z0 = hΨ(pk )i, so pk − ℓσ ∈ Ker(σ) for some
ℓ prime to p.
Now, rk(A) < p by Lemma 4.7(d) and since Ψ(σ) ∈ Z0 ≤ Fr(A). Also, Z = Z0 hΨ(σ)i = Z0
by Lemma 4.7(a), and m ≡ 1 (mod p − 1) by Lemma 4.7(f) and since Ψ(σ) ∈ Z0 . So
′
′
|A| = |A/Z|·|Z| = pm = pk (p−1)+1 , and A/℧k (A) has exponent k ′ , rank at most p − 1, and
′
′
′
order pk (p−1) . This proves that A/℧k (A) is homocyclic of rank p − 1 and exponent pk , and
hence that A ∼
= (Cpk′ )p−2 × Cpk′+1 . Thus k ′ = k − 1 (recall A has exponent pk ). This also
shows that hpk−1, σi has index p in Ker(Ψ), and hence (since pk−1 − ℓσ is in the kernel) that
Ker(Ψ) = hpk , pk−1 − ℓσi.
If k = 2 and ℓ ≡ 1 (mod p), then Ker(Ψ) = hp2 , p − σi, and (u − 1)p−1 ∈ Ker(Ψ) by
Lemma 4.8. But then Ψ((u − 1)p−2) ∈ Z where Z = Z0 = ℧1 (A), and this is impossible
since A/℧1 (A) has rank p − 1.
Finally, EF = {A} ∪ H0 by Lemma 4.7(e) and since EF ∩ H =
6 ∅. We are thus in the
situation of case (c).
26
BOB OLIVER AND ALBERT RUIZ
′
Case 3.2: ℧k (A) = 1. Since Ψ(σ) ∈ Z0 r 1 or EF ∩ H = ∅ by (1), we have m ≡ 1 (mod
′
p −1) by Lemma 4.7(f). Thus m−1 = k ′ (p −1), so 1 6= Z0 = hΨ((1 − u)m−1 )i = hΨ(pk −1 σ)i
′
′
by Lemma 4.8. Since pk −1 σ ∈
/ Ker(Ψ), we have |Z| ≥ |Ψ(σ)| = pk , and |A| = |A/Z|·|Z| ≥
′
′
′
pm−1+k = pk p . Hence A = Im(Ψ) is homocyclic of rank p and exponent pk , contradicting
our assumption.
We now give some examples to show that all cases listed in Proposition 4.9 can occur.
Example 4.10. We list here some examples of pairs (G, A) satisfying the hypotheses of
Theorem 4.5. In all cases, we assume that G ∈ Gp∧ and U ∈ Sylp (G). By Lemma 3.2(b),
Ω1 (A) and A/Fr(A) must be minimally active and indecomposable.
(a): Each Qp G-module of dimension p − 1 whose restriction to U is isomorphic to the
canonical action on Qp (ζ) can be used to construct homocyclic examples of arbitrary
exponent, by adding scalars as needed to meet one of the conditions in Table 4.1.
More interesting are examples where A is not homocyclic. By Proposition 4.9,
Ω1 (A) and A/Fr(A) must be not only minimally active and indecomposable of dimension p − 1, but also not simple. By Table 6.1, this occurs only when Ap ≤ G ≤
Σp ×Cp−1 , SL2 (p) ≤ G ≤ GL2 (p), or PSL2 (p) ≤ G ≤ PGL2 (p)×Cp−1 . By Proposition
3.8(a), for each minimally active, indecomposable Fp G-module V of dimension p − 1,
there is a Zp G-lattice Λ such that Λ/pΛ ∼
= V . If 0 6= V0 < V is a nontrivial proper
Fp G-submodule, and Λ0 < Λ is such that pΛ < Λ0 and Λ0 /pΛ ∼
= V0 , then we can take
A∼
= Λ0 /pk Λ for arbitrary k ≥ 2.
(b): These are homocyclic, and there are many such examples, obtained from the Fp Gmodules in Table 6.1 of dimension at least p (all of them are reductions of lattices in
Qp G-modules). Lemma 3.2(c) together with Theorem 4.5 imply, very roughly, that
each Fp G-module that yields simple fusion systems with elementary abelian A and
with EF ⊆ {A} ∪ B will also give simple fusion systems with A of exponent pk for
arbitary k > 1. Some of the resulting fusion systems are realizable (see Table 4.2),
while “most” are exotic.
(c): Fix an odd prime p, k ≥ 2, and ℓ prime to p such that ℓ 6≡ 1 (mod p) if k = 2. Set
′
G = Σp × Cp−1 and G0 = O p (G) ∼
= Ap , and set U = (1 2 · · · p) ∈ Sylp (G). Let
p
Λ∼
= (Zp ) be the Zp G-lattice upon which Σp acts by permuting a Zp -basis {e1 , . . . , ep },
and where the factor Cp−1 acts via multiplication by (p − 1)-st roots of unity in Z×
p.
Now define
A = Λ pk Λ, pk−1 ei − ℓ(e1 + · · · + ep ) 1 ≤ i ≤ p ,
and let ei ∈ A be the class of ei ∈ Λ. This defines a finite Zp G-module of rank p − 1
and exponent pk , as described in the last column in Table 4.3, where Ψ : Zp U −→ A
is defined by setting Ψ(ξ) = ξ·e1 .
Set Z = CA (U). Note that |pk−1A| = p, and A/pk−1 A ∼
=
= Λ/hpk−1Λ, e1 + · · · + ep i ∼
k−1
k−1
k−2
Zp [ζ]/(p ). Hence Z ≥ p A, and |Z| > p only if p (e1 + 2e2 + · · · + pep ) ∈ Z.
But this last is the case only if pk−2 ((e1 + · · · + ep ) − pe1 ) = 0, which is not possible
since we assumed that either k ≥ 3 or ℓ 6≡ 1 (mod p). Thus |Z| = p in all cases.
Now set S = A ⋊ U, and set G∨ = NG (U). (Note that Z = Z0 in the notation of
2.1.) Define µA : G∨ −→ ∆ as in Notation 2.9. One easily checks that µA (G∨ ) = ∆.
Set G = G0 µ−1
A (∆−1 ). This now defines an action which satisfies the conditions in
Theorem 4.5, including condition (d.iii′′ ).
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
27
We now combine Theorem 4.5 with Proposition 4.9 to prove our main result on simple
fusion systems over finite p-groups with an abelian subgroup of index p and exponent at
least p2 . Recall that EF is the set of essential subgroups in a fusion system F (Definition
1.2), and that Gp∧ is a certain class of finite groups (Definition 3.1).
Theorem A. Fix an odd prime p.
(a) Let F be a simple fusion system over a finite nonabelian p-group S with an abelian subgroup A < S of index p such that A ∈ EF . Let k be such that A has exponent pk , and assume k ≥ 2 (A is not elementary abelian). Set G =AutF (A), U = AutS (A) ∈ Sylp (G),
V = Ω1 (A), and V0 = ℧k−1 (A) ≤ V . Let G∨(V ) = α ∈ NG (U) [α, CV (U)] ≤ [U, V ]
and µV : G∨(V ) −→ ∆ be as in Notation 2.9. Let Ψ : Zp U −→ A be as in Notation 4.6.
Then G ∈ Gp∧ , restriction defines an isomorphism G ∼
= AutF (V ), and
V is a faithful, minimally active, indecomposable Fp G-module, and
[G, V0 ] = V0 . Also, one of the cases in Table 4.4 holds, where V has
no 1-dimensional submodule in cases (iv′ ) and (iv′′ ).
Tbl.4.1 Tbl.4.3 dim(V ) r = dim(V0 )
(iv′′ )
(b)
≥p
V0 = V
Ker(Ψ)
(i)
(ii)
V0 = V
(a)
p−1
(iii′′ )
hpk , σ,
pk−1 (u − 1)r i
r =p−2
(iii′ )
1≤r ≤p−1
(c)
p−1
r=1
G=
≥ ∆0
O p (G)µ−1
V (∆0 )
EF r{A}
hpk i
1≤r ≤p−1
(iv′ )
µV (G∨
(V ) )
(∗fin )
∆
i∈I
Bi
H0 ∪ B ∗
p′
B 0 ∪ H∗
S
i∈I Hi
O (G)·G∨
(V )
O (G)·G∨
(V )
≥ ∆−1
Op (G)µ−1
V (∆−1 )
′
S
p′
∆
hpk , pk−1 − ℓσi
B0
′
H0
Table 4.4. The sets in the last column are as defined in Notation 2.1 and 4.2.
(b) Conversely, assume that G ∈ Gp∧ and U ∈ Sylp (G), and that V is a faithful, minimally
active, indecomposable Fp G-module satisfying (∗fin ) for some submodule 0 6= V0 ≤ V ,
where G∨ is the subgroup of all elements g ∈ NG (U) such that [g, CV (U)] ≤ [U, V ].
Then for each k ≥ 2, there is a simple fusion system F over a finite p-group S containing
an abelian subgroup A of index p, and such that G ∼
= AutF (A), A has exponent pk ,
k−1
∼
∼
Ω1 (A) = V and ℧ (A) = V0 as Fp G-modules, and with EF r {A} as described in
Table 4.1. Furthermore, any other simple fusion system with these properties, and with
the same essential subgroups, is isomorphic to F .
Proof. (a) Set U = AutS (A) ∈ Sylp (G) and Z = CA (U).
Under the above assumptions, G ∈ Gp∧ by Lemma 3.2(a), V is faithful, minimally active,
and indecomposable by Lemma 3.2(b), and dim(V ) = rk(A) ≥ p − 1 by Proposition 4.9. In
particular, restriction induces a monomorphism G = AutF (A) −→ AutF (V ), and this is an
isomorphism by the extension axiom (Definition 1.1) and since A = CS (V ). From now on,
we identify G = AutF (V ).
28
BOB OLIVER AND ALBERT RUIZ
Set G∨(A) = α ∈ NG (U) [α, CA (U)] ≤ [U, A] . Thus G∨(A) ≤ G∨(V ) , and µA : G∨(A) −→ ∆
′
is as in Notation 2.9. For some ∆x ∈ {∆, ∆0 , ∆−1 }, µA (G∨(A) ) ≥ ∆x and G = O p (G)µ−1
A (∆x )
∨
∨
by Theorem 4.5(d). So the same holds if we replace G(A) by G(V ) and µA by µV .
Assume we are in case (a) of Table 4.3. Then Z = Z0 has order p, so |A| = pm , and m ≡ 0
(mod p − 1) if and only if A is homocyclic. Also, Ψ(σ) = 1. Thus we are in case (i), (iii′′ ),
(iv′ ), or (iv′′ ) of Table 4.1 if A is homocyclic, or in case (ii), (iii′ ), (iii′′ ), or (iv′′ ) if A is not
homocyclic. Since m ≡ 0 (mod p − 1) and Ψ(σ) = 1 when A is homocyclic, (iv′′ ) is a special
case of (iv′ ). The other information follows from the two earlier tables.
Now assume we are in case (b) or (c) in Table 4.3. Then Ψ(σ) ∈
/ Fr(Z), so this corresponds
to case (iii′′ ) or (iv′′ ) in Table 4.1. Since EF r {A} = H0 in cases (c) and (iii′′ ), and
EF r {A} = B0 in cases (b) and (iv′′ ), these are the only possible correspondences. In case
(b), A is homocyclic, and so we have V0 = V in Table 4.4. In case (c), A ∼
= (Cpk−1 )p−2 × Cpk ,
so V = Ω1 (A) has a submodule V0 = ℧k (A) of rank 1.
There is a surjective homomorphism ψ : A −−−→ V0 of Zp G-modules, defined by setting
k−1
ψ(a) = ap . Since [G, A] = A by Theorem 4.5(c), we have [G, V0 ] = V0 .
If V = Ω1 (A) has a 1-dimensional Fp G-submodule W , then W ∈ CV (U) ≤ CA (U), so
W = Z0 = CA (U) ∩ [U, A] by Theorem 4.5(b), which is impossible in cases (iv′ ) and (iv′′ )
of Theorem 4.5(d).
(b) Fix G ∈ Gp∧ and U ∈ Sylp (G), and let V be an Fp G-module that satisfies (∗fin ), where
G∨ is the subgroup of all elements g ∈ NG (U) such that [g, CV (U)] ≤ [U, V ]. Fix k ≥ 2.
Assume we have chosen a finite Zp G-module A such that Ω1 (A) ∼
= V as Fp G-modules, and
such that either dim(V ) = p − 1 and |CA (U)| = p, or dim(V ) ≥ p and A is homocyclic. Let
G∨(A) ≤ G∨(V ) be as in the proof of (a). If rk(V ) = rk(A) = p − 1, then since CA (U) ∼
= CV (U)
∨
∨
∨
has order p, G(A) = G(V ) = NG (U) and µA = µV . So the properties of G(V ) and µV in Table
4.4 also hold for G∨(A) and µA .
′
If rk(V ) = rk(A) ≥ p and A is homocyclic, then µV (G∨(V ) ) ≥ ∆0 , and G = O p (G)µ−1
V (∆0 ).
−1
∨
∨
In such cases, we could have G(A) < G(V ) , but for each α ∈ µV (∆0 ) of order prime to p,
α acts trivially on CV (U) ∼
= Ω1 (CA (U)) by definition of µV (and Lemma A.1), and hence
−1
also acts trivially on CA (U) (see [G, Theorem 5.2.4]). Thus µ−1
V (∆0 ) = µA (∆0 ), and so the
∨
∨
information in Table 4.4 still holds if we replace G(V ) and µV by G(A) and µA .
Cases 4.9(a,b): Assume that we are in Case (a) or (b) in Proposition 4.9. By Proposition
3.8(a), there is a Zp G-lattice Λ such that Λ/pΛ ∼
= V as Fp G-modules. Let Λ0 ≤ Λ be such
that Λ0 ≥ pΛ and Λ0 /pΛ ∼
= V0 , and set A = Λ0 /pk Λ. Set S = A ⋊ U, Z = Z(S) = CA (U),
S ′ = [S, S] = [U, A], and Z0 = Z ∩ S ′ .
We first check that conditions (a)–(d) in Theorem 4.5 all hold. Condition (d) follows
immediately from (∗fin ). Condition (a) (|Z0 | = p) follows from Lemma 3.7(c) and since A is
defined to be a quotient group of a Zp G-lattice.
Assume 1 6= B ≤ Z is G-invariant; we claim that B = Z0 . If rk(A) = p − 1, then Z = Z0 ,
and there is nothing to prove. If not, then rk(A) ≥ p, we are in case (iv′′ ), and so V has
no 1-dimensional Fp G-submodule. Thus rk(Z) ≥ rk(B) ≥ 2, and hence dim(V ) ≥ p + 1. If
dim(V ) = p + 1, then by Lemma 3.5, every nontrivial Fp G-submodule has nontrivial action
of U, contradicting the assumption B ≤ Z. If dim(V ) ≥ p + 2, then V is simple by [CrOS,
Proposition 3.7(c)]. So Condition 4.5(b) holds in all cases.
If rk(A) ≥ p, then V0 = V , A/Fr(A) ∼
= V0 , and so [G, A] = A since [G, V0 ] = V0 . If
rk(A) = p − 1, then by Lemma A.5(a,b,c), Λ0 |U ∼
= Zp [ζ] as Zp U-modules, and the radical of
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
29
Λ0 |U has index p. Thus pΛ is contained in the radical, and in [G, A] = A since V0 ∼
= Λ0 /pΛ
and [G, V0 ] = V0 . This proves 4.5(c).
By Theorem 4.5, there is a unique simple fusion system F over A ⋊ U such that G =
AutF (A), A ∈ EF , and EF r {A} is as described in Table 4.4. Since A is unique (up
to isomorphism of Zp G-modules) by Proposition 3.9(a) (in case 4.9(b)) or 3.9(b) (in case
4.9(a)), this shows that F is uniquely determined by V .
Case 4.9(c): Assume that we are in Case (c) in Proposition 4.9. In particular, dim(V ) =
p−1 and dim(V0 ) = 1. By Lemma 3.4(c), there is a projective, minimally active Fp G-module
W > V such that dim(W ) = p and thus dim(W/V ) = 1.
By Proposition 3.8(b), there is a Zp G-lattice Λ such that Λ/pΛ ∼
= W , and such that Λ
has a Zp G-submodule Λ0 = CΛ (G0 ) of rank 1. In particular, Λ is free as a Zp U-module
since W is free as an Fp U-module. Let ΛV < Λ be the Zp G-sublattice of index p such that
ΛV /pΛ ∼
= V . Define
b = Λ pk Λ + pk−1ΛV + pΛ0 ∼
A
= Cpk × (Cpk−1 )p−2 × Cp .
b is a p-dimensional Fp G-module, and contains a 2-dimensional submodule
Then Ω1 (A)
k
b0 = pk−1 Λ + Λ0
A
p Λ + pk−1 ΛV + pΛ0
∼
= pk−1Λ/(pk Λ + pk−1 ΛV ) ⊕ Λ0 /pΛ0 ∼
= (W/V ) ⊕ V0 .
Now, V0 ∼
= W/V as Fp [NG (U)]-modules by Lemma 3.4(a) and since dim(W ) = p. Also,
′
O (G) acts trivially on each of them and G = O p (G)NG (U) by the Frattini argument, so
V0 and W/V are isomorphic as Fp G-modules, and any Fp -linear isomorphism is Fp G-linear.
Hence for fixed a ∈ Λ r ΛV and fixed ℓ prime to p such that ℓ 6≡ 1 (mod p) if k = 2, the
quotient group
b A
b1
b1 = [pk−1 a − ℓσ·a] ≤ A
b0
A = A/
where
A
p′
b where the two summands of A
b0 have been identified, and hence is a
is a quotient group of A
P
b1 is independent of the choice
Zp G-module. Here, as usual, σ = u∈U u ∈ Zp U. Note that A
of a, and depends on ℓ only modulo p.
b A
b0 ∼
Since A/
= Λ/(pk−1Λ + Λ0 ), where Λ/Λ0 ∼
= Zp [ζ] as Zp U-modules by Lemma A.5(c),
k−2
we have that |CA/
(u − 1)p−2 a for 1 6= u ∈ U.
b0 (U)| = p, and is generated by the class of p
b A
b
The class of pk−2 (u − 1)p−2 a in A is fixed by U if and only if as classes in A,
b1
[pk−2 (u − 1)p−1 a] = [pk−2 (σ − p)a] = [pk−2 σ·a] − [pk−1 a] ∈ A
(where the second equality holds by Lemma 4.8). But this fails to hold under our hypotheses:
b or k = 2 and ℓ 6≡ 1 (mod p),
either k > 2, in which case [pk−2 σ·a] = 0 and [pk−1 a] 6= 0 in A;
b1 . Thus in all such cases, CA (U) = A
b0 /A
b1 and has order p.
in which case [σ·a] − [pa] ∈
/A
b0 /A
b1 = CA (U) as Fp G-modules, and since V and Ω1 (A) are
Since V0 ∼
= A
= Λ0 /pΛ0 ∼
both (p − 1)-dimensional minimally active Fp G-modules, Lemma 3.4(d) applies to show that
V ∼
= Ω1 (A) as Fp G-modules.
By construction, (G, A) satisfies all of the conditions in case (c) of Proposition 4.9, as
well as condition (d.iii′′ ) of Theorem 4.5. Since |Z| = |CA (U)| = p, conditions (a) and (b)
in 4.5 also hold: |Z0 | = p, and no nontrivial subgroup of Z is G-invariant except possibly
Z0 = Z. Finally, 4.5(c) holds ([G, A] = A) since A/[U, A] = A/Ωk−1 (A) has order p, and
has nontrivial action of G since G acts nontrivially on Z0 = ℧k−1 (A) (since µA (G∨ ) ≥ ∆−1 ).
30
BOB OLIVER AND ALBERT RUIZ
By Theorem 4.5, there is a unique simple fusion system F over A ⋊ U such that G =
AutF (A), A ∈ EF , and EF = {A}∪H. Since A is unique up to isomorphism of Zp G-modules
by Proposition 3.9(b), this shows that F is uniquely determined by V .
5. Simple fusion systems over nonabelian discrete p-toral groups with
abelian subgroup of index p
We now focus on the case where A and S are infinite. Since most of the results in Section
2 assume Notation 2.1, and in particular that S contains a unique abelian subgroup of index
p, we begin by proving that this always holds when Op (F ) = 1.
Lemma 5.1. Let F be a saturated fusion system over an infinite discrete p-toral group S
with an abelian subgroup A of index p. Assume also that Op (F ) = 1. Then |A/Z(S)| = ∞.
Proof. Let S0 E S denote the identity component of S. If |A/Z(S)| < ∞, then S0 ≤ Z(S),
so S0 is contained in (and is characterisitic in) each P ∈ EF . Thus S0 E F by Proposition
1.3(c), so S0 ≤ Op (F ) = 1, contradicting the assumption that |A| = ∞.
Corollary 5.2. If F is a saturated fusion system over an infinite discrete p-toral group S
with an abelian subgroup A of index p, and Op (F ) = 1, then A is the only abelian subgroup
of index p in S.
Proof. Since |A/Z(S)| = ∞ by Lemma 5.1, this follows from Lemma A.3.
Lemma 5.3. Assume Notation 2.1. If |A| = ∞ and Op (F ) = 1, then A/Z and S ′ are both
discrete p-tori of rank p − 1.
Proof. Since |A/Z| = ∞ by Lemma 5.1 and CA/Z (U) = Z2 /Z has order p by Lemma 2.4,
A/Z ∼
= (Z/p∞ )p−1 by Lemma A.5(d). Also, A/Z ∼
= S ′ by Lemma A.3.
Lemma 5.4. Assume Notation 2.1. If |A| = ∞ and Op (F ) = 1, then A = ZS ′ . As one
consequence, each of the sets B and H consists of one S-conjugacy class.
Proof. Fix a generator u ∈ U, and define χ : A/Z −−−→ A/Z by setting χ(aZ) = [a, u]Z.
Then Im(χ) = ZS ′ /Z, and Ker(χ) = Z2 /Z has order p by Lemma 2.4. Since A/Z is a
discrete p-torus by Lemma 5.3, χ must be onto, and hence ZS ′ = A.
Thus if P = Zhxi and Q = Zhyi are two members of H, where yx−1 ∈ A, then there are
z ∈ Z and a ∈ A such that yx−1 = axa−1 x−1 z. Then y ∈ axZ, so Q = aP , and P and Q are
S-conjugate. A similar argument shows that all members of B are S-conjugate.
Part of the next lemma follows from Lemma 2.7 when p is odd. But since we also need it
here when p = 2, we prove it independently of the earlier lemma.
Lemma 5.5. Assume Notation 2.1, and also that |A| = ∞, Op (F ) = 1, and A ∈
/ EF . Then
Z = Z0 , EF = H, and A = S ′ is a discrete p-torus of rank p − 1.
Proof. Since A ∈
/ EF , Lemmas 2.2 and 2.3 imply that EF = B or EF = H. If EF = B, then
since Z = Z(S) is normalized by AutF (S) and by AutF (P ) for each P ∈ B (Z is characteristic
in P by Lemma 2.5(b)), Z E F by Proposition 1.3(c), contradicting the assumption that
Op (F ) = 1. Thus EF = H.
′
For P = Zhxi ∈ H, by Lemma 2.5(a), P = P1 × P2 , where P1 = CP (O p (AutF (P ))) < Z,
Z = P1 × Z0 , and Z0 < P2 ∼
= Cp × Cp , and where each factor Pi is normalized by AutF (P ).
If P ∗ ∈ H is another member, then P ∗ = gP for some g ∈ S by Lemma 5.4, and P1 =
′
g
P1 = CgP (O p (AutF (gP ))) is also normalized by AutF (gP ). Finally, P1 is normalized by
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
31
AutF (S) since AutF (S) = Inn(S)·NAutF (S) (P ) by the Frattini argument, and thus P1 E F
by Proposition 1.3(c). So P1 ≤ Op (F ) = 1, and hence Z = Z0 .
Since Z = Z0 , we have A = S ′ by Lemma 5.4, and so A is a discrete p-torus of rank p − 1
by Lemma 5.3.
When A is finite and p is odd, it was shown in [Ol, Lemma 2.4] that Op (F ) = 1 and
A∈
/ EF imply Z = Z0 . When A is finite and p = 2, this is not true: for each odd prime
p, the 2-fusion system of PSL2 (p2 )hφi, where φ is a field automorphism of order 2, is a
counterexample. Note that in this case, S ∼
= D × C2 for some dihedral 2-group D (whose
order depends on p).
The case p = 2 is now very easy to handle.
Theorem 5.6. Let S be an infinite nonabelian discrete 2-toral group with abelian subgroup
A < S of index 2. Let F be a saturated fusion system over S such that O2 (F ) = 1. Then F
is isomorphic to the 2-fusion system of SO(3) (if A is not F -essential) or of PSU(3) (if A
is F -essential).
Proof. Recall that A is the unique abelian subgroup of index 2 in S by Corollary 5.2. So we
can use Notation 2.1. Also, |Z0 | = 2 by Lemma 2.4, and EF r {A} = H or B by Lemmas
2.2 and 5.4 (and since A E F if EF ⊆ {A}).
Case 1: Assume first that A ∈
/ EF . Then by Lemma 5.5, EF = H, Z = Z0 , and A = S ′ is
a discrete p-torus of rank 1. Thus A ∼
= Z/2∞ , where this group is inverted by the action of
S/A. Also, for each P ∈ EF = H, P ∼
= Σ3 . Thus
= C2 × C2 and hence AutF (P ) = Aut(P ) ∼
there is a unique choice of fusion system F on S, and it must be isomorphic to the fusion
system of SO(3).
Case 2: Now assume that A is F -essential, and set G = AutF (A). Then AutS (A) ∈ Syl2 (G)
has order 2, so |G| = 2m for some odd m. By Proposition A.7, we can write G = G1 × G2
and A = A1 × A2 , where Gi acts faithfully on Ai and trivially on A3−i for i = 1, 2, where
|G1 | is odd, G2 ∼
= Σ3 , and A2 ∼
= C2k × C2k for some 1 ≤ k ≤ ∞.
Now, |A2 | = ∞ since A1 ≤ Z and |A/Z| = ∞ (Lemma 5.1). Hence Z0 is not a direct
factor in Z, so H ∩ EF = ∅ by Lemma 2.5(a), and EF = {A} ∪ B. Each α ∈ AutF (S)
normalizes A and hence normalizes A1 . For each P ∈ B and each α ∈ AutF (P ), α(Z) = Z
since Z = Z(P ), α|Z ∈ AutF (Z) extends to α ∈ AutF (S), and thus α(A1 ) = A1 . So A1 E F ,
and A1 ≤ O2 (F ) = 1.
Thus A ∼
= Σ3 . For each P ∈ B = EF r {A}, P ∈
= (Z/2∞ )2 and AutF (A) = G ∼
∞
(Z/2 ) ×C2 Q8 , and the subgroup isomorphic to Q8 is unique. Hence OutF (P ) ∼
= Σ3 is
uniquely determined, and F is determined uniquely by AutF (A). So F is the 2-fusion
system of PSU(3).
We now focus on the cases where p is odd.
Proposition 5.7. Assume Notation 2.1. Assume also that p is odd, |A| = ∞, and Op (F ) =
1. Then A is a discrete p-torus. If rk(A) ≥ p, then Z is also a discrete p-torus, and has
rank rk(A) − p + 1.
Proof. We first apply Lemma 2.8, with A1 = A and A2 the identity component of A. Since
S ′ is a discrete p-torus by Lemma 5.3, we have A2 Z ≥ ZS ′ = A = A1 by Lemma 5.4. So
by Lemma 2.8, A ≤ A2 Z0 = A2 , the last equality since Z0 ≤ S ′ ≤ A2 . Thus A is a discrete
p-torus.
32
BOB OLIVER AND ALBERT RUIZ
Set G = AutF (A) and V = Ω1 (A), and choose 1 6= u ∈ U ∈ Sylp (G). By Lemma 3.2(a,b),
G ∈ Gp∧ (so |U| = p), and V is faithful, minimally active, and indecomposable as an Fp Gmodule. So if dim(V ) = rk(A) ≥ p, then by Lemma 3.3(a,b), the action of u on V has a
Jordan block of length p, and hence dim(Ω1 (Z)) = dim(CV (U)) = rk(A) − p + 1.
Let Z1 , Z2 ≤ Z be such that Z1 is a discrete p-torus, Z2 is a finite abelian p-group, and
Z = Z1 × Z2 . Since rk(A/Z) = p − 1 by Lemma 5.3, rk(Z1 ) = rk(A) − p + 1 = dim(Ω1 (Z)).
So Z2 = 1, and Z = Z1 is a discrete p-torus.
Lemma 5.8. Assume Notation 2.1, and also that p is odd, |A| = ∞, and Op (F ) = 1. Then
no proper nontrivial subgroup of S is strongly closed in F . Thus F is simple if and only if
it contains no proper normal subsystem over S.
Proof. Assume that 1 6= Q ≤ S is strongly closed in F . If Q ≤ Z, then Q is contained in all
F -essential subgroups, so Q E F by Proposition 1.3(c), contradicting the assumption that
Op (F ) = 1. Thus Q Z.
Now, (QZ/Z) ∩ Z(S/Z) 6= 1 since Q E S, so Q ∩ Z2 Z. Fix g ∈ (Q ∩ Z2 ) r Z. Then
Q ≥ [g, S] = Z0 since Q E S.
Fix P ∈ EF r {A} ⊆ B ∪ H (recall A 5 F ). If P ∈ B, then the AutF (P )-orbit of
g ∈ (Q ∩ Z2 ) r Z is not contained in A. If P ∈ H, then the AutF (P )-orbit of Z0 ≤ Q is not
contained in A. So in either case, Q A. Hence Q ≥ [Q, S] ≥ [U, A] = S ′ .
′
Set G0 = O p (G). Since Q ∩ A is normalized by the action of G, and contains [U, A]
where G0 is the normal closure of U in G, Q ≥ [G0 , A]. Since CA (G0 ) ≤ CA (U) = Z and
CA (G0 ) is normalized by G, CA (G0 ) ≤ Z0 ≤ [U, A] by Lemma 2.7. So by Lemma A.2,
[G0 , A] ≥ CA (U) = Z. Thus Q ≥ ZS ′ = A, and so Q = S since Q A.
The last statement is immediate.
Lemma 5.9. Assume Notation 2.1 and 2.9, and also that |A| = ∞ and S splits over A.
(a) The kernel of µ : Aut∨ (S) −→ ∆ does not contain any elements of finite order prime to
p.
(b) Fix Q ∈ B∪H, and set t = 0 if Q ∈ B, t = −1 if Q ∈ H. Assume that µ(Aut∨F (S)) ≥ ∆t .
e ≥ Q ∩ S ′ which are normalized by
Then there are unique subgroups Ze ≤ Z and Q
NAutF (S) (Q), and are such that the following hold.
e and Q
e∼
(i) If Q ∈ H, then Q = Ze × Q
= Cp × Cp .
e Z∩Q
e = Z0 = Z(Q),
e and Q
e is extraspecial of
(ii) If Q ∈ B, then Ze = Z, Q = Z Q,
3
order p and exponent p.
e ∼
(iii) Thus in all cases, Out(Q)
= GL2 (p). If α ∈ Aut(Q) is such that α|Ze = Id,
′
e then α extends to some
e
e
α(Q) = Q, α(Q ∩ A) = Q ∩ A, and α| e ∈ O p (Aut(Q)),
Q
α ∈ NAut∨F (S) (Q).
Proof. (a) Fix α ∈ Ker(µ) of finite order prime to p. Then α induces the identity on Z/Z0
since α ∈ Aut∨F (S), and on Z0 and S/A since µ(α) = (1, 1). In particular, α|Z = Id by
Lemma A.1, and α|A is Zp U-linear.
Fix x ∈ S r A, and let ψ ∈ End(A) be the homomorphism ψ(g) = [g, x]. Then ψ
commutes with α|A since α(x) ∈ xA, and ψ induces an injection from Zi (S)/Zi−1(S) into
Zi−1 (S)/Zi−2(S) for each i ≥ 2. Since α|Z = Id, this shows that α induces the identity on
Zi (S)/Zi−1(S) for each i, and hence that α|Zi(S) = Id for each i by Lemma A.1 again. Thus
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
33
α|Ω1 (A) = Id since Ω1 (A) ≤ Zi (S) for some i, so α|Ωm (A) = Id for each m ≥ 1 by [G, Theorem
5.2.4]. So α|A = Id, and α = IdS by Lemma A.1 again.
(b) This proof is essentially the same as that of [CrOS, Lemma 2.6(b)] (a similar result but
with |A| < ∞). We sketch an alternative argument here.
Set K = NAut∨F (S) (Q) for short. By the Frattini argument and since all members of
the Aut(S)-orbit of Q are S-conjugate to Q (Lemma 5.4), Aut∨F (S) = Inn(S)·K. Hence
µ(K) ≥ ∆t . Also, |NInn(S) (Q)| ≤ |NS (Q)/Z| < ∞ since |NS (Q)/Q| = p by Lemma 2.2, and
NInn(S) (Q) = AutNS (Q) (S) is normal of index prime to p in K since Inn(S) E AutF (S) has
finite index prime to p by the Sylow axiom.
By the Schur-Zassenhaus theorem, there is K0 < K of order prime to p such that K =
e = [K0 , Q].
K0 ·AutNS (Q) (S). Set Ze = CQ (K0 ) and Q
• If Q ∈ H and µ(K0 ) = µ(K) ≥ ∆−1 , then K0 acts nontrivially on Q/Z and on Z0 , and
e ∩ A = Z0 . Also, Q = Ze × Q
e and Z = Ze × Z0
trivially on Z/Z0. Hence Ze ≤ Z and Q
e∼
by [G, Theorem 5.2.3] (applied to the subgroups Ωm (Q) for m ≥ 1), and Q
= Cp ×Cp
∼
e
since S splits over A. In particular, Out(Q) = GL2 (p).
If β ∈ K0 is such that µ(β) generates ∆−1 , then µ(β) = (r, r −1) for some generator
0
e∼
for an appropriate
r of (Z/p)× , so β|Qe acts on Q
= Cp × Cp as the matrix 0r r−1
p′
e
e ∼
choice of basis. Thus AutS (Q)hβ|Qe i = NOp′ (Aut(Q))
e (Z0 ) where O (Aut(Q)) = SL2 (p),
proving (iii) in this case.
• If Q ∈ B and µ(K0 ) = µ(K) ≥ ∆0 , then α acts nontrivially on Q/Z2 , and trivially on
Z0 and Z/Z0. Hence α acts trivially on Z by Lemma A.1, nontrivially on Z2 /Z, and
e ∩ A = Z2 ∩ S ′ . Also, Q
e is extraspecial of order p3 and (since S
so Ze = Z and Q
e ∼
splits over A) exponent p. In particular, Out(Q)
= GL2 (p). By [G, Theorem 5.2.3],
e and Ze ∩ Q
e = Z0 .
applied to the abelian p-groups Ωm (Q/Z0 ) for m ≥ 1, Q = ZeQ
If β ∈ K0 is such that µ(β) generates ∆0 , then µ(β) = (r, 1) for some gener′
e
e ∼
Q)
ator r of (Z/p)× , so [β|Qe ] has order (p−1) in O p (Aut(Q))/Inn(
= SL2 (p). So
e
AutS (Q)hβ|
e i = N p′
e (Q ∩ A) in this case, again proving (iii).
Q
O (Aut(Q))
e ≥ Q∩S ′ ≥ [NS (Q), Q] in all cases, we have Ze = CQ (K)
Since Ze ≤ Z ≤ CQ (NS (Q)) and Q
e = [K, Q]. Thus Ze and Q
e are independent of the choice of K0 , and are normalized by
and Q
∨
K = NAut∨F (S) (Q). Since AutF (S) is normal in AutF (S) (the kernel of a homomorphism to
e are also normalized
Aut(Z/Z0)), we see that K is normal in NAutF (S) (Q), and hence Ze and Q
by NAutF (S) (Q). These are easily seen to be the unique subgroups that satisfy the required
conditions.
When F is a fusion system over a discrete p-toral group S, then for each Q ≤ S, we define
another subgroup Q• ≤ S as follows. Let T be the identity component of S. If m ≥ 0 is
m
m
the smallest integer such that g p ∈ T for each g ∈ S, and Q[m] = ℧m (Q) = hg p | g ∈ Qi,
def
then Q• = Q·I(Q[m] )0 , where I(Q[m] ) = CT (CAutF (T ) (Q[m] )) and I(Q[m] )0 is its identity
component. Thus Q ≤ Q• ≤ QT for each Q. See [BLO3, Definition 3.1] or [BLO6, Definition
3.1] for more detail, as well as the motivation for this construction.
Lemma 5.10. Assume the notation and hypotheses of 2.1, and also that |A| = ∞. For each
Q ∈ B ∪ H ∪ {A, S}, Q• = Q.
34
BOB OLIVER AND ALBERT RUIZ
Proof. By Proposition 5.7, A is the identity component of S. Thus A and G play the role
of T and W = AutF (T ) in [BLO6, Definition 3.1]. Since Q ≤ Q• ≤ QA for each Q ≤ S, we
have A• = A and S • = S.
Now assume Q ∈ H∪B. By assumption, S/A has exponent p = p1 . Since Q/Z ∼
= Cp or Cp2
def
def
(Lemma 2.5), Q[1] = hg p | g ∈ Qi ≤ Z. Hence CG (Q[1] ) ≥ U, and I(Q[1] ) = CA (CG (Q[1] )) ≤
CA (U) = Z. It follows that Q• ≤ Q·I(Q[1] ) = Q.
We are now ready to prove our main theorem used to construct simple fusion systems over
infinite discrete p-toral groups with abelian subgroup of index p.
Theorem 5.11. Fix an odd prime p. Let S be an infinite nonabelian discrete p-toral group
which contains an abelian subgroup A E S of index p, and let F be a simple fusion system
over S. Assume Notation 2.1 and 2.9 (where the uniqueness of A follows from Corollary
5.2). Then the following hold:
(a) U ∈ Sylp (G) and S splits over A.
(b) A is a discrete p-torus of rank at least p − 1, Z0 = CA (U) ∩ [U, A] has order p, and
A = CA (U)·[U, A].
(c) There are no non-trivial G-invariant subgroups of Z = CA (U), aside (possibly) from
Z0 .
(d) Either
′
(i) EF r {A} = H, rk(A) = p − 1, µA (Aut∨F (A)) ≥ ∆−1 , and G = O p (G)·µ−1
A (∆−1 );
or
(ii) EF r {A} = B, rk(A) ≥ p − 1, µA (Aut∨F (A)) ≥ ∆0 , µA (Aut∨F (A)) = ∆0 if
′
rk(A) ≥ p, G = O p (G)·µ−1
A (∆0 ), and Z0 is not G-invariant.
Here, we regard µA as a homomorphism defined on Aut∨F (A).
Conversely, let S be an infinite discrete p-toral group containing a unique abelian subgroup
A E S of index p, let G ≤ Aut(A) be such that AutS (A) ∈ Sylp (G), and adopt the notation in
2.1 and 2.9. Assume that (a)–(d) hold, with Aut∨F (A) replaced by G∩Aut∨ (A) and EF r{A}
replaced by E0 = H or B in (d). Then there is a unique simple fusion system F over S such
that G = AutF (A) and EF r {A} = E0 .
Proof. We prove in Steps 1 and 3 that conditions (a)–(d) are necessary, and prove the
converse in Step 2.
Step 1: Assume that F is a simple fusion system over S. We must show that conditions
(a)–(d) hold. By Corollary 5.2, A is the unique abelian subgroup of index p in S.
(a,b,c) Point (a) holds by Corollary 2.6 and since A is fully automized. The last two
statements in (b) hold by Lemmas 2.4 and 5.4, and (c) holds by Lemma 2.7 and since
Op (F ) = 1. Finally, A is a discrete p-torus by Proposition 5.7, and rk(A) ≥ p − 1 by Lemma
5.3.
(d) Since A 5 F , there is P ∈ EF ∩ (H ∪ B). Set t = 0 if P ∈ B, and t = −1 if P ∈ H.
′
Set H = AutF (P ) and H0 = O p (H). By Lemma 2.5, H0 /Inn(P ) ∼
= SL2 (p), and acts
trivially on Z/Z0. Since AutS (P ) ∈ Sylp (H) = Sylp (H0 ), we can choose α ∈ NH0 (AutS (P ))
of order p−1 in H/Inn(P ). By the extension axiom, α extends to an element of AutF (NS (P )),
and since P is maximal among F -essential subgroups by Lemmas 2.2 and 2.3, α = α
b|P for
some α
b ∈ AutF (S). Set α0 = α
b|A ∈ G.
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
35
Now, α0 induces the identity on Z/Z0 since α does, and α0 ∈ NG (U) since it extends to
S. Thus α0 ∈ Aut∨F (A). By Lemma 2.5, α acts as an element of SL2 (p) on P/Z ∼
= Cp2 (if
2
×
P ∈ B) or on P/P1 ∼
= Cp where Z = P1 × Z0 (if P ∈ H). Hence for some s ∈ (Z/p) of order
p − 1, µA (α0 ) = (s, s−1 ) if P ∈ H, or µA (α0 ) = (s, 1) if P ∈ B. Thus µA (Aut∨F (A)) ≥ ∆t in
either case.
Assume rk(A) ≥ p. By Proposition 5.7, Z is a discrete p-torus. Hence each element of
Aut∨F (A) induces the identity on Z0 since it induces the identity on Z/Z0, and µA (Aut∨F (A)) ≤
∆0 .
′
∨
Set G0 = O p (G)·µ−1
A (∆t ). Since Ker(µA |AutF (A) ) = U by Lemma 5.9(a), and since
∨
µA (AutF (A)) ≥ ∆t by assumption, we have G ≥ µ−1
A (∆t ), and hence G ≥ G0 . We will
show in Step 3 (with the help of the constructions in Step 2) that G = G0 , thus finishing
the proof of (d).
Step 2: Now assume that S, A, and G are as above, and set G∨ = G ∩ Aut∨ (A). Assume
that (a) and (b) hold, and also that µA (G∨ ) ≥ ∆t for some t ∈ {0, −1}. (Note that
G∨ ≤ NG (U) since each element is the restriction of an automorphism of S.) We must show
that these are realized by a unique saturated fusion system F , which is simple if (c) and (d)
hold. Set E0 = H if t = −1, or E0 = B if t = 0.
Set Γ = A ⋊ G, and identify S = A ⋊ U ∈ Sylp (Γ). Choose a generator x ∈ U < S. Set
Z = Z(S), Z2 = Z2 (S), as in Notation 2.1.
Set Q = Zhxi if E0 = H, or Q = Z2 hxi if E0 = B. Thus Q ∈ E0 , and each member of E0
is S-conjugate to Q by Lemma 5.4. Set K = AutΓ (Q). By assumption, there is α ∈ NG∨ (U)
such that µA (α) generates ∆t , and α extends to some α ∈ Aut∨ (S) such that α(x) ∈ U. In
particular, α(Q) = Q, and α ∈ AutΓ (S).
By Lemma 5.9(b), applied with FS (Γ) in the role of F , there are unique subgroups Ze ≤ Z
e ≥ Q∩S ′ , both normalized by NAut (S) (Q), and such that Q = Ze × Q
e and Q
e∼
and Q
= Cp ×Cp
Γ
3
e
e
e
e
if Q ∈ H; or Z = Z, Q = Z Q, Z ∩ Q = Z0 , and Q is extraspecial of order p and exponent
p if Q ∈ B. Let Θ ≤ Aut(Q) be the unique subgroup containing Inn(Q) that acts trivially
e normalizes Q,
e and is such that Θ/Inn(Q) ∼
on Z,
= SL2 (p).
We next claim that
(1) each α ∈ AutΓ (Q) extends to some α ∈ NAutΓ (S) (Q);
(2) AutΓ (Q) normalizes Θ;
(3) AutS (Q) ∈ Sylp (Θ) = Sylp (ΘAutΓ (Q)); and
(4) NΘ (AutS (Q)) ≤ AutΓ (Q).
Point (1) holds since S = QA where A E Γ, and hence NΓ (Q) ≤ NΓ (S). By assumption,
e and hence normalizes Θ. Thus (2) follows
each element of NAutΓ (S) (Q) normalizes Ze and Q,
from (1). Since |OutS (Q)| = |NS (Q)/Q| = p by Lemma 2.2, this acts trivially on Z ≥ Ze
e and Out(Q)
e ∼
and normalizes Q,
= GL2 (p) where Sylp (GL2 (p)) = Sylp (SL2 (p)), we see that
OutS (Q) ∈ Sylp (Θ/Inn(Q)) and hence that AutS (Q) ∈ Sylp (Θ). Also, Θ has index prime
to p in ΘAutΓ (Q) since AutS (Q) ∈ Sylp (AutΓ (Q)), and this proves (3). Finally, (4) follows
from Lemma 5.9(b.iii).
Set F = hFS (Γ), Θi: the smallest fusion system over S which contains FS (Γ) and such
that AutF (Q) ≥ Θ. Set K = {S, A} ∪ E0 . Then K is invariant under F -conjugacy, and
is closed in the space of all subgroups of S [BLO6, Definition 1.11]. Thus condition (i) in
36
BOB OLIVER AND ALBERT RUIZ
[BLO6, Theorem 4.2] holds for K; and condition (iii) holds (P ∈ K and P ≤ Q ≤ P • imply
Q ∈ K) since P = P • for each P ∈ K (Lemma 5.10).
By Lemma 2.3, if E0 = B, then the members of H are not F -centric. So in all cases, if
P ≤ S is F -centric and P ∈
/ K, then P is not contained in any member of EF = E0 ∪ {A},
and hence OutS (P ) E OutF (P ). This proves condition (iv) in [BLO6, Theorem 4.2]:
Op (OutF (P )) ∩ OutS (P ) 6= 1
whenever P ≤ S is F -centric and not in K.
We refer to [BLO6, Definition 1.11] for the definitions of “K-generated” and “K-saturated”.
By construction, F is K-generated. To show that F is K-saturated, we must prove that each
P ∈ K is fully automized and receptive in F (Definition 1.1). If P = A or P = S, then
AutF (P ) = AutΓ (P ), and this is easily checked. So it remains to show this when P = Q.
By (2), AutF (Q) = Θ·AutΓ (Q). So Q is fully automized by (3). If α ∈ NAutF (Q) (AutS (Q)),
then α ∈ AutΓ (Q) by (4), and hence extends to some α ∈ AutΓ (S) by (1). So Q is also
receptive. This finishes the proof of condition (ii) in [BLO6, Theorem 4.2], and hence F is
saturated by that theorem.
Now assume (c) and (d) hold; we must prove that F is simple. By (c), there are no
non-trivial G-invariant subgroups of Z except possibly for Z0 . Also, EF ⊇ H in case (d.i),
and Z0 is not G-invariant in case (d.ii). Hence Op (F ) = 1 by Lemma 2.7. By Lemma 5.8,
F is simple if there are no proper normal fusion subsystems in F over S.
Assume F0 E F is a normal fusion subsystem over S, and set G0 = AutF0 (A). Then G0 E
′
G, and G0 ≥ O p (G) since it is the normal closure of U = AutS (A). Also, µA (Aut∨F0 (A)) ≥ ∆t
by Step 1, applied with F0 in the role of F . Since µA is injective on G∨ /AutS (A) by Lemma
′
5.9(a), we have G0 ≥ O p (G)·µ−1
A (∆t ) = G. Thus G0 = G, and AutF0 (S) = AutF (S) by the
extension axiom, so F0 = F by the Frattini condition on a normal subsystem (see Definition
1.4). This finishes the proof that F is simple.
e in Lemma 5.9(b).
The uniqueness of F follows from the uniqueness of Ze and Q
Step 3: We return to the situation of Step 1, where it remains only to prove that G0 = G.
By Step 2, there is a unique saturated fusion subsystem F0 ≤ F over S such that EF0 = EF
and AutF0 (A) = G0 . The invariance condition on F0 ≤ F (Definition 1.4) holds by the
′
uniqueness of F0 , and the Frattini condition holds since G = O p (G)NG (U) ≤ G0 NG (U)
(where each element of NG (U) extends to an element of AutF (S) by the extension axiom).
Thus F0 E F , so F0 = F since F is simple, and hence G0 = G.
As a special case, we next show that for each prime p, there is (up to isomorphism) a
unique simple fusion system over an infinite discrete p-toral group with abelian subgroup of
index p which is not essential.
Theorem 5.12. For each odd prime p, there is, up to isomorphism, a unique simple fusion
system F over an infinite nonabelian discrete p-toral group S which contains an abelian
subgroup A < S of index p that is not F -essential. The following hold for each such p, F ,
S, and A:
(a) The group A is a discrete p-torus of rank p − 1, and S splits over A. Also, AutF (A) ∼
=
Cp ⋊ Cp−1, OutF (S) ∼
= Cp−1, and EF = H (defined as in Notation 2.1).
e < Γ be the subgroup of classes of diagonal
(b) Fix a prime q 6= p, set Γ = PSLp (Fq ), let A
e xi ∈ Sylp (Γ) for some permutation matrix
matrices of p-power order, and set Se = Ahe
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
37
x
e of order p. Then there is an isomorphism S ∼
= Se that restricts to an isomorphism
∼
∼
e and induces isomorphisms AutF (S) = AutΓ (S)
e and AutF (A) ∼
e
A = A,
= AutNΓ (S)
e (A).
(c) If p = 3, then F is isomorphic to the 3-fusion system of PSU(3), and also to the 3-fusion
system of PSL3 (Fq ) for each prime q 6= 3. For p ≥ 5, F is not realized by any compact
Lie group, nor by any p-compact group.
Proof. We use the notation of 2.1 and 2.9. In particular, G = AutF (A).
(a,b) Assume F is a simple fusion system over an infinite discrete p-toral group S with an
abelian subgroup A < S of index p such that A ∈
/ EF . By Lemma 5.5, Z = Z0 , EF = H,
′
and A = S is a discrete p-torus of rank p − 1. In particular, |Z| = |Z0 | = p. Also, S splits
over A by Corollary 2.6.
It remains to describe G = AutF (A) and AutF (S) and prove (b). Since A ∈
/ EF , U E G,
and OutF (S) ∼
G/U.
(Each
α
∈
Aut
(A)
extends
to
Aut
(S)
by
the
extension
axiom.)
=
F
F
Since S splits over A, each α ∈ NAut(A) (U) extends to an automorphism of S. Since
Z = Z0 , this implies that NAut(A) (U) = Aut∨ (A). Also, µA (G) = ∆−1 by Theorem 5.11(d)
′
and since O p (G) = U ≤ Ker(µA ).
Let R = Zp [ζ] and p = (1 − ζ)R be as in Notation 4.6, regarded as Zp U-modules. By
Proposition A.4, A ∼
= HomZp (Qp /Zp , A) for some (p−1)-dimensional
= (Qp /Zp )⊗Zp Λ and Λ ∼
∼
Zp G-lattice Λ, and Λ|U = R as Zp U-modules by Lemma A.5(c). These isomorphisms induce
isomorphisms of automorphism groups
Aut∨ (A) = NAut(A) (U) ∼
= CAut(R) (U) ⋊ Gal(Qp (ζ)/Qp )
= NAut(R) (U) ∼
∼
= (1 + p) × F×
= R× ⋊ Gal(Qp (ζ)/Qp ) ∼
p ⋊ Gal(Qp (ζ)/Qp ),
and these send Ker(µA ) ≤ Aut∨ (A) (the group of automorphisms of A that commute with
∨
U and are the identity on Z = Z0 ) onto 1 + p, and send the subgroup sc(F×
p ) ≤ Aut (A) of
scalar multiplication by (p − 1)-st roots of unity onto F×
p . Thus
Aut∨ (A) = NAut(A) (U) = Ker(µA ) × sc(F×
p) ⋊W
∨
for a certain subgroup W ∼
= Cp−1 . Set G = (U × sc(F×
p ))·W < Aut (A): a subgroup of order
p(p − 1)2 .
Since Aut∨ (A)/Ker(µA ) has order (p − 1)2 , and since Ker(µA ) ∼
= (1 + p) is an abelian prop-group and hence uniquely m-divisible for each m prime to p, we have H i (H; Ker(µA )) = 0
for each H ≤ Aut∨ (A)/Ker(µA ) and each i > 0. Hence for each subgroup K < Aut∨ (A) of
order prime to p, K ∩ Ker(µA ) = 1 since Ker(µA ) is a pro-p-group, Ker(µA )·K splits over
Ker(µA ) with a splitting unique up to conjugacy, and thus K is conjugate by an element of
Ker(µA ) to a subgroup of sc(F×
p )·W . In particular, G is conjugate to a subgroup of G, and we
can assume (without changing the isomorphism type of F ) that G ≤ G. Finally, one easily
sees that µA sends G onto ∆ with kernel U, and hence that G = (µA |G )−1 (∆−1 ) ∼
= Cp ⋊ Cp−1
is uniquely determined.
e
e is most easily seen by identifying Γ = PGLp (Fq ), so that A
A natural isomorphism A ∼
=A
∞ p
∞ ∼
is the quotient of (Z/p ) by the diagonal Z/p = Op (Z(Γ)), with the permutation action
e xi, and
e ∼
of AutΓ (A)
= A ⋊ U with Se = Ahe
= Σp . This then extends to an isomorphism of S ∼
e
of AutF (A) with AutNΓ (S)
e (A).
Existence and uniqueness of F: Let A, U E G, and S = A ⋊ hxi be as described in the
proof of (b). Since µA (G) = ∆−1 , conditions (a)–(d) in Theorem 5.11 all hold with EF = H.
38
BOB OLIVER AND ALBERT RUIZ
So such an F exists by that theorem. It is unique up to isomorphism by the uniqueness in
the theorem and by the restrictions shown in the proof of (b).
(c) If F is realized by a compact Lie group or a p-compact group, then by Proposition 1.5
and since all elements in S are F -conjugate to elements in A, F is realized by a connected,
simple p-compact group, and the action of the Weyl group AutF (A) on Q⊗Z Hom(A, Qp /Zp )
is generated by pseudoreflections. But if p ≥ 5, then AutF (A) ∼
= Cp ⋊ Cp−1 contains no
pseudoreflections other than the identity. So p = 3, and we easily check that F is realized
by PSU(3), or by PSL3 (Fq ) for q 6= 3.
We can now describe the simple fusion systems over discrete p-toral groups with discrete
p-torus of index p in terms of the classification of certain faithful, minimally active, indecomposable modules carried out in [CrOS].
Theorem B. Fix an odd prime p.
(a) Let F be a simple fusion system over an infinite nonabelian discrete p-toral group S
with an abelian subgroup A < S of index p. Assume also that A is F -essential. Set
G = AutF (A) and V = Ω1 (A), let H and B be as in Notation 2.1, and let G∨ = Aut∨F (A)
and µA : G∨ −→ ∆ be as in Notation 2.9. Then A is a discrete p-torus, S splits over
A, G ∈ Gp∧ , and for some t ∈ {0, −1},
V is a faithful, minimally active, indecomposable Fp G-module. Either
′
dim(V ) = p − 1, µA (G∨ ) ≥ ∆t and G = O p (G)µ−1
A (∆t ); or dim(V ) ≥
′
p, t = 0, µA (G∨ ) = ∆0 , and G = O p (G)·G∨. Also, EF = {A} ∪ H if
t = −1, while EF = {A} ∪ B if t = 0. If t = 0, then V contains no
1-dimensional Fp G-submodule.
(∗∞ )
(b) Conversely, assume that G ∈ Gp∧ , U ∈ Sylp (G), and t ∈ {0, −1}, and that V is an Fp Gmodule that satisfies (∗∞ ), where G∨ is the subgroup of all elements α ∈ NG (U) such
that [α, CV (U)] ≤ [U, V ]. Then there are a discrete G-p-torus A and a simple fusion
system F over S = A ⋊ U such that AutF (A) = AutG (A) ∼
=V
= G, such that Ω1 (A) ∼
as Fp G-modules, and such that EF = {A} ∪ H if t = −1, or EF = {A} ∪ B if t = 0.
Furthermore, any other simple fusion system with these properties is isomorphic to F .
(c) Among the fusion systems specified in (b), the only ones that are realized as fusion
systems of compact Lie groups or of p-compact groups are those listed in Table 5.1.
Proof. (a) Set U = AutS (A) ∈ Sylp (G) and Z = CA (U).
Under the above assumptions, A is a discrete p-torus by Proposition 5.7, G ∈ Gp∧ by
Lemma 3.2(a), V is faithful, minimally active, and indecomposable by Lemma 3.2(b), and
rk(V ) = rk(A) ≥ p − 1 by Lemma 5.3. Also, for some t ∈ {0, −1}, µA (G∨ ) ≥ ∆t and
′
G = O p (G)µ−1
A (∆t ), and EF is as described in (∗∞ ), by Theorem 5.11(d). Since A 5 F , S
splits over A by Corollary 2.6. If t = 0 and V0 < V is a 1-dimensional Fp G-submodule, then
V0 ≤ CV (U) ≤ Z, which is impossible by Theorem 5.11(c,d.ii).
Assume rk(V ) = rk(A) ≥ p. Since V is minimally active and indecomposable, V |U is the
direct sum of a free module Fp U and an Fp -vector space with trivial U-action by Lemma
3.3, and hence Ω1 (Z) = CV (U) has rank rk(A) − p + 1. Also, Z is a discrete p-torus by
Proposition 5.7, so for each α ∈ G∨ = Aut∨F (A), α acts via the identity on Z0 since it acts
via the identity on Z/Z0 (see Notation 2.9). Thus µA (α) ∈ ∆0 by definition of µA , so t = 0
and µA (G∨ ) = ∆0 in this case, finishing the proof of (∗∞ ).
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
p
conditions
rk(A) G = AutF (A) E0
p
p≥5
p−1
Σp
H
PSU(p)
PSLp (Fq )
p
p < n < 2p
n−1
Σn
B
PSU(n)
PSLn (Fq )
p p ≤ n < 2p, n ≥ 4
n
C2n−1 ⋊ Σn
B
PSO(2n)
P Ω2n (Fq )
p
2 < m | (p − 1)
p ≤ n < 2p, n ≥ 4
n
(Cm )n−1 ⋊ Σn
B
X(m, m, n)
5
n = 6, 7
n
W (En )
B
En
En (Fq )
7
n = 7, 8
n
W (En )
B
En
En (Fq )
3
2
GL2 (3)
B
X12
CF4 (K) (γ)
5
4
(4 ◦ 21+4 ).Σ5
B
X29
5
4
(4 ◦ 21+4 ).Σ6
B
X31
7
6
61 ·PSU4 (3).22
B
X34
39
p-cpct. gp. tors. lin. gp.
E8 (K)
Table 5.1. The sixth column lists a compact Lie group or a p-compact group
that realizes the fusion system F described in the first five columns. Here,
X(m, m, n) denotes the p-compact group with Weyl group G(m, m, n) in the
notation of [ST, § 2], and Xk the one with Weyl group number k in [ST, Table
VII]. In the last column, we give, in some cases, a torsion linear group that
realizes F : q 6= p is prime, K ⊆ F2 is the union of the odd degree extensions
of F2 , and γ ∈ Aut(F4 (K)) is a graph automorphism of order 2. In the fourth
column, B.C means an extension of B by C, and the subscripts in the entry
61 ·PSU4 (3).22 are Atlas notation [At, p. 52].
(b) Fix G ∈ Gp∧ , U ∈ Sylp (G), and t ∈ {0, −1}, and let V be an Fp G-module that satisfies
(∗∞ ), where G∨ is the subgroup of all elements g ∈ NG (U) such that [g, CV (U)] ≤ [U, V ].
By Proposition 3.8(a), there is a Zp G-lattice Λ such that Λ/pΛ ∼
= V as Fp G-modules. Set
V
as
F
G-modules
(see Proposition
A = (Qp /Zp ) ⊗Zp Λ: a discrete G-p-torus where Ω1 (A) ∼
=
p
A.4). To simplify notation, we identify V = Ω1 (A). Set S = A⋊U. Set Z = Z(S) = CA (U),
S ′ = [S, S] = [U, A], and Z0 = Z ∩ S ′ .
We next check that conditions (a)–(d) in Theorem 5.11 all hold. Conditions (a) and (d)
follow immediately from (∗∞ ), and (b) (|Z0| = p) was shown in Lemma 3.7(b).
Assume 1 6= B ≤ Z is G-invariant. If rk(A) = p − 1, then Z = Z0 has order p, so
B = Z0 . Otherwise, by (∗∞ ), t = 0, and V contains no 1-dimensional Fp G-submodule. Thus
dim(Ω1 (B)) ≥ 2, so dim(V ) = rk(A) ≥ p + 1. If dim(V ) ≥ p + 2, then V is simple by [CrOS,
Proposition 3.7(c)], while if dim(V ) = p + 1, then V contains no nontrivial Fp G-submodule
with trivial U-action by Lemma 3.5. Thus B = Z0 , and this proves condition 5.11(c).
By Theorem 5.11, there is a unique simple fusion system F over S such that G = AutF (A),
and EF r {A} = H (if t = −1) or B (if t = 0). Since A is unique (up to isomorphism of
Zp G-modules) by Lemma 3.10, this shows that F is uniquely determined by V .
(c) If F is realized by a compact Lie group or a p-compact group, then by Proposition 1.5 and
since all elements in S are F -conjugate to elements in A, F is realized by a connected, simple
p-compact group, and the action of the Weyl group G = AutF (A) on Q ⊗Z Hom(A, Qp /Zp )
40
BOB OLIVER AND ALBERT RUIZ
is irreducible as a group generated by pseudoreflections. Using the list of pseudoreflection
groups and their realizability over Qp compiled by Clark & Ewing [CE], as well as the
assumption that vp (|G|) = 1, we see that G must be one of the groups listed in Table 5.1,
or else one of the other groups G(m, d, n) (of index d in Cm ≀ Σn ) for d | m | (p − 1) with
′
d < m. The latter are eliminated by the condition G = O p (G)·µ−1
A (∆0 ) in (d.ii) (i.e., the
fusion systems of the corresponding p-compact groups are not simple), and so we are left
with the groups listed in the table.
Since a p-compact group is determined by its Weyl group by [AGMV, Theorem 1.1], it
remains only to check, when rk(A) = p − 1 and based on the constructions of these groups,
whether B ⊆ EF or H ⊆ EF . This situation occurs only in the last four cases listed in the
table, in which cases the p-compact group was constructed by Aguadé [Ag, §§ 5–7, 10], and
the use of SU(p) in his construction shows that extraspecial groups of order p3 and exponent
p appear as essential subgroups.
In those cases where a torsion linear group is given in Table 5.1, it is a union of a sequence
of finite groups that by Table 4.2 realize a sequence of finite fusion subsystems of F .
The different situations handled in Theorem B are partly summarized in Table 5.2.
dim(V ) EF r {A} µA (G∨ )
p−1
≥p
Condition
G=
′
H
≥ ∆−1 O p (G)·µ−1
A (∆−1 )
B
≥ ∆0
B
= ∆0
′
O p (G)µ−1
A (∆0 )
−
V contains no 1-dimensional
Fp G-submodule
Table 5.2.
6. Examples
Recall Definition 3.1: for a given prime p, Gp is the class of finite groups G with U ∈
Sylp (G) of order p and not normal, and Gp∧ is the class of those G ∈ Gp such that AutG (U) =
Aut(U). It remains now to describe explicitly which finite groups G ∈ Gp∧ and Fp G-modules
V can appear in Theorems A and B. This follows immediately from the work already done
in [CrOS], and is stated in Proposition 6.1 and Table 6.1. As in [CrOS], when p is a fixed
prime, we define, for each odd integer i prime to p,
∆i/2 = {(r 2 , r i ) | r ∈ (Z/p)× } .
(Compare with the definition of ∆i in Notation 2.9.)
Proposition 6.1. Assume that G ∈ Gp∧ , and that V is a faithful, minimally active, indecomposable Fp G-module such that dim(V ) ≥ p − 1. If dim(V ) ≥ p, then assume also
that µV (G∨ ) ≥ ∆0 ; and if dim(V ) = p, then assume that V contains no 1-dimensional
Fp G-submodule. Then either
(a) the image of G in PGL(V ) is not almost simple, and G ≤ G with the given action on
V for one of the pairs (G, V ) listed in Table 6.1 with no entry G0 ; or
(b) the image of G in PGL(V ) is almost simple, and G0 ≤ G ≤ G with the given action on
V for one of the triples (G0 , G, V ) listed in Table 6.1.
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
41
In all cases, the entry under dim(V ) gives the dimensions of the composition factors of V ;
thus a single number means that V is simple.
p
G0
p
SL2 (p) or PSL2 (p)
(p ≥ 5)
p
Ap
p
Ap+1
dim(V )
GL2 (p) or
(p−n−1)/n PGL2 (p) × Cp−1
p − 1, p
p An (p+2 ≤ n ≤ 2p−1)
µV (G∨
0)
∆
1
2 ∆0
{(u2 , un−1 )}
∆
∆−1/2 ,
p
Σp+1 × Cp−1
∆
n−1
Σn × Cp−1
∆0
1
2 ∆0
1/(p−2)
(p ≥ 5)
µV (G∨ )
1
2 ∆0
1
2 ∆−1
1
2 ∆0
(p−2)/1
(p ≥ 5)
G
Σp × Cp−1
∆
∆
p
—
n
Cp−1 ≀ Sn (n ≥ p)
∆
—
3
—
2
GL2 (3)
∆
—
5
2·A6
4
4·S6
∆
∆1/2
5
—
4
(C4 ◦ 21+4 ).S6
∆
—
5
Sp6 (2) = W (E7 )′
7
G0 × 4
∆0
1
2 ∆0
1
2 ∆0
7
6·PSL3 (4)
6
G0 .21
∆
×
F×2
p × Fp
7
61 ·PSU4 (3)
6
G0 .22
∆
×
F×2
p × Fp
7
PSU3 (3)
6
G0 .2 × 6
∆
7
PSU3 (3)
7
G0 .2 × 6
∆
7
SL2 (8)
7
G0 :3 × 6
∆
1
2 ∆1
1
2 ∆0
1
3 ∆1
7
G0 × 6
∆
∆3
5
7
PSp4 (3) = W (E6
Sp6 (2) = W (E7
)′
)′
6
W (E6 ) × 4
∆0 .2
7
′
2·Ω+
8 (2) = W (E8 )
8
W (E8 ) × 3
∆0 .2
∆3
11
PSU5 (2)
10
G0 .2 × 10
∆
1
2 ∆2
11
2·M12
10, 10
G0 .2 × 5
∆
∆1/2 , ∆7/2
11
2·M22
10, 10
G0 .2 × 5
∆
∆1/2 , ∆7/2
13
PSU3 (4)
12
G0 .4 × 12
∆
1
3 ∆1
Table 6.1. Pairs (G, V ), where G ∈ Gp∧ , G ≤ G, G ≥ G0 when a quasisimple
group G0 is given, and where V is a minimally active indecomposable module
of dimension at least p − 1, such that µV (G∨ ) ≥ ∆0 if dim(V ) ≥ p, and
such that V does not have a 1-dimensional submodule if dim(V ) = p. In all
cases, dim(V ) gives the dimensions of the composition factors in V . Also,
2
×
F×2
p = {r | r ∈ Fp }. The notation B.C, B:C, and B·C for extensions is as in
the Atlas [At, p. xx], as well as the subscripts used to make precise certain
central extensions or automorphism groups.
Proof. We take as starting point the information in [CrOS, Table 4.1]. We drop from that
table those cases where dim(V ) < p − 1, and also those cases where dim(V ) ≥ p and
µV (G∨ ) ∆0 , or where dim(V ) = p and V contains a 1-dimensional Fp G-submodule.
42
BOB OLIVER AND ALBERT RUIZ
Since the table in [CrOS] is restricted to representations of dimension at least 3, we must
add those representations of dimension 2 that appear. Since dim(V ) ≥ p − 1, this occurs
only for p = 3, and thus G ≤ GL2 (3). Since this group is solvable, the image of G in PGL(V )
cannot be almost simple, and so this case is covered by the unique row of the table restricted
to p = 3.
We now give two examples, in terms of the pairs (G, V ) that appear in Table 6.1, to
illustrate how this table can be used to list explicit fusion systems as described by Theorems
A and B. When V is an Fp -vector space, we set Autsc (V ) = Z(Aut(V )) ∼
= F×
p : the group of
automorphisms given by scalar multiplication.
Example 6.2. Fix an odd prime p ≥ 5 and a finite group G ∈ Gp∧ , and choose U ∈ Sylp (G).
Let V be a simple, (p − 1)-dimensional, minimally active Fp G-module, and assume that
AutG (V ) ≥ Autsc (V ). Then µV (G∨ ) = ∆ by [CrOS, Proposition 3.13(a)]. Let Λ be a
Zp G-lattice such that Λ/pΛ ∼
= V (see Proposition 3.8(a)).
(a) By case (i–a) in Table 4.4, for each k ≥ 2, there is a unique simple fusion systems F
over (Λ/pk Λ) ⋊ U, with AutF (Λ/pk Λ) ∼
= G, and such that EF = {A} ∪ H0 ∪ B∗ .
(b) By case (iv′ –a) in Table 4.4, for each k ≥ 2 and each ∅ 6= I ⊆ {0, 1, . . . , p − 1}, there is
−1
k
k
a unique simple fusion systemSFI over
(Λ/p Λ) ⋊ U, with AutFI (Λ/p Λ) = G0 µV (∆0 ),
and such that EFI = {A} ∪
i∈I Bi .
(c) By case (iii′′ –a) in Table 4.4, for each k ≥ 2, there is a unique simple fusion system F
over (Λ/pk Λ) ⋊ U, with AutF (Λ/pk Λ) = G0 µ−1
V (∆−1 ), and such that EF = {A} ∪ H0 .
(d) Set A = (Qp /Zp ) ⊗Zp Λ, regarded as a discrete G-p-torus. By Theorem B, there are
′
unique simple fusion systems FB and FH over A ⋊ U, with AutFB (A) ∼
= O p (G)µ−1
V (∆0 )
−1
p′
∼
and EFB = {A} ∪ B, and AutFH (A) = O (G)µV (∆−1 ) and EFH = {A} ∪ H.
Since V is simple (since there is no (p − 2)-dimensional submodule), none of the cases (ii–
a), (iii′ –a), or (iii′′ –c) in Table 4.4 can occur with G0 ≤ G ≤ G and V ∼
= Ω1 (A). Since
′′
dim(V ) < p, case (iv –b) in Table 4.3 cannot occur.
′
The last column in Table 6.1 can be used to help determine the subgroups O p (G)µ−1
V (∆t )
for i = 0, −1. For example:
• When p = 5, G0 ∼
= 2·A6 , and dim(V ) = 4, we have µV (G∨0 ) = ∆1/2 : the subgroup of
order 4 in ∆ = (Z/5)× × (Z/5)× generated by the class of (4, 2). Since ∆1/2 ∆t = ∆
for t = 0, −1, we have G0 µ−1
V (∆t ) = G.
• When p = 7, G0 ∼
= 6·PSL3 (4) or 6·PSU4 (3), and dim(V ) = 6, we have that µV (G∨ ) has
0
G0 µ−1
V (∆t )
index 2 in ∆ and does not contain ∆t for any t. So in all cases,
= G: an
extension of the form G0 .2.
• If p = 7, G0 ∼
= PSU3 (3), and dim(V ) = 6, then µV (G∨0 ) = 21 ∆1 : a subgroup of order 3
that intersects trivially with ∆t for t = 0, −1. So in this case, G0 µ−1
V (∆t ) has the
form G0 .2 × 3 (where the precise extension depends on t).
We now look at one case where the Fp G-module V is not simple.
Example 6.3. Let p, G, U, V , and Λ be as in Example 6.2, except that we assume that V
is indecomposable but not simple, and contains a (p − 2)-dimensional submodule V0 < V .
Let Λ0 < Λ be a Zp G-sublattice of index p such that Λ0 /pΛ ∼
= V0 .
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
43
• There are simple fusion systems exactly as described in cases (a), (b), (c), and (d) in
Example 6.2. In addition, we have:
(e) By case (ii–a) in Table 4.4, for each k ≥ 2, there is a unique simple fusion systems F
over (Λ0 /pk Λ) ⋊ U, with AutF (Λ0 /pk Λ) ∼
= G, and such that EF = {A} ∪ B0 ∪ H∗ .
(f) By case (iii′ –a) in Table 4.4, for each k ≥ 2 and each ∅ 6= I ⊆ {0, 1, . . . , p − 1},
there is a unique simple fusion system FI over (Λ0 /pk Λ) ⋊ U, with AutFI (Λ0 /pk Λ) =
S
′
O p (G)µ−1
V (∆−1 ), and such that EFI = {A} ∪
i∈I Hi .
Since there is no 1-dimensional submodule, case (iii′′ –c) in Table 4.4 cannot occur with
G0 ≤ G ≤ G and V ∼
= Ω1 (A). Since dim(V ) < p, case (iv′′ –b) in Table 4.3 cannot occur.
If we chose to restrict the above examples to the case dim(V ) = p − 1, this is because
when dim(V ) is larger, there are far fewer possibilities. By Table 4.4, A ∼
= Λ/pk Λ for some
k ≥ 2, and EF = {A} ∪ B0 . Similarly, by Table 5.2, there is only one possibility for F when
A is a discrete p-torus with Ω1 (A) ∼
=V.
Appendix A. Background on groups and representations
We collect here some miscellaneous group theoretic results which were needed earlier. We
begin with a few elementary properties of discrete p-toral groups that are easily reduced to
the analogous statements about finite p-groups.
Lemma A.1. Fix a prime p, a discrete p-toral group P , and a finite group G ≤ Aut(P )
of automorphisms of P . Let 1 = P0 E P1 E · · · E Pm = P be a sequence of subgroups, all
normal in P and normalized by G. Let H ≤ G be the subgroup of those g ∈ G which act
via the identity on Pi /Pi−1 for each 1 ≤ i ≤ m. Then H is a normal p-subgroup of G, and
hence H ≤ Op (G).
Proof. See, e.g., [BLO3, Lemma 1.7(a)].
Lemma A.2. Fix an abelian group A each of whose elements has p-power order. Let G ≤
Aut(A) be a finite group of automorphisms, and choose U ∈ Sylp (G). Then
CA (U) ≤ [G, A] ⇐⇒ CA (G) ≤ [G, A] ⇐⇒ CA (G) ≤ [U, A] .
Proof. This is shown in [CrOS, Lemma 1.9] when A is a finite abelian p-group, and the proof
given there also applies when A is infinite and p-power torsion.
Lemma A.3. Let S be a nonabelian discrete p-toral group, with abelian subgroup A < S of
index p, and set Z = Z(S) = CS (A) and S ′ = [S, S] = [S, A]. Then S ′ ∼
= A/Z. Also, A is
the unique abelian subgroup of index p in S if and only if |S ′| = |A/Z| > p.
Proof. Choose 1 6= u ∈ AutS (A), and define ϕ : A −→ A by setting ϕ(a) = a − u(a). Then
Z = Ker(ϕ) and S ′ = Im(ϕ), so A/Z ∼
= S ′.
If |A/Z| = p, then S/Z ∼
= Cp × Cp (it cannot be cyclic since S is nonabelian), and each
subgroup of index p in S containing Z is abelian. Conversely, if B is a second abelian
subgroup of index p, then Z = A ∩ B since S = AB, so |A/Z| = p.
We now turn attention to discrete p-toral groups and discrete G-p-tori. We start with the
well known equivalence between discrete G-p-tori and Zp G-lattices (see Definition 3.6).
44
BOB OLIVER AND ALBERT RUIZ
Proposition A.4. Fix a prime p and a finite group G. Then there is a natural bijection
∼
= /
isomorphism classes of Zp G-lattices
isomorphism classes of
o
in finitely generated Qp G-modules
discrete G-p-tori
A✤
/
HomZp (Qp /Zp , A)
(Qp /Zp ) ⊗Zp Λ o
✤
Λ
If A is a discrete G-p-torus and Λ = HomZp (Qp /Zp , A), then for each n ≥ 1, evaluation at
∼
=
[1/pn ] ∈ Qp /Zp defines an Fp G-linear isomorphism Λ/pn Λ −−−→ Ωn (A).
Proof. If Λ is a Zp G-lattice, then (Qp /Zp ) ⊗Zp Λ is a discrete G-p-torus, and if A is a discrete
G-p-torus, then HomZp (Qp /Zp , A) is a Zp G-lattice. It is an easy exercise to show that the
natural homomorphisms
eval
(Qp /Zp ) ⊗Zp HomZp (Qp /Zp , A) −−−−−−−→ A
and
λ7→(r7→r⊗λ)
Λ −−−−−−−−−−→ HomZp Qp /Zp , (Qp /Zp ) ⊗Zp Λ
are isomorphisms for each Zp G-lattice Λ and each discrete G-p-torus A. The last statement
now follows from the short exact sequence
incl
(x7→pn x)
0 −−−→ (p−n Zp )/Zp −−−−−→ Qp /Zp −−−−−−−→ Qp /Zp −−−→ 0.
The next lemma is mostly a well known result in elementary number theory.
Lemma A.5. Fix a prime p, and let U be a group of order p. Let ζ be a primitive p-th
root of unity, and regard Qp (ζ) and Zp [ζ] as Zp U-modules via some choice of isomorphism
U∼
= hζi.
(a) There are exactly two irreducible Qp U-modules up to isomorphism: a 1-dimensional
module with trivial U-action, and a (p − 1)-dimensional module isomorphic to Qp (ζ).
(b) The ring Zp [ζ] is a local ring with maximal ideal p = (1 − ζ)Zp [ζ]. Also, pZp [ζ] = pp−1 .
(c) Let M be a (p−1)-dimensional irreducible Qp U-module, and let Λ < M be a Zp U-lattice.
Then Λ ∼
= Zp [ζ] as Zp U-modules, and hence Λ/pΛ ∼
= Zp [ζ]/pp−1 is indecomposable as
an Fp U-module.
(d) Let B be an infinite abelian discrete p-toral
Q group (written additively), upon which U
acts with |CB (U)| = p. Assume also that u∈U u(x) = 1 for each x ∈ B. Then B is a
discrete p-torus of rank p − 1, and B ∼
= Qp (ζ)/Zp [ζ] as Zp U-modules.
Proof. (a,b) By [Gd, Proposition 6-2-6], (1−ζ)Z[ζ] is the only prime ideal in Z[ζ] containing
pZ[ζ] = (1 − ζ)p−1Z[ζ]. Hence Qp ⊗Q Q(ζ) = Qp (ζ), so dimQp (Qp (ζ)) = p − 1, and Zp [ζ]
is a local ring with maximal ideal p = (1 − ζ)Zp [ζ] where pp−1 = pZp [ζ]. This proves (b),
and also that Qp (ζ) is an irreducible (p − 1)-dimensional Qp U-module. So the only other
irreducible Qp U-module is Qp with the trivial action.
(c) By (a), we can assume that M = Qp (ζ). Thus Λ is a Zp U-lattice in Qp (ζ), so pm Λ ≤
Zp [ζ] is an ideal for m large enough. Hence pm Λ = pk = (1 − ζ)k Zp [ζ] for some k, and Λ is
isomorphic to Zp [ζ] as a Zp U-module.
Q
(d) Since u∈U u(x) = 1 for each x ∈ B, we can regard B as a Zp [ζ]-module. For each n ≥ 1,
Ωn (B)/(1−ζ)Ωn(B) = |CΩn(B) (U)| = |CB (U)| = p, and so by (b), there is rn ∈ Ωn (B) that
generates Ωn (B) as a Zp [ζ]-module. Let Rn be the set of all such generators of Ωn (B), and
REDUCED FUSION SYSTEMS OVER p-GROUPS WITH ABELIAN SUBGROUP OF INDEX p: III
45
let ϕn : Rn −→ Rn−1 be the map ϕn (rn ) = (rn )p . The (Rn , ϕn ) thus form an inverse system
of nonempty finite sets. An element (rn )n≥1 in the inverse limit defines an isomorphism
Qp (ζ)/Zp [ζ] ∼
= (Qp /Zp ) ⊗Zp Zp [ζ] −−−−→ B of Zp [ζ]-modules (hence of Zp U-modules), where
n
(1/p ) ⊗ ξ is sent to ξ·rn .
Lemma A.6. Fix a prime p, an abelian discrete p-toral group A and a finite group of
automorphisms G ≤ Aut(A). Assume, for S ∈ Sylp (G), that S 5 G and |A/CA (S)| = p.
Then |S| = p.
Proof. This is shown in [Ol, Lemma 1.10] when A is finite, and the general case follows since
G acts faithfully on Ωk (A) for k large enough.
Proposition A.7. Fix an abelian discrete p-toral group A, and a subgroup G ≤ Aut(A).
Assume the following.
(i) Each Sylow p-subgroup of G has order p and is not normal in G.
(ii) For each x ∈ G of order p, [x, A] has order p, and hence CA (x) has index p.
′
Set H = O p (G), A1 = CA (H), and A2 = [H, A]. Then G normalizes A1 and A2 , A =
A1 × A2 , and H ∼
= Cp2 . There are groups of automorphisms
= SL2 (p) acts faithfully on A2 ∼
Gi ≤ Aut(Ai ) (i = 1, 2), such that p ∤ |G1 |, G2 ≥ AutH (A2 ) ∼
= SL2 (p), and G E G1 × G2 (as
a subgroup of Aut(A)) with index dividing p − 1.
Proof. This is shown in [Ol, Lemma 1.11] when A is finite. The general case then follows by
regarding A as the union of the groups Ωk (A) for k ≥ 1.
References
[Ag]
[Al]
[AG]
J. Aguadé, Constructing modular classifying spaces, Israel J. Math. 66 (1989), 23–40
J. Alperin, Local representation theory, Cambridge Univ. Press (1986)
K. Andersen & J. Grodal, The classification of 2-compact groups, Journal Amer. Math. Soc. 22
(2009), 387–436
[AGMV] K. Andersen, J. Grodal, J. Møller, & A. Viruel, The classification of p-compact groups for p odd,
Ann. of Math. 167 (2008), 95–210.
[AOV1] K. Andersen, B. Oliver, & J. Ventura, Reduced, tame, and exotic fusion systems, Proc. London
Math. Soc. 105 (2012), 87–152
[AOV2] K. Andersen, B. Oliver, & J. Ventura, Fusion systems and amalgams, Math. Z. 274 (2013), 1119–
1154
[AKO] M. Aschbacher, R. Kessar, & B. Oliver, Fusion systems in algebra and topology, Cambridge Univ.
Press (2011)
[Ben] D. Benson, Representations and cohomology I, Cambridge Univ. Press (1995)
[BLO2] C. Broto, R. Levi, & B. Oliver, The homotopy theory of fusion systems, J. Amer. Math. Soc. 16
(2003), 779–856
[BLO3] C. Broto, R. Levi, & B. Oliver, Discrete models for the p-local homotopy theory of compact Lie
groups and p-compact groups, Geometry and Topology 11 (2007), 315–427
[BLO6] C. Broto, R. Levi, & B. Oliver, An algebraic model for finite loop spaces, Algebr. Geom. Topol. 14
(2014), 2915–2981
[CE]
A. Clark, & J. Ewing, The realization of polynomial algebras as cohomology rings, Pacific J. Math.
50 (1974), 425–434.
[At]
J. Conway, R. Curtis, S. Norton, R. Parker, & R. Wilson, Atlas of finite groups, Oxford Univ. Press
(1985)
[Cr]
D. Craven, The theory of fusion systems, Cambridge Univ. Press (2011)
[CrOS] D. Craven, B. Oliver, & J. Semeraro, Reduced fusion systems over p-groups with abelian subgroup
of index p: II, Advances in Math. 322 (2017), 201–268
[DW] W. Dwyer & C. Wilkerson, Homotopy fixed-point methods for Lie groups and finite loop spaces,
Annals of math. 139 (1994), 395–442
[Gd]
L. Goldstein, Analytic number theory, Prentice-Hall (1971)
46
BOB OLIVER AND ALBERT RUIZ
[Gon] A. González, The structure of p-local compact groups of rank 1, Forum Math. 28 (2016), 219–253
[GLR] A. González, T. Lozano, & A. Ruiz, On some new examples of p-local compact groups, (preprint)
arXiv:1512.00284v3 [math.AT].
[G]
D. Gorenstein, Finite groups, Harper & Row (1968)
[LO]
R. Levi & B. Oliver, Construction of 2-local finite groups of a type studied by Solomon and Benson,
Geometry & Topology 6 (2002), 917–990
[Mø]
J. Møller, N -determined 2-compact groups. I, Fund. math. 195 (2007), 11–84
[Ol]
B. Oliver, Simple fusion systems over p-groups with abelian subgroup of index p: I, J. Algebra 398
(2014), 527–541
[Pg]
L. Puig, Frobenius categories, J. Algebra 303 (2006), 309–357
[RV]
A. Ruiz & A. Viruel, The classification of p-local finite groups over the extraspecial group of order
p3 and exponent p, Math. Z. 248 (2004), 45–65
[ST]
G. Shephard & J. Todd, Finite unitary reflection groups, Canadian J. Math. 6 (1954), 274–304
LAGA, Institut Galilée, Av. J-B Clément, 93430 Villetaneuse, France
E-mail address: [email protected]
Departament de Mathemàtiques, Edifici C, Universitat Autònoma de Barcelona, 08193
Bellaterra, Spain
E-mail address: [email protected]
| 4math.GR
|
Gaussian Process Landmarking on Manifolds
Tingran Gao ∗
Shahar Z. Kovalsky †
Doug M. Boyer ‡
Ingrid Daubechies §
arXiv:1802.03479v1 [stat.ME] 9 Feb 2018
February 13, 2018
Abstract
As a means of improving analysis of biological shapes, we propose a greedy algorithm for sampling a Riemannian
manifold based on the uncertainty of a Gaussian process. This is known to produce a near optimal experimental design
with the manifold as the domain, and appears to outperform the use of user-placed landmarks in representing geometry
of biological objects. We provide an asymptotic analysis for the decay of the maximum conditional variance, which
is frequently employed as a greedy criterion for similar variance- or uncertainty-based sequential experimental design
strategies; to our knowledge this is the first result of this type for experimental design. The key observation is to link
the greedy algorithm with reduced basis methods in the context of model reduction for partial differential equations.
We apply the proposed landmarking algorithm to geometric morphometrics, a branch of evolutionary biology focusing
on the analysis and comparisons of anatomical shapes, and compare the automatically sampled landmarks with the
“ground truth” landmarks manually placed by evolutionary anthropologists; the results suggest that Gaussian process
landmarks perform equally well or better, in terms of both spatial coverage and downstream statistical analysis. We
expect this approach will find additional applications in other fields of research.
Keywords: Gaussian Process, Experimental Design, Active Learning, Manifold Learning, Reduced Basis Methods, Geometric Morphometrics, Procrustes Distance
1
Introduction
This paper applies principles of the statistics field of optimal experimental design to geometric morphometrics. In
contrast to methods for statistical estimation and inference, which typically focus on studying the error made by
estimators with respect to a finite collection of sample observations (generated either randomly or deterministically)
and selecting the estimator so as to minimize this error, given a (fixed) collection of samples, the paradigm of optimal
experimental design is to minimize the empirical risk by an “optimal” choice of sample locations, while the estimator
itself and the number of samples are kept fixed [110, 6]. Finding an optimal design amounts to choosing sample points
that are most informative for a class of estimators so as to reduce the number of observations; this is most desirable
when even one observation is expensive to acquire (e.g. in spatial analysis (geostatistics) [137, 33] and computationally
demanding computer experiments [122]), but similar ideas have long been exploited in the probabilistic analysis of
more classical numerical analysis problems (see e.g. [133, 150, 114]).
Geometric morphometrics is a subfield of evolutionary biology that focuses on quantifying the (dis-)similarities
between pairs of two-dimensional anatomical surfaces based on their spatial configurations. These analyses typically
start with manually identifying an equal number of geometrically or semantically meaningful feature points, or landmarks, on each specimen in a potentially large collection of anatomical surfaces; the landmarks are certified by domain
experts to be in consistent one-to-one correspondences. Pairwise or group-wise Procrustes analysis [54, 38, 55] is then
applied to quantitatively compare the specimens in the collection; these quantitative comparison results are then subject
to further statistical analysis to draw biologically relevant conclusions for the dataset.
In this paper, we adopt the methodology of optimal experimental design for discretely sampling Riemannian
manifolds, and illustrate the effectiveness of our approach on anatomical surfaces stemming from evolutionary anthropology. We consider this application as prototypical; i.e. we expect that the approach we propose here can be
∗ Department
of Statistics, The University of Chicago, Chicago, IL 60637, USA ([email protected]).
of Mathematics, Duke University, Durham, NC 27708, USA ([email protected]).
‡ Department of Evolutionary Anthropology, Duke University, Durham NC 27708, USA ([email protected]).
§ Department of Mathematics and Department of Electrical and Computer Engineering, Duke University, Durham NC 27708, USA ([email protected]).
† Department
1
applied more widely, to other application domains. In contexts different from evolutionary biology, closely related
(continuous or discretized) manifold sampling problems are addressed in [5, 87, 59], where smooth manifolds are
discretized by optimizing the locations of (a fixed number of) points so as to minimize a Riesz functional, or by [102],
studying surface simplification via spectral subsampling. These approaches, however, tend to distribute points uniformly when applied to two-dimensional surfaces, whereas our motivation is rooted in the application of geometric
morphometrics, for which informative geometric features needed for effective downstream statistical shape analysis
are typically neither uniformly distributed nor naturally associated with an energy potential. Instead, our work is inspired by recent research on active learning with Gaussian processes [31, 113, 66] as well as related applications in
finding landmarks along a manifold [75]. Our approach considers a Gaussian process on the manifold whose covariance structure is specified by the heat kernel. In turn, we design a greedy landmarking strategy which aims to produce
a set of geometrically-significant samples.
The main objects of study in geometric morphometrics are anatomical surfaces of extinct and extant animals;
we will focus on the surfaces of bones and teeth of particular interest for evolutionary biologists. With the advancement of digitization technologies, discrete triangular mesh representations of continuous biological surfaces
are readily available and are used for high throughput analysis for large collections of specimens. We will thus focus
on two-dimensional surfaces represented by discrete triangular meshes. We denote a triangular mesh with notation
G = (V, E, F), where V, E, F stand for the set of vertices, edges, and faces, respectively. To see the link between
landmark identification and active learning with uncertainty sampling [74, 126], let us consider the regression problem
of estimating a function f : V → R defined over the vertices of a triangular mesh G. However, rather than construct the
estimator from random sample observations, we adopt the point of view of active learning, in which one is allowed to
sequentially query the values of f at user-picked vertices x ∈ V. In order to minimize the empirical risk of an estimator
fˆ within a given number of iterations, the simplest and most commonly used strategy is to first evaluate (under reasonable probabilistic model assumptions) the informativeness of the vertices on the mesh that have not been queried, and
then greedily choose to inquire the value of f at the vertex x at which the response value fˆ (x)—inferred from all previous queries—is most “uncertain” in the sense of attaining highest conditional variance (other uncertainty measures
could be used as well, such as the Shannon entropy); these sequentially obtained points of highest uncertainties will
be treated as morphometrical landmarks in our proposed algorithm.
This straightforward application of an active learning strategy relies upon selecting a regression function f of rich
biological information. In the absence of a natural candidate regression function f , we seek to reduce in every iteration
the maximum “average uncertainty” of a class of regression functions, e.g., specified by a Gaussian process prior [113].
Throughout this paper we will denote GP (m, K) for the Gaussian process on a smooth, compact Riemannian manifold
M with mean function m : M → R and covariance function K : M × M → R. If we interpret choosing a single
most “biologically meaningful” function f as a manual “feature handcrafting” step, the specification of a Gaussian
process prior can be viewed as a less restrictive and more stable “ensemble” version; the geometric information
can be conveniently encoded into the prior by specifying an appropriate covariance function K. We construct such
a covariance function in Section 3 by reweighting the heat kernel of the Riemannian manifold M, adopting (but
in the meanwhile also appending further geometric information to) the methodology of Gaussian process optimal
experimental design [120, 122, 39] and sensitivity analysis [121, 101] from the statistics literature.
Our Contributions. The main contribution in this paper is three-fold: First, we propose to automatically landmark
an anatomical surface using a greedy procedure in Gaussian process active learning, and demonstrate the efficacy
and informativeness of the resulting landmarks in analyzing real datasets in geometric morphometrics. Second, our
approach utilizes a Gaussian process defined by a reweighted heat kernel, which has not been studied in the existing manifold learning literature; nonetheless, we show that, in the continuous limit, the reweighted kernel converges
pointwise to a Fokker-Planck operator determined by the weight function, akin to a properly reweighted kernel in diffusion maps. Third, we provide a convergence rate analysis for the greedy algorithm of uncertainty-based sequential
experimental design, which amounts to estimating the uniform rate of decay for the conditional variance of a Gaussian
process as the number of greedily picked design points approaches infinity; on a C ∞ -manifold we deduce that the convergence is faster than any inverse polynomial rate, which is the optimal rate any non-greedy landmarking algorithm
can attain on a generic smooth manifold. This analysis makes use of recent results in the analysis of reduced based
methods. To our knowledge, this is the first analysis of this type for greedy algorithms in optimal experimental design;
the convergence results obtained from this analysis can also be used to bound the number of iterations in Gaussian
process active learning [31, 66, 75] and maximum entropy design [122, 73, 100].
The rest of this paper is organized as follows. Section 2 sets notations and provides background materials for Gaus2
sian processes and geometric morphometrics; Section 3 details the construction of heat kernels on triangular meshes,
as well as the “reweighted kernel” constructed from these discretized heat kernels; Section 4 presents the unsupervised
landmarking algorithm for anatomical surfaces inspired by recent work on uncertainty sampling in Gaussian process
active learning [75]; Section 5 provides deeper theoretical insights for the mechanism of the landmarking algorithm
by connecting it with manifold learning, maximum entropy experimental design, and reduced based method; Section 6 demonstrates several applications of the proposed algorithm in automated geometric morphometrics; Section 7
summarizes the current paper with a discussion and a brief sketch of potential future directions.
2
Background
This section introduces the notations that will be used throughout this paper. We describe Gaussian processes on
a Riemannian manifold with the heat kernel as covariance functions, and provide background information for the
geometric morphometrics application that motivated this paper.
2.1
Heat Kernels and Gaussian Processes on Riemannian Manifolds: A Spectral Embedding Perspective
Let (M, g) be an orientable compact Riemannian manifold of dimension d ≥ 1 with finite volume, where g is the
Riemannian metric on M. Denote dvol M for the canonical volume form M with coordinate representation
p
dvol M (x) = |g (x)| dx1 ∧ · · · ∧ dxd .
The finite volume will be denoted as
Vol (M) =
Z
dvol M (x) =
M
Z
p
|g (x)| dx1 ∧ · · · ∧ dxd < ∞,
M
and we will fix the canonical normalized volume form dvol M /Vol (M) as reference. Throughout this paper, all distributions on M are absolutely continuous with respect to dvol M /Vol (M).
The single-output regression problem on the Riemannian manifold M will be described as follows. Given independent and identically distributed observations {(Xi , Yi ) ∈ M × R | 1 ≤ i ≤ n} of a random variable (X, Y) on the product
probability space M × R, the goal of the regression problem is to estimate the conditional expectation
f (x) := E (Y | X = x)
(1)
which is often referred to as a regression function of Y on X [140]. The joint distribution of X and Y will always be
assumed absolutely continuous with respect to the product measure on M × R for simplicity. A Gaussian process (or
Gaussian random field) on M with mean function m : M → R and covariance function K : M × M → R is defined
as the stochastic process of which any finite marginal distribution on n fixed points x1 , · · · , xn ∈ M is a multivariate
Gaussian distribution with mean vector
m ({x1 , · · · , xn }) := (m (x1 ) , · · · , m (xn )) ∈ Rn
and covariance matrix
K (x1 , x1 ) · · ·
..
K ({x1 , · · · , xn }) :=
.
K (xn , x1 ) · · ·
K (x1 , xn )
..
∈ Rn×n .
.
K (xn , xn )
A Gaussian process with mean function m : M → R and covariance function K : M × M → R will be denoted as
GP (m, K), as stated in Section 1.
We will frequently make use of the Reproducing Kernel Hilbert Space (RKHS) framework (see e.g. [34, 92]) HK
on M with kernel function K. For any positive semi-definite symmetric kernel function K : M × M → R, Mercer’s
Theorem [34, Theorem 3.6] states that K admits a uniformly convergent expansion of the form
K (x, y) =
∞
X
e−λi φi (x) φi (y) ,
i=0
3
∀x, y ∈ M,
2
2
2
where {φi }∞
i=0 ⊂ L (M) are the eigenfunctions of the integral operator T K : L (M) → L (M) defined by
Z
T K f (x) :=
K (x, y) f (y) dvol M (y) , ∀ f ∈ L2 (M)
M
and e−λi , i = 0, 1, · · · , ordered so that e−λ0 ≥ e−λ1 ≥ e−λ2 ≥ · · · , are the eigenvalues of this integral operator corresponding to the eigenfunctions φi , i = 0, 1, · · · , respectively. Regression under this framework amounts to restricting
the regression function to lie in the Hilbert space
∞
∞
X
X
λi 2
(2)
f =
αi φi αi ∈ R,
e αi < ∞
HK :=
i=0
i=0
on which the inner product is defined as
h f, giHK =
∞
X
eλi h f, φi iL2 (M) hg, φi iL2 (M) .
(3)
i=0
The reproducing property is reflected in the identity
hK (·, x) , K (·, y)iHK = K (x, y)
∀x, y ∈ M.
(4)
Borrowing terminologies from kernel-based learning methods (see e.g. [34] and [123]), the eigenfunctions and eigenvalues of T K define a feature mapping
M 3 x 7−→ Φ (x) := e−λ0 /2 φ0 (x) , e−λ1 /2 φ1 (x) , · · · , e−λi /2 φi (x) , · · · ∈ `2
such that the kernel value K (x, y) at an arbitrary pair x, y ∈ M is given exactly by the inner product of Φ (x) and Φ (y)
in the feature space `2 , i.e.
K (x, y) = hΦ (x) , Φ (y)i`2 , ∀x, y ∈ M,
and we have
∞
X
−λi /2
2
(β
)
HK =
f
=
β
·
e
φ
=
Φi
β
=
,
β
,
·
·
·
,
β
,
·
·
·
∈
`
.
hβ,
2
i
i
0
1
i
`
i=0
In words, the RKHS framework embeds the Riemannian manifold M into an infinite dimensional Hilbert space `2 , and
converts the (generically) nonlinear regression problem on M into a linear regression problem on a subset of `2 .
On Riemannian manifolds, there is a natural choice for the kernel function: the heat kernel of the Laplace-Beltrami
operator. Denote ∆ : C 2 (M) → C 2 (M) for the Laplace-Beltrami operator on M with respect to the metric g, i.e.
p
1
∆ f = p ∂i |g| gi j ∂ j f ,
|g|
∀ f ∈ C ∞ (M)
where the sign convention is such that −∆ is positive semidefinite. If the manifold M has no boundary, the spectrum
of −∆ is well-known to be real, non-negative, discrete, with eigenvalues satisfying 0 = λ0 < λ1 ≤ λ2 ≤ · · · % ∞,
with ∞ the only accumulation point of the spectrum; when M has non-empty boundary we assume Dirichlet boundary
condition so the same conclusion holds for the eigenvalues. If we denote φi for the eigenfunction of ∆ corresponding
to the eigenvalue λi , then the set {φi | i = 0, 1, · · · } constitutes an orthonormal basis for L2 (M) under the standard inner
product
Z
h f1 , f2 i M :=
f1 (x) f2 (x) dvol M (x) .
M
The heat kernel kt (x, y) := k (x, y; t) ∈ C 2 (M × M) × C ∞ ((0, ∞)) is the fundamental solution of the heat equation on
M:
∂t u (x, t) = −∆u (x, t) ,
x ∈ M, t ∈ (0, ∞) .
That is, if the initial data is specified as
u (x, t = 0) = v (x)
4
then
u (x, t) =
Z
kt (x, y) v (y) dvol M (y) .
M
In terms of the spectral data of ∆ as (see e.g. [116, 13]), the heat kernel can be written as
kt (x, y) =
∞
X
e−λi t φi (x) φi (y) ,
∀t ≥ 0, x, y ∈ M.
(5)
i=0
For any fixed t > 0, the heat kernel defines a Mercer kernel on M by
(x, y) 7→ kt (x, y)
∀ (x, y) ∈ M × M
and the feature mapping takes the form
M 3 x 7−→ Φt (x) := e−λ0 t/2 φ0 (x) , e−λ1 t/2 φ1 (x) , · · · , e−λi t/2 φi (x) , · · · ∈ `2 .
(6)
Note in particular that
kt (x, y) = hΦt (x) , Φt (y)i`2 .
(7)
√
d
n+2
In fact, up to a multiplicative constant c (t) = 2 (4π) 4 t 4 , the feature mapping Φt : M → `2 has long been studied in
spectral geometry [12] and is known to be an embedding of M into `2 ; furthermore, with the multiplicative correction
by c (t), the pullback of the canonical metric on `2 is asymptotically equal to the Riemannian metric on M.
In this paper we focus on Gaussian processes on Riemannian manifolds with heat kernels (or “reweighted” counterparts thereof; see Section 3.2) as covariance functions. There are at least two reasons for heat kernels to be considered
as natural candidates for covariance functions of Gaussian processes on manifolds. On the one hand, as argued in
[25, §2.5], the abundant geometric information encoded in the Laplace-Beltrami operator makes the heat kernel a
canonical choice for Gaussian processes; Gaussian processes defined this way impose natural geometric priors based
on randomly rescaled solutions of the heat equation. On the other hand, by (7), a Gaussian process on M with heat
kernel is equivalent to a Gaussian process on the embedded image of M into `2 under the feature mapping (6) with a
dot product kernel; this is reminiscent of the methodology of extrinsic Gaussian process regression (eGPR) [76] on
manifolds — in order to perform Gaussian process regression on a nonlinear manifold, eGPR first embeds the manifold into a Euclidean space using an arbitrary embedding, then perform Gaussian process regression on the embedded
image following standard procedures for Gaussian process regression. This spectral embedding interpretation also
underlies recent work constructing Gaussian priors, by means of the graph Laplacian, for uncertainty quantification of
graph semi-supervised learning [15].
2.2
Geometric Morphometrics: Old and New
Statistical shape analysis, often termed geometric morphometrics in comparative biology, is the quantitative analysis
of variations and correlations among biological forms through the Cartesian coordinates of “landmarks”—biologically
informative, repeatable, and in some sense corresponding anatomical loci—on surfaces representing anatomy of biological organisms [151, 1]. In order for the comparisons across specimens to be meaningful, practitioners in this field
often require that the landmarks be consistently annotated on each specimen in a manner reflecting the “operational
homology1 ” or “biological correspondence” (as discussed in [20]) across individuated traits inherited from a common
ancestry [89]. For instance, in the generalized Procrustes analysis (GPA) framework [54, 38, 55, 2], the Procrustes
distance between two surfaces S 1 , S 2 is computed using the following procedure:
n
o n
o
(i) Specify two sets of operationally homologous landmarks x`(1) | 1 ≤ ` ≤ L , x`(2) | 1 ≤ ` ≤ L on S 1 , S 2 , respectively;
(ii) Compute the distance between S 1 and S 2 by minimizing the energy functional
L
1 X (1)
dcP (S 1 , S 2 ) = inf
T x` − x`(2)
T ∈E(3) L
`=1
12
2
where E (3) is the group of rigid motions in R3 .
1 The
term “homology” in the context of evolutionary theory bears a different meaning than in modern topology; see e.g. [115, Part IV].
5
(8)
This idea can be generalized to analyze a collection of consistently landmarked shapes, either assuming each set of
landmarks on the same shape is centered at the origin so the variational problem is defined on a product space of
orthogonal groups [55, 98, 136, 95], or estimate the optimal orthogonal and translation group elements jointly without
the overall centering assumption [26].
The key to successfully applying the Procrustes framework in statistical shape analysis is to obtain an equal number of consistent, operationally homologous landmarks on every shape in a potentially enormous collection of shapes.
Consistently landmarking a collection of shapes relies crucially upon domain knowledge and tedious manual labor,
and the skill to perform it “correctly” typically requires years of professional training; even then the “correctness” can
be subject to debate among experts (see e.g. [19] for an example on Lepilemur teeth). In the first place, extracting a
finite number of landmarks from a continuous surface inevitably loses geometric information, unless when the shapes
under consideration are easily seen to be uniquely determined by the landmarks (e.g. planar polygonal shapes, as considered in [68][40]), which is rarely the case for geometric morphometricians in biology; this problem of “inadequate
coverage” motivated the introduction of semilandmarks—additional points along curves containing critical curvature
information about the morphology—to compensate for the loss of geometry in the landmarking process. Unfortunately, the essential arbitrariness of the semilandmarks along a curve induces additional uncertainty that needs to be
quantified and reduced [151, §2], especially in the absence of sharp anatomical features; the constraint of picking an
equal number of landmarks on each shape also turns out to be far too artificial when the anatomical forms undergo
complex evolutionary and developmental processes.
To mitigate both the scalability and the subjectivity issues in the existing Procrustes analysis framework, a recent
trend of research in geometric morphometrics advocates automated workflows to bypass the repetitive, laborious,
and time-consuming process of manual landmark placement; see e.g. [2, 19, 109, 79, 18, 80, 20, 42, 143, 70, 62]
and the references therein. This type of technique works directly with digitized anatomical surfaces encoded in the
data structure of discrete triangular meshes; numerical algorithms are combined with computer graphics and geometry processing to provide high throughput, landmark-free approaches for precise phenotyping [107, 60, 61] on the
discrete triangular meshes on their entirety (often consisting of thousands to millions of vertices in R3 ), without the
manual landmarking stage to filter down the number of variables using a priori domain knowledge. As much as
the proponents of landmark-based morphometrics claim that landmark coordinates contain the information utilized in
more traditional, measurement-based morphometrics as a proper subset [151, §2], the precursors of automated geometric morphometrics believe that using whole surfaces as input to their analysis passes even more information to the
downstream analysis than the few dozens of landmarks.
Despite the capability of generating high quality pairwise shape registrations, the automated geometric morphometric methods suffer from interpretability problems: since all comparisons are performed merely pairwise, composing
the obtained correspondences along a closed loop does not give rise to an identity map in general. This lack of
transitivity (see e.g. [45, 43]) demands additional post-processing steps to translate the pairwise results into familiar input date to standard downstream phylogenetic analysis [103, 46], which bears a strong similarity with recent
studies in synchronization problems [26, 44]. The loop inconsistency of pairwise correspondences also challenges
the interpretability of automated geometric morphometrics, since it becomes virtually impossible to identify functionally equivalent regions across distinct anatomical structures in a consistent manner. Until fully automated geometric
morphometric algorithms reach the maturity with comparable explanatory power to a human practitioner of landmarkbased geometric morphometrics, deeper and more systematic understanding of the landmarking process still seem of
great interest and value.
The methodology we propose in this paper incorporates an algorithmic landmarking procedure into the automated
pairwise registration algorithms. As will be detailed in Section 6, we will first generate a set of candidate landmarks
on each of the anatomical surfaces based on the “uncertainty” modeled by a Gaussian process, then apply a matching scheme that filters out non-corresponding candidate landmarks between a pair of surfaces based upon bounded
conformal distortion [78]. Detecting morphometrically meaningful landmarks on anatomical surfaces in a completely
unsupervised manner is a daunting task, since some of the most reliable landmarks are determined by patterns of
juxtapositions of tissues—termed “Type 1 landmarks” by Bookstein [117]—which is almost always absent on the
triangular meshes input to automated algorithms [17]. The selection process is further complicated by the requirement of the consistency of relative landmark positions across the data collected, as well as the specific functionality
of the biological organism being studied [151]. While geometry processing algorithms (e.g. [8, 82, 24, 22, 139]) are
capable of detecting sharp geometric features dictated by metric or topology (Bookstein’s “Type 2 landmarks” [117]),
as well as producing high quality pairwise registrations for accurate determination of operationally homologous loci,
semilandmarks and a majority of “Type 3 landmarks” in Bookstein’s typology of landmarks are marked simply for
6
adequate and/or comprehensive coverage of the anatomical forms [117]. For instance, some Type 3 landmarks are
included in the analysis for being “furthest away” from sharp geometric or functional features [151]. These observations motivated us to consider algorithmic analogies of geometric morphologists’ daily practice beyond the scope
of computational geometry, shedding light upon landmark identification from the perspective of Bayesian statistics
quantifying the “uncertainty” of morphometric analysis.
3
3.1
Reweighted Heat Kernels
Discrete Heat Kernels on Point Clouds
When the Riemannian manifold M is a submanifold embedded in an ambient Euclidean space RD (D d) and
sampled only at finitely many number of points {x1 , · · · , xn }, we know from the literature of Laplacian eigenmaps
[10, 11] and diffusion maps [32, 130, 131] that the extrinsic squared exponential kernel matrix
xi − x j 2
(9)
K = Ki j
= exp −
1≤i, j≤n
t
1≤i, j≤n
is a consistent estimator (up to a multiplicative constant) of the heat kernel of the manifold M if {xi | 1 ≤ i ≤ n} are
sampled uniformly and i.i.d. on M with appropriately adjusted bandwidth parameter t > 0 as n → ∞; similar results
holds when the squared exponential kernel is replaced with any anisotropic kernel, and additional renormalization
techniques can be used to adjust the kernel if the samples are i.i.d. but not uniformly distributed on M, see e.g. [32]
for more details. These theoretical results in manifold learning justifies using extrinsic kernel functions in a Gaussian
process regression framework when the manifold is an embedded submanifold of an ambient Euclidean space; the
kernel (9) is also used in [148] for Gaussian process regression on manifolds in a Bayesian setting.
3.2
Reweighted Kernels
The heat kernel of the Riemannian manifold M defines covariance functions for a family of Gaussian processes on M,
but this type of covariance functions only depend on the spectral properties of M, whereas in practice we would often
like to incorporate prior information addressing relative high/low confidence of the selected landmarks. For example,
the response variables might be measured with higher accuracy (or equivalently the influence of random observation
noise is damped) where the predictor falls on a region on the manifold M with lower curvature. We encode this type of
prior information regarding the relative importance of different locations on the domain manifold in a smooth positive
weight function w : M → R+ defined on the entire manifold, of which the higher values of w (x) indicates a relatively
higher importance if a predictor variable is sampled near x ∈ M. Since we assume M is closed, w is bounded below
away from zero. To “knit” the weight function into the heat kernel, notice that by the reproducing property we have
Z
kt (x, y) =
kt/2 (x, z) kt/2 (z, y) dvol M (z)
(10)
M
and we can naturally apply the weight function to deform the volume form, i.e. define
Z
ktw (x, y) =
kt/2 (x, z) kt/2 (z, y) w (z) dvol M (z) .
(11)
M
Obviously, ktw (·, ·) = kt (·, ·) on M × M if we pick w ≡ 1 on M, using the expression (5) for heat kernel kt (·, ·) and
the orthonormality of the eigenfunctions {φi | i = 0, 1, · · · }. Intuitively, (11) reweighs the mutual interaction between
different regions on M such that the portions with high weights have a more significant influence on the covariance
structure of the Gaussian process on M. The rest of the analysis for spectral Gaussian process in Section 3.1 carry
through after replacing GP (m, kt ) with GP m, ktw .
In practice, when the manifold is sampled only at finitely many i.i.d. points {x1 , · · · , xn } on M, the reweighted
kernel can be calculated from the discrete extrinsic kernel matrix in (9) as
n
!
X
xk − x j 2
kxi − xk k2
w
w
K = Ki j
= exp −
· w (xk ) · exp −
= K > WK
(12)
1≤i, j≤n
t/2
t/2
k=1
1≤i, j≤n
7
where W is a diagonal matrix of size n × n with w (xk ) at its k-th diagonal entry, for all 1 ≤ k ≤ n, and K is the discrete
squared exponential kernel matrix in (9). It is worth point out that the reweighted kernel K w no longer equals the
kernel K in (9) when we set w ≡ 1 at this discrete level. The form of the reweighted kernel (11) has also appeared in
[27] as the symmetrization of an asymmetric anisotropic kernel.
Though the reweighting step appears to be a straightforward implementation trick, it turns out to be crucial in
the application of automated geometric morphometrics—the landmarking algorithm that will be presented in Section 4 produces biologically much more representative features on anatomical surfaces when the reweighted kernel is
adopted. We shall illustrate this in greater detail in Section 6.
4
Gaussian Process Landmarking as Sequential Experimental Design
We present in this section an algorithm motivated by [75] that automatically places “landmarks” on a compact Riemannian manifold using a Gaussian process active learning strategy. Let us begin with an arbitrary nonparametric
regression model in the form of (1). Unlike in standard supervised learning in which a finite number of samplelabel pairs are provided, an active learning algorithm can iteratively decide, based on memory of all previously inquired sample-label pairs, which sample to inquire for label in the next step. In other words, given sample-label pairs
(X1 , Y1 ) , (X2 , Y2 ) , · · · , (Xn−1 , Yn−1 ) observed up to the n-th step, an active learning algorithm can decide which sample
Xn to query for the label information Yn = f (Xn ) of the regression function f to be estimated; typically, the algorithm
assumes full knowledge of the sample domain, has access to the regression function f as a black box, and strives to
optimize its query strategy so as to estimate f in as few steps as possible. With a Gaussian process prior GP (m, K) on
the regression function class, the joint distribution of a finite collection of response values (Y1 , · · · , Yn ) is assumed to
follow a multivariate Gaussian distribution Nn (m (X1 , · · · , Xn ) , K (X1 , · · · , Xn )) where
m (X1 )
K (X1 , X1 ) · · · K (X1 , Xn )
..
..
n×n
m (X1 , · · · , Xn ) = ... ∈ Rn , K (X1 , · · · , Xn ) =
(13)
∈ R .
.
.
(X
)
(X
)
(X
)
m n
K n , X1 · · · K n , Xn
For simplicity of statement, introduce the following short-hand notations:
1
Xn−1
=
X1
..
.
Xn−1
,
1
Yn−1
Y1
= ... ,
Yn−1
K (X, X1 )
..
1
,
=
K X, Xn−1
.
K (X, Xn−1 )
Kn−1
K (X1 , X1 ) · · ·
..
=
.
K (Xn−1 , X1 ) · · ·
K (X1 , Xn−1 )
..
.
.
K (Xn−1 , Xn−1 )
Given (n − 1) observed samples (X1 , Y1 ) , · · · , (Xn−1 , Yn−1 ), at any X ∈ M, the conditional probability of the response
1
value Y (X) | Yn−1
follows a normal distribution N (ξn−1 (X) , Σn−1 (X)) where
>
1
−1 1
ξn−1 (X) = K X, Xn−1
Kn−1
Yn−1 ,
>
1
−1
1
Σn−1 (X) = K (X, X) − K X, Xn−1
Kn−1
K X, Xn−1
.
(14)
In our landmarking algorithm, we simply choose Xn to be the location on the manifold M with the largest variance, i.e.
Xn := argmax Σn−1 (X)
X∈M
>
1
−1
1
= argmax K (X, X) − K X, Xn−1
Kn−1
K X, Xn−1
.
(15)
X∈M
Notice that this successive procedure of “landmarking” X1 , X2 , · · · on M is independent of the specific choice of
regression function in GP (m, K) since we only need the covariance function K : M × M → R.
We now describe the main algorithm of this paper, an unsupervised landmarking procedure for anatomical surfaces.
Let M be a 2-dimensional compact surface isometrically embedded in R3 , and denote κ : M → R, η : M → R for
the Gaussian curvature and (scalar) mean curvature of M. Define a family of weight function wλ,ρ : M → R≥0
parametrized by λ ∈ [0, 1] and ρ > 0 as
λ |κ (x)|ρ
wλ,ρ (x) = Z
ρ
(1 − λ) |η (x)|ρ
+Z
ρ
|κ (ξ)| dvol M (ξ)
|η (ξ)| dvol M (ξ)
M
M
8
,
∀x ∈ M.
(16)
This weight function seeks to emphasize theinfluence
of high wcurvature locations on the surface M on the covariance
w
structure of the Gaussian process prior GP m, kt λ,ρ , where kt λ,ρ is the reweighted heat kernel defined in (11). For
all applications discussed in this paper, we set m ≡ 0 but occasionally alter the parameters λ and ρ to fine-tune the
landscape of the weight function (nevertheless, unless otherwise specified, we set by default λ = 1/2 and ρ = 1). For
all practical purposes, we only concern ourselves with M being a piecewise linear surface, represented as a discrete
triangular mesh T = (V, E) with vertex set V = x1 , · · · , x|V| ⊂ R3 and edge set E. We calculate the mean and Gaussian
curvature functions η, κ on the triangular mesh (V, E) using standard algorithms in computational geometry [30, 3].
The weight function wλ,ρ can then be calculated at each vertex xi by
wλ,ρ (xi ) =
λ |κ (xi )|ρ
|V|
X
(1 − λ) |η (xi )|ρ
,
|V|
X
ρ
|η (xk )| ν (xk )
+
ρ
|κ (xk )| ν (xk )
k=1
∀xi ∈ V
(17)
k=1
w
where ν (xk ) is the area of the Voronoi cell centered at xi on the triangular mesh T . The reweighted heat kernel kt λ,ρ is
then defined on V × V as
|V|
X
w
kt λ,ρ xi , x j =
kt/2 (xi , xk ) kt/2 xk , x j wλ,ρ (xk ) ν (xk )
(18)
k=1
where the (unweighted) heat kernel kt is calculated as in (9). Until a fixed total number of landmarks are collected, at
each step k the algorithm computes the uncertainty score Σ(k) on V from the existing (k − 1) landmarks ξ1 , · · · , ξk−1 by
wλ,ρ
> wλ,ρ
kt (xi , ξ1 ) kt (ξ1 , ξ1 ) · · ·
w
..
..
Σ(k) (xi ) = kt λ,ρ (xi , xi ) −
wλ,ρ .
wλ,ρ .
kt (xi , ξk−1 )
kt (ξk−1 , ξ1 ) · · ·
−1 wλ,ρ
kt (xi , ξ1 )
..
,
.
wλ,ρ
wλ,ρ
kt (ξk−1 , ξk−1 )
kt (xi , ξk−1 )
w
kt λ,ρ (ξ1 , ξk−1 )
..
.
∀xi ∈ V (19)
and pick the k-th landmark ξk according to the rule
ξk = argmax Σ(k) (xi ) .
xi ∈V
If there are more than one maximizers of Σ(k) , we just randomly pick one; at step 1 the algorithm simply picks the
w
vertex maximizing x 7→ kt λ,ρ (x, x) on V. See Algorithm 1 for a comprehensive description.
Remark 1. We require the inputs to be triangular meshes with edge connectivity only for the computation of discrete
curvatures. This is not a hard constraint, though — many algorithms are readily available for computing curvatures
on point clouds where the connectivity information is not present (see e.g. [119, 35]). Algorithm 1 can be easily
adapted to use curvatures computed on point clouds as weights in the covariance function construction (11), which
makes it applicable to a much wider range of input data accessible in geometric morphometrics; see Section 6.
Remark 2. Note that, according to (19), each step adds only one new row and one new column to the inverse covariance matrix, which enables us to perform rank-1 updates to the covariance matrix according to the block matrix
inversion formula (see e.g. [113, §A.3])
!−1
!
−1
−1
−1
K
P
Kn−1
In−1 + µPP> Kn−1
−µKn−1
P
Kn−1 = n−1
=
−1
P> K (Xn , Xn )
−µP> Kn−1
µ
where
K (X1 , Xn )
..
∈ Rn−1 ,
P =
.
K (Xn−1 , Xn )
−1
−1
µ = K (Xn , Xn ) − P> Kn−1
P ∈ R.
This simple trick significantly improves the computational efficiency as it avoids directly inverting the covariance
matrix when the number of landmarks becomes large as the iteration progresses.
Before we delve into the theoretical aspects of Algorithm 1, let us present a few typical instances of this algorithm
in practical use. A more comprehensive evaluation of the applicability of Algorithm 1 to geometric morphometrics
9
Algorithm 1 Gaussian Process Landmarking with Reweighted Heat Kernel
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
procedure GPL(T , L, λ ∈ [0, 1], ρ > 0, > 0)
. Triangular Mesh T = (V, E), number of landmarks L
κ, η ← DiscreteCurvatures(T )
. calculate discrete Gaussian curvature κ and mean curvature η on T
ν ← VoronoiAreas(T )
. calculate the area of Voronoi cells around each vertex xi
wλ,ρ ← CalculateWeight(κ,
η,
λ,
ρ,
ν)
. calculate weight function wλ,ρ according to (17)
2
∈ R|V|×|V|
W ← exp − xi − x j
1≤i,
j≤|V|
Λ ← diag wλ,ρ (x1 ) ν (x1 ) , · · · , wλ,ρ x|V| ν x|V| ∈ R|V|×|V|
ξ1 , · · · , ξ L ← ∅
Ψ←0
`←1
Kfull ← W > ΛW ∈ R|V|×|V|
Ktrace ← diag (Kfull ) ∈ R|V|
while ` < L + 1 do
if ` = 1 then
Σ ← Ktrace
else
Σ ← Ktrace − diag Ψ> Ψ ξ1 , · · · , ξ` , : Ψ ∈ R|V|
end if
ξ` ← argmax Σ
Ψ ← Kfull :, ξ1 , · · · , ξ`
` ←`+1
end while
return ξ1 , · · · , ξL
end procedure
. initialize landmark list
. calculate uncertainty scores by (19)
will be deferred to Section 6. In a nutshell, the Gaussian process landmarking algorithm picks the landmarks on the
triangular mesh successively, according to the uncertainty score function Σ at the beginning of each step; at the end of
each step the uncertainty score function gets updated, with the information of the newly picked landmark incorporated
into the inverse convariance matrix defined as in (14). Figure 1 illustrates the first few successive steps on a triangular
mesh discretization of a fossil molar of primate Plesiadapoidea. Empirically, we observed that the updates on the
uncertainty score function are mostly local, i.e. no abrupt changes of the uncertainty score are observed away from a
small geodesic neighborhoods centered at the new landmark. Guided by uncertainty and curvature-reweighted covariance function, the Gaussian process landmarking often identifies landmarks of abundant biological information—for
instance, the first Gaussian process landmarks are often highly biologically informative, and demonstrate comparable
level of coverage with the observer landmarks manually picked by human experts. See Figure 2 for a visual comparison between the automatically generated landmarks with the observer landmarks manually placed by evolutionary
anthropologists on a different digitized fossil molar.
5
Convergence Analysis: Manifold Learning, Maximum Entropy Design,
and Reduced Basis in Reproducing Kernel Hilbert Spaces
In this section, we provide some insights for the manifold landmarking algorithm proposed in Section 4 through several
different lenses: Section 5.1 provides a pointwise convergence analysis for the reweighted kernel to the heat kernel of
a Schoedinger operator; Section 5.2 relates Algorithm 1 to the near-optimal greedy strategy employed in maximum
entropy experimental design; Section 5.3 establishes the rate of convergence for Algorithm 1 by bounding the decay
of the conditional covariance in terms of the number of iterations. These analyses reveal the interplay among manifold
learning, optimal experimental design, and reduced basis methods in partial differential equations.
10
Figure 1: The first 8 landmarks picked successively by Gaussian Process Landmarking (Algorithm 1) on a digitized fossil molar of
Plesiadapoidea (extinct mammals from the Paleocene and Eocene of North America, Europe, and Asia [128]), with the uncertainty
scores at the end of each step rendered on the triangular mesh as a heat map. In each subfigure, the pre-existing landmarks are
colored green, and the new landmark is colored red. At each step, the algorithm picks the vertex on the triangular mesh with the
highest uncertainty score (computed according to (14)), then update the score function.
5.1
Pointwise Convergence of the Reweighted Kernel with Normalization
The reweighted kernel (11) is essential to the application of Algorithm 1 to geometric morphometrical applications.
Intuitively, the reweighting step modifies the Euclidean heat kernel (9) by amplifying the influence of locations with
relatively high weights, as suggested in (11) as well as its discretization (18). To investigate the role played by the
reweighted kernel in greater detail, we study the behavior of the reweighed kernel in the asymptotic regime when
the number of samples increases to infinity. To motivate Theorem 1 below, let us consider i.i.d. (with respect to the
N
standard volume measure) samples {xi }i=1
on a closed Riemannian manifold (M, g), as well as an arbitrary function
2
f ∈ C (M). To simplify the discussion, assume for the moment that the samples are uniformly distributed on M with
respect to the normalized volume form on M. It is well known in manifold learning (see e.g. [130, 32]) that, for any
x ∈ M,
!
Z
N
x − x j 2
1 X
1
kx − yk2
exp −
f (y) dVol (y) as n → ∞.
(20)
exp −
f x j −→
N j=1
2
Vol (M) M
2
Denote the weight function used in (11) by
w (x) = e−V(x) ,
V (x) ≥ 0
∀x ∈ M
(21)
since in our practice only non-negative weight functions w ∈ C 2 (M) are used, and typically we pick a normalization
(e.g. in (17)) to ensure 0 ≤ w ≤ 1 for numerical stability. Repeatedly using the “law of large numbers” argument (20),
we have the following convergence for the reweighted kernel:
!
N
N
X
xk − x j 2
1 X
kx − xk k2
f x j −→
exp −
exp [−V (xk )]
exp −
2
2
N 2 k=1
j=1
!
!
Z Z
1
kx − zk2
kz − yk2
exp −
exp [−V (z)] exp −
f (y) dVol (z) dVol (y)
2
2
[Vol (M)]2 M M
11
as N → ∞.
Figure 2: Left: Sixteen observer landmarks on a digitized fossil molar of a Teilhardina (one of the oldest known fossil primates
closely related with living tarsiers and anthropoids [9]) identified manually by evolutionary anthropologists as ground truth, first
published in [19]. Right: The first 22 landmarks picked by Gaussian Process Landmarking (Algorithm 1). The numbers next to
each landmark indicate the order of appearance. The Gaussian process landmarks strikingly resembles the observer landmarks: the
red landmarks (Number 1-5, 7, 8, 10, 11, 16, 19) signals geometric sharp features (cusps or saddle points corresponding to local
maximum/minimum Gaussian curvature); the blue landmarks sit either along the curvy cusp ridges and grooves (Number 13, 18,
20, 22) or at the basin (Number 9), serving the role often played by semilandmarks; the four green landmarks (Number 6, 12, 15,
17) approximately delimit the “outline” of the tooth in occlusal view.
This motivates us to consider the asymptotic behavior of the integral operator on the right hand side in the asymptotic
regime → ∞, a recurring theme in manifold learning (see e.g. [10, 32, 131, 43]).
Theorem 1. Let M be a d-dimensional closed manifold. For any f, V ∈ C ∞ (M) and V ≥ 0,
!
!
Z Z
kx − zk2
kz − yk2
exp [−V (z)] exp −
f (y) dVol (z) dVol (y)
exp −
2
2
M M
!
!
Z Z
kx − zk2
kz − yk2
exp −
exp [−V (z)] exp −
dVol (z) dVol (y)
2
2
M M
3
−→ f (x) + ∆ f (x) − ∇ f (x) · ∇V (x) + O 2
(22)
as → 0.
The proof of Theorem 1 can be found in Appendix A. Theorem 1 indicates that a proper normalization of the
reweighted kernel (11) gives rise to an approximation to the heat kernel of the backward Kolmogorov operator
L = −∆ + ∇V · ∇.
Note that this operator is also the infinitesimal generator of the diffusion process determined by stochastic differential
equation
√
dXt = −∇V (Xt ) dt + 2 dWt
where Wt is the standard Wiener process defined on M. In particular, Theorem 1 suggests that our construction of the
reweighted kernel encodes information from the weight function into the dynamics of the stochastic process on M.
12
Remark 3. It is interesting to compare the result of Theorem 1 with the related but different kernel construction in
diffusion maps [32, 94]. The integral operator in (22) is a properly normalized version of integration a smooth function
against the kernel
!
!
Z
kz − yk2
kx − zk2
e
exp [−V (z)] exp −
dVol (z)
KV (x, y) :=
exp −
2
2
M
which is obtained from sandwiching exp [−V (·)] with the squared exponential kernel. If we pick a different order of
sandwiching, namely, construct the kernel as
!
kx − yk2
KV (x, y) := exp [−V (x)] exp −
exp −V (y) ,
2
then a quick computation in the same spirit as [14, Theorem 1] leads to
Z
!
KV (x, y) f (y) dvol M (y)
3
1
M
Z
−→ f (x) + ∆ f (x) − ∇ f (x) · ∇V (x) + O 2
2
KV (x, y) dvol M (y)
as → 0.
M
In other words, the infinitesimal generator differs from the one calculated in (22) only by an additional scalar factor
eV with
in front of the Laplace-Beltrami operator. Indeed, we observe very similar numerical results when replacing K
e
KV in Algorithm 1. Empirical evidence motivates us to conjecture that the subtlety between using KV and KV in
the landmarking algorithm lies mainly at the choice of the bandwidth parameter, but a deeper investigation of the
statistical performance with respect to the choice of bandwidth is beyond the scope of the current paper.
The behavior of the kernel becomes even more interesting if we let the weight function depend on the bandwidth
parameter > 0 as well. For instance, if we replace (21) by
w (x) = e−V(x)/ ,
V (x) ≥ 0
∀x ∈ M
(23)
then the integral operator considered in Theorem 1 is replaced with
!
"
#
!
Z Z
1
V (z)
kz − yk2
kx − zk2
T f (x) :=
exp
−
exp
−
f (y) dVol (z) dVol (y) , ∀ f ∈ C 2 (M) .
exp
−
2
2
[Vol (M)]2 M M
(24)
An almost identical calculation as in the proof of Theorem 1 leads to
Corollary 1. Under the same assumptions as in Theorem 1,
!
"
#
!
Z Z
V (z)
kx − zk2
kz − yk2
exp −
exp −
exp −
f (y) dVol (z) dVol (y)
2
2
M M
!
"
#
!
Z Z
V (z)
kz − yk2
kx − zk2
exp −
exp −
dVol (z) dVol (y)
exp −
2
2
M M
3
1
−→ f (x) + ∆ f (x) − ∇ f (x) · ∇V (x) + O 2
(25)
as → 0.
Similar to Theorem 1, Corollary 1 shows that a proper normalization of the kernel reweighted by (23) approximates
the heat kernel of the operator
1
L := −∆ + ∇V · ∇.
The dependence on is of particular interest. It is well known in the literature of semi-classical analysis that L is
conjugate to the Witten Laplacian on 0-forms:
∆V, = e−V/2 (L ) eV/2
where
∆V, = − 2 ∆ +
1
|∇V|2 − ∆V.
4
2
13
Since the eigenfunctions of ∆V, corresponding to the leading small eigenvalues concentrate near the critical points
of the potential function V for sufficiently small > 0 (see e.g. [64] or [105, Theorem 3.9]), the eigenfunctions
corresponding to the leading small eigenvalues of L also concentrate near the critical points of V after being multiplied
by e−V/2 , which, by the proof of Theorem 1, can be approximated by the square root of the denominator of (25). Note
that the matrix Kfull in Algorithm 1 corresponds to the integral kernel in the numerator of (25) if we choose Λ to be
the diagonal matrix with exp [−V (vi ) /] at its i-th diagonal entry, where vi is the i-th vertex on the triangular mesh;
setting D to be the diagonal matrix with the i-th row sum of Kfull at its i-th diagonal entry, the Witten Laplacian ∆V,
can then be approximated using
D−1/2 Kfull D−1/2
up to a scalar multiplication by the bandwidth . We would thus expect the first few eigenfunctions corresponding to
the smallest eigenvalues of D−1/2 Kfull D−1/2 to concentrate near the critical points of V. This can be easily verified with
numerical experiments; see Figure 3 for an example.
Figure 3: Concentration of the eigenfunctions of the Witten Laplacian with respect to a potential function V. The left boxed
subplot illustrates the potential function V using a heat map; we constructed this potential to have 4 wells centered around 4 cusps
manually picked on a Lorisidae mandibular molar. The 4 eigenfunctions corresponding to the 4 smallest eigenvalues of the Witten
Laplacian with potential V are depicted on the 2 × 2 panel on the right; it can be read off from the colorbars that the support of each
of these 4 eigenfunctions are concentrated around critical points of V. See Section 5.1 for more details.
Remark 4. The localization effects on eigenfunctions can also be achieved by adding a diagonal matrix, which
represents a potential function on the domain of interest, to the kernel matrix; this idea has found applications in
biomedical data analysis [36] and computer graphics [88]. The insight we gained from relating the reweighted kernel
to Witten Laplacian reveals that the localization can be equivalently realized through a simple nonlinear procedure
motivated. This coincidence can be explained as follows: while adding a diagonal potential to the kernel provides
a direct discretization of the Schoedinger operator LV = −∆ + V, the reweighted kernel (12) can be thought of as a
discretization of the heat kernel of LV by taking only “one-hop” paths in the Feynmann-Kac formula for e−tLV . An
extensive discussion along this direction is beyond the scope of the current paper; interested readers may find useful
the works [134, 135, 58] and the references therein.
The connection between the reweighted kernel (11) and the semi-classical analysis of the Witten Laplacian provides insights for the behavior of the Gaussian process landmarks generated from Algorithm 1. Given a Gaussian
14
process GP (m, K) defined on a manifold M, the eigenfunctions of K—properly reweighted by their corresponding eigenvalues— gives rise to the Karhunen-Loève basis for GP (m, K), with respect to which the sample paths
of GP (m, K) adopt expansions with i.i.d. standard normal coefficients. If the low-frequency components of these
expansions tend to concentrate at certain regions on M, when fitting an unknown function in GP (m, K) using an active
learning procedure, it could be much more efficient if one begins with the inquiry for the function value at those regions
of concentration. After information at regions of concentration are collected to some extent, it becomes necessary to
search for clues filling up the complement on M of those concentration regions. Therefore, spreading landmarks on
M using a Gaussian process with reweighted kernel balances out the information contained in the landscape of the
weight function and the space-filling experimental design strategies. Another advantage of not selecting landmarks
solely based on the weight function is that the weight function can be too spurious to produce reliable, semantically
meaningful landmarks, at least for our application in geometric morphometrics; see Figure 4 for an example for such
a comparison.
Figure 4: Left: Gaussian process landmarks generated using Algorithm 1 with a exponential squared kernel function reweighted
by weight function w; Middle: The local maxima of the weight function used in the reweighed kernel, which appears semantically
less meaningful from the perspective of geometric morphometrics; Right: Points produced by the geodesic farthest point sampling
(see e.g. [52, 91]), a greedy algorithm common used for generating uniformly sampled (or approximately space-filling) points on a
triangular mesh. Comparing these three sampling approaches, Gaussian process landmarks has the advantage of tending to fill up
the manifold while prioritizing the choice of semantically meaningful features for geometric morphometrics.
5.2
Near-Optimality of Gaussian Process Landmarking as Maximum Entropy Design
Algorithm 1 greedily picks the point on the manifold M that maximizes the conditional variance as the next landmark.
This sequential optimization approach is reminiscent of a popular approximation scheme in entropy-based experimental design [73, 100, 122]. The key observation here is the equivalence between minimizing the conditional entropy (or
maximizing the information gain) and maximizing the determinant of the conditional covariance matrix; see e.g. [122,
§6.2.1], [73, §3.1], [100, §2.2]. More concretely, if we denote the maximum entropy of a Gaussian process GP (m, K)
on a manifold M with respect to any n observations as
OPT (n) :=
max
{x1 ,··· ,xn }⊂M
det K (x1 , · · · , xn )
where K (x1 , · · · , xn ) is defined in (13), and
GPL (n) := det K (X1 , · · · , Xn )
where X1 , · · · , Xn are generated using Algorithm 1
then by the submodularity of the entropy function [69] we conclude, using classic results [97], that
OPT (n) ≥ GPL (n) ≥ (1 − 1/e) OPT (n) .
In other words, the entropy of the greedy algorithm is equivalent to the maximum entropy up to a multiplicative
constant. We point out that this justifies the information-theoretic asymptotic near-optimality of Gaussian Process
Landmarking (Algorithm 1).
15
5.3
Rate of Convergence: A Connection with Reduced Basis Methods in Reproducing Kernel Hilbert Spaces
In this subsection we analyze the rate of convergence of our main Gaussian process landmarking algorithm in Section 4. While the notion of “convergence rate” in the context of Gaussian process regression (i.e. kriging [93, 137]) or
scattered data approximation (see e.g. [145] and the references therein) refers to how fast the interpolant approaches
the true function, our focus in this paper is the rate of convergence of Algorithm 1 itself, for instance, the number of
steps the algorithm takes before it terminates. In practice, unless a maximum number of landmarks is predetermined,
a natural criterion for terminating the algorithm is to specify a threshold for the ∞-norm of the conditional covariance
(15) over the manifold. We emphasize again that, although this greedy approach is motivated by the probabilistic
model of Gaussian processes, the conditional covariance function as well as the greedy algorithm of manifold landmarking is completely deterministic once the kernel function is known. Our analysis is thus centered around bounding
the uniform rate at which the conditional covariance decays with respect to number landmarks greedily selected.
To this end, we observe the connection between Algorithm 1 and a greedy algorithm studied thoroughly for reduced basis methods in [16, 37] in the context of model reduction. While the analyses in [16, 37] assume general
Hilbert and Banach spaces, we apply their result to a reproducing kernel Hilbert space (RKHS), denoted as HK ,
naturally associated with a Gaussian process GP (m, K); as will be demonstrated below, the conditional covariance
function with respect to n selected landmarks can be interpreted as a distance function between elements of HK to
an n-dimensional subspace of HK determined by the selected landmarks. We emphasize that, though the connection
between Gaussian process and RKHS is well known (see e.g. [142] and the references therein), we are not aware of
existing literature addressing the resemblance between the two classes of greedy algorithms widely used in Gaussian
process experimental design and reduced basis methods.
We begin with a brief summary of the greedy algorithm in reduced basis methods for a general Banach space
(X, k·k). The algorithm strives to approximate all elements of X using a properly constructed linear subspace spanned
by (as few as possible) selected elements from a compact subset F ⊂ X; thus the name “reduced” basis. A popular
greedy algorithm for this purpose generates successive approximation spaces by choosing the first basis f1 ∈ F
according to
f1 := argmax k f k
(26)
f ∈F
and, successively, when f1 , · · · , fn−1 are picked already, choose
fn+1 := argmax dist ( f, Vn )
(27)
f ∈F
where
Vn = span { f1 , f2 , · · · , fn }
and
dist ( f, Vn ) := inf k f − gk .
g∈Vn
In words, at each step we greedily pick the function that is “farthest away” from the set of already chosen basis elements. Intuitively, this is a Banach space analogy of the farthest point sampling algorithm [52, 91] already mentioned
in Section 5.1. To fit Gaussian process landmarking into this framework, let us first specialize this construction to the
case when X is the reproducing kernel Hilbert space HK ⊂ L2 (M), where M is a compact Riemannian manifold and
K is the reproducing kernel. A natural choice for K is the heat kernel kt (·, ·) : M × M → R with a fixed t > 0 as in
Section 2.1, but for a submanifold isometrically embedded into an ambient Euclidean space it is common as well to
choose the kernel to be the restriction to M of a positive (semi-)definite kernel in the ambient Euclidean space such as
(9) or (12), for which Sobolev-type error estimates are known in the literature of scattered data approximation [96, 41].
It follows from standard RKHS theory that
X
ai K (·, xi ) | ai ∈ R, xi ∈ M, card (I) < ∞
HK = span
i∈I
and, by the compactness of M and the regularity of the kernel function, we have for any x ∈ M
hK (·, x) , K (·, x)iHK = K (x, x) ≤ kKk∞,M×M < ∞
16
(28)
which justifies the compactness of
F := span {K (·, x) | x ∈ M}
(29)
as a subset of HK . In fact, since we only used the compactness of M and the boundedness of K on M× M, the argument
above for the compactness of F can be extended to any Gaussian process defined on a compact metric space with a
bounded kernel. The initialization step (26) now amounts to selecting K (·, x) from F that maximizes
kK (·, x)k2HK = hK (·, x) , K (·, x)iHK = K (x, x)
which is identical to (15) when n = 1 (or equivalently, Line 14 in Algorithm 1); furthermore, given n ≥ 1 previously
selected basis functions K (·, x1 ) , · · · , K (·, xn ), the (n + 1)-th basis function will chosen according to (27), i.e. fn+1 =
K (·, xn ) maximizes the infimum
inf
g∈span{K(·,x1 ),··· ,K(·,xn )}
kK (·, x) −
=
inf
a1 ,··· ,an ∈R
gk2HK
=
inf
a1 ,··· ,an ∈R
K (x, x) − 2
n
X
K (·, x) −
n
X
2
ai K (·, xi )
HK
i=1
ai K (x, xi ) +
i=1
n X
n
X
ai a j K xi , x j
i=1 j=1
−1
K (x1 , x1 ) · · · K (x1 , xn ) K (x, x1 )
(∗)
..
..
..
= K (x, x) − 2 K (x, x1 ) , · · · , K (x, xn )
.
.
.
K (xn , x1 ) · · · K (xn , xn )
K (x, xn )
−1
K (x1 , x1 ) · · · K (x1 , xn ) K (x, x1 )
..
..
..
+ K (x, x1 ) , · · · , K (x, xn )
.
.
.
K (xn , x1 ) · · · K (xn , xn )
K (x, xn )
−1
K (x1 , x1 ) · · · K (x1 , xn ) K (x, x1 )
..
..
..
= K (x, x) − K (x, x1 ) , · · · , K (x, xn )
.
.
.
K (xn , x1 ) · · · K (xn , xn )
K (x, xn )
(30)
where the equality (∗) follows from the observation that, for any fixed x ∈ M, the minimizing vector (a1 , · · · , an ) ∈ Rn
satisfies
K (x, x1 ) K (x1 , x1 ) · · · K (x1 , xn ) a1
.
..
..
..
=
..
.
.
.
K (x, xn )
K (xn , x1 ) · · · K (xn , xn ) an
−1
a1 K (x1 , x1 ) · · · K (x1 , xn ) K (x, x1 )
..
..
..
.
⇔ ... =
.
.
.
K (xn , x1 ) · · · K (xn , xn )
K (x, xn )
an
It is clear at this point that maximizing the rightmost quantity in (30) is equivalent to following the greedy landmark
selection criterion (15) at the (n + 1)-th step. We thus conclude that Algorithm 1 is equivalent to the greedy algorithm
for reduced basis method in HK , a reproducing kernel Hilbert space modeled on the compact manifold M. The
following lemma summarizes this observation for future reference.
Lemma 1. Let M be a compact Riemannian manifold, and let K : M × M → R be a positive semidefinite kernel function. Consider the reproducing kernel Hilbert space HK ⊂ L2 (M) as defined in (28). For any x ∈ M and a collection
of n points Xn = {x1 , x2 , · · · , xn } ⊂ M, the orthogonal projection Pn from HK to Vn = span {K (·, xi ) | 1 ≤ i ≤ n} is
Pn (K (·, x)) =
n
X
a∗i (x) K (·, xi )
i=1
where a∗i : M → R is the inner product of (K (x, x1 ) , · · · , K (x, xn )) with the i-th row of
−1
K (x1 , x1 ) · · · K (x1 , xn )
..
..
.
.
.
K (xn , x1 ) · · · K (xn , xn )
17
In particular, a∗i has the same regularity as the kernel Φ, for all 1 ≤ i ≤ n. Moreover, the squared distance between
K (·, x) and the linear subspace Vn ⊂ HK has the closed-form expression
PK,Xn (x) : = kK (·, x) − Pn (K (·, x))k2HK
=
min
a1 ,··· ,an ∈R
K (·, x) −
n
X
2
ai K (·, xi )
HK
i=1
K (x1 , x1 ) · · ·
..
= K (x, x) − K (x, x1 ) , · · · , K (x, xn )
.
K (xn , x1 ) · · ·
−1
K (x1 , xn ) K (x, x1 )
..
..
.
.
.
K (xn , xn )
K (x, xn )
(31)
Consequently, for any Gaussian process defined on M with covariance structure given by the kernel function K, the
conditional covariance of the Gaussian process conditioned on the observations at x1 , · · · , xn ∈ M equals to the
distance between K (·, x) and the subspace Vn spanned by K (·, x1 ) , · · · , K (·, xn ).
The function PK,Xn : M → R≥0 defined in (31) is in fact the squared power function in the literature of scattered
data approximation; see e.g. [145, Definition 11.2].
The convergence rate of greedy algorithms for reduced basis methods has been investigated in a series of works
[23, 16, 37]. The general paradigm is to compare the maximum approximation error incurring after the n-th greedy
step, denoted as
σn := dist ( fn+1 , Vn ) = max dist ( f, Vn ) ,
f ∈F
with the Kolmogorov width (c.f. [83]), a quantity characterizing the theoretical optimal error of approximation using
any n-dimensional linear subspace generated from any greedy or non-greedy algorithms, defined as
dn := inf sup dist ( f, Y)
Y f ∈F
where the first infimum is taken over all n-dimensional subspaces Y of X. When n = 1, both σ1 and d1 degenerate to
the ∞-bound of the kernel function on M × M, i.e. kKk∞,M×M . In [37] the following comparison between {σn | n ∈ N}
and {dn | n ∈ N} was established:
Theorem 2 ([37], Theorem 3.2 (The γ = 1 Case)). For any N ≥ 0, n ≥ 1, and 1 ≤ m < n, there holds
n
Y
σ2N+` ≤
`=1
n m
m
n n−m 2m 2n−2m
σN+1 dm
.
n−m
This result establishes the asymptotic optimality of the greedy basis selection procedure. In particular, setting
N = 0 and taking advantage of the monotonicity of the sequence {σn | n ∈ N}, one has from Theorem 2 that
√
m
n−m
n
dmn
σn ≤ 2 min kKk∞,M×M
1≤m<n
for all n ∈ N. Using the monotonicity of {σn | n ∈ N}, by setting m = bn/2c we have the even more compact inequality
σn ≤
√
1
1
2
2
dbn/2c
2 kKk∞,M×M
for all n ∈ N.
(32)
If we have an asymptotic bound for {dn | n ∈ N} (in the regime n → ∞), inequality (32) can be invoked to establish an
asymptotic bound for {σn | n ∈ N}. We emphasize here that the definition of {dn | n ∈ N} only involves those elements
in a compact subset F of the ambient Hilbert space HK ; in our setting, the compact subset (29) consists of only
functions of the form K (·, x) for some x ∈ M, thus
dn = inf sup dist K (·, x) , span {K (·, xi ) | 1 ≤ i ≤ n}
x1 ,··· ,xn ∈M x∈M
K (x1 , x1 ) · · ·
..
= inf sup K (x, x) − K (x, x1 ) , · · · , K (x, xn )
.
x1 ,··· ,xn ∈M x∈M
K (xn , x1 ) · · ·
18
−1
K (x1 , xn ) K (x, x1 )
..
..
.
.
.
K (x , x )
K (x, x )
n
n
n
(33)
To ease notation, we will always denote Xn := {x1 , · · · , xn } as in Lemma 1. Write the maximum value of the function
PK,Xn over M as
ΠK,Xn := sup PK,Xn (x) .
(34)
x∈M
The Kolmogorov width dn can be put in these notations as
dn =
inf
x1 ,··· ,xn ∈M
ΠK,Xn .
(35)
The problem of bounding {dn | n ∈ N} thus reduces to bounding the infimum of ΠK,Xn over all n-dimensional linear
subspaces of F .
When M is an open, bounded subset of a standard Euclidean space, upper bounds for ΠK,Xn are often established—
in a kernel-adaptive fashion—using the fill distance [145, Chapter 11]
hXn := sup min x − x j
x∈M x j ∈Xn
(36)
where k·k is the Euclidean norm of the ambient space. For instance, when K is a squared exponential kernel (9) and
the domain is a cube (at least compact and convex, as pointed out in [144, Theorem 1]) in a Euclidean space, [145,
Theorem 11.22] asserts that
#
"
log hXn
∀hXn ≤ h0
(37)
ΠK,Xn ≤ exp c
hXn
for some constants c > 0, h0 > 0 depending only on M and the kernel bandwidth t > 0 in (9). Similar bounds have been
established in [147] for Matérn kernels, but the convergence rate is only polynomial. In this case, by the monotonicity
of the function x 7→ log x/x for x ∈ (0, e), we have, for all sufficiently small hXn ,
"
#
log hn
dn = inf ΠK,Xn ≤ exp c
x1 ,··· ,xn ∈M
hn
where
hn :=
inf
Xn ⊂M, |Xn |=n
hXn
(38)
is the minimum fill distance attainable for any n sample points on M. We thus have the following theorem for the
convergence rate of Algorithm 1 for any compact, convex set in a Euclidean spaces:
Theorem 3. Let Ω ⊂ RD be a compact and convex subset of the D-dimensional Euclidean space, and consider a
Gaussian process GP (m, K) defined on Ω, with the covariance function K being of the squared exponential form (9)
with respect to the ambient D-dimensional Euclidean distance. Let X1 , X2 , · · · , denote the sequence of landmarks
greedily picked on Ω according to Algorithm 1, and define for any n ∈ N the maximum conditional covariance on Ω
with respect to the first n landmarks X1 , · · · , Xn as
−1
K (X1 , X1 ) · · · K (X1 , Xn ) K (x, X1 )
..
..
..
.
σn = max K (x, x) − K (x, X1 ) , · · · , K (x, Xn )
.
.
.
x∈Ω
K (Xn , X1 ) · · · K (Xn , Xn )
K (x, Xn )
Then
log hn
σn = O β hn
as hn → 0
(39)
for some positive constant β > 1 depending only on the geometry of the domain Ω and the bandwidth of the squared
exponential kernel K; hn is the minimum fill distance of n arbitrary points on Ω (c.f. (38)).
Proof. By the monotonicity of the sequence {σn | n ∈ N}, it suffices to establish the convergence rate for a subsequence.
Using directly (32), (35), (37), and the definition of hn in (38), we have the inequality
"
#
√
√
1
1
log hn
c log hn
2
2
σ2n ≤ 2 kKk∞,
exp
= 2 kKk∞,
β hn
where β := exp (c/2) > 1
Ω×Ω
Ω×Ω
2 hn
for all N 3 n ≥ N. Here the positive constants N = N (Ω, t) > 0 and c = c (Ω, t) > 0 depend only on the geometry of
Ω and the bandwidth of the squared exponential kernel. This completes the proof.
19
Convex bodies in RD are far too restricted as a class of geometric objects for modeling anatomical surfaces in our
main application (c.f. Section 6). The rest of this section will be devoted to generalizing the convergence rate for
squared exponential kernels (9) to their reweighted counterparts (12), and more importantly, for submanifolds of the
Euclidean space. The crucial ingredient is an estimate of the type (37) bounding the sup-norm of the squared power
function using fill distances, tailored for restrictions of the squared exponential kernel
!
1
2
(40)
K (x, y) = exp − kx − yk , x, y ∈ M
2
as well as the reweighted version
Kw (x, y) =
Z
"
#
1
w (z) exp −
kx − zk2 + kz − yk2 dvol M (z) ,
2
M
x, y ∈ M
(41)
where w : M → R≥0 is the non-negative weight function. Note that when w (x) ≡ 1, ∀x ∈ M the reweighted kernel
(41) does not coincide with the squared exponential kernel (40), no even up to normalization, since the domain of
integration is M instead of the entire RD . It is not straightforward either to naı̈vely enclose the compact manifold M
with a compact, convex subset Ω of the ambient space and reuse Theorem 3 by extending/restricting functions to/from
M to Ω, since the samples are constrained to lie on M but the convergence will be in terms of fill distances in Ω.
Nevertheless, the desired bound can be established using local parametrizations of the manifold, i.e., working within
each local Lipschitz coordinate chart and taking advantage of the compactness of M.
We will henceforth impose no additional assumptions, other than compactness and smoothness, on the geometry
of the Riemannian manifold M. In the first step we refer to a known uniform estimate from [145, Theorem 17.21] for
power functions on a compact Riemannian manifold.
Lemma 2. Let M be a d-dimensional C ` compact manifold isometrically embedded in RD (where D > d), and let
Φ ∈ C 2k (M × M) be any positive definite kernel function on M × M with 2k ≤ `. There exists a positive constant
h0 = h0 (M) > 0 depending only on the geometry of the manifold M such that, for any collection of n distinct points
Xn = {x1 , · · · , xn } on M with hXn ≤ h0 , the following inequality holds:
ΠΦ,Xn = sup PΦ,Xn (x) ≤ Ch2k
Xn
x∈M
where C = C (k, M, Φ) > 0 is a positive constant depending only on the manifold M and the kernel function Φ. This of
course further implies for all hn ≤ h0
inf
ΠΦ,Xn ≤ Ch2k
n
Xn ⊂M, |Xn |=n
where hn is the minimum fill distance of n arbitrary points on Ω (c.f. (38)).
Proof. This is essentially [145, Theorem 17.21], with the only adaptation that the definition of the power function
throughout [145] is the square root of the PΦ,Xn in our definition (34).
Lemma 2 suggests that the convergence of Algorithm 1 is faster than any polynomial of hn . The dependence on hn
can be made more direct in terms of the number of samples n by the following geometric lemma.
Lemma 3. Let M be a d-dimensional C ` compact Riemannian manifold isometrically embedded in RD (where D > d).
Denote ωd−1 for the surface measure of the unit sphere in Rd , and Vol (M) for the volume of M induced by the
Riemannian metric. There exists a positive constant N = N (M) > 0 depending only on the manifold M such that
! d1
1
2d+1 d
Vol (M) · n− d
for any N 3 n ≥ N.
hn ≤
ωd−1
Proof. For any r > 0 and x ∈ M, we denote BrD (x) for the (extrinsic) D-dimensional Euclidean ball centered at x ∈ M,
and set Br (x) := BrD (x) ∩ M. In other words, Br (x) is a ball of radius r centered at x ∈ M with respect to the “chordal”
metric on M induced from the ambient Euclidean space RD . Define the covering number and the packing number for
M with respect to the chordal metric balls by
n
[
N (r) := N (M, k·kD , r) := min
M⊂
Br (xi ) | xi ∈ M, 1 ≤ i ≤ n
,
N∈N
i=1
n
[
P (r) := P (M, k·kD , r) := max
Br/2 (xi ) ⊂ M, Br/2 (xi ) ∩ Br/2 x j = ∅ for all 1 ≤ i , j ≤ n xi ∈ M, 1 ≤ i ≤ n
.
N∈N
i=1
20
By the definition of fill distance and hn (c.f. (38)), the covering number N (hn ) is lower bounded by n; furthermore,
by the straightforward inequality P (r) ≥ N (r) for all r > 0, we have
n < N (hn ) ≤ P (hn ) ,
i.e. there exist a collection of n points x1 , · · · , xn ∈ M such that the n chordal metric balls Bhn /2 (xi ) | 1 ≤ i ≤ n form
a packing of M. Thus
n
X
Vol Bhn /2 (xi ) ≤ Vol (M) < ∞
i=1
where the last inequality follows from the compactness of M. The volume of each Bhn /2 (xi ) can be expanded asymptotically for small hn as (c.f. [67])
!d
!
2 kBk2x − kHk2x hn 2
ωd−1 hn
+ O hd+3
as hn → 0
(42)
Vol Bhn /2 (x) =
1 +
n
d
2
8 (d + 2)
2
where ωd−1 is the surface measure of the unit sphere in Rd , B is the second fundamental form of M, and H is the
mean curvature normal. The compactness of M ensures the boundedness of all these extrinsic curvature terms. Pick n
sufficiently large so that hn is sufficiently small (again by the compactness of M) to ensure
!d
ωd−1 hn
.
Vol Bhn /2 (x) ≥
2d 2
It then follows from (42) that
nωd−1 hn
2d
2
! d1
1
2d+1 d
hn ≤
Vol (M) · n− d .
ωd−1
!d
≤ Vol (M)
⇒
We are now ready to conclude that Algorithm 1 converges faster than any inverse polynomials in the number of
samples with our specific choice of kernel functions, regardless of the presence of reweighting.
Theorem 4. Let M be a d-dimensional C ∞ compact manifold isometrically embedded in RD (where D > d), and
let Φ ∈ C ∞ (M × M) be any positive definite kernel function on M. For any k ∈ N, there exist positive constants
N = N (M) > 0 and Ck = Ck (M, Φ) > 0 such that
k
σn ≤ Ck n− d for all n ≥ N.
k
In words, Algorithm 1 converges at rate O n− d for all k ∈ N.
Proof. Use Lemma 2, Lemma 3 and the regularity of the kernel function Φ.
Though it is natural to conjecture that exponential rate of convergence holds at least for the Euclidean radial
basis kernel (40), Theorem 4 is about as far as we can get with our current techniques, unless we impose additional
assumptions on the regularity of the manifolds of interest. It is tempting to proceed directly as in [145, Theorem 17.21]
by working locally on coordinate charts and citing the exponential convergence result for radial basis kernels in [145,
Theorem 11.22]; unfortunately, even though kernel K is of radial basis type in the ambient space RD , it is generally
no longer of radial basis type in local coordinate charts, unless one imposes additional restrictive assumptions on the
growth of the derivatives of local parametrization maps (e.g. all coordinate maps are affine). We will not pursue the
theoretical aspects of these additional assumptions in this paper.
Remark 5. The asymptotic optimality of the rate established in Theorem 4 for Gaussian process landmarking follows
from Theorem 2. In other words, the Gaussian process landmarking algorithm leads to a rate of decay of the ∞-norm
of conditional covariances that is at least as fast as any other landmarking algorithms, including random or uniform
sampling on the manifold. In our application of comparative biology that motivated this paper, it is more important
that Gaussian process landmarking is capable of identifying biologically meaningful and operationally homologous
points across the anatomical surfaces even when the number of landmarks in small (n ∞); see e.g. Figure 6 in
Section 6.1. A more thorough theory explaining this advantageous aspect of Gaussian process landmarking will be
left for future work.
21
6
Application in Automated Geometric Morphometrics
This section is divided into three parts. Section 6.1 compares Gaussian process landmarks generated from Algorithm 1
with “ground truth” landmarks manually picked by comparative biologists on a real dataset of anatomical surfaces,
demonstrating comparable levels of coverage of biologically significant features. The results presented in this section
provide quantitative evidence that Gaussian process landmarks are capable of capturing geometric features encoding
important information for comparative biologists on individual anatomical surfaces. Section 6.2 adapts the image
feature matching algorithm of [81] for the registration of pairs of anatomical surface via matching Gaussian process
landmarks computed on each individual surface. We compare the resulting shape correspondence maps with a baseline obtained using previously developed continuous Procrustes analysis in [2, 19]; the results suggest that, though
Gaussian process landmarks are generated on each individual shape separately, they implicitly encode operationally
homologous features that can be compared and contrasted across shapes, and such pairwise comparison results are
comparable with those obtained from standard Procrustes shape analysis based on “ground truth” observer landmarks.
The pairwise shape distances induced from matching landmarks are turned into ordination plots (two-dimensional
embeddings of distance matrices, commonly employed for visualizing the “morphospace” characterizing shape variances) in Section 6.3; it turns out that Gaussian process landmarks lead to a favorable ordination, which is in better
agreement with observations made by comparative biologists and paleontologists in existing literature.
6.1
Unsupervised Landmarking on Individual Anatomical Surfaces
Gaussian process landmarks are generated on each anatomical surface individually, regardless of the size of the shape
collection; the landmarks manually selected by human experts (observers), however, may well depend on the information gradually exposed to the human expert as he/she moves through a collection of surfaces. It is thus surprising
that the individually generated, “local” Gaussian process landmarks bear an unignorable amount of similarity with the
observer landmarks selected with certain extent of collection-wise or “global” knowledge, as we have already illustrated in Figure 2 through an example fossil molar. This subsection is devoted to a more thorough and quantitative
comparison between Gaussian process landmarks and “ground truth” observer landmarks placed by human experts.
We begin by collecting in Figure 5 examples applying the Gaussian Process Landmarking algorithm to other types
anatomical surfaces than molars, namely, a subset of the astragalus and calcaneus bones of tarsiers first published in
[20]. Though observer landmarks are not available for this dataset and thus we can not directly compare automatically
generated landmarks with manually placed ones, it can be recognized from Figure 5 that the algorithm is capable of
consistently capturing both geometric features (“Type 3 landmarks” [117]) and semilandmarks (delineating ridges and
grooves), which is consistent with the results for the molar dataset.
To quantitatively validate the similarity between Gaussian process landmarks and observer landmarks, we calculate, on each anatomical surface, the median geodesic distance from an observer landmark to its closest Gaussian
process landmark. This median geodesic distance can obviously be calculated for any other type of landmarks in place
of the Gaussian process landmarks; we shall refer to it as the median observer-to-automatic landmark distance, for
the sake of simplicity. Since the number of automatic landmarks is typically difficult to fix a priori, we compute the
median observer-to-automatic landmark distances for a varying number of automatic landmarks, obtaining a curve that
encodes the rate with which the median observer-to-automatic landmark distances decay to zero as the number of automatic landmarks increases. Comparing curves obtained from different automatic landmark generation methods then
provides us with a way to compare how closely each type of automation “mimics” the observer landmarks. When we
perform such curve comparisons for a large collection of surfaces in a dataset, one can in principle construct statistics
(e.g. “mean” or “standard deviation” in the “space of curves”) along the lines of Functional Data Analysis (FDA, see
e.g. [112, 111] and the references therein), but we will have to defer such a statistically systematic treatment to future
work.
Using this strategy, we compare Gaussian process landmarks with landmarks obtained from Geodesic Farthest
Point Sampling (GFPS) [52, 91], a popular downsampling technique in automated geometric morphometrics, on a
dataset of 116 second mandibular molars of prosimian primates and closely related non-primates first published in
[19]. GFPS is known to produce approximately uniformly distributed points on surfaces, with respect to the canonical
surface volume measure. In our numerical experiments in this subsection, we choose the first point in GFPS to be the
same as the first point obtained by Gaussian process landmarking, to eliminate the effects of random initialization. As
a baseline, we also calculate the median observer-to-automatic landmark distances for completely randomly picked
vertices on the triangular meshes in this dataset. The results are presented in Figure 6, in which each curve is obtained
22
Figure 5: Gaussian process landmarks on several different types of anatomical surfaces. All triangular meshes are acquired from
µCT scans. Top Row: Six second mandibular molars of prosimian primates and non-primate close relatives, from a dataset of
116 molars first published in [19]. Middle Tow: Six astragalus bones of tarsiers from a dataset of 40 astragali first published in
[20]. Bottom Row: Six calcaneus bones of tarsiers from a dataset of 40 calcanei first published in [20]. On all three types of
anatomical surfaces, Gaussian process landmarks tend to play the roles of both landmarks and semilandmarks (c.f. [151, §2]): the
curvature-reweighted kernel promotes automatically selecting sharp peaks or saddle points on the anatomical surface; after most of
the prominent geometric features—normally recognized as Type 2 landmarks [151]—are captured, the uncertainty-based criterion
encourages the identification of semilandmark-type points along ridges and grooves.
by averaging individual curves over the entire shape collection; confidence intervals of one standard deviation are also
plotted in shades of transparency. Figure 6 suggests that Gaussian process landmarks consistently outperforms GFPS
landmarks and the random baseline in terms of coverage of observer landmarks.
6.2
Pairwise Surface Registration by Matching Gaussian Process Landmarks
We demonstrate in this subsection the benefit of using Gaussian process landmarks for establishing correspondences
between pairs of surfaces. In particular, although Gaussian process landmarks are obtained independently on each
surface, they turn out to encode geometrically significant features — shared more often among similar or related
shapes — that are recognized as “operational homologous” [20] loci by comparative biologists. Statistical analysis
also suggest that these correspondence maps reach comparable explanatory power to observer landmarks placed by
human experts, in terms of induced shape distances. The comparison between the morphospaces characterized by
these shape distances is deferred to Section 6.3.
23
Figure 6: Median geodesic distance from an observer landmark to its nearest automatically generated landmark, with respect to
different numbers of automatic landmarks, on a collection of 116 second mandibular molars of prosimian primates and closely
related non-primates (see [19] for a more detailed description of this dataset and the observer landmark acquisition). Each point
on any of the three solid curves is obtained by averaging the 116 median observer-to-automatic landmark distances over the entire
dataset, and the transparent bands represent confidence intervals of one standard deviation. The three types of automatic landmarks
are Gaussian process landmarks (GP), geodesic farthest point sampling landmarks (GFPS), and random landmarks uniformly
selected from the vertices of each triangular mesh. The random landmarks are only used as a baseline in this experiment.
6.2.1
Experimental Setup and Methodology
Bounded Distortion Gaussian Process Landmark Matching. Let S 1 , S 2 be two-dimensional disk-type surfaces
(conformally equivalent to planar disks by the Uniformization Theorem; see e.g. [2, 79, 80]) and let ξ1(1) , . . . , ξL(1)1 ∈
S 1 and ξ1(2) , . . . , ξL(2)2 ∈ S 2 be two sets of Gaussian process landmarks computed using Algorithm 1 on S 1 and S 2 ,
respectively. Note that the algorithm we present in this subsection works equally well for L1 , L2 , though we choose
L1 = L2 = 40 throughout this paper to simplify the discussion. Adopting the approach suggested in [81] for featurebased image matching, we devise the following two-step approach for establishing geometrically-consistent matchings
between the two sets of Gaussian process landmarks:
(1) Parametrization: For each surface S j ( j = 1, 2) we follow [132, 72] to compute an as-isometric-as-possible
(AIAP) two-dimensional parametrization, which is a diffeomorphism Φ j : S j → Ω j ⊂ R2 from S j to a connected
planar domain Ω j minimizing the (discretization of the) isometric distortion energy
Z
2
F (φ) :=
(43)
|∇φ (x)|2 + ∇φ−1 (x) dvol M (x) .
Sj
Each landmark ξ`( j) is mapped to a unique corresponding point ζ`( j) = Φ j ξ`( j) ∈ Ω j ⊂ R2 , where
Φ j := argmin F (φ) .
φ:S j →R2
(2) Bounded Distortion Matching: Following [81], we search within the set of planar diffeomorphisms between Ω1
and Ω2 with conformal distortion [78] bounded by a pre-fixed constant K ≥ 1. This algorithm strives to find
a maximal subset of geometrically-consistent correspondences within an initial set of candidate matches. In the
extreme case of K = 1, the search is constrained within the set of strictly angle-preserving (conformal) maps for
the continuous isometric distortion energy and the set of planar rigid transformations for the discretized isometric
distortion energy; we select K = 1.5 in this experiment to slightly enlarge the search space of candidate maps.
As input to this matching algorithm, for each Gaussian process landmark ζ`(1) on S 1 , we choose T ≥ 2 Gaus(2)
(2)
sian process landmarks ζ`→1
, . . . , ζ`→T
as initial putative candidate matches from the Gaussian process landmarks
24
n (2)
o
ζ j | 1 ≤ j ≤ L2 on S 2 ; the algorithm then searches for the bounded distortion map Ψ : Ω1 → Ω2 that minimizes
the mismatch count
L1 X
T
X
(2) 0
Ψ ζ`(1) − ζ`→k
,
(44)
`=1 k=1
where, following the notations of [81], k·k0 denotes the mixed (2, 0)-norm defined as
(1)
(2)
(1)
1 if Ψ ζ` , ζ`→k ,
(2) 0
Ψ ζl − ζl→k =
0 otherwise.
The initial candidate matches in our experiments are generated by comparing the Wave Kernel Signature (WKS)
[7] of the Gaussian process landmarks — the T Gaussian process landmarks
n (2) on S 2 with omost similar WKS’s
(1)
(measured in Euclidean distances) with that of each ζ` ∈ S 1 are selected as ζ`→t | 1 ≤ t ∈ T . Ideally, “incorrect”
initial matches potentially leading to large conformal distortions will be filtered out by minimizing (44) under the
bounded conformal distortion constraint. Our implementation uses T = 2 which turned out to be an empirically
satisfactory choice.
The pairwise registration algorithm outputs a subset of 1 ≤ L ≤ min {L1 , L2 } one-to-one correspondences ξ̃`(1) ↔
` = 1, . . . , L, from the initial L1 T candidate Gaussian process landmark matches; the number of matched landmarks, L, is automatically determined by the final bounded distortion map and could be vary between different pairs.
A final step in this algorithm pipeline “interpolates” the L pairs of matched Gaussian process landmarks to obtain a
diffeomorphism between the surfaces S 1 and S 2 . For this purpose, we use the technique developed in [132, 72] which
computes a map
Ψ̃ : Ω1 → Ω2 that minimizes the isometric distortion energy (43) subject to the L linear equality constraints Ψ̃ ζ̃`(1) = ζ̃`(2) (1 ≤ ` ≤ L) representing the “sparse” correspondences between Gaussian process landmarks.
The composition f12 = Φ2 −1 ◦ Ψ̃ ◦ Φ1 finally produces the desired map f : S 1 → S 2 that approximately minimizes the
mismatch count (44). This last interpolation step is indispensable for the purpose of interpretability (producing fullsurface registrations for visual comparisons) as well as evaluation (inducing shape Procrustes distances for ordination,
see Section 6.3). Figure 7 outlines the complete workflow of landmark matching and surface registration for a pair of
molars surfaces from [19]; a few examples are shown in Figure 8.
ξ̃`(2) ,
Evaluation Metrics and the Baseline. We demonstrate the rich taxonomic information captured by the maps output
from the proposed bounded distortion Gaussian process landmark matching (GP-BD) algorithm, by comparing the
shape Procrustes distances induced by maps computed from GP-BD with those obtained by the same feature matching
methodology but alternative choices of landmarking schemes. For any inter-surface map f : S 1 → S 2 , we define the
Procrustes distance induced by f as
! 12
Z
2
dP ( f ) = min
(45)
k f (x) − R (x)k dvolS 1 (x) ,
R∈E3
S1
where E3 stands for the rigid motion group in R3 . The distance dP ( f ) measures the spatial registration error induced
by the map f between the two surfaces. To ensure that the Procrustes distances induced by the various methods listed
in Table 1 is comparable, we normalize each surface to have unit surface area. When the map f is the continuous
Procrustes map produced from the algorithm presented in [2], dP gives exactly the continuous Procrustes distance
between S 1 and S 2 .
The alternatives considered in this section include two other different types of Gaussian process landmarks (GPnW BD and GPEuc -BD) with alternative kernel functions (different from (18)), two different strategies of utilizing the
“ground truth” user-placed observer landmarks (GT-BD and GT2 -BD), as well as the baseline continuous Procrustes
maps (CPM) reported in [2, 19]; Table 1 provides a summary of these variants. Specifically, GPEuc -BD and GPnW BD are based on Gaussian processes with the standard squared exponential kernel (9) and the trivially weighted
kernel (10) (or equivalently, setting w ≡ 1 in (12)), respectively. GT-BD simply replaces the GP landmarks with
“ground truth” (GT) landmarks, which then follows through with candidate selection via WKS and BD-filtering;
GT2 -BD also takes ground truth landmarks as inputs, but skips the candidate selection by setting the ground truth
correspondences between observer landmarks as candidate matches, though BD-filtering still applies and prunes out
potential “geometrically incompatible” correspondences leading to large conformal distortions. A detailed description
of observer landmark acquisition can be found in [19].
25
(a)
(b)
(c)
(d)
(e)
Figure 7: The outline of our approach for landmark matching and surface registration. (a) Gaussian process landmarks computed
using Algorithm 1 on two surfaces; (b) planar parameterizations of the surfaces that minimize (43), overlaid with putative landmark matches (indicated by purple lines); (c) Bounded-Distortion correspondences: a subset of geometrically-consistent matches
computed by approximately minimizing (44); (d) Paris of corresponding Gaussian process landmarks found by bounded distortion matching, illustrated by spheres with matching colors; (e) Texture-map visualization of an inter-surface map interpolating the
landmark correspondences depicted in (d).
The shape distances will be compared in several different ways. We will first compare cumulative distributions
of the pairwise shape distance values, followed by two statistical tests addressing (i) the correlation between each
automatic shape distance and the observer-determined landmarks Procrustes distance (ODLP), and (ii) the capability of each shape distance at distinguishing taxonomic groups. A qualitative but more intuitive comparison of the
morphospaces [90] characterized by these shape distances will be deferred to Section 6.3, in the form of ordination
plots (two dimensional embedding of the shape distance matrices as visual representations for shape variances across
species groups). The discussion in Section 6.3 is however oriented slightly more towards comparative biologists.
6.2.2
Comparison Results
We compared the pairwise surface registration methods listed in Table 1 on a dataset consisting of 116 triangular
meshes representing surfaces of the second mandibular molars of prosimian primates and closely related non-primates
[19]. We computed correspondences between each pair of surfaces within this dataset, totaling in over 13, 000 intersurface registrations.
Figure 8 shows example pairs of surfaces, visualizing inter-surface maps maps induced by GP-BD correspondences
as well as the baseline CPM maps. These examples illustrate typical differences between maps computed from GP-BD
and CPM along with their Procrustes distances; GP-BD maps often offer an improvement over CPM in terms of both
visual quality and their ability to relate biologically meaningful and operationally homologous regions.
Comparing Cumulative Distributions of Distance Values. We first provide a crude comparison across the various
methods listed in Table 1. Figure 9 plots the cumulative distributions of Procrustes distances induced by each type
of pairwise registration method in Table 1, which serves as a direct comparison of the proportions of pairwise intersurface correspondences for which the Procrustes distance dP (45) is below a given threshold. The figure further
includes a curve for the Observer-Determined Landmarks Procrustes (ODLP) distances, computed using the standard
Procrustes analysis between sets of corresponding landmarks manually placed by human experts [19]. Recall from
Section 6.2.1 that the surface areas are normalized so as to ensure that Procrustes distances induced by different shape
correspondences are comparable.
26
Abbreviation
GP-BD
GPEuc -BD
GPnW -BD
GT-BD
GT2 -BD
CPM
Description of the pairwise surface registration method
Bounded Distortion (BD) filtering for Gaussian process (GP) landmarks computed with Algorithm 1
Bounded Distortion (BD) filtering for Gaussian process (GP) landmarks computed with the standard Euclidean heat kernel (9)
Bounded Distortion (BD) filtering for Gaussian process (GP) landmarks computed with a nonweighted kernel (10)
Bounded Distortion (BD) filtering for ground truth (GT), observer-placed landmarks
Bounded Distortion (BD) filtering for ground truth (GT), observer-placed landmarks, also using
the ground truth correspondences as candidate initial matches
Continuous Procrustes inter-surface Maps (CPM) computed using the method of [19, 2]
Table 1: The pairwise surface registration methods compared in Section 6.2 and Section 6.3.
Noticeably, the cumulative distribution of GP-BD distances most closely resemble that of GT-BD, obtained by
replacing the Gaussian process landmarks with the ground-truth observer landmarks (but otherwise using exactly the
same algorithm involving WKS and BD-filtering). Also, GP-BD outperforms GPnW -BD (Gaussian process with a
trivially weighted kernel) and GPEuc -BD (Gaussian process with the standard Euclidean heat kernel). This comparison
suggests that the Gaussian process landmarks computed with Algorithm 1 provide a good proxy for geometrically
significant features needed to determine meaningful correspondences between surfaces.
Expectedly, GP-BD falls short compared to distances that rely on both the ground-truth observer landmarks and
their true correspondences (GT2 -BD and ODLP). Comparing GP-BD to the baseline continuous Procrustes maps
(CPM) in terms of distance distributions is equivocal as the two curves in Figure 9 cross each other when the Procrustes distance threshold is about 0.15. Nonetheless, comparing GP-BD to CPM indicates that the former produces
more pairwise correspondences with shape distances less than 0.1 (at which the vertical gap between the two curves
reaches its maximum). This, along with the correlation reported in [45] between smaller cP distances and better morphometric interpretability of the associated maps, implies that Gaussian process landmarks potentially lead to more
stable and interpretable comparative biological analysis if combined with other globally transitive geometric morphometric methods. Further validation will have to be left for future exploration.
Statistical Tests Comparing the Distance Matrices. We now provide a more thorough comparison — using nonparametric statistical tests — for the shape distance matrices induced by the various methods listed in Table 1. Treating
the OPLD (Observer-Determined Landmarks Procrustes distance) matrix as a proxy for the “ground truth” accepted
among geometric morphometricians (c.f. [19]), we first compare the correlation between each automatically computed
shape distance matrix with ODLP using the Mantel correlation test [86]; the explanatory power of the shape distances
(in terms of the ability to distinguish taxonomic groups) are then compared using a variant of the multivariate analysis
of variance (MANOVA) for distance matrices.
Due to the dependence of entries in a distance matrix (e.g. constrained by the triangular inequality), assessments of
correlation between distance matrices, or between a distance matrix and a (continuous or categorical) variable, often
involves repeatedly permuting the rows and columns of the distance matrix. In fact, permutation test (or randomization
test) has become an increasingly common statistical tool in the nonparametric setting; see e.g. [104, 99, 53] and the
references therein. Table 2 demonstrates the results of Mantel correlation analysis [86] between the distance matrices
computed in this paper against the ODLP distance matrix. GT-BD and GT2 -BD correlate best with ODLP, which
is as expected since they both rely directly on the same set of observer-determined landmarks used for computing
the ODLP distances; their high correlation with ODLP justifies the use of bounded distortion maps for capturing
biologically meaningful and corresponding geometric features. GP-BD outperforms CPM, which is consistent with
our observation in Figure 9 that a majority (over 80%) of the distance values computed from GP-BD is more similar in
distribution to the distances computed with ground truth landmarks. The advantageously higher correlation of GP-BD
over GPEuc and GPnW illustrates the importance of reweighting in the kernel construction.
We also perform Permutational Multivariate Analysis of Variance (PERMANOVA) [4] for the distance matrices
computed in this paper, with the taxonomic groups shown in the ordination plots (Figure 10, see detailed explanations
in Section 6.3) as the treatment effects. The result is presented in Table 3. The purpose of this test is to quantitatively compare the power of differentiating taxonomic groups for these distance matrices. The pseudo F-statistics in
27
Source
GP-BD
0.059
0.083
0.185
0.185
0.092
0.099
0.078
CPM
0.096
(a)
(b)
(c)
(d)
Figure 8: Inter-surface maps and induced Procrustes distances: texture overlaid on the Source surface (top row) is mapped by
f : S 1 → S 2 so as to visualize correspondence. The second row (GP-BD) shows maps induced by correspondences established
between Gaussian process landmarks computed with Algorithm 1. The bottom row (CPM) compares the baseline continuous
procrustes maps. GP-BD outperforms CPM in examples (a)-(c), wherein the map it induces establishes a better correspondence
between morphological features of the surfaces; example (d) shows a case in which GP-BD is outperformed by CPM. The inset
values are the the Procrustes distance dP associated with each map.
PERMANOVA is a properly normalized ratio between among-group and within-group sum of squared distances. The
statistical significance is then calculated from the fraction—among sufficiently many shuffles of the rows and columns
of the distance matrix—of the permutation instances that produce a higher pseudo F-ratio. While all distance matrices
demonstrate statistically significant pseudo F-ratios in this test, the GT-BD distance matrix leads the board of pseudo
F-ratios, indicating its superior ability of separating taxonomic groups; the GP-BD distance matrix and its two variants
consistently outperform CPM in terms of pseudo F-ratios, verifying again the improved quality of pairwise anatomical
surface registrations as well as the induced shape Procrustes distances.
6.3
Ordination and Comparative Biological Interpretation for Shape Collections
This subsection provides a qualitative but visually more straightforward way to compare the shape distance matrices
obtained from the various pairwise registration methods in Table 1 on a collection of shapes. The results presented here
suggests that a proper mechanism matching Gaussian process landmarks can potentially capture more morphometrical
28
100
% Surface Correspondences
90
80
70
60
GP-BD
50
GP Euc -BD
GP -BD
40
nW
GT-BD
30
GT 2 -BD
ODLP
CPM
20
10
0
0
0.05
0.10
0.15
Procrustes Distance
0.20
0.25
Figure 9: Comparing GP-BD to the alternative methods listed in Table 1. Each curve represents the proportion of correspondences,
among a collection of 116 surfaces, for which the Procrustes distance dP falls below a given threshold. The cumulative distribution
of GP-BD distances most closely resembles that of GT-BD which utilizes the ground truth observer landmarks. It falls short of
GT2 -BD and the Observer-Determined Landmarks Procrustes (ODLP), both of which rely additionally on the “ground truth” correspondences between observer landmarks. In terms of producing more relatively smaller distances between the shape pairs (which is
favorable by the correlation between smaller distances and enhanced morphometric interpretability reported in [45]; also suggested
by the cumulative distribution curves of ODLP, GT2 -BD, and GT-DB), GP-BD outperforms the baseline continuous Procrustes
maps (CPM), as well as variants of Gaussian process landmarks with alternative kernel constructions, GPnW -BD (Gaussian process
landmarks with a non-weighted kernel) and GPEuc -BD (Gaussian process landmarks with a standard Euclidean heat kernel).
corr. coeff.
P value
GT-BD
0.7042
**
GT2 -BD
CPM
GP-BD GPEuc -BD
0.7563
0.6647 0.6870
0.6135
**
**
**
**
Significance codes: **P < 0.01.
GPnW -BD
0.6257
**
Table 2: Results of Mantel correlation analysis for the various distances computed in this paper versus ODLP distances. The
high correlation of GT-BD and GT2 -BD versus ODLP is not surprising due to their direct dependence on the observer-determined
landmarks. The distance computed from GP-BD has higher correlation with ODLP than the other distances, which is consistent with
the CDF plot Figure 9 and the ordination plots in Figure 10. We set the number of permutations to 9999 since it is recommended in
[4] that at least 5000 permutations should be done for tests with an α-level of 0.01.
information faithfully in the induced shape distances, thus providing better characterizations of the shape variation
across species groups.
Ordination and Taxonomic Classification. Comparative biologists, ecologists, and other natural and social scientists frequently rely on ordination or gradient analysis to summarize and emphasize patterns in multivariate datasets
[28, 29, 4]. The equivalent but more familiar term for statisticians and applied mathematicians is dimension reduction.
Generally speaking, it is assumed or hypothesized that groups that are distantly related or have different ecologies will
display some physical differences related to these variables, but is difficult to know which physical traits will be most
affected, or whether certain traits are independent of each other. Ordination methods such as Principal Components
Analysis (PCA), Multi-Dimensional Scaling (MDS), and Correspondence Analysis (CA) reveal the nature of those
physical differences by summarizing multivariate datasets in lower dimensional graphical form.
In Figure 10 we visualize ordinations induced from four different distance matrices using classical multi-dimensional
scaling (MDS). The dataset consists of 116 second mandibular molars of prosimian primates and non-primate close
relatives, first published in [19]. The distances include the continuous Procrustes distance [2] and three bounded distortion map based distances with different matching criteria. In each ordination MDS plot, points corresponding to the
same species group are enclosed in a polygonal, and the color pattern of the polygons indicates shape similarity among
29
pseudo F-ratio
P value
ODLP
12.264
**
GT-BD GT2 -BD
CPM
GP-BD
16.6999 13.9038 9.42259 16.2604
**
**
**
**
Significance codes: **P < 0.01.
GPEuc -BD
11.1461
**
GPnW-BD
12.9873
**
Table 3: PERMANOVA test results for various distance matrices computed in this paper for a collection of 116 second mandibular
molars of prosimian primates and non-primate close relatives. The number of group is 30, equaling to the number of polygonal
regions in the ordination plot in Figure 10. The first column refers to observer-determined landmarks Procrustes (ODLP) distances
calculated in [19] as a baseline. All distance matrices are shown to be statistically significant in terms of their powers of separating
species groups; the significant of GP-BD is comparable to that of GT-BD which has the highest pseudo F-ratio over all distances
matrices considered in this paper. The number of permutations is set to 9999 for the same reason as in Table 2.
generic specimens across species groups, based on visual inspection and traditional comparative analyses. Similar ordination plots have been used for comparing different algorithms designed for automatically quantifying the geometric
similarity of anatomical surfaces; see e.g. [20, 141, 45]. This type of plots can also be considered as visualizations of
morphospaces in evolutionary and developmental biology [90].
An important criterion for good ordination is the extent to which specimens belonging to the same species group
cluster near each other, and specimens from different species groups are separated from each other. A popular summary statistic is to measure the ratio between within-group distances and among-group distances, which we quantitatively calculate and present in Table 3. Nevertheless, it is already evident from qualitatively inspecting Figure 10 that
bounded distortion maps (Figure 10a, Figure 10b, Figure 10d) in general better distinguish species groups than continuous Procrustes maps (Figure 10c). Considering taxonomic group membership and previous interpretations regarding
shape affinities of different taxa, the ordination induced from GP-BD distances (Figure 10a) is visually most appealing.
We explain some compelling examples in detail, referencing previous observations by biologists and paleontologists.
(i) Figure 10a has a relatively low within/between-group distance ratio (see Table 3 for more details), with remarkable distinctiveness of the Indriidae, Tupaia, and Cynocephalus groups from the other groups, which is a faithful
reflection of the shape dissimilarity based on visual inspections;
(ii) Among all different ordinations presented in Figure 10, Figure 10a best separates out fossil (the brown-colored
groups — Plesiadapoidea, Plesiolestes, Purgatorius, Leptacodon, and Altanius) and living non-primates (the
yellow-colored groups — Tupaia and Ptilocercus (Scandentia or “treeshrews”) and Cynocephalus (Dermoptera))
from strepsirrhine primates and most early fossil primates (light and dark groups of lemurs — Eulemur, Lemur,
Varecai, Cheirogaleus, Hapalemur, Prolemur, and Cheirogaleidae; light and dark green groups of Lorisidae —
Loris, Galago, Arctocebus, Nycticebus, and Perodicticus; the red group of the extinct adapiforms and omomyiforms — Adapis, Cantius, and Donrussellia);
(iii) The thought-provoking patterns of overlapping reflected in Figure 10a are in accordance with views established
in existing comparative biological literature. For instance, Tarsius (a haplorhine primate) and plesiadapoid nonprimates overlap, reflecting the idea by previous authors that they might be united in a common group called
“Plesiotarsiiformes” [47, 48]. As another example, Teilhardina, Donrussellia, and Cantius all overlap, reflecting
the fact that all of them represent very primitive members of primates: Donrussellia was originally thought to be
a new species of Teilhardina [118]; Cantius was originally thought to be an omomyiform like Teilhardina instead
of an adapiform [129], and [50] observed additional features of the premolar teeth uniquely relating Cantius torresi to Teilhardina. Though Cantius is now reocognised as an adapiform, the wide separation between it and
the other sampled adapiform (Adapis) also matches previous qualitative discussions and analyses of independent
datasets. In particular, the fact that Adapis overlaps Lemur, Eulemur, and Lepilemur is reminiscent of arguments
by [49] that Adapis has special affinities to the living strepsirrhines to the exclusion of other adapiforms, as well
as that Lepilemur belongs to a group called “Megaladapidae” in part because of the claim that some members of
the group have teeth extremely similar to those of Adapis ([124]). Anaylsis of other skeletal regions such as the
ankle bones [125] also link Adapis more closely to living lemurs than to other adapiforms. Finally, the overlap of
Eosimias with treeshrews matches the suggestion by [51] that this taxon, otherwise thought of as an anthropoid,
is in fact treeshrew-like in its dentition. Again analyses of ankle bones return a similar pattern; see e.g. [149, 21].
30
(a) GP-BD
(b) GT-BD
(c) CPM
(d) GT2 -BD
Figure 10: Multi-Dimensional Scaling (MDS) plots visualizing shape similarities characterized by four alternative distance matrices for a collection of 116 second mandibular molars of prosimian primates and non-primate close relatives. The distances
investigated here include (a) Procrustes distance induced by GP-BD maps that match as many pairs of Gaussian process landmarks
as possible within distortion bound K = 1.5; (b) Procrustes distance induced by GT-BD maps that match as many pairs of observer
placed landmarks as possible within distortion bound K = 1.5, but without “ground truth” correspondences between observer
placed landmarks; (c) continuous Procrustes distance [2]; (d) Procrustes distance induced by GT2 -BD maps that match as many
observed placed landmarks as possible within distortion bound K = 1.5, with “ground truth” correspondences between observer
placed landmarks as initial matching candidates. Each polygon encloses a collection of individual specimens of closely related
species expected to be similar based on visual inspection and traditional comparative analyses. Similar ordination plots for GPEuc
or GPnW are omitted in this figure because the ordination shown in those two plots are much worse than any of the plots (a)–(d).
According to the PERMANOVA test results in Table 3, pairwise surface registration with GP-BD achieves comparable power of
separating taxonomic groups as matching “ground truth” landmarks placed by professional comparative biologists.
31
7
Discussion and Future Work
This paper discusses a greedy algorithm for automatically selecting representative points on compact manifolds, motivated by the methodology of experimental design with Gaussian process prior in statistics. With a carefully modified
heat kernel specified as the covariance function in the Gaussian process prior, our algorithm is capable of producing
biologically highly meaningful feature points on some anatomical surfaces. The algorithmically produced landmarks
resemble biologists’ landmarks selected with expert knowledge, providing adequate coverage for both geometric features and semantically “uncertain” regions on the anatomical surfaces of practical interest. We tested the applicability
of this landmarking procedure for various tasks on real datasets.
A future direction of interest is to build theoretical analysis for the optimal experimental design aspects of manifold
learning: Whereas existing manifold learning algorithms estimate the underlying manifold from discrete samples, our
algorithm concerns economical strategies for encoding geometric information into discrete samples. The landmarking
procedure can also be interpreted as a compression scheme for manifolds; correspondingly, the standard manifold
learning algorithms may be understood as a decoding mechanism.
The Gaussian process landmarking algorithm presented in this paper takes one anatomical surface as input at one
time, which is not exactly consistent with the working style of geometric morphometricians. This special group of biologists are trained to identify semantically meaningful features not only based upon each individual shape in hand but
also with respect to the consistency of “global information” presented over the entire shape collection. In the standard
practice, landmarking a new anatomical surface typically involve repeated comparisons with all the other surfaces
in the collection, and the already placed landmarks on a surface are still subject to change upon future knowledge
acquired from landmarking more surfaces. It would be highly interesting to adapt our algorithm to accommodate for
this type of group-wise comparison strategies as well.
Though geometric approaches in general lack the ability to recognize Type 1 landmarks, we think that the similarities between Gaussian process landmarks and user selected Type 2 and Type 3 landmarks, over a number of tests
presented in this work, call into question whether user-based selection can be justified when “overall shape” is of
interest. In particular, we found the ordination plots to be as successful and in some ways superior to the user-based
landmarks in reflecting previous ideas about shape affinities. Given this, whether the ground truth of user-based landmarks is “better” or not becomes subjective. While F-ratios were highest in the user-determined sample, we note that
it is highly likely that, for any underdetermined landmarks (e.g. the Type 1 landmarks, Type 3 landmarks, or when
trying to place Type 2 landmarks on eroded features like blunted cusp tips) biologists will tend to minimize variance
within species. In other words, if there is not enough geometry to allow consistent placement of a point, users are
likely to unconsciously choose a positioning that visually maximizes similarity to other members of the species (actually, it would be hard to justify doing anything else). Given the likely bias towards minimizing within group error
by biologists during landmarking, it is quite remarkable that the Gaussian process approach comes so close to the
“ground truth”, and exceeds by so much the other automated methods in terms of F-ratio. In user placed landmarks,
there is also the question of how many and which landmarks were chosen, not to mention whether one user has more
error in placing the same landmarks than another. Finally, there is the limitation that any traditional landmark used
must be present in every specimen of the sample. Given all the issues with the traditional user placed approach, and
the demonstrated ability of Gaussian process to emulate qualitatively and quantitatively Type 2 and Type 3 landmarks,
we recommend avoiding hand-placement of such landmarks.
Last but not the least, this work stems from an attempt to impose Gaussian process priors on diffeomorphisms
between distinct but comparable biological structures, with which a rigorous Bayesian statistical framework for biological surface registration may be developed. The motivation is to measure the uncertainty of pairwise bijective
correspondences automatically computed from geometry processing and computer vision techniques. We hope this
conditional covariance based landmarking algorithm will shed light upon generalizing covariance structures from a
single shape to pairs or even collections of shapes.
A
Proof of Theorem 1
Proof of Theorem 1. Recall from [130](or [32, Lemma 8], [131, Lemma B.10]) that
!
Z
3
1
kx − yk2
exp
−
f (y) dVol (y) = f (x) + E (x) f (x) + ∆ f (x) + O 2
d
2
2
(2π) 2 M
32
as → 0
(46)
where E is a scalar function of the curvature of M at x ∈ M. Thus
!
Z
3
d
kz − yk2
exp [−V (z)]
exp −
f (y) dVol (y) = (2π) 2 exp [−V (z)] f (z) + E (z) f (z) + ∆ f (z) + O 2 .
2
2
M
(47)
Replacing the function f in (46) with the right hand side of (47), we have
!
!
Z Z
kx − zk2
kz − yk2
exp −
exp [−V (z)] exp −
f (y) dVol (z) dVol (y)
2
2
M M
(
3
= (2π)d exp [−V (x)] f (x) + E (x) f (x) + ∆ f (x) + O 2
2
3
+ E (x) exp [−V (x)] f (x) + E (x) f (x) + ∆ f (x) + O 2
2
2
3)
3
+O 2
+ ∆ exp [−V (x)] f (x) + E (x) f (x) + ∆ f (x) + O 2
2
2
(
= (2π)d f (x) exp [−V (x)]
+
o
3
n
E (x) f (x) + ∆ f (x) exp [−V (x)] + f (x) E (x) exp [−V (x)] + ∆ f (x) exp [−V (x)] + O 2
2
)
as → 0.
Plugging the right hand side above into the numerator of the left hand side of (22), and replacing the denominator
with the same expression as the numerator except for setting f ≡ 1 in (22), we obtain, as → 0,
!
!
Z Z
kx − zk2
kz − yk2
exp −
exp [−V (z)] exp −
f (y) dVol (z) dVol (y)
2
2
M M
!
!
Z Z
kx − zk2
kz − yk2
exp −
exp [−V (z)] exp −
dVol (z) dVol (y)
2
2
M M
o
3
n
f (x) exp [−V (x)] +
E (x) f (x) + ∆ f (x) exp [−V (x)] + f (x) E (x) exp [−V (x)] + ∆ f (x) exp [−V (x)] + O 2
2
=
o
3
n
exp [−V (x)] + E (x) exp [−V (x)] + E (x) exp [−V (x)] + ∆ exp [−V (x)] + O 2
2
(
i
3)
h
(x)
(x)
(x)
(x)]
(x)
(x)]
(x)
f
+ ∆f
+ exp [V
∆ f
exp [−V
+O 2 ×
= f
+ 2E
2
(
3)
1 − 2E (x) + exp [V (x)] ∆ exp [−V (x)] + O 2
2
i
3
h
= f (x) + 2E (x) f (x) + ∆ f (x) + exp [V (x)] ∆ f (x) exp [−V (x)] − 2E (x) f (x) − f (x) exp [V (x)] ∆ exp [−V (x)] + O 2
2
3
= f (x) + ∆ f (x) + exp [V (x)] ∆ f (x) exp [−V (x)] − f (x) exp [V (x)] ∆ exp [−V (x)] + O 2
2
3
3
= f (x) + 2∆ f (x) − 2∇ f (x) · ∇V (x) + O 2 = f (x) + ∆ f (x) − ∇ f (x) · ∇V (x) + O 2 .
2
This completes the proof.
Acknowlegement TG would also like to thank Peng Chen ([email protected]) at The University of Texas at
Austin for pointers to the reduced basis method literature, and Chen-Yun Lin ([email protected]) at Duke University for many useful discussions on heat kernel estimates. The authors would also like to thank Shaobo Han
([email protected]), Rob Ravier ([email protected]), and Shan Shan ([email protected]) for inspirational discussions. This work is supported by Simons Math+X Investigators Award #400837 and NSF CAREER
Award BCS-1552848.
33
References
[1] D. C. Adams, F. J. Rohlf, and D. E. Slice, A Field Comes of Age: Geometric Morphometrics in the 21st Century,
Hystrix, 24 (2013), pp. 7–14.
[2] R. Al-Aifari, I. Daubechies, and Y. Lipman, Continuous Procrustes Distance Between Two Surfaces, Communications on Pure and Applied Mathematics, 66 (2013), pp. 934–964.
[3] P. Alliez, D. Cohen-Steiner, O. Devillers, B. Lévy, and M. Desbrun, Anisotropic Polygonal Remeshing, ACM
Trans. Graph., 22 (2003), pp. 485–493.
[4] M. J. Anderson, A New Method for Non-Parametric Multivariate Analysis of Variance, Austral Ecology, 26
(2001), pp. 32–46.
[5] M. Atiyah and P. Sutcliffe, The Geometry of Point Particles, in Proceedings of the Royal Society of London A:
Mathematical, Physical and Engineering Sciences, vol. 458, The Royal Society, 2002, pp. 1089–1115.
[6] A. Atkinson, A. Donev, and R. Tobias, Optimum Experimental Designs, with SAS, vol. 34, Oxford University
Press, 2007.
[7] M. Aubry, U. Schlickewei, and D. Cremers, The Wave Kernel Signature: A Quantum Mechanical Approach to
Shape Analysis, in Computer Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on,
IEEE, 2011, pp. 1626–1633.
[8] T. F. Banchoff, Critical Points and Curvature for Embedded Polyhedral Surfaces, The American Mathematical
Monthly, 77 (1970), pp. 475–485.
[9] C. Beard, Teilhardina. The International Encyclopedia of Primatology. 1–2.,
10.1002/9781119179313.wbprim0444.
2017.
DOI:
[10] M. Belkin and P. Niyogi, Laplacian Eigenmaps for Dimensionality Reduction and Data Representation, Neural
Comput., 15 (2003), pp. 1373–1396.
[11]
, Towards a Theoretical Foundation for Laplacian-Based Manifold Methods, in Learning Theory,
Springer, 2005, pp. 486–500.
[12] P. Bérard, G. Besson, and S. Gallot, Embedding Riemannian Manifolds by Their Heat Kernel, Geometric &
Functional Analysis GAFA, 4 (1994), pp. 373–398.
[13] N. Berline, E. Getzler, and M. Vergne, Heat Kernels and Dirac Operators, Springer, 1992 ed., 12 2003.
[14] T. Berry and J. Harlim, Variable Bandwidth Diffusion Kernels, Applied and Computational Harmonic Analysis,
40 (2016), pp. 68–96.
[15] A. Bertozzi, X. Luo, A. Stuart, and K. Zygalakis, Uncertainty Quantification in the Classification of High
Dimensional Data. submitted, 2017.
[16] P. Binev, A. Cohen, W. Dahmen, R. DeVore, G. Petrova, and P. Wojtaszczyk, Convergence Rates for Greedy
Algorithms in Reduced Basis Methods, SIAM journal on mathematical analysis, 43 (2011), pp. 1457–1472.
[17] F. L. Bookstein, Morphometric Tools for Landmark Data: Geometry and Biology, Cambridge University Press,
1991.
[18] D. M. Boyer, L. Costeur, and Y. Lipman, Earliest Record of Platychoerops (Primates, Plesiadapidae), a new
species from Mouras Quarry, Mont de Berru, France, American Journal of Physical Anthropology, 149 (2012),
pp. 329–346.
[19] D. M. Boyer, Y. Lipman, E. St. Clair, J. Puente, B. A. Patel, T. Funkhouser, J. Jernvall, and I. Daubechies,
Algorithms to Automatically Quantify the Geometric Similarity of Anatomical Surfaces, Proceedings of the
National Academy of Sciences, 108 (2011), pp. 18221–18226.
34
[20] D. M. Boyer, J. Puente, J. T. Gladman, C. Glynn, S. Mukherjee, G. S. Yapuncich, and I. Daubechies, A New
Fully Automated Approach for Aligning and Comparing Shapes, The Anatomical Record, 298 (2015), pp. 249–
276.
[21] D. M. Boyer, S. Toussaint, and M. Godinot, Postcrania of the Most Primitive Euprimate and Implications for
Primate Origins, Journal of Human Evolution, 111 (2017), pp. 202–215.
[22] A. M. Bronstein, Spectral Descriptors for Deformable Shapes, arXiv preprint arXiv:1110.5015, (2011).
[23] A. Buffa, Y. Maday, A. T. Patera, C. Prud’homme, and G. Turinici, A Priori Convergence of the Greedy Algorithm for the Parametrized Reduced Basis Method, ESAIM: Mathematical Modelling and Numerical Analysis,
46 (2012), pp. 595–603.
[24] U. Castellani, M. Cristani, S. Fantoni, and V. Murino, Sparse Points Matching by Combining 3D Mesh
Saliency with Statistical Descriptors, Computer Graphics Forum, 27 (2008), pp. 643–652.
[25] I. Castillo, G. Kerkyacharian, and D. Picard, Thomas Bayes’ Walk on Manifolds, Probability Theory and
Related Fields, 158 (2014), pp. 665–710.
[26] K. N. Chaudhury, Y. Khoo, and A. Singer, Global Registration of Multiple Point Clouds using Semidefinite
Programming, SIAM Journal on Optimization, 25 (2015), pp. 468–501.
[27] X. Cheng, A. Cloninger, and R. R. Coifman, Two-Sample Statistics Based on Anisotropic Kernels, arXiv
preprint arXiv:1709.05006, (2017).
[28] K. Clarke and R. Green, Statistical Design and Analysis for a ’Biological Effects’ Study, Marine Ecology
Progress Series, (1988), pp. 213–226.
[29] K. R. Clarke, Non-Parametric Multivariate Analyses of Changes in Community Structure, Austral Ecology, 18
(1993), pp. 117–143.
[30] D. Cohen-Steiner and J.-M. Morvan, Restricted Delaunay Triangulations and Normal Cycle, in Proceedings
of the nineteenth annual symposium on Computational geometry, ACM, 2003, pp. 312–321.
[31] D. A. Cohn, Z. Ghahramani, and M. I. Jordan, Active Learning with Statistical Models, Journal of Artificial
Intelligence Research, 4 (1996), pp. 129–145.
[32] R. R. Coifman and S. Lafon, Diffusion Maps, Applied and Computational Harmonic Analysis, 21 (2006), pp. 5–
30. Special Issue: Diffusion Maps and Wavelets.
[33] N. Cressie, Statistics for Spatial Data, Wiley Series in Probability and Statistics, John Wiley & Sons, Inc.,
2015.
[34] N. Cristianini and J. Shawe-Taylor, An Introduction to Support Vector Machines and Other Kernel-Based
Learning Methods, Cambridge University Press Cambridge, 2000.
[35] L. Cuel, J.-O. Lachaud, Q. Mérigot, and B. Thibert, Robust Geometry Estimation Using the Generalized
Voronoi Covariance Measure, SIAM Journal on Imaging Sciences, 8 (2015), pp. 1293–1314.
[36] W. Czaja and M. Ehler, Schroedinger Eigenmaps for the Analysis of Biomedical Data, IEEE Transactions on
Pattern Analysis and Machine Intelligence, 35 (2013), pp. 1274–1280.
[37] R. DeVore, G. Petrova, and P. Wojtaszczyk, Greedy Algorithms for Reduced Bases in Banach Spaces, Constructive Approximation, 37 (2013), pp. 455–466.
[38] I. L. Dryden and K. V. Mardia, Statistical Shape Analysis, vol. 4, John Wiley & Sons New York, 1998.
[39] T. E. Fricker, J. E. Oakley, and N. M. Urban, Multivariate Gaussian Process Emulators with Nonseparable
Covariance Structures, Technometrics, 55 (2013), pp. 47–56.
[40] D. S. Fry, Shape Recognition Using Metrics on the Space of Shapes, PhD thesis, Harvard University, Cambridge, MA, USA, 1993. UMI Order No. GAX94-12337.
35
[41] E. Fuselier and G. B. Wright, Scattered Data Interpolation on Embedded Submanifolds with Restricted Positive
Definite Kernels: Sobolev Error Estimates, SIAM Journal on Numerical Analysis, 50 (2012), pp. 1753–1776.
[42] T. Gao, Hypoelliptic Diffusion Maps and Their Applications in Automated Geometric Morphometrics, PhD
thesis, Duke University, 2015.
[43]
, The Diffusion Geometry of Fibre Bundles, submitted, (2016). arXiv:1602.02330.
[44] T. Gao, J. Brodzki, and S. Mukherjee, The Geometry of Synchronization Problems and Learning Group Actions,
arXiv preprint arXiv:1610.09051, (2016).
[45] T. Gao, G. S. Yapuncich, I. Daubechies, S. Mukherjee, and D. M. Boyer, Development and Assessment of Fully
Automated and Globally Transitive Geometric Morphometric Methods, With Application to a Biological Comparative Dataset With High Interspecific Variation, The Anatomical Record, (2017), pp. n/a–n/a. Submitted.
[46] L. Z. Garamszegi, Modern Phylogenetic Comparative Methods and Their Application in Evolutionary Biology,
Concepts and Practice. London, UK: Springer, (2014).
[47] J. W. Gidley, Paleocene Primates of the Fort Union, with Discussion of Relationships of Eocene Primates,
Proceedings of the United States National Museum, 63 (1923), pp. 1–38.
[48] P. D. Gingerich, Cranial Anatomy and Evolution of Early Tertiary Plesiadapidae (Mammalia, Primates), University of Michigan Papers on Paleontology, 15 (1976), pp. 1–141.
[49]
[50]
, Radiation of Eocene adapidae in Europe, Geobios, 10 (1977), pp. 165–182.
, Early Eocene Cantius torresi—oldest Primate of Modern Aspect from North America, Nature, 319
(1986), pp. 319–321.
[51] M. Godinot, Primate Origins: A Reappraisal of Historical Data Favoring Tupaiid Affinities, Primate Origins:
Adaptations and Evolution. Springer, New York, (2007), pp. 83–142.
[52] T. F. Gonzalez, Clustering to Minimize the Maximum Intercluster Distance, Theoretical Computer Science, 38
(1985), pp. 293–306.
[53] P. I. Good, Permutation, Parametric, and Bootstrap Tests of Hypotheses (Springer Series in Statistics), SpringerVerlag New York, Inc., Secaucus, NJ, USA, 2004.
[54] J. C. Gower, Generalized Procrustes Analysis, Psychometrika, 40 (1975), pp. 33–51.
[55] J. C. Gower and G. B. Dijksterhuis, Procrustes Problems, vol. 3 of Oxford Statistical Science Series, Oxford
University Press Oxford, 2004.
[56] A. Grigor’yan, Heat Kernels on Weighted Manifolds and Applications, Cont. Math, 398 (2006), pp. 93–191.
[57] V. Guillemin and A. Pollack, Differential Topology, vol. 370, American Mathematical Soc., 2010.
[58] B. Güneysu, The Feynman-Kac Formula for Schrödinger Operators on Vector Bundles over Complete Manifolds, Journal of Geometry and Physics, 60 (2010), pp. 1997–2010.
[59] D. Hardin and E. Saff, Minimal Riesz Energy Point Configurations for Rectifiable d-Dimensional Manifolds,
Advances in Mathematics, 193 (2005), pp. 174–204.
[60] E. Harjunmaa, A. Kallonen, M. Voutilainen, K. Hämäläinen, M. L. Mikkola, and J. Jernvall, On the Difficulty
of Increasing Dental Complexity, Nature, 483 (2012), pp. 324–327.
[61] E. Harjunmaa, K. Seidel, T. Häkkinen, E. Renvoisé, I. J. Corfe, A. Kallonen, Z.-Q. Zhang, A. R. Evans,
M. L. Mikkola, I. Salazar-Ciudad, et al., Replaying Evolutionary Transitions From the Dental Fossil Record,
Nature, 512 (2014), pp. 44–48.
36
[62] B. R. Hassett and T. Lewis-Bale, Comparison of 3D Landmark and 3D Dense Cloud Approaches to Hominin
Mandible Morphometrics Using Structure-From-Motion, Archaeometry, 59 (2017), pp. 191–203. ARCH-050070-2015.R2.
[63] E. Hebey, Nonlinear Analysis on Manifolds: Sobolev Spaces and Inequalities, vol. 5, American Mathematical
Soc., 2000.
[64] B. Helffer and J. Sjöstrand, Puits Multiples en Mecanique Semi-Classique iv Etude du Complexe de Witten,
Communications in partial differential equations, 10 (1985), pp. 245–340.
[65] E. Hsu, Estimates of Derivatives of the Heat Kernel on a Compact Riemannian Manifold, Proceedings of the
american mathematical society, 127 (1999), pp. 3739–3744.
[66] A. Kapoor, K. Grauman, R. Urtasun, and T. Darrell, Active Learning with Gaussian Processes for Object
Categorization, in Computer Vision, 2007. ICCV 2007. IEEE 11th International Conference on, IEEE, 2007,
pp. 1–8.
[67] L. Karp and M. Pinsky, Volume of a Small Extrinsic Ball in a Submanifold, Bulletin of the London Mathematical
Society, 21 (1989), pp. 87–92.
[68] D. G. Kendall, Shape Manifolds, Procrustean Metrics, and Complex Projective Spaces, Bulletin of the London
Mathematical Society, 16 (1984), pp. 81–121.
[69] C.-W. Ko, J. Lee, and M. Queyranne, An Exact Algorithm for Maximum Entropy Sampling, Operations Research, 43 (1995), pp. 684–691.
[70] P. Koehl and J. Hass, Landmark-Free Geometric Methods in Biological Shape Analysis, Journal of The Royal
Society Interface, 12 (2015), p. 20150795.
[71] S. Z. Kovalsky, N. Aigerman, R. Basri, and Y. Lipman, Large-Scale Bounded Distortion Mappings, ACM Trans.
Graph., 34 (2015), pp. 191–1.
[72] S. Z. Kovalsky, M. Galun, and Y. Lipman, Accelerated quadratic proxy for geometric optimization, ACM
Transactions on Graphics (TOG), 35 (2016), p. 134.
[73] A. Krause, A. Singh, and C. Guestrin, Near-Optimal Sensor Placements in Gaussian Processes: Theory,
Efficient Algorithms and Empirical Studies, Journal of Machine Learning Research, 9 (2008), pp. 235–284.
[74] D. D. Lewis and W. A. Gale, A Sequential Algorithm for Training Text Classifiers, in Proceedings of the 17th
annual international ACM SIGIR conference on Research and development in information retrieval, SpringerVerlag New York, Inc., 1994, pp. 3–12.
[75] D. Liang and J. Paisley, Landmarking Manifolds with Gaussian Processes., in ICML, 2015, pp. 466–474.
[76] L. Lin, M. Niu, P. Cheung, and D. Dunson, Extrinsic Gaussian Process (EGPS) for Regression and Classification on Manifolds. private communication, 2017.
[77] L. Lin, B. St Thomas, H. Zhu, and D. B. Dunson, Extrinsic Local Regression on Manifold-Valued Data, Journal
of the American Statistical Association, (2016). to appear.
[78] Y. Lipman, Bounded Distortion Mapping Spaces for Triangular Meshes, ACM Transactions on Graphics (TOG),
31 (2012), p. 108.
[79] Y. Lipman and I. Daubechies, Conformal Wasserstein Distances: Comparing Surfaces in Polynomial Time,
Advances in Mathematics, 227 (2011), pp. 1047–1077.
[80] Y. Lipman, J. Puente, and I. Daubechies, Conformal Wasserstein Distance: II. Computational Aspects and
Extensions., Math. Comput., 82 (2013).
[81] Y. Lipman, S. Yagev, R. Poranne, D. W. Jacobs, and R. Basri, Feature Matching with Bounded Distortion, ACM
Transactions on Graphics (TOG), 33 (2014), p. 26.
37
[82] Y.-S. Liu, M. Liu, D. Kihara, and K. Ramani, Salient Critical Points for Meshes, in Proceedings of the 2007
ACM symposium on Solid and physical modeling, ACM, 2007, pp. 277–282.
[83] G. G. Lorentz, M. von Golitschek, and Y. Makovoz, Constructive Approximation: Advanced Problems,
vol. 304, Springer Berlin, 1996.
[84] W. Madych and S. Nelson, Bounds on Multivariate Polynomials and Exponential Error Estimates for Multiquadric Interpolation, Journal of Approximation Theory, 70 (1992), pp. 94–114.
[85] P. Malliavin and D. W. Stroock, Short Time Behavior of the Heat Kernel and Its Logarithmic Derivatives,
Journal of Differential Geometry, 44 (1996), pp. 550–570.
[86] N. Mantel, The Detection of Disease Clustering and a Generalized Regression Approach, Cancer research, 27
(1967), pp. 209–220.
[87] A. Martınez-Finkelshtein, V. Maymeskul, E. Rakhmanov, and E. Saff, Asymptotics for Minimal Discrete Riesz
Energy on Curves in Rd , Canad. J. Math, 56 (2004), pp. 529–552.
[88] S. Melzi, E. Rodolà, U. Castellani, and M. M. Bronstein, Localized Manifold Harmonics for Spectral Shape
Analysis, in Computer Graphics Forum, Wiley Online Library, 2017.
[89] P. Mitteroecker and P. Gunz, Advances in Geometric Morphometrics, Evolutionary Biology, 36 (2009),
pp. 235–247.
[90] P. Mitteroecker and S. M. Huttegger, The Concept of Morphospaces in Evolutionary and Developmental
Biology: Mathematics and Metaphors, Biological Theory, 4 (2009), pp. 54–67.
[91] C. Moenning and N. A. Dodgson, Fast Marching Farthest Point Sampling, tech. rep., University of Cambridge,
Computer Laboratory, 2003.
[92] M. Mohri, A. Rostamizadeh, and A. Talwalkar, Foundations of Machine Learning, MIT press, 2012.
[93] S. Molnar, On the Convergence of the Kriging Method, in Annales Univ Sci Budapest Sect Comput, vol. 6,
1985, pp. 81–90.
[94] B. Nadler, S. Lafon, R. R. Coifman, and I. G. Kevrekidis, Diffusion Maps, Spectral Clustering and Reaction
Coordinates of Dynamical Systems, Applied and Computational Harmonic Analysis, 21 (2006), pp. 113–127.
[95] A. Naor, O. Regev, and T. Vidick, Efficient Rounding for the Noncommutative Grothendieck Inequality, in
Proceedings of the forty-fifth annual ACM symposium on Theory of computing, ACM, 2013, pp. 71–80.
[96] F. J. Narcowich, J. D. Ward, and H. Wendland, Sobolev Error Estimates and a Bernstein Inequality for Scattered Data Interpolation via Radial Basis Functions, Constructive Approximation, 24 (2006), pp. 175–186.
[97] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher, An Analysis of Approximations for Maximizing Submodular
Set Functions-I, Mathematical Programming, 14 (1978), pp. 265–294.
[98] A. Nemirovski, Sums of Random Symmetric Matrices and Quadratic Optimization under Orthogonality Constraints, Mathematical programming, 109 (2007), pp. 283–317.
[99] T. E. Nichols and A. P. Holmes, Nonparametric Permutation Tests for Functional Neuroimaging: A Primer
with Examples, Human brain mapping, 15 (2002), pp. 1–25.
[100] S. Niranjan, A. Krause, S. M. Kakade, and M. Seeger, Gaussian Process Optimization in the Bandit Setting:
No Regret and Experimental Design, in Proceedings of the 27th International Conference on Machine Learning,
2010.
[101] J. E. Oakley and A. O’Hagan, Probabilistic Sensitivity Analysis of Complex Models: A Bayesian Approach,
Journal of the Royal Statistical Society: Series B (Statistical Methodology), 66 (2004), pp. 751–769.
[102] A. C. Öztireli, M. Alexa, and M. Gross, Spectral sampling of manifolds, ACM Transactions on Graphics
(TOG), 29 (2010), p. 168.
38
[103] E. Paradis, Analysis of Phylogenetics and Evolution with R, Springer Science & Business Media, 2011.
[104] F. Pesarin, Multivariate Permutation Tests: with Applications in Biostatistics, vol. 240, Wiley Chichester, 2001.
[105] D. L. Peutrec, F. Nier, and C. Viterbo, Precise Arrhenius Law for p-Forms: The Witten Laplacian and MorseBarannikov Complex, Annales Henri Poincaré, 14 (2013), pp. 567–610.
[106] U. Pinkall and K. Polthier, Computing Discrete Minimal Surfaces and Their Conjugates, Experimental mathematics, 2 (1993), pp. 15–36.
[107] I. Plyusnin, A. R. Evans, A. Karme, A. Gionis, and J. Jernvall, Automated 3D Phenotype Analysis Using Data
Mining, PLoS One, 3 (2008), p. e1742.
[108] K. Polthier, Computational Aspects of Discrete Minimal Surfaces, Proc. of the Clay Summer School on Global
Theory of Minimal Surfaces, to appear, (2002).
[109] J. Puente, Distances and Algorithms to Compare Sets of Shapes for Automated Biological Morphometrics, PhD
thesis, Princeton University, 2013.
[110] F. Pukelsheim, Optimal Design of Experiments, SIAM, 2006.
[111] J. Ramsay and B. Silverman, Functional Data Analysis, Springer Series in Statistics, Springer, 2005.
[112] J. O. Ramsay and B. W. Silverman, Applied Functional Data Analysis: Methods and Case Studies, vol. 77,
Springer New York, 2002.
[113] C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for Machine Learning, Adaptive Computation and
Machine Learning, The MIT Press, 2006.
[114] K. Ritter, Average-Case Analysis of Numerical Problems, Springer, 2007.
[115] F. J. Rohlf and F. L. Bookstein, Proceedings of the Michigan Morphometrics Workshop, University of Michigan
Museum of Zoology, 1990.
[116] S. Rosenberg, The Laplacian on a Riemannian Manifold: An Introduction to Analysis on Manifolds, no. 31 in
London Mathematical Society Student Texts, Cambridge University Press, 1997.
[117] V. L. Roth, On Three-Dimensional Morphometrics, and on the Identification of Landmark Points, in Contributions to Morphometrics, L. F. Marcus, E. Bello, and G.-V. A., eds., Museo Nacional de Ciencias Naturales,
Madrid, 1993, pp. 41–61.
[118] D. E. Russell, P. Louis, and D. E. Savage, Primates of the French Early Eocene, University of California
Publications in the Geological Sciences, 73 (1967), pp. 1–46.
[119] R. B. Rusu and S. Cousins, 3D is Here: Point Cloud Library (PCL), in IEEE International Conference on
Robotics and Automation (ICRA), Shanghai, China, May 9-13 2011.
[120] J. Sacks, S. B. Schiller, and W. J. Welch, Designs for Computer Experiments, Technometrics, 31 (1989),
pp. 41–47.
[121] A. Saltelli and S. Tarantola, On the Relative Importance of Input Factors in Mathematical Models: Safety
Assessment for Nuclear Waste Disposal, Journal of the American Statistical Association, 97 (2002), pp. 702–
709.
[122] T. J. Santner, B. J. Williams, and W. I. Notz, The Design and Analysis of Computer Experiments, Springer
Series in Statistics, Springer Science & Business Media, 2013.
[123] B. Scholkopf and A. J. Smola, Learning with Kernels: Support Vector Machines, Regularization, Optimization,
and Beyond, MIT press, 2001.
39
[124] J. H. Schwartz and I. Tattersall, Evolutionary Relationships of Living Lemurs and Lorises (Mammalia, Primates) and Their Potential Affinities with European Eocene Adapidae., Anthropological papers of the AMNH;
v. 60, pt. 1, (1985).
[125] E. R. Seiffert, L. Costeur, and D. M. Boyer, Tarsal Morphology of Caenopithecus, a Large Adapiform Primate
from the Middle Eocene of Switzerland, PeerJ, 3 (2015), p. e1036.
[126] B. Settles, Active Learning Literature Survey, University of Wisconsin, Madison, 52 (2010), p. 11.
[127] S.-J. Sheu, Some Estimates of the Transition Density of a Nondegenerate Diffusion Markov Process, The Annals
of Probability, (1991), pp. 538–561.
[128] M. T. Silcox, Plesiadapiform. The International Encyclopedia of Primatology. 1–2., 2017.
10.1002/9781119179313.wbprim0038.
DOI:
[129] E. L. Simons, A New Eocene Primate Genus, Cantius, and a Revision of Some Allied European Lemuroids,
vol. 7, British Museum, 1962.
[130] A. Singer, From Graph to Manifold Laplacian: The Convergence Rate, Applied and Computational Harmonic
Analysis, 21 (2006), pp. 128–134.
[131] A. Singer and H.-T. Wu, Vector Diffusion Maps and the Connection Laplacian, Communications on Pure and
Applied Mathematics, 65 (2012), pp. 1067–1144.
[132] J. Smith and S. Schaefer, Bijective parameterization with free boundaries, ACM Transactions on Graphics
(TOG), 34 (2015), p. 70.
[133] K. Smith, On the Standard Deviations of Adjusted and Interpolated Values of an Observed Polynomial Function
and its Constants and the Guidance They Give Towards a Proper Choice of the Distribution of Observations,
Biometrika, 12 (1918), pp. 1–85.
[134] O. Smolyanov, H. Weizsäcker, and O. Wittich, Brownian Motion on a Manifold as Limit of Stepwise Conditioned Standard Brownian Motions, Stochastic Processes, Physics and Geometry: New Interplays, II, 29 (2000),
pp. 589–602.
[135] O. G. Smolyanov, H. v Weizsäcker, and O. Wittich, Chernoff’s Theorem and Discrete Time Approximations of
Brownian Motion on Manifolds, Potential Analysis, 26 (2007), pp. 1–29.
[136] A. M.-C. So, Moment Inequalities for Sums of Random Matrices and Their Applications in Optimization, Mathematical Programming, 130 (2011), pp. 125–151.
[137] M. L. Stein, Interpolation of Spatial Data: Some Theory for Kriging, Springer Science & Business Media,
2012.
[138] D. W. Stroock and J. Turetsky, Upper Bounds on Derivatives of the Logarithm of the Heat Kernel, Communications in Analysis and Geometry, 6 (1998), pp. 669–685.
[139] G. K. Tam, Z.-Q. Cheng, Y.-K. Lai, F. C. Langbein, Y. Liu, D. Marshall, R. R. Martin, X.-F. Sun, and P. L.
Rosin, Registration of 3D Point Clouds and Meshes: A Survey from Rigid to Nonrigid, IEEE Transactions on
Visualization and Computer Graphics, 19 (2013), pp. 1199–1217.
[140] A. B. Tsybakov, Introduction to Nonparametric Estimation, Springer Publishing Company, Incorporated,
1st ed., 2008.
[141] K. Turner, S. Mukherjee, and D. M. Boyer, Persistent homology transform for modeling shapes and surfaces,
Information and Inference: A Journal of the IMA, 3 (2014), pp. 310–344.
[142] A. W. van der Vaart and J. H. van Zanten, Reproducing Kernel Hilbert Spaces of Gaussian Priors, in Pushing
the Limits of Contemporary Statistics: Contributions in Honor of Jayanta K. Ghosh, Institute of Mathematical
Statistics, 2008, pp. 200–222.
40
[143] N. Vitek, C. Manz, T. Gao, J. Bloch, S. Strait, and D. M. Boyer, Semi-supervised Determination of Pseudocryptic Morphotypes Using Observer-free Characterizations of Anatomical Alignment and Shape. Accepted,
2017.
[144] W. Wang, R. Tuo, and C. J. Wu, Universal Convergence of Kriging, arXiv preprint arXiv:1710.06959, (2017).
[145] H. Wendland, Scattered Data Approximation, vol. 17, Cambridge University Press, 2004.
[146] H.-T. Wu, Embedding Riemannian Manifolds by the Heat Kernel of the Connection Laplacian, Advances in
Mathematics, 304 (2017), pp. 1055–1079.
[147] Z.-M. Wu and R. Schaback, Local Error Estimates for Radial Basis Function Interpolation of Scattered Data,
IMA Journal of Numerical Analysis, 13 (1993), pp. 13–27.
[148] Y. Yang and D. B. Dunson, Bayesian Manifold Regression, The Annals of Statistics, 44 (2016), pp. 876–905.
[149] G. S. Yapuncich, E. R. Seiffert, and D. M. Boyer, Quantification of the Position and Depth of the Flexor
Hallucis Longus Groove in Euarchontans, with Implications for the Evolution of Primate Positional Behavior,
American Journal of Physical Anthropology, 163 (2017), pp. 367–406.
[150] D. Ylvisaker, Designs on Random Fields, A Survey of Statistical Design and Linear Models, 37 (1975),
pp. 593–607.
[151] M. L. Zelditch, D. L. Swiderski, and H. D. Sheets, eds., Geometric Morphometrics for Biologists: A Primer,
Academic Press, San Diego, second ed., 2012.
41
| 1cs.CV
|
Provable Dynamic Robust PCA or Robust Subspace Tracking
Praneeth Narayanamurthy and Namrata Vaswani
{pkurpadn, namrata} @iastate.edu
Iowa State University, Ames, IA, USA
arXiv:1705.08948v2 [cs.IT] 22 Oct 2017
October 24, 2017
Abstract
Dynamic robust PCA refers to the dynamic (time-varying) extension of the robust PCA (RPCA)
problem. It assumes that the true (uncorrupted) data lies in a low-dimensional subspace that can change
with time, albeit slowly. The goal is to track this changing subspace over time in the presence of sparse
outliers. This work provides the first guarantee for dynamic RPCA that holds under weakened versions
of standard RPCA assumptions and a few other simple assumptions. We analyze a novel algorithm
based on the recently introduced Recursive Projected Compressive Sensing (ReProCS) framework. Our
result is significant because (i) it removes the strong assumptions needed by the two previous complete
guarantees for ReProCS-based algorithms; (ii) it shows that, it is possible to achieve significantly
improved outlier tolerance by exploiting slow subspace change and a lower bound on most outlier
magnitudes; and (iii) it proves that the proposed algorithm is online (after initialization), fast, and, has
near-optimal storage complexity.
1
Introduction
Principal Components Analysis (PCA) is a widely used dimension reduction technique in a variety of
scientific applications. Given a set of data vectors, PCA tries to finds a smaller dimensional subspace that
best approximates a given dataset. According to its modern definition [1], robust PCA (RPCA) is the
problem of decomposing a given data matrix into the sum of a low-rank matrix (true data) and a sparse
matrix (outliers). The column space of the low-rank matrix then gives the desired principal subspace
(PCA solution). In recent years, the RPCA problem has been extensively studied, e.g., [1, 2, 3, 4, 5, 6]. A
common application of RPCA is in video analytics in separating video into a slow-changing background
image sequence (modeled as a low-rank matrix) and a foreground image sequence consisting of moving
objects or people (sparse) [1]. Dynamic RPCA refers to the dynamic (time-varying) extension of RPCA.
It assumes that the true (uncorrupted) data lies in a low-dimensional subspace that can change with time,
albeit slowly. This is a more appropriate model for long data sequences, e.g., surveillance videos. The goal
is to track this changing subspace over time in the presence of sparse outliers. Hence this problem can
also be referred to as robust subspace tracking. We study an algorithm based on the Recursive Projected
Compressive Sensing (ReProCS) framework [7] that is significantly simpler compared to those studied in
earlier works [7, 8, 9].
Dynamic RPCA or Robust Subspace Tracking Problem Statement. At each time t, we observe data
vectors yt ∈ Rn that satisfy
yt := `t + xt , for t = 1, 2, . . . , d
(1)
where xt is the sparse outlier vector and `t is the true data vector that lies in a fixed or slowly changing
low-dimensional subspace of Rn . To be precise, `t = Qt at where Qt is an n × r basis matrix1 with r n
1
tall matrix with mutually orthonormal columns
1
and with k(I − Qt−1 Qt−1 0 )Qt k small compared to kQt k = 1. Here and below, 0 denotes matrix transpose
and k · k refers to the l2 norm of a vector or the induced l2 norm of a matrix. We use Tt to denote the
support set of xt .
Given an initial subspace estimate, P̂0 , the goal is to track span(Qt ) within a short delay of each
subspace change. The initial estimate can be obtained by applying any static (batch) RPCA technique,
e.g., PCP [1] or AltProj [4], to the first ttrain data frames, Y[1,ttrain ] . Here and below, [a, b] refers to all
integers between a and b, inclusive and [a, b) := [a, b − 1]. Also, MT denotes a sub-matrix of M formed
by columns indexed by entries in the set T . A by-product of our solution approach is that the true data
vectors `t , the sparse outliers xt , and their support sets Tt can also be tracked on-the-fly. In many practical
applications, in fact, xt or Tt is often the quantity of interest.
For basis matrices P̂ , P , we use SE(P̂ , P ) := k(I − P̂ P̂ 0 )P k to quantify the subspace error (SE)
between their respective column spans. This measures the principal angle between the subspaces. Thus
slow subspace change means SE(Qt−1 , Qt ) 1.
We assume that (i) |Tt |/n is upper bounded, (ii) Tt changes enough over time so that any one index is
not part of the outlier support for too long, and (iii) the columns of Qt are dense (non-sparse). These
assumptions are quantified in Sec. 3. For the purpose of obtaining guarantees, we will further assume
that the subspace span(Qt ) is piecewise constant with time, i.e., that Qt = Qtj for all t ∈ [tj , tj+1 ) and
that there are a total of J changes. Thus t0 = 1 and tJ+1 = d. Also, at each change time, tj , the change
is “small”. Let Pj := Qtj . This will mean two things. First, only one direction changes at each tj (the
changing direction could be different for different change times), and second that SE(Pj−1 , Pj ) ≤ ∆ 1.
We will also study noisy dynamic RPCA. This is the case where
yt := `t + xt + vt , for t = 1, 2, . . . , d
(2)
and vt is small bounded noise.
Relation to original (static) RPCA. To connect with the original RPCA problem [1, 3, 4], define
the n × d data matrix Y := [y1 , y2 , . . . yd ] := L + X where L, X are similarly defined. Let rmat denote
the rank of L and use outlier-frac-col and outlier-frac-col to denote the maximum fraction of outliers per
column and per row of Y . Also let outlier-frac = max(outlier-frac-row, outlier-frac-col). RPCA results
bound outlier-frac. For dynamic RPCA solutions, we will define outlier-frac-row slightly differently. It
will be the fraction per row of any n × α sub-matrix of Y where α denotes the number of frames used in
each subspace update. Since α is large enough (see (10)), the two definitions are only a little different.
The dynamic RPCA problem assumes a bound on maxt |Tt |/n. This is equivalent to bounding on
outlier-frac-col. The requirement of Tt ’s changing enough is equivalent to a bound on outlier-frac-row.
The denseness assumption on the Qt ’s is similar to denseness (incoherence) of left singular vectors of L
assumed by RPCA.
With the piecewise constant subspace assumption, another way to understand dynamic RPCA is as
follows. Split the matrix L into sub-matrices L = [L0 , L1 , L2 , . . . LJ ] with Lj := L[tj ,tj+1 ) . Then, each Lj
has rank r, and is such that the column spaces of Lj−1 and Lj differ by only a little (as quantified by the
slow subspace change model summarized above). Also, the rank of L is rmat ≤ r + J. Given an estimate
of a part of L0 (obtained using PCP or AltProj), the goal is to recursively estimate the `t ’s (which are
columns of L) from sequentially arriving sparse outlier corrupted data vectors yt .
Contributions. The most important contribution of this work is that it obtains the first correctness
guarantee for solving dynamic RPCA that holds under weakened versions of standard RPCA assumptions,
slow subspace change, and a few simple assumptions (it needs that outlier magnitudes are either large
enough or very small, and it needs `t ’s to be mutually independent). We say “weakened” because our
guarantee implies that our proposed algorithm can tolerate an order-wise larger fraction of outliers per
row than existing static RPCA approaches without requiring the outlier support to be uniformly randomly
generated or without needing any other model on support change. For the video application, this implies
that it tolerates slow moving and occasionally static foreground objects much better than the static
approaches. This fact is also backed up by comparisons on real videos shown in Sec. 8.
2
A second contribution is the algorithm itself which we call simple-ReProCS (s-ReProCS). Unlike
previous provably correct ReProCS-based methods [8, 9], s-ReProCS ensures that the estimated subspace
dimension is bounded by (r + 1) at all times without needing the complicated cluster-EVD step. Moreover,
to our best knowledge, s-ReProCS is the first RPCA or dynamic RPCA solution that has near-optimal
storage complexity while allowing the subspace dimension to be greater than one. We also develop an
offline extension of s-ReProCS.
The current work gives a significantly shorter and simpler proof than the earlier guarantees for
ReProCS-based methods. It does this, in part, by first separately proving a result for the problem of
“correlated-PCA” or “PCA in data-dependent noise” [10, 11] with partial subspace knowledge. This result
may also be of independent interest.
Paper Organization. We discuss the related work below in Sec. 2 and explain how our guarantee
compares with other provable results on RPCA or dynamic RPCA from the literature. The proposed
algorithm, simple-ReProCS, and its performance guarantees, Theorem 3.1 and Corollary 3.2, are given in
Sec. 3. This section also explains the notation used in the paper. We discuss the guarantees in detail
in Sec. 4. Sec. 5 provides the main ideas that lead to the proof of Theorem 3.1. We prove Theorem 3.1
under the assumption that the subspace change times are known in Sec. 6. This proof helps illustrate
all the ideas of the actual proof but with minimal extra notation. The general proof of Theorem 3.1 is
given in Appendix C. Theorem 3.1 relies on a guarantee for PCA in data-dependent noise [10, 11] when
partial subspace knowledge is available. This result is proved in Sec. 7. We provide a summary of our
empirical evaluation of simple-ReProCS in Sec. 8. The details of the experimental comparisons are given
in Appendix G. We conclude and discuss future directions in Sec. 9.
2
Discussion of Related Work
In terms of other solutions for provably correct dynamic RPCA, there is very little work. This includes an
early partial guarantee [7] and recent correctness results for [8, 9] ReProCS-based algorithms. This also
includes modified-PCP, which is a solution for RPCA with partial subspace knowledge, and which can be
shown to also solve dynamic RPCA [12]. We discuss these below. To place our work in context we also
provide a detailed comparison of offline s-ReProCS with the provably correct offline solutions for RPCA
[1, 2, 3, 4, 5, 6].
Other than the above, there is work on online or streaming algorithms for the original (static) RPCA
problem. A very recent streaming solution for it is [13]. This comes with a provable guarantee but only
for rmat = r = 1 (one-dimensional RPCA). In earlier work, an online stochastic optimization based solver
for PCP (ORPCA) was developed [14]. This came with only a partial guarantee, i.e., the guarantee
required assumptions on intermediate algorithm estimates. The result assumed that the basis matrix for
the subspace estimate at each t was full rank. Other than the above methods, many heuristics have been
developed to solve RPCA in an online fashion, e.g., GRASTA [15], and older work, e.g., [16, 17]. From a
practical standpoint, any online algorithm will implicitly also provide a tracking solution. However, as our
experimental comparisons with ORPCA and GRASTA show, the solution may not be as good as that of
ReProCS which explicitly exploits slow subspace change. We do not discuss these works here in detail
since our work mainly focuses on provable guarantees.
Limitations of earlier ReProCS-based guarantees [7, 8, 9]. In [7], the ReProCS idea was first introduced
and a partial guarantee for a ReProCS-based algorithm was proved. We call it a partial guarantee because
it needed to assume something about the intermediate subspace estimates returned by the algorithm.
However, this work is important because it developed a nice framework for proving guarantees for dynamic
RPCA solutions. Both the later complete guarantees [8, 9] as well as our current result build on this
framework.
Our work is a significant improvement over the complete guarantees obtained in [8, 9] for two other
ReProCS-based algorithms. (i) These needed very specific assumptions on how the outlier support could
change. Our result removes such a requirement and instead only needs a bound on the fraction of outliers
3
Table 1: Comparing s-ReProCS and offline s-ReProCS with other RPCA methods. See note below the table for details.
Assumptions
PCP(C)
PCP(H)[3]/AltProj[4]/
GD
Proposed alg.: s-ReProCS/
S-RPCA
mod-PCP
[1]
NO-RMC[6]
[5]
offline s-ReProCS
[13]
[12]
online (recursive)
No
No
No
Yes
Yes
Yes
unif. random support
Yes
No
No
No
No
Yes
outlier-frac-row ≤
c
c/rmat
c/rmat
3/2
c
c/rmat
c
outlier-frac-col ≤
c
c/rmat
c/rmat
3/2
c/r
c/rmat
c
for rmat = r = 1
rmat := rank(L) ≤
c logn2 n
other assumptions
(i) strong incoh.,
(ii) right incoh
(i) right incoh
right-incoh
(ii) NO-RMC also needs d ≈ n
Time/Storage/
PCP
c logn2 n
1
(i) slow subspace change, (ii) outlier
AltProj
magnitudes large or very small,
(i) slow subspace change,
right incoh
(ii) right incoh, (iii) strong incoh (weaker),
(iii) at ’s bounded, Y[1,ttrain ] satisfies
Y[1,ttrain ] satisfies
outlier-frac ≤ c/r & right incoh
outlier-frac ≤ c/r & right incoh
GD
NO-RMC
no. algo params
Proposed alg.: s-ReProCS
S-RPCA
offline s-ReProCS
Time
O(nd2 1 )
2
O(ndrmat
log 1 )
O(ndrmat log 1 )
3
O(nrmat
log2 n)
O(ndr log 1 )
Memory
O(nd)
O(nd)
O(nd)
O(nd)
O(nr(log n))
O(ndrmat log 1 )
for rmat = r = 1
O(nr(log n) log
no. algo. params
1
2
5
4
4
O(nrmat )
1
)
for rmat = r = 1
4
The top table compares assumptions required by various methods. All approaches need left incoherence
and hence that is not compared. The comparison also ignores all dependence on condition numbers
(treats it a constant). Notation:
“strong incoh” refers to the strong incoherence assumption needed by
q
PCP(C) [1]: kU V 0 kmax ≤
µrmat
nd
where k.kmax is the maximum magnitude entry, and U , V are the left
and right singular vectors of L; “left incoh” means maxi kIi 0 U k2 ≤ µrmat /n and “right incoh” means
maxi kIi 0 V k2 ≤ µrmat /d; “outlier mag. either large or very small” refers to the lower bound on
xmin /15 − bv given in Corollary 3.2 (see first two paras in Sec. 4). The bottom table compares
computational and storage complexity. Here denotes final desired error.
per column of the data matrix and on the fraction per row of an α-consecutive-column sub-matrix of the
data matrix (for α large enough). (ii) The subspace change model assumed in the earlier papers can be
interpreted as the current model (given below in Sec. 3) with θj = 90◦ or equivalently with ∆min = ∆ = 1.
This is an unrealistic model for slow subspace change, e.g., in 3D, it implies that the subspace changes
from the x-y plane to the y-z plane. Instead, our current model allows changes from x-y plane to a slightly
tilted x-y plane as shown in Fig. 1. This modification is more realistic and it allows us to replace the
upper bound on λchd required by the earlier results by a similar bound on λchd ∆2 . Since ∆ quantifies
rate of subspace change, this new requirement is much weaker. It can be satisfied by assuming that ∆ is
small. (iii) The required minimum delay between subspace change times in the earlier results depended
on 1/2 where is the desired final subspace error after a subspace update is complete. This is a strong
requirement. Our current result removes this unnecessarily strong dependence. The delay now only
depends on (− log ) which makes it much smaller. (iv) Unlike [8, 9], we analyze a simple ReProCS-based
algorithm that ensures that the estimated subspace dimension is bounded by (r + 1), without needing the
complicated cluster-SVD algorithm. This is why our guarantee allows outlier fractions per column to be
below c/r. The work of [8] needed this to be below c/rmat while [9] needed an extra assumption (clustered
eigenvalues). For long data sequences, c/r can be much larger than c/rmat .
Complete guarantees for other dynamic RPCA or RPCA solutions. Another approach that solves
dynamic RPCA, but in a piecewise batch fashion, is modified-PCP (mod-PCP) [12]. The guarantee
for mod-PCP was proved using ideas borrowed from [1] for PCP. Thus, like [1], it also needs uniformly
randomly generated support sets which is an unrealistic requirement. For the video application, this
4
requires that foreground objects are single pixel wide and move around the entire image completely
randomly over time. This is highly impractical. We compare our guarantee with those of modified-PCP
and of streaming-RPCA [13] (which works only for rmat = r = 1) in Table 1. In the same table, we also
compare our guarantee for an offline extension of s-ReProCS with the key results for offline and batch
methods for static RPCA - [1] (referred to as PCP(C)), [3] (PCP(H), this strictly improves upon [2]),
AltProj [4], RPCA via gradient descent (GD) [5] and nearly-optimal robust matrix completion (NO-RMC)
[6]. The table also contains a speed and memory complexity comparison.
From Table 1, it is clear that for data that satisfies slow subspace change and the assumption that
outlier magnitudes are either large or very small, and that is such that its first ttrain frames, Y[1,ttrain ] ,
satisfy AltProj (or PCP) assumptions, s-ReProCS and offline s-ReProCS have the following advantages
over other methods.
1. For the data matrix after ttrain , i.e., for Y[ttrain +1,d] , ReProCS needs the weakest bound on outlier-frac-row
without requiring strong incoherence or uniformly randomly generated outlier support sets. This is
comparable to the bound needed by PCP(C) or mod-PCP but both assume uniform random outlier
supports which is a very strong requirement. PCP(C) also needs strong incoherence.
2. For the data matrix after ttrain , i.e., for Y[ttrain +1,d] , the ReProCS bound of c/r on outlier-frac-col is
also weaker than the bound of c/rmat which is needed by all the other RPCA guarantees that do
not assume uniformly random outlier supports. Recall that rmat = r + J. When r is small, e.g.,
r ∈ O(log n), but J is much larger, e.g., J = cd/(r log n), then rmat is of the same order as J whereas
r is much smaller. Thus, ReProCS works as long as outlier-frac-col ≤ c/r = c/ log n which is a pretty
weak requirement. However, all others require outlier-frac-row ≈ outlier-frac-col ∈ O(log2 n/d). This
is a very strong requirement since d can be as large as n (or even more). Moreover, S-RPCA result
also requires rmat = r = 1.
3. The storage complexity of s-ReProCS is better than that of all other methods that provably work
for ranks larger than one and it is nearly optimal. Observe that it is within logarithmic factors of nr
which equals the degrees of freedom of an r-dimensional subspace in Rn .
4. Both in terms of time complexity order (Table 1) and experimentally (see Sec. 8), s-ReProCS and
its offline counterpart are among the fastest, while having the best, or nearly the best, performance
experimentally as well. Order-wise, only NO-RMC [6] is faster than s-ReProCS. However, NO-RMC
cannot recover the outlier matrix X or its support. The reason is that it deliberately under-samples
the data matrix Y by randomly throwing away some of its entries and using only the rest even when
all are available. In other words, it always solves the robust matrix completion (RPCA with missing
entries) problem and this is what results in a significant speed-up. This is also why it cannot recover
all entries of X.
Notice that ReProCS assumes bounded-ness of the principal subspace coefficients at . This is also needed
by all the other methods when they impose incoherence of right singular vectors of L (right incoh in the
table). In fact, right incoherence is a stronger requirement than just bounded-ness of the at ’s.
Another way to compare ReProCS with the batch solutions is to assume that these are applied on
pieces of data Yj := Y[tj ,tj+1 ) . If this were done, rmat in the table will get replaced by r, and d will get
replaced by (tj+1 − tj ), both of which are smaller. With this, the second limitation discussed above
disappears but the first one still remains. The speed and memory limitations become less serious. From
the assumptions required (which are sufficient conditions), there seems to be no disadvantage to using
fewer columns for the batch methods. But in practice, there is some loss in performance especially when
outlier fractions are larger.
5
Algorithm 1 Simple-ReProCS (with tj known). We state this first for simplicity. The actual automatic version is
given later in Algorithm 4. Let L̂t;α := [`ˆt−α+1 , `ˆt−α+2 , . . . , `ˆt ].
1:
2:
3:
4:
5:
6:
7:
Input: P̂0 , yt , Output: x̂t , `ˆt , P̂t , Parameters: ωsupp , K, α, ξ, r, tj ’s
Q̂ttrain ← P̂0 ; j ← 1, k ← 1
for t > ttrain do
(x̂t , T̂t ) ← ProjCS(Q̂t−1 , yt )
% Algorithm 2
ˆ
`t ← yt − x̂t .
(Q̂t , P̂j , j, k) ← SubUp(L̂t;α , P̂j−1 , t, tj , j, k, Q̂t−1 )
% Algorithm 3
end for
Algorithm 2 Projected CS (ProjCS)
function ProjCS(Q̂t−1 , yt )
Ψ ← I − Q̂t−1 Q̂t−1 0
ỹt ← Ψyt
x̂t,cs ← arg minx̃ kx̃k1 s.t kỹt − Ψx̃k ≤ ξ
T̂t ← {i : |x̂t,cs | > ωsupp }
x̂t ← IT̂t (ΨT̂t 0 ΨT̂t )−1 ΨT̂t 0 ỹt
return x̂t , T̂t
end function
3
Algorithm and Main Result
In this section, we give the notation, explain and summarize the proposed algorithm simple-ReProCS
(s-ReProCS), and finally give the precise assumptions needed and the main result. We discuss the result
in Sec. 4.
3.1
Notation
We use the interval notation [a, b] to mean all of the integers between a and b, inclusive. We will often
use J to denote a time interval and J α to denote a time interval of length α. We use 1S to denote the
indicator function for statement S, i.e. 1S = 1 if S holds and 1S = 0 otherwise. We use k · k without a
subscript to denote the l2 norm of a vector or the induced l2 norm of a matrix. For other lp norms, we use
k · kp . For a set T , we use IT to refer to an n × |T | matrix of columns of the identity matrix indexed by
entries in T . With this notation, Ii refers to the i-th column of the identity matrix2 . For a matrix A, A0
denotes its transpose and AT := AIT is the sub-matrix of A that contains the columns of A indexed by
entries in T . We use λmin (.) (σmin (.)) to denote the minimum eigen (singular) value of a matrix. Similarly
for λmax (.) and σmax (.).
A matrix with mutually orthonormal columns is referred to as a basis matrix. For basis matrices P̂ , P ,
we use SE(P̂ , P ) := k(I − P̂ P̂ 0 )P k to quantify the subspace error (SE) between their respective column
spans. This measures the sine of the principal angle between the subspaces.
The letters c and C denote different numerical constants in each use.
3.2
Simple-ReProCS (s-ReProCS)
Simple-ReProCS proceeds as follows. The initial subspace is assumed to be accurately known (obtained
using AltProj or PCP). At time t, if the previous subspace estimate, Q̂t−1 , is accurate enough, because
of slow subspace change, projecting yt = xt + `t onto its orthogonal complement will nullify most of `t .
2
Many papers use ei to denote Ii , however, in this paper we use et for something else and hence do not use ei for the i-th
column of I.
6
Algorithm 3 Subspace Update (SubUpd).
function SubUp(L̂t;α , P̂j−1 , t, tj , j, k, Q̂t−1 )
if t = tj + uα for u = 1, 2, · · · , K + 1 then
B ← (I − P̂j−1 P̂j−1 0 )L̂t;α
P̂j,chd,k ← SV D1 [B]
% subspace addition: via K steps of projection-SVD
Q̂t ← [P̂j−1 , P̂j,chd,k ], k ← k + 1.
if k = K + 1 then
P̂j ← SV Dr [L̂t;α ]
% subspace deletion: via subspace re-estimation using simple SVD
Q̂t ← P̂j , j ← j + 1, k ← 1.
end if
else
Q̂t ← Q̂t−1
end if
return Q̂t , P̂j , j, k
end function
We compute ỹt := Ψyt where Ψ := I − Q̂t−1 Q̂t−1 0 . Thus, ỹt = Ψxt + bt where bt := Ψ`t and kbt k is
small. Recovering xt from ỹt is thus a traditional compressive sensing (CS) / sparse recovery problem in
small noise [18]. This is solvable because incoherence (denseness) of Qt ’s and slow subspace change implies
that Ψ satisfies the restricted isometry property [7, Lemma 3.7]. We compute x̂t,cs using l1 minimization
followed by thresholding based support estimation to get T̂t . A Least Squares (LS) based debiasing step
on T̂t returns the final x̂t . We then estimate `t as `ˆt = yt − x̂t .
The `ˆt ’s are used for the subspace update step which involves (i) detecting subspace change; (ii)
obtaining improved estimates of the changed direction(s) by K steps of projection-SVD [7], each done with
a new set of α frames of `ˆt ; and (iii) a simple SVD based subspace re-estimation step, done with another
new set of α frames. The subspace update steps are designed assuming a piecewise constant subspace
change model; however, as can be seen from our experiments, the algorithm itself works even without
this assumption (it works for real videos as well). The approach works because, every time the subspace
changes, the change can be detected within a short delay, and after that, the K projection-SVD steps help
get progressively improved estimates of the changed subspace direction. The simple SVD step re-estimates
the entire subspace and, in the process, deletes the deleted subspace direction from the estimate.
We summarize a basic version of simple-ReProCS in Algorithm 1. This assumes that the change times
tj are known. The actual algorithm that detects changes automatically is longer and is given as Algorithm
4 later. The estimates of the subspace or of `t ’s can be improved in offline mode as follows. At t = t̂j + Kα,
the K projection-SVD steps are done and hence the subspace estimate at this time is accurate enough
with high probability (whp). At this time, offline s-ReProCS (last line of Algorithm 4) goes back and sets
Q̂t = [P̂j−1 , P̂j,chd,K ] for all t ∈ [t̂j−1 + Kα, t̂j + Kα). It also uses this to get improved estimates of x̂t
and `ˆt for all these times t.
3.3
Assumptions and Main Result
Subspace change assumption. We assume that Qt is piecewise constant with time, i.e.,
Qt = Qtj for all t ∈ [tj , tj+1 ), j = 0, 1, . . . , J,
(3)
with t0 = 1 and tJ+1 = d. Henceforth let Pj := Qtj . Consider the change from Pj−1 to Pj . Let Pj−1,chg
denote the direction from span(Pj−1 ) that “changes” at tj and let Pj,chd denote its “changed” version.
Thus span(Pj−1 ) = span([Pj−1,fix , Pj−1,chg ]) and span(Pj ) = span([Pj−1,fix , Pj,chd ]) where Pj−1,fix is an
n × (r − 1) matrix that denotes the “fixed” part of the subspace at tj . Assume that
SE(Pj−1 , Pj ) = SE(Pj−1,chg , Pj,chd ) ≤ ∆.
7
(4)
Pj−1,new
span (Pj )
θj
Pj−1,ch
Pj−1,fix
span (Pj−1 )
Figure 1: Subspace change example in 3D with r = 2.
The equality is trivial: it only implies that Pj,chd is orthogonal to Pj−1,fix . The inequality with ∆ 1
implies slow subspace change.
In this work we obtain a guarantee with assuming the simplest possible model on subspace change
where, at a time, only one direction can change. Observe though that, at different change times, the
changing direction could be different and hence, over a long period of time, the entire subspace could
change.
It is possible to easily replace one direction changing by rchg > 1 directions changing and a simple
model on how the changes occur (with rchg being small compared to r). We explain this in Appendix
E. It is also possible to study the most general case where rchg = r and hence no model is assumed for
subspace change (only a bound on how much the change is). This requires significant changes to the
algorithm and the guarantee and is being studied in follow-up work. Finally, it is even possible to let
the subspace dimension r change at each change time, although this will require some changes to the
statistical assumption on the coefficients at given below.
Understanding the subspace change assumption: equivalent generative model. To understand the
above assumption better, let the new direction added to the subspace at time tj be Pj,new , i.e., let
Pj,new :=
(I − Pj−1,chg Pj−1,chg 0 )Pj,chd
SE(Pj−1,chg , Pj,chd )
(5)
(the denominator helps ensure that kPj,new k = 1). Also, let the angle between Pj−1,chg and Pj,chd be θj ,
i.e., let
θj := cos−1 |Pj−1,chg 0 Pj,chd |
and without loss of generality, assume that 0 ≤ θj ≤ π/2. Then, the following facts are immediate.
1. | sin θj | = sin θj = SE(Pj−1,chg , Pj,chd ),
2. Pj,chd = Pj,new sin θj + Pj−1,chg cos θj ,
3. Pj,new is orthogonal to Pj−1 ,
4. span(Pj ) ⊆ span([Pj−1 , Pj,new ]),
5. the direction that gets deleted at tj is Pj,del := Pj−1,chg sin θj − Pj,new cos θj .
Thus, our subspace change assumption is exactly equivalent to the following generative model on Pj : for a
θj satisfying sin θj ≤ ∆, and for a direction Pj,new that is orthogonal to Pj−1 ,
Pj = [(Pj−1 Rj )[1,r−1] , Pj,chd ], where Pj,chd := (Pj−1 Rj )r cos θj + Pj,new sin θj
| {z }
|
{z
}
Pj−1,chg
Pj−1,fix
and Rj is an r × r rotation matrix. For a simple example in 3D (n = 3), see Fig. 1.
8
(6)
With the above model, observe that span(Pj ) equals the span of left singular vectors of Lj := L[tj ,tj+1 ) .
The span of left singular vectors of L is contained in, or equal to, span([P0 , P1,new , P2,new , . . . , PJ,new ]).
Equality holds if each Pj,new is orthogonal to all others and to P0 .
Incoherence (denseness) of columns of Pj . In order to separate the `t ’s from the sparse outliers xt ,
we need to assume that the `t ’s are themselves not sparse. This is ensured if we can assume that column
vectors of Pj are dense enough. To quantify this, we use the incoherence assumption [1]. Define the
incoherence (non-denseness) parameter [1], µ, as the smallest real number so that
r
r
µr
µ
0
0
max
max kIi Pj k ≤
, and max
max kIi Pj,new k ≤
.
(7)
j=1,2,...,J i=1,2,...,n
j=1,2,...,J i=1,2,...,n
n
n
where Pj,new is defined above in (5). Incoherence means that µ is upper bounded. This assumption is
similar to the assumption on left singular vectors of L needed by all solutions for RPCA. RPCA solutions
also need incoherence of right singular vectors which we do not need. Instead ReProCS only needs
boundedness of at ’s (explained below). This is a weaker requirement.
Assumption on principal subspace coefficients at . We assume that the at ’s are zero mean, mutually
independent, and element-wise bounded random variables (r.v.) with diagonal covariance matrix Λ. Since
the at ’s are element-wise bounded, there exists an η < ∞, such that
max max
j=1,2,...r
t
(at )2j
≤ η.
λj (Λ)
For most bounded distributions, η is a little more than one, e.g., if the entries of at are zero mean uniform,
then η = 3. In the rest of the paper, η will be assumed to be a numerical constant.
Bound on outlier fractions. Similar to earlier RPCA works, we also need outlier fractions to be bounded.
However, we need different bounds on this fraction per column and per row. The row bound can be much
larger3 . Let outlier-frac-col := maxt |Tt |/n denotes the maximum outlier fraction in any column of Y .
Because ReProCS is an online approach that updates the subspace estimate every α frames, we need the
fraction of outliers per row of a sub-matrix of Y with α consecutive columns to be bounded. To quantify
this, for a time interval, J , define
γ(J ) :=
1 X
1{i∈Tt } .
i=1,2,...,n |J |
(8)
max
t∈J
where 1S is the indicator function for statement S. Thus γ(J ) is the maximum outlier fraction in any row
of the sub-matrix YJ of Y . Let J α denote a time interval of duration α. We will bound
outlier-frac-row :=
max γ(J α ).
(9)
J α ⊆[t1 ,d]
Definitions. Define the following.
1. Use λ− and λ+ to denote the minimum and maximum eigenvalues of Λ and let f :=
condition number.
λ+
λ−
be its
2. Split at as at = [at,fix 0 , at,chd 0 ]0 where at,chd is the scalar coefficient corresponding to the changed
direction. Similarly split its diagonal covariance matrix as Λ = diag(Λfix , λchd ).
3. Define ∆min := minj SE(Pj−1,chg , Pj,chd ). Using the generative model, this is also equal to minj sin θj .
4. Let xmin := mint mini∈Tt |(xt )i | denote the minimum outlier magnitude.
3
One practical application where this is useful is for slow moving or occasionally static video foreground moving objects.
For a stylized example of this, see Model G.26 given in Appendix G.
9
5. Let ε̃ denote the bound on initial subspace error, i.e., let
SE(P̂0 , P0 ) ≤ ε̃.
6. For numerical constants C that are re-used to denote different numerical values, define
K() := dC log(∆/)e, and
α > α∗ := Cf 2 (r log n).
(10)
Main Result. We can now state our main result.
Theorem 3.1 (Dynamic RPCA). Consider simple-ReProCS given in Algorithm 4. Let K = K(ε̃) and α
be as defined above. Assume that
1. (initial SE) SE(P̂0 , P0 ) ≤ ε̃ for ε̃ satisfying ε̃f ≤ 0.01 and ε̃f ≤ 0.8∆min ;
2. (subspace change) (3) and (4) hold with tj+1 − tj > (K + 3)α, ∆ satisfies
xmin /15;
√
√
√
η(ε̃ rλ+ + ∆ λchd ) <
0.01
3. (outlier fractions and denseness) outlier-frac-col ≤ min( 0.09
2µr , 2µ ) where µ is defined in (7); and
outlier-frac-row ≤ fb02 where b0 := 0.01 and outlier-frac-row is defined in (9);
4. (statistical assumptions) assumption on at ’s holds;
5. (algo. parameters) set K = K(ε̃) and α as in (10), set ξ = xmin /15, ωsupp = xmin /2, ωevals = 5ε̃2 f λ+ .
Then, with
probability √at least 1 − 12dn−12 , at all times, t, T̂t = Tt , and kx̂t − xt k = k`ˆt − `t k ≤
√ √ +
1.2 η(ε̃ rλ + 0.11∆ λchd ); tj ≤ t̂j ≤ tj + 2α; and, for all t ∈ [tj + (K + 3)α, tj+1 ), SE(Q̂t , Qt ) ≤ ε̃.
Detailed bounds are given in Corollary 4.4.
Consider offline s-ReProCS (last line of Algorithm 4). At all times t,
SE(Q̂offline
, Qt ) ≤ 2ε̃, and kxt − x̂offline
k = k`ˆoffline
− `t k ≤ 2.4ε̃k`t k.
t
t
t
Corollary 3.2 (Noisy Dynamic RPCA). In the setting of Theorem 3.1, suppose that yt = xt + `t + vt
−
with kvt k ≤ bv , kE[vt vt 0 ]k ≤ λ+
v = 0.03∆λ , and with vt ’s being zero mean, mutually independent, and
independent of the xt ’s and the `t ’s. If the RHS of the bound on ∆ is replaced by (xmin /15) − bv and
2
v
if α ≥ max(α∗ , Cf λb−
log n), then, conclusions of Theorem 3.1 hold with the bounds on k`ˆt − `t k and
offline
k`ˆt
− `t k replaced by their respective values plus 1.2bv and with Tt defined as set of indices i for which
|(xt )i | ≥ xmin .
Proof. We explain the ideas leading to the proof of Theorem 3.1 in Sec. 5. We prove it under the
assumption t̂j = tj in Sec. 6. The proof without this assumption is given in Appendix C. We prove
Corollary 3.2 in Appendix D.
Remark 3.3. The first condition (accurate initial estimate) can be satisfied by applying PCP or AltProj
on the first ttrain data frames. This requires assuming that ttrain ≤ t1 , Y[1,ttrain ] has outlier fractions in any
row or column bounded by outlier-frac ≤ c/r, and that the right singular vectors of L[1,ttrain ] also satisfy
incoherence [4, 3] (its left singular vectors are P0 which are already assumed to be incoherent). Also see
Table 1.
10
4
Discussion
In this section, we explain what our result means and its implications, discuss the speed and memory
guarantees, explain how to set algorithm parameters, state the other conclusions of Theorem 3.1, and
finally discuss its limitations.
Understanding the result. Theorem 3.1 shows that, with high probability (whp), the subspace change
gets detected within a delay of at most 2α = Cf 2 (r log n) frames, and the subspace gets estimated
accurately within at most (K + 2)α = Cf 2 (r log n)(− log ε̃) frames. Each column of the low rank matrix
is recovered with a small time-invariant bound without any delay. If offline processing is allowed, with a
delay of at most (K + 2)α, we can guarantee all recoveries within normalized error ε̃, or, in fact, with minor
modifications, within any = cε̃ for c < 1 (we discuss this point below when we talk about limitations).
Notice also that the required delay between subspace change times is more than r by only logarithmic
factors. Since the previous subspace is not exactly known (is known within error at most ε̃), at each
update step, we do need to estimate an r-dimensional subspace, and not a one-dimensional one. Hence it
is not clear if the required delay can be reduced any further.
Consider the bound on ∆ in the subspace change assumption. This requires that a function of the rate
of subspace change ∆ and initial error ε̃ be upper bounded by xmin (the minimum outlier magnitude).
This assumption is implicitly requiring that xmin is lower bounded. This seems counter-intuitive since
sufficiently small magnitude corruptions should not be problematic. This is indeed true. Observe that,
using Corollary 3.2, this requirement can be relaxed by classifying the small
Pmagnitude entries of xt as
vt . Then one needs the same lower bound on xmin − 15bv where b2v = maxt i:|(xt )i |<xmin ((xt )i )2 . To be
precise, one needs the xt ’s to be such that there exists an xmin > 0 for which xmin − 15bv > 0. In words,
one needs the outlier magnitudes to be either large (easy to detect) or very small (do not affect subspace
updates in a significant way). This is a weaker and more realistic requirement.
Subspace and outlier assumptions’ tradeoff. When there are fewer outliers in the data or when
outliers are easy to detect, one would expect to need weaker assumptions on the true data subspace or
its rate of change. This is indeed true. For the static RPCA results, this is encoded in the condition
outlier-frac ≤ c/(µr) where µ quantifies not-denseness of both left and right singular vectors. Thus, when
fewer outliers are present, the subspace dimension, r, can be larger and µ can be larger (the singular
vectors can be less dense). From Theorem 3.1, this is also how outlier-frac-col, µ (not-denseness of only
left singular vectors) and r are related for dynamic RPCA.
On the other hand, for our result, outlier-frac-row and the lower bound on (xmin − 15bv ) govern the
allowed rate of subspace change. The latter relation is easily evident from the bound on ∆. If (xmin − 15bv )
is larger, a larger ∆ can be tolerated. The relation of outlier-frac-row to rate of change is not evident from
the way the guarantee is stated above. The reason is we have assumed outlier-frac-row ≤ b0 with b0 = 0.01
and used that to get a simple expression for K. If we did not do this, we would need K to satisfy
p
c1 ∆(c2 b0 )K + 0.2ε̃ ≤ ε̃.
1∆
With this, K needs to be K = d − log(c1 √b ) log( c0.8ε̃
)e. Recall that we need tj+1 − tj ≥ (K + 2)α. Thus, a
2
0
smaller b0 means one of two things: either a larger ∆ (more change at each subspace change time) can be
tolerated while keeping K, and hence the lower bound on delay between change times, the same; or, for ∆
fixed, a smaller lower bound on delay between change times is needed. The above can be understood by
carefully checking the proof4 of Theorem 6.5 and its application to show correctness of the projection-SVD
steps.
Dependence on f . Observe that f appears in our guarantee in the bound on outlier-frac-row and
in the expression for α. The outlier-frac-row bound is stated that way only for simplicity. Actually,
for all time instants except the α-length period when the subspace deletion step is run, we only need
4
The
√ multiplier 0.37 of qchd in its first claim is obtained by setting b0 = 0.01. If we do not do this, 0.37 will get replaced
by c2 b0 .
11
outlier-frac-row ≤ b0 = 0.01. We need the tighter bound outlier-frac-row ≤ b0 /f 2 only for the simple SVD
based deletion step to work (i.e., only for t ∈ [t̂j + Kα, t̂j + Kα + α)). Thus, if offline ReProCS were being
used to solve a static RPCA type problem (where rmat is nicely bounded), one could choose to never run
the subspace deletion step. This will mean that the resulting algorithm (s-ReProCS-no-delete) will need
outlier-frac-col < c/µrmat but outlier-frac-row < 0.01 will suffice (the bound would not depend on f 2 ).
A second point to note is that f is the condition number of E[Lj Lj 0 ] for any j. The condition number
of the entire matrix L can be much larger when slow subspace change holds (∆ is small). To see this, let
κ2 denote the condition number of E[LL0 ], so that, whp, κ is approximately the condition number of L.
f
≈ c ∆f2 when ∆ is small.
It is not hard to to see that5 , in the worst case (if λ− = λchd ), κ2 ≥ 1−√1−2c∆
2
Storage and time complexity. Observe that the s-ReProCS algorithm needs memory of order nα in
online mode and Knα in offline mode. Assuming α = α∗ , even in offline mode, its storage complexity is
only O(nr log n log(1/ε̃)). This differs from the optimal storage complexity of O(nr) by only log factors.
Also, it is not hard to see that the time complexity of s-ReProCS is O(ndr(− log ε̃)). We explain this in
Appendix F.
Algorithm parameters. Observe from Theorem 3.1 that we need knowledge of only 4 model parameters
- r, λ+ , λ− and xmin - to set our algorithm parameters. The initial dataset used for estimating P̂0 (using
PCP/AltProj) can also be used to get an accurate estimate of r, λ− and λ+ using standard techniques.
Thus one really only needs to set xmin . If continuity over time is assumed, a simple heuristic is to let it be
time-varying and use mini∈T̂t−1 |(x̂t−1 )i | as its estimate at time t.
Other conclusions of Theorem 3.1. While proving Theorem 3.1, we also prove the following corollary
which shows that SE(Q̂t , Qt ) decays roughly exponentially after each projection-SVD step and finally
decays to below ε̃ by t̂j + Kα + α. The same is true also for ket k (the error in recovering `t or xt ). Fig. 2
illustrates the subspace error decay visually.
Corollary 4.4. Under the assumptions of Theorem 3.1,
1. the subspace change is detected in at most 2α frames and there are no false detects, i.e., tj ≤ t̂j ≤
t̂j + 2α,
2. for t ∈ [tj , t̂j + α), SE(Q̂t , Qt ) ≤ 2ε̃ + ∆,
3. for t ∈ [t̂j + (k − 1)α, t̂j + kα), SE(Q̂t , Qt ) ≤ 1.2ε̃ + (0.45)k−1 · 0.06∆, for k = 1, 2, . . . , K
4. for t ∈ [t̂j + Kα, t̂j + Kα + α), SE(Q̂t , Qt ) ≤ 2ε̃
5. for t ∈ [t̂j + Kα + α, t̂j+1 ), SE(Q̂t , Qt ) ≤ ε̃
(the bound on ket k = k`ˆt − `t k = kx̂t − xt k will also decrease in the same proportion as SE; see the last
claim of Lemma 6.13)
Limitations. ReProCS needs a few extra assumptions beyond slow subspace change and what static
RPCA solutions need: (i) instead of a bound on outlier fractions per row of the entire data matrix (which
is what static RPCA methods assume), it needs such a bound for every sub-matrix of α consecutive
columns; (ii) it puts statistical assumptions (mutual independence) on the principal subspace coefficients
at ; (iii) it needs to lower bound (xmin − 15bv ) where bv is a bound on the small noise or modeling error vt ;
and (iv) we use ε̃ to denote both the initial subspace error as well as the final recovery error achieved after
a subspace update is complete.
5
To understand this, suppose that there is only one subspace change and suppose that the intervals are equal,
0
0
0
0 0
i.e., d − t1 = t1 − t0 .
Then, E[LL
]/d = P0,fix Λfix P0,fix + [P0,chg P1,new ]B[P0,chg P1,new ] where B =
(0.5 + 0.5 cos2 θ1 ) −0.5 sin θ1 cos θ1
λchd
. The maximum eigenvalue of E[LL0 ]/d is λ+ . Its minimum eigenvalue is the
−0.5 sin θ1 cos θ1
0.5 sin2 θ1
minimum eigenvalue of B which can p
be computed as (1 − cos θ1 )λchd . In the worst case λchd = λ− . When the intervals are
p
not equal, this gets replaced by (1 − 1 − 2c sin2 θ1 )λ− . This is upper bounded by (1 − 1 − 2c∆2 ).
12
Here (i) is needed because ReProCS is an online algorithm that uses α frames at a time to update
the subspace, and one needs to show that each update step provides an improved estimate compared to
the previous one. However, since α is large enough, requiring a bound b on outlier-frac-row defined in (9)
is not too much stronger than requiring the same bound b on the outlier fractions per row of the entire
n × d matrix Y . In fact, if we compare the various RPCA solutions with storage complexity fixed at
O(nα) = O(nr log n), i.e., if we implement the various static RPCA solutions for every new batch of α
frames of data, then, the static RPCA solutions will also need to bound outlier-frac-row defined in (9). As
discussed earlier, these will require a much tighter bound of c/r though.
(iv) is assumed for simplicity. What we can actually prove is something slightly stronger: if the initial
error is ε̃, and if = cε̃ for a constant c which may be less than one, then, without any changes, we can
guarantee the final subspace error to be below such an . More generally, as long as the initial error ε̃ ≤ ∆,
it is possible to achieve final error for any > 0 if we assume that t1 − ttrain > Kα, assume a larger lower
bound on x2min , and if we modify our initialization procedure. This idea is being studied in follow-up work.
Limitations (ii) and (iii) are artifacts of our proof techniques. It is not hard to replace (ii) by an
autoregressive model on the at ’s by borrowing similar ideas from [9]. (iii) is needed because our proof
first tries to show exact outlier support recovery by solving a CS problem to recover the outliers from the
projected measurements, followed by thresholding. It should be possible to relax this by relaxing the exact
support recovery requirement. This will require some other changes in the proof. It may be possible to
also completely eliminate it if we replace the CS step by thresholding with carefully decreasing thresholds
in each iteration (borrow the idea of AltProj); however, we may then require the same tight bound on
outlier-frac-row that AltProj needs. By borrowing the stagewise idea of AltProj or NO-RMC, it should
also be possible to remove all dependence on f .
5
5.1
Main ideas of the proof: Why ReProCS works
Why ReProCS with tj known works
To understand things simply, first assume that t̂j = tj , i.e., the subspace change times are known. Consider
Algorithm 1. At each time t this consists of three steps - projected Compressive Sensing (CS) to estimate
xt , estimating `t by subtraction, and subspace update. Consider projected CS. This is analyzed in Lemma
6.13. At time t, suppose that we have access to Q̂t−1 which is a good estimate of the previous subspace,
span(Qt−1 ). Because of slow subspace change, this is also a good estimate of span(Qt ). Its first step
projects yt orthogonal to Q̂t−1 to get ỹt . Recall that ỹt = Ψxt + bt where bt := Ψ`t is small. Using the
incoherence (denseness) assumption and span(Qt−1 ) being a good estimate of span(Qt ), it can be argued
that the restricted isometry constant (RIC) of Ψ := I − Q̂t−1 Q̂t−1 0 will be small. Using [18, Theorem 1.2],
this, along with kbt k being small, ensures that l1 minimization will produce an accurate estimate, x̂t,cs , of
xt . The support estimation step with a carefully chosen threshold, ωsupp = xmin /2, and a lower bound on
xmin − 15bv then ensures exact support recovery, i.e., T̂t = Tt . With this, the LS step output, x̂t , satisfies
x̂t = xt + et with
et := ITt (ΨTt 0 ΨTt )−1 ΨTt 0 `t = ITt (ΨTt 0 ΨTt )−1 ITt 0 Ψ`t
(11)
and with ket k being small.
Computing `ˆt := yt − x̂t , then also gives a good estimate of `t that satisfies `ˆt = `t − et with et as
above.
The subspace update step uses `ˆt ’s to update the subspace. Since `ˆt = `t − et with et satisfying
(11), et depends on `t . Because of this, the subspace update does not involve a PCA or an incremental
PCA problem in the traditionally studied setting (data and corrupting noise/error being independent
or uncorrelated). It is, in fact, an instance of PCA when the the noise/error et in the observed data `ˆt
depends on the true data `t . Thus `t and the noise/error et are correlated. This problem was studied in
[10, 11] where it was referred to as “correlated-PCA” or “PCA in data-dependent noise”.
13
Using this terminology, our subspace update problem (estimating Pj using P̂j−1 ) is a problem of PCA
in data-dependent noise with partial subspace knowledge. To simplify the analysis of this part, we first
study this problem and obtain a guarantee for it in Theorem 6.5 in Sec. 6.2 below. This theorem along
with Lemma 6.13 (that analyzes the projected-CS step discussed above) help obtain a guarantee for the
k-th projection-SVD step in Lemma 6.14. The k = 1 and k > 1 cases are handled separately. The main
assumption required for applying Theorem 6.5 holds because et is sparse with support Tt that changes
enough (outlier-frac-row bound of Theorem 3.1 holds). The subspace deletion via simple SVD step of
subspace update is studied in Lemma 6.15. This step solves a problem of PCA in data-dependent noise
and so it directly uses the results from [11].
To understand the flow of the proof, consider the interval [tj , tj+1 ). Assume that, before tj , the previous
subspace has been estimated with error ε̃, i.e., we have P̂j−1 with SE(P̂j−1 , Pj−1 ) ≤ ε̃. We explain below
that this implies that, under the theorem’s assumptions, we get SE(P̂j , Pj ) ≤ ε̃ before tj+1 . We remove
the subscripts j in some of this discussion. Define the interval Jk := [tj + (k − 1)α, tj + kα).
1. Before the first projection-SVD step (which is done at t = tj + α), i.e., for t ∈ J1 , we have no
estimate of Pnew , and hence only a crude estimate of Pchd . In particular, we can only get the bound
SE(Q̂t , Pchd ) = SE(P̂j−1 , Pchd ) ≤ ε̃ + | sin θ| for this interval.
• As a result, the bound on the√“noise”, bt , seen
√ by the projected-CS step is also the largest for this
interval, we have kbt k ≤ C(ε̃ rλ+ +| sin θ| λchd ). Using the CS guarantee, followed by ensuring
exact support recovery (as explained above), this implies
√ that et satisfies√(11) and that we get
a similar bound on the final CS step error: ket k ≤ C(ε̃ rλ+ + 0.11| sin θ| λchd ). The factor of
0.11 in the second term of this bound is obtained because, for this interval, Ψ = I − P̂j−1 P̂j−1 0
and so ΨPnew ≈ Pnew and Pnew is dense. Thus one can show that kITt 0 ΨPnew k2 ≤ 0.11.
• This bound on et , along with using the critical fact that et := `ˆt − `t satisfies (11) (is sparse)
and its support Tt changes enough (the outlier-frac-row bound of Theorem 3.1 holds), ensures
that we get a better estimate of Pchd after the first projection-SVD step. This is what allows us
to apply Theorem 6.5. Using it we can show that SE(Q̂t , Pchd ) = SE([P̂j−1 , P̂j,chd,1 ], Pchd ) ≤
0.1ε̃ + 0.06| sin θ| for t ∈ J2 . See proof of k = 1 case of Lemma 6.14 and Fact 6.12.
2. Thus we have a much better estimate of Pchd for t ∈ J2 than for J1 . Because of this, kbt k is
smaller, and hence ket k is smaller for t ∈ J2 . This, along with the sparsity and changing support,
Tt , of et , ensures an even better estimate at the second projection-SVD step. We can show that
SE(Q̂t , Pchd ) = SE([P̂j−1 , P̂j,chd,2 ], Pchd ) ≤ 0.1ε̃ + 0.45 · 0.06| sin θ| for t ∈ J3 . See proof of k > 1
case of Lemma 6.14 and Fact 6.12.
3. Proceeding this way, we show that SE(Q̂t , Pchd ) = SE([P̂j−1 , P̂j,chd,k ], Pchd ) ≤ 0.1ε̃+0.45k−1 (0.06| sin θ|)
after the k-th projection-SVD step. Picking K appropriately, gives SE(Q̂t , Pchd ) ≤ ε̃ after K steps,
i.e., at t = tj + Kα. In all the above intervals, SE(Q̂t , Qt ) ≤ ε̃ + SE(Q̂t , Pchd ). Thus at t = tj + Kα,
SE(Q̂t , Qt ) ≤ 2ε̃.
4. At t = tj + Kα, Q̂t contains (r + 1) columns. The subspace re-estimation via simple SVD step
re-estimates Pj in order to delete the deleted direction, Pdel , from Q̂t . The output of this step is P̂j
(the final estimate of span(Pj )). Thus, at t = tj + Kα + α, Q̂t = P̂j and we can show that it satisfies
SE(Q̂t , Qt ) = SE(P̂j , Pj ) ≤ ε̃. See Lemma 6.15. The re-estimation is done at this point because, for
times t in this interval, k`ˆt − `t k = ket k ≤ 2.4ε̃k`t k. For PCA in data-dependent noise, simple SVD
needs α ≥ (q/)2 (r log n) where q is the error/noise to signal ratio and is the final desired error
level. For our problem, the “noise” is et and thus q = 2.4ε̃ and = ε̃. Since q/ is a constant and
hence α ≥ α∗ suffices when simple SVD is applied at this time.
14
5.2
Why automatic subspace change detection and Automatic Simple-ReProCS works
The subspace change detection approach is summarized in Algorithm 4. The algorithm toggles between
the “detect” phase and the “update” phase. It starts in the “detect” phase. If the j-th subspace change is
detected at time t, we set t̂j = t. At this time, the algorithm enters the “update” (subspace update) phase.
We then repeat the K projection-SVD steps and the one subspace re-estimation via simple SVD step
from Algorithm 1 with the following change: the k-th projection-SVD step is now done at t = t̂j + kα − 1
(instead of at t = tj + kα − 1) and the subspace re-estimation is done at t = t̂j + Kα + α − 1 := t̂j,f in .
Thus, at t = t̂j,f in , the subspace update is complete. At this time, the algorithm enters the “detect” phase
again. To understand the change detection strategy, consider the j-th subspace change. Assume that
the previous subspace Pj−1 has been accurately estimated by t = t̂j−1,f in and that t̂j−1,f in < tj . Let
P̂∗ := P̂j−1 denote this estimate. At this time, the algorithm enters the “detect” phase in order to detect
the next (j-th) change. Let Bt := (I − P̂∗ P̂∗ 0 )[`ˆt−α+1 , . . . , `ˆt ]. For every t = t̂j−1,f in + uα, u = 1, 2, . . . ,
√
we detect change by checking if the maximum singular value of Bt is above a pre-set threshold, ωevals α,
or not.
We claim that, whp, under Theorem 3.1 assumptions, this strategy has no false detects and correctly
detects change within a delay of at most 2α frames. The former is true because, for any t for which
√
[t − α + 1, t] ⊆ [t̂j−1,f in , tj ), all singular values of the matrix Bt will be close to zero (will be of order ε̃)
√
and hence its maximum singular value will be below ωevals α. Thus, whp, t̂j ≥ tj . To understand why
t −t̂
in
eα := tj,∗ . Since
the change is correctly detected within 2α frames, first consider t = t̂j−1,f in + d j j−1,f
α
we assumed that t̂j−1,f in < tj (the previous subspace update is complete before the next change), tj lie
in the interval [tj,∗ − α + 1, tj,∗ ]. Thus, not all of the `t ’s in this interval will contain the newly added
direction. Depending on where in the interval tj lies, the algorithm may or may not detect the change at
this time. However, in the next interval, i.e., for t ∈ [tj,∗ + 1, tj,∗ + α], all of the `t ’s will contain the newly
changed direction. We can prove that Bt for this time t will have maximum singular value that is above
the threshold. This proof is where the assumption that ε̃f ≤ 0.8∆min is needed. Thus, if the change is not
detected at tj,∗ , whp, it will get detected at tj,∗ + α. Hence one can show that, whp, either t̂j = tj,∗ , or
t̂j = tj,∗ + α, i.e., tj ≤ t̂j ≤ tj + 2α.
6
Proving Theorem 3.1 with assuming t̂j = tj
To give a less notation-ally intensive proof, we first prove Theorem 3.1 under the assumption that t̂j = tj .
The proof without this assumption is given in Appendix C. With assuming t̂j = tj , we are studying
Algorithm 1. Recall from Sec. 5 that the subspace update step involves solving a problem of PCA in
data-dependent noise when partial subspace knowledge is available. We provide a guarantee for this
problem in Sec. 6.2 and use it in our proof in Sec. 6.4.
6.1
Few more definitions
The following points should be noted.
1. For the entire proof we will use the equivalent subspace change model described in (6). To use this,
let θ− and θ+ denote lower and upper bounds on θj , i.e., let θ− := minj θj ≤ θ+ . Then
∆min ≡ | sin θ− |, and ∆ ≡ | sin θ+ |
(12)
The proof obtains guarantees that contain | sin θ+ | or | sin θ− |. Using (12), we get our main result.
2. Define s := (outlier-frac-col)n. This bounds |Tt | and will be used below when we analyze the
projected-CS step.
3. Simplified-ReProCS will be called just ReProCS.
15
4. In the writing below αchd and αdel may appear a few times. Both will take the value α∗ given in
(10).
6.2
PCA in data-dependent noise with partial subspace knowledge
The Problem. We are given a set of α frames of observed data yt := `t + wt , with wt = Mt `t , `t = P at
and with P satisfying
P = [P∗,fix , Pchd ], where Pchd := (P∗,chg cos θ + Pnew sin θ),
P∗,fix = (P∗ R0 )I[1,r−1] , P∗,chg = (P∗ R0 )Ir , and R0 is an r × r rotation matrix. Also, the at ’s are zero
mean, mutually independent, and bounded r.v.’s as assumed before. The matrices Mt are unknown.
Let J α denote the α-frame time interval for which the yt ’s are available. We also have access to
a partial subspace estimate P̂∗ that satisfies SE(P̂∗ , P∗ ) ≤ ε̃, and that is computed using data that is
independent of the `t ’s (and hence of the yt ’s) for t ∈ J α . The goal is to estimate span(P ) using P̂∗ and
the yt ’s for t ∈ J α .
Projection-SVD / Projection-EVD. Let Φ := I − P̂∗ P̂∗ 0 . A natural way to estimate P is to first
compute P̂chd as the top eigenvector of
Dobs :=
1 X
Φyt yt 0 Φ.
α
α
t∈J
and set P̂ = [P̂∗ , P̂chd ]. We refer to this strategy as “projection-EVD” or “projection-SVD” 6 . In this
paper, we are restricting ourselves to only one changed direction and hence we compute only the top
eigenvector (or left singular vector) of Dobs . In general if there were rchg > 1 directions, we would compute
all eigenvectors with eigenvalues above a threshold, see [7, 9].
The Guarantee. We can prove the following about projection-SVD.
Theorem 6.5. Consider the above setting for an α ≥ α0 where
α0 := Cη max(f (r log n), ηf 2 (r + log n)).
Assume that the Mt ’s can be decomposed as Mt = M2,t M1,t where M2,t is such that kM2,t k ≤ 1 but
1 X
M2,t M2,t 0 ≤ b0 = 0.01.
α
α
(13)
t∈J
Let q0 denote a bound on maxt kM1,t P∗ k and let qchd denote a bound on maxt kM1,t Pchd k, i.e., we have
kM1,t P∗ k ≤ q0 and kM1,t Pchd k ≤ qchd for all t ∈ J α . Assume that
q0 ≤ 2ε̃, qchd ≤ 0.2| sin θ|, and ε̃f ≤ 0.01| sin θ|.
(14)
Define the event E∗ := {SE(P̂∗ , P∗ ) ≤ ε̃}. The following hold.
1. Conditioned on E∗ , w.p. at least 1 − 12n−12 , SE(P̂ , P ) ≤ ε̃ + SE(P̂ , Pchd ) and
SE(P̂ , Pchd ) ≤ (ε̃ + | sin θ|)
0.36qchd + 0.09ε̃
0.36qchd + 0.09ε̃
≤ 1.01| sin θ|
| sin θ|
| sin θ|
< 0.37qchd + 0.1ε̃.
6
For the rest of the paper, we use the term “projection-SVD” rather than “projection-EVD” to refer to the above since
SVD is a more commonly understood term in modern literature (EVD was common in older literature, e.g., [19].
16
2. Conditioned on E∗ , w.p. at least 1 − 12n−12 ,
λmax (Dobs ) ≥ (0.97 sin2 θ − 0.37qchd | sin θ| − 0.1ε̃| sin θ|)λchd .
Proof. The proof follows using a careful application of the Davis-Kahan sin θ theorem [19] followed by
using matrix Bernstein [20] to bound the numerator terms in the sin θ theorem bound and Vershynin’s
sub-Gaussian result [21] to bound the extra terms in its denominator. While the overall approach is similar
to that used by [11] for the basic correlated-PCA problem, this proof requires significantly more work. We
give the proof in Section 7.
Remark 6.6. Theorem 6.5 holds even when E∗ is replaced by E0 := E∗ ∩ Ẽ(Z) where Ẽ(Z) is an event that
depends on a r.v. Z that is such that the pair {P̂∗ , Z} is still independent of the `t ’s (and hence of the
yt ’s) for t ∈ J α .
In the result above, the bounds assumed in (14) are not critical. They only help to get a simple
expression for the subspace error bound. As will be evident from the proof, we can also get a guarantee
without assuming (14), and with any value of b0 .
The main assumption needed by Theorem 6.5 is (13) on the data-dependency matrices Mt . This
is required because the noise wt depends on the true data `t and hence the instantaneous values of
both the noise power and of the signal-noise correlation (even after being projected orthogonal to P̂∗ )
can be large if λchd is large. However, (13) helps ensure that the time-averaged noise power and the
time-averaged projected-signal-noise correlation are much smaller. Using the definitions of q0 and qchd ,
2 λ
0
+
kE[wt wt 0 ]k ≤ q02 λ+ + qchd
λchd := cwl . By approprichd := cw and kE[Φ`t wt ]k ≤ ε̃q0 λ + (ε̃ + | sin θ|)qchd
√
1 P
0
ately applying
Cauchy-Schwartz
(Theorem
A.22),
it
is
not
hard
to
see
that
k
b0 cw
t∈J α E[wt wt ]k ≤
α
√
P
and k α1 t∈J α E[Φ`t wt 0 Φ]k ≤ b0 cwl . Since b0 = 0.01, both bounds are 0.1 times their instantaneous
values. These bounds are derived in the proof of Lemma 7.20.
For our problem, (13) holds because we can let M2,t = ITt and M1,t as the rest of the matrix multiplying
`t in
(11). Then, using the bound on outlier fractions per row from Theorem 3.1, it is easy to see that
1 P
0 ≤ b . We state this formally next in Lemma 6.7.
0
t M2,t M2,t
α
Lemma 6.7. Assume that the outlier-frac-row bound of Theorem 3.1 holds. Then, for any α-length
interval J α ⊆ [t1 , d],
1 X
b0
ITt ITt 0 = γ(J α ) ≤ outlier-frac-row ≤ 2 ≤ b0 .
α
f
α
t∈J
Proof. The proof is straightforward. It is given in Appendix B.
6.3
Two simple lemmas from [7]
The following two lemmas taken from [7] will be used in proving Theorem 3.1.
Lemma 6.8. [7, Lemma 2.10] Suppose that P , P̂ and Q are three basis matrices. Also, P and P̂ are of
the same size, Q0 P = 0 and k(I − P̂ P̂ 0 )P k = ζ∗ . Then,
1. k(I − P̂ P̂ 0 )P P 0 k = k(I − P P 0 )P̂ P̂ 0 k = k(I − P P 0 )P̂ k = k(I − P̂ P̂ 0 )P k = ζ∗
2. kP P 0 − P̂ P̂ 0 k ≤ 2k(I − P̂ P̂ 0 )P k = 2ζ∗
3. kP̂ 0 Qk ≤ ζ∗
p
4. 1 − ζ∗2 ≤ σi (I − P̂ P̂ 0 )Q ≤ 1
Lemma 6.9. [7, Lemma 3.7] For an n × r basis matrix P ,
17
1. max|T |≤s kIT 0 P k2 ≤ s maxi=1,2,...,r kIi 0 P k2
2. δs (I − P P 0 ) = max|T |≤s kIT 0 P k2
3. If P = [P1 , P2 ] then kIT 0 P k2 ≤ kIT 0 P1 k2 + kIT 0 P2 k2 .
6.4
Proof of Theorem 3.1 with t̂j = tj
Definition 6.10. We will use the following definitions in our proof.
1. Let θ := θj , P∗ := Pj−1 , Pnew := Pj,new , Pchd := Pj,chd := (P∗,chg cos θ + Pnew sin θ), P := Pj .
Similarly define P̂∗ := P̂j−1 , P̂ := P̂j , and let P̂chd,k := P̂j,chd,k denote the k-th estimate of Pchd .
2. ζchd,0 := SE(P̂∗ , Pchd ) and ζchd,k := SE([P̂∗ , P̂chd,k ], Pchd ) for k = 1, 2, . . . , K
3. The scalars
+
ζchd,0
:= ε̃ + | sin θ|,
+
ζchd,1
:= 0.37 · 1.2((0.1 + ε̃)| sin θ| + ε̃) + 0.1ε̃ and
+
+
ζchd,k := 0.37 · (1.2ζchd,k−1
) + 0.1ε̃ for k = 2, 3, . . . , K.
4. The events
+
Γ0 := {SE(P̂∗ , P∗ ) ≤ ε̃}: clearly Γ0 implies that SE(P̂∗ , Pchd ) ≤ ζchd,0
:= ε̃ + | sin θ|,
+
Γk := Γk−1 ∩ {SE([P̂∗ , P̂chd,k ], Pchd ) ≤ ζchd,k
} for k = 1, 2, . . . , K, and
ΓK+1 := ΓK ∩ {SE(P̂ , P ) ≤ ε̃}.
5. The time intervals:
Jk := [tj + (k − 1)α, tj + kα) for k = 1, 2, . . . , K: the projection-SVD intervals,
JK+1 := [tj + Kα, tj + Kα + α): the subspace re-estimation interval,
JK+2 := [tj + Kα + α, tj+1 ): the interval when the current subspace update is complete and before
the next change.
We first prove the SE bounds of Theorem 3.1. With these, the other bounds follow easily. To obtain
the SE bounds, we will be done if we prove the following claim.
Claim 6.11. Given SE(P̂∗ , P∗ ) ≤ ε̃, w.p. at least 1 − (K + 1)12n−12 ,
+
+
1. SE(P̂∗ , Pchd ) ≤ ζchd,0
, SE([P̂∗ , P̂chd,1 ], Pchd ) ≤ ζchd,1
,
+
2. for k > 1, SE([P̂∗ , P̂chd,k ], Pchd ) ≤ ζchd,k
,
+
3. and so SE([P̂∗ , P̂chd,K ], Pchd ) ≤ ζchd,K
≤ ε̃ (using definition of K) and SE([P̂∗ , P̂chd,K ], P ) ≤ 2ε̃.
4. Further, after the deletion step, SE(P̂ , P ) ≤ ε̃.
Proving the above claim is equivalent to showing that Pr(ΓK+1 |Γ0 ) ≥ 1 − (K + 1)12n−12 .
This claim is an easy consequence of three key lemmas and Fact 6.12 given below. Fact 6.12 provides
+
simple upper bounds on ζchd,k
that will be used at various places. The first lemma, Lemma 6.13, shows
that, assuming that the “subspace estimates so far are good enough”, the projected CS step “works” for
the next α frames, i.e., for all t ∈ Jk , T̂t = Tt ; et is sparse and supported on Tt and satisfies (11), and ket k
is bounded. The second lemma, Lemma 6.14, uses Lemma 6.13 and Theorem 6.5 to show that, assuming
that the “subspace estimates so far are good enough”, with high probability (whp), the subspace estimate
at the next projection-SVD step is even better than the previous ones. Applying Lemma 6.14 for each
k = 1, 2, . . . , K proves the first two parts of Claim 6.11. The third part follows easily from the first two
and the definition of K. The fourth part follows using Lemma 6.15, which shows that, assuming that
the K-th projection-SVD step produces a subspace estimate that is within 2ε̃ of the true subspace, the
subspace re-estimation step returns an estimate that is within ε̃ of the true subspace.
18
Fact 6.12. Using ε̃ ≤ ε̃f ≤ 0.01| sin θ|,
+
1. ζchd,0
:= ε̃ + | sin θ| ≤ 1.01| sin θ|,
+
2. ζchd,1
:= 0.37 · 1.2((0.1 + ε̃)| sin θ| + ε̃) + 0.1ε̃ ≤ 0.06| sin θ|,
+
+
+
0.1
:= 0.37 · 1.2ζchd,k−1
+ 0.1ε̃ ≤ 0.45k−1 ζchd,1
+ 1−0.45
ε̃ ≤ 0.45k−1 (0.06| sin θ|) + 0.16ε̃ ≤ 0.03| sin θ|
3. ζchd,k
Lemma 6.13 (Projected CS). Recall from Sec. 6.1 that s is an upper bound on |Tt |. Under assumptions
of Theorem 3.1, the following holds for k = 1, 2, . . . , K + 2. Let Ψ1 := I − P̂∗ P̂∗ 0 , Ψk := I − P̂∗ P̂∗ 0 −
P̂chd,k−1 P̂chd,k−1 0 for k = 2, 3, . . . , K + 1, and ΨK+2 := I − P̂ P̂ 0 . From Algorithm 1,
Ψ = Ψk for t ∈ Jk , k = 1, 2, . . . , K + 2.
Assume that Γk−1 holds. Then,
1. max|T |≤2s kIT 0 P̂∗ k ≤ 0.3 + ε̃ ≤ 0.31.
2. max|T |≤2s kIT 0 P̂chd,k−1 k ≤ 0.1 + ε̃ +
+
ζchd,k−1
+ε̃
| sin θ|
≤ 0.1 + 0.01 + 0.04 < 0.15.
3. δ2s (Ψ1 ) ≤ 0.312 < 0.12, δ2s (Ψk ) ≤ 0.312 + 0.152 < 0.12 for k = 2, 3, . . . , K + 2,
4. for all t ∈ Jk , k(ΨTt 0 ΨTt )−1 k ≤ φ+ := 1.2
5. for all t ∈ Jk , T̂t = Tt
6. for all t ∈ Jk , et := x̂t − xt = `t − `ˆt satisfies (11) with Ψ = Ψk for t ∈ Jk
√
√ √
7. for t ∈ J1 , ket k ≤ 1.2 η(ε̃ rλ+ + 0.11| sin θ+ | λchd );
√
√
√
+
for t ∈ Jk , ket k ≤ 1.2 η(ε̃ rλ+ + ζchd,k−1
λchd ) for k = 2, 3, . . . K;
√
√
for t ∈ JK+1 , ket k ≤ 2.4 η ε̃ rλ+ .
√ √
for t ∈ JK+2 , ket k ≤ 1.2 η ε̃ rλ+ .
Proof. The first claim of Lemma 6.9 along with the outlier-frac-col bound of Theorem 3.1 implies that for
any set T with |T | ≤ 2s
r
r
2sµr
2sµ
0
0
kIT P∗ k ≤
≤ 0.3 and kIT Pnew k ≤
≤ 0.1
(15)
n
n
Using (15), for any set T with |T | ≤ 2s,
kIT 0 P̂∗ k ≤ kIT 0 (I − P∗ P∗ 0 )P̂∗ k + kIT 0 P∗ P∗ 0 P̂∗ k ≤ k(I − P∗ P∗ 0 )P̂∗ k + kIT 0 P∗ k
= k(I − P̂∗ P̂∗ 0 )P∗ k + 0.3
≤ ε̃ + 0.3 ≤ 0.31.
The second row used (15) and the following: since P̂∗ and P∗ have the same dimension, SE(P∗ , P̂∗ ) =
SE(P̂∗ , P∗ ) (follows by Lemma 6.8, item 1). The third row follows using the definition of event Γk−1 .
Proceeding similarly for P̂chd,k−1 and Pnew (both have the same dimension),
kIT 0 P̂chd,k−1 k ≤ k(I − Pnew Pnew 0 )P̂chd,k−1 k + kIT 0 Pnew k
= k(I − P̂chd,k−1 P̂chd,k−1 0 )Pnew k + 0.1
≤ k(I − P̂∗ P̂∗ 0 − P̂chd,k−1 P̂chd,k−1 0 )Pnew k + kP̂∗ 0 Pnew k + 0.1
≤
+
ζchd,k−1
+ ε̃
| sin θ|
+ ε̃ + 0.1 ≤ 0.04 + 0.01 + 0.1 ≤ 0.15.
19
The second row used item 1 of Lemma 6.8 and (15). The third row used triangle inequality. The last row
P −P
cos θ
follows using Pnew = chd sin∗,chg
and the definition of event Γk−1 . Using this, triangle inequality and
θ
| cos θ| ≤ 1, we bound the first term. Using item 3 of Lemma 6.8, we bound the second term. The final
bounds use Fact 6.12.
To get the above bound, we use Pnew (and not Pchd ) because (i) kP̂∗ 0 Pnew k ≤ ε̃ since P∗ is orthogonal
to Pnew (but we do not have a small upper bound on kP̂∗ 0 Pchd k) and because (ii) we have assumed
denseness of Pnew with a tighter bound in (7) (but we do not have this tighter bound on kIi 0 Pchd k).
The third claim follows using the first two claims and Lemma 6.9. The fourth claim follows from the
third claim as follows:
Ψ Tt 0 Ψ Tt
−1
≤
1
1
1
≤
≤
< 1.2 = φ+ .
1 − δs (Ψ)
1 − δ2s (Ψ)
1 − 0.12
The last three claims follow the approach of the proof of [7, Lemma 6.4]. There are minor differences
because we set ξ a little differently now. We provide the proof in Appendix B.
Lemma 6.14 (Projection-SVD). Under the assumptions of Theorem 3.1, the following holds for k =
+
1, 2, . . . K. Conditioned on Γk−1 , w.p. at least 1 − 12n−12 , SE([P̂∗ , P̂chd,k ], Pchd ) ≤ ζchd,k
, i.e., Γk holds.
Proof. Assume that Γk−1 holds. The proof first uses Lemma 6.13 to get an expression for et = `t − `ˆt and
then applies Theorem 6.5 with the modification given in Remark 6.6. Using Lemma 6.13, for all t ∈ Jk ,
`ˆt = `t − et = `t − ITt (ΨTt 0 ΨTt )−1 ITt 0 Ψ`t
where Ψ = I − P̂∗ P̂∗ 0 − P̂chd,k−1 P̂chd,k−1 0 with P̂chd,0 = [.].
In the k-th projection-SVD step, we use these `ˆt ’s and P̂∗ to get a new estimate of Pchd using
projection-SVD. To bound SE([P̂∗ , P̂chd,k ], Pchd ), we apply Theorem 6.5 (Remark 6.6)7 with E0 ≡ Γk−1 ,
yt ≡ `ˆt , wt = et , α ≥ α0 ≡ α∗ , and J α ≡ Jk . We can let M2,t = −ITt and M1,t = (ΨTt 0 ΨTt )−1 ITt 0 Ψ.
Using the outlier-frac-row bound of Theorem 3.1 and Lemma 6.7, the main assumption needed by Theorem
6.5, (13), holds. With P = Pj satisfying (6), and α∗ defined in (10), all the key assumptions of Theorem
6.5 hold. The simpler expression of α∗ suffices because we treat η as a numerical constant and so
f 2 (r log n) > f 2 (r + log n) for large n, r.
We now just need to compute q0 and qchd for each k, ensure that they satisfy (14), and apply the
result. The computation for k = 1 is different from the rest. When k = 1, Ψ = I − P̂∗ P̂∗ 0 . Thus, using
item 4 of Lemma 6.13 and the definition of event Γk−1 , kM1,t P∗ k ≤ φ+ ε̃ = 1.2ε̃ = q0 , q0 < 2ε̃, and
kM1,t Pchd k ≤ φ+ (kITt 0 (I − P̂∗ P̂∗ 0 )Pnew k| sin θ| + ε̃)
≤ 1.2(kITt 0 Pnew k + kP̂∗ 0 Pnew k)| sin θ| + 1.2ε̃
≤ 1.2((0.1 + ε̃)| sin θ| + ε̃) = qchd .
The third row follows using (15) (which, in turn, relies on the outlier-frac-col bound of Theorem 3.1) and
kP̂∗ 0 Pnew k ≤ ε̃ (item 3 of Lemma 6.8). Using ε̃ ≤ ε̃f ≤ 0.01| sin θ|, clearly qchd < 0.2| sin θ|. Applying
+
Theorem 6.5, SE([P̂∗ , P̂chd,1 ], Pchd ) ≤ 0.37qchd + 0.1ε̃ = 0.37 · 1.2((0.1 + ε̃)| sin θ| + ε̃) + 0.1ε̃ = ζchd,1
.
Consider k > 1. Now Ψ = I − P̂∗ P̂∗ 0 − P̂chd,k−1 P̂chd,k−1 0 . With this, we still have kM1,t P∗ k ≤ φ+ ε̃ =
1.2ε̃ = q0 and q0 < 2ε̃. But, to bound kM1,t Pchd k we cannot use the approach that worked for k = 1.
The reason is that k[P̂∗ , P̂chd,k−1 ]0 Pnew k is not small. However, instead, we can now use the fact that
+
[P̂∗ , P̂chd,k−1 ] is a good estimate of Pchd , with SE([P̂∗ , P̂chd,k−1 ], Pchd ) ≤ ζchd,k−1
(from definition of event
Γk−1 ). Thus,
+
kM1,t Pchd k ≤ φ+ SE([P̂∗ , P̂chd,k−1 ], Pchd ) ≤ 1.2ζchd,k−1
= qchd .
7
We use Remark 6.6 with E∗ ≡ Γ0 , Z ≡ {`ˆ1 , `ˆ2 , . . . , `ˆtj +(k−1)α−1 }, and Ẽ(Z) = Γk−1 \ Γ0 .
20
By Fact 6.12, qchd < 0.2| sin θ|. Applying Theorem 6.5, SE([P̂∗ , P̂chd,k ], Pchd ) ≤ 0.37qchd + 0.1ε̃ =
+
+
0.37 · 1.2ζchd,k−1
+ 0.1ε̃ = ζchd,k
.
Lemma 6.15 (Simple SVD based subspace re-estimation). Under the assumptions of Theorem 3.1, the
following holds. Conditioned on ΓK , w.p. at least 1 − 12n−12 , SE(P̂ , P ) ≤ ε̃, i.e., ΓK+1 holds.
Proof. Assume that ΓK holds. Using Lemma 6.13, for all t ∈ JK+1 ,
`ˆt = `t − et = `t − ITt (ΨTt 0 ΨTt )−1 ITt 0 Ψ`t
where Ψ = I − P̂∗ P̂∗ 0 − P̂chd,K P̂chd,K 0 . Re-estimating the entire subspace using simple SVD applied to
these `ˆt ’s is an instance of correlated-PCA with yt ≡ `ˆt , wt ≡ et . We can apply the following result
for
[11, Theorem 2.13] to bound SE(P̂ , P ). Recall P̂ contains the top r eigenvectors of
P correlated-PCA
ˆt `ˆ0 . The following is a simplified version of [11, Theorem 2.13]. It follows by upper bounding
`
t
t∈JK+1
−
+
λv,P ,P⊥ and λ+
v,rest by λv and lower bound λv,P by zero in [11, Theorem 2.13].
Theorem 6.16. For t ∈ J α , we are given data vectors
P yt := `t + wt + vt where wt = Mt `t and `t = P at .
Let P̂ be the matrix of top r eigenvectors P
of α1 t∈J α yt yt 0 . Assume that Mt can be decomposed as
Mt = M2,t M1,t so that kM2,t k ≤ 1 but α1 t M2,t M2,t 0 ≤ b for a b < 1. Let q be an upper bound on
2
+
maxt∈J α kM1,t P k. When λ+
v > 0, let rv := bv /λv . For an εSE > 0, define
q2
λ+
q2
v
2
2
α0 := Cη max f (r log n) 2 , − 2 f max(rv , r)(log n) 2 , ηf (r log 9 + 10 log n) .
εSE λ εSE
εSE
√
If α ≥ α0 , and 3 bqf +
λ+
v
λ−
≤ 0.46εSE , then, w.p. at least 1 − 12n−12 , SE(P̂ , P ) ≤ εSE . .
Here we use the vt = 0 version of the above theorem. Thus, λ+
v = 0 and bv = 0 and so rv = 0. The
above general version is used for proving Corollary 3.2 (proved in Appendix). Apply the above result
with yt ≡ `ˆt , wt ≡ et , α ≥ α∗ , and J α ≡ JK+1 . From the expression for et , we can let M2,t ≡ −ITt ,
+
M1,t ≡ (ΨTt 0 ΨTt )−1 ITt 0 Ψ. Next we compute q. Since ΓK holds, SE([P̂∗ , P̂chd,K ], P ) ≤ ε̃ + ζchd,K
≤ 2ε̃.
Thus, kM1,t P k ≤ φ+ SE([P̂∗ , P̂chd,K ], P ) ≤ 1.2 · 2ε̃ = q. The final desired error is εSE = ε̃. Using Lemma
6.7 and the outlier-frac-row bound from Theorem 3.1, the bound on the time-average of M2,t M2,t 0 holds
0.462
with b ≡ b0 = 0.01
< (3·2.4f
. Applying the above result with εSE = ε̃, and q = 2.4ε̃, we conclude the
f2
)2
following: for α ≥ α∗ , w.p. at least 1 − 12n−12 , SE(P̂ , P ) ≤ ε̃. The simpler expression of α∗ suffices
because η is treated as a numerical constant and so f 2 (r log n) > f 2 (r + log n) for large n, r. Also of course
λ+
v = 0 so the second term in the α0 expression above is zero.
Proof of Claim 6.11. Lemma 6.14 tells us that Pr(Γk |Γk−1 ) ≥ 1 − 12n−12 . Lemma 6.15 tells us that
Pr(ΓK+1 |ΓK ) ≥ 1−12n−12 . Thus, Pr(ΓK+1 |Γ0 ) = Pr(ΓK+1 , ΓK , . . . Γ1 |Γ0 ) = Pr(Γ1 |Γ0 ) Pr(Γ2 |Γ1 ) . . . Pr(ΓK+1 |ΓK ) ≥
(1 − 12n−12 )K (1 − 12n−12 ). since ΓK+1 ⊆ ΓK ⊆ ΓK−1 · · · ⊆ Γ0 . The result follows since (1 − 12n−12 )K (1 −
12n−12 ) ≥ 1 − (K + 1)12n−12 .
Proof of Theorem 3.1 with t̂j = tj . Define the events Γ1,0 := {SE(P̂0 , P0 ) ≤ ε̃}, Γj,k := Γj,k−1 ∩{SE([P̂j−1 , P̂j,chd,k ]) ≤
+
ζchd,k
}, for k = 1, 2, . . . , K, Γj,K+1 := Γj,K ∩ {SE(P̂j , Pj ) ≤ ε̃} and Γj+1,0 := Γj,K+1 . We can state and
prove Lemmas 6.14 and 6.15 with Γk replaced by Γj,k . Then Claim 6.11 implies that Pr(Γj,K+1 |Γj,0 ) ≥
1 − 12n−12 . Using ΓJ,K+1 ⊆ ΓJ−1,K+1 · · · ⊆ Γ1,K+1 ⊆ Γ1,0 and Γj+1,0 := Γj,K+1 , Pr(ΓJ,K+1 |Γ1,0 ) =
Pr(ΓJ,K+1 , ΓJ−1,K+1 , . . . Γ1,K+1 |Γ1,0 ) = Pr(Γ1,K+1 |Γ1,0 ) Pr(Γ2,K+1 |Γ2,0 ) . . . Pr(ΓJ,K+1 |ΓJ,0 ) ≥ (1 − (K +
1)12n−12 )J ≥ 1 − J(K + 1)12n−12 ≥ 1 − dn−12 .
Event ΓJ,K+1 implies that Γj,k holds for all j and for all k. Thus, all the SE bounds given in Corollary
4.4 hold. Using Lemma 6.13, it implies that T̂t = Tt for all the time intervals of interest and that the
bounds on ket k hold.
21
The proof of Theorem 3.1 without assuming t̂j = tj is given in Appendix C. This studies the automatic
algorithm, Algorithm 4. Most of the key ideas remain the same, it just needs more notation.
Algorithm 4 Simplified-Auto-ReProCS. Let L̂t;α := [`ˆt−α+1 , `ˆt−α+2 , . . . , `ˆt ].
1: Input: P̂0 , yt , Output: x̂t , `ˆt , P̂t
2:
3:
4:
5:
6:
7:
8:
9:
10:
Parameters: ωsupp , K, α, ξ, r, ωevals
Let L̂t;α := [`ˆt−α+1 , `ˆt−α+2 , . . . , `ˆt ].
Q̂ttrain ← P̂0 ; j ← 1, k ← 1
for t > ttrain do
(x̂t , T̂t ) ← ProjCS(Q̂t−1 ,yt ) % Algorithm 2
`ˆt ← yt − x̂t .
(Q̂t , P̂j , t̂j , k, j, phase) ← AutoSubUpd(L̂t;α , P̂j−1 , t, t̂j−1 , j, k, phase, Q̂t−1 ) % Algorithm 5
end for
Offline ReProCS: At t = t̂j + Kα, for all t ∈ [t̂j−1 + Kα, t̂j + Kα − 1], set Q̂offline
← [P̂j−1 , P̂j,chd,K ];
t
offline
0
−1
0
0
0
x̂t
← IT̂t (ΨT̂t ΨT̂t ) ΨT̂t yt where Ψ := I − P̂j−1 P̂j−1 − P̂j,chd,K P̂j,chd,K ; and `ˆoffline
← yt −
t
offline
x̂t
.
Algorithm 5 Automatic Subspace Update
function AutoSubUpd(L̂t;α , P̂j−1 , t, t̂j−1 , j, k, phase, Q̂t−1 )
t̂j−1,f in ← t̂j−1 + Kα + α − 1
if phase = detect and t = t̂j−1,f in + uα then
B ← (I − P̂j−1 P̂j−1 0 )L̂t,α
√
if σmax (B) ≥ αωevals then
phase ← update, t̂j ← t,
end if
Q̂t ← Q̂t−1
end if
if phase = update then
(Q̂t , P̂j , k) ← SubUp(L̂t;α , Q̂t−1 , t, t̂j−1 , j, k, phase, Q̂t−1 ) % Algorithm 3
end if
if k = K + 1 then
phase ← detect
end if
return Q̂t , P̂j , t̂j , j, k, phase
end function
7
Proof of Theorem 6.5: PCA in data-dependent noise with partial
subspace knowledge
We prove Theorem 6.5 with the modification given in Remark 6.6. Thus we condition on E0 defined in the
remark. Recall that Φ := I − P̂∗ P̂∗ 0 . Let
QR
ΦPchd = Echd Rchd
(16)
denote the reduced QR decomposition of (ΦPchd ). Here, and in the rest of this proof, we write things in a
general fashion to allow Pchd to contain more than one direction. This makes it easier to understand how
our guarantees extend to the more general case (Pchd being an n × rchg basis matrix with rchg > 1) easier.
The proof uses the following lemma at various places.
22
Lemma 7.17. The following results follow from definitions. Assume that E0 holds. Then,
1. kM1,t P∗,fix k ≤ q0 , kM1,t P∗,chd k ≤ q0 and kM1,t Pchd k ≤ qchd
2. kΦP∗,fix k ≤ ε̃, kΦP∗,chd k ≤ ε̃, kΦPnew k ≤ 1,
3. kRchd k = kΦPchd k ≤ ε̃| cos θ| + | sin θ| ≤ ε̃ + | sin θ|
p
4. σmin (Rchd ) = σmin (ΦPchd ) ≥ sin2 θ(1 − ε̃2 ) − 2ε̃| sin θ|
p
√
5. kΦ`t k ≤ 2ε̃ ηrλ+ + | sin θ| ηλchd .
Proof. The proof is given in Appendix B. Most items are straightforward. The fourth one needs a few
lines of proof.
Proof of Theorem 6.5. We have
SE(P̂ , P ) = SE([P̂∗ , P̂chd ], [P∗,fix , Pchd ]) ≤ SE([P̂∗ , P̂chd ], P∗,fix ) + SE([P̂∗ , P̂chd ], Pchd )
≤ ε̃ + SE([P̂∗ , P̂chd ], Pchd )
where the last inequality used Lemma 7.17. Consider SE([P̂∗ , P̂chd ], Pchd ).
SE([P̂∗ P̂chd ], Pchd ) ≤ (I − P̂chd P̂chd 0 )Echd kRchd k ≤ SE(P̂chd , Echd )(ε̃ + | sin θ|)
(17)
The last inequality used Lemma 7.17. To bound SE(P̂chd , Echd ), we use the Davis-Kahan sin θ theorem
[19] given below.
Theorem 7.18 (Davis-Kahan sin θ theorem). Consider n × n Hermitian matrices, D and D̂ such that
0
A 0
E
D = E E⊥
0 Arest E⊥ 0
0
Λ 0
F
D̂ = F F⊥
0 Λrest F⊥ 0
where [E, E⊥ ] and [F , F⊥ ] are orthogonal matrices and rank(F ) = rank(E). Let
H = D̂ − D.
If λmin (A) − λmax (Arest ) − kHk > 0 and rank(E) = rank(F ), then,
kHk
.
(18)
λmin (A) − λmax (Arest ) − kHk
P
To use this result to bound SE(P̂chd , Echd ), let D̂ := Dobs = α1 t Φyt yt 0 Φ. Its top eigenvector is
P̂chd . We need to define a matrix D that is such that its top eigenvector is Echd and the gap between its
first and second eigenvalues is more than kHk. Consider the matrix
I − FF0 E ≤
D := Echd AEchd 0 + Echd,⊥ Arest Echd,⊥ 0 where
X
X
1
1
0
0
0
0
A := Echd
Φ`t `t Φ Echd , Arest := Echd,⊥
Φ`t `t Φ Echd,⊥ .
α
α
If λmax (Arest ) < λmin (A), then Echd is the top eigenvector of D. Moreover, if λmax (Arest ) < λmin (A)−kHk,
then the gap requirement holds too. Thus, by the sin θ theorem,
kHk
SE(P̂chd , Echd ) = I − P̂chd P̂chd 0 Echd ≤
(19)
λmin (A) − λmax (Arest ) − kHk
23
Here again, we should point out that, in the simple case that we consider where Pchd is a vector (only one
direction changes), A is a non-negative scalar and λmin (A) = A. However the above discussion applies
even in the general case when rchg > 1.
The rest of the proof obtains high probability bounds on the terms in the above expression. kHk =
D − D̂ can be bounded as follows.
P
Echd Echd 0 Φ`t `t 0 ΦEchd,⊥ Echd,⊥ 0 . Then,
Lemma 7.19. Let term11 = α1
kHk ≤ 2
1X
1X
Φ`t wt 0 Φ + 2 kterm11k +
Φwt wt 0 Φ .
α
α
Proof. The proof is simple. It is given in Appendix B.
The next lemma obtains high probability bounds on the three terms in the above lemma and the two
other terms from (19).
Lemma 7.20. Assume that the assumptions of Theorem 6.5 with the modification given in Remark
2
6.6 hold. Let 0 = 0.01| sin θ|(ε̃ + qchd ), 1 = 0.01(qchd
+ ε̃2 ), and v = 0.01. For an α ≥ α0 :=
2
Cη max f r log n, ηf (r + log n) , conditioned on E0 , all the following hold w.p. at least 1 − 12n−12 :
√
P
1. α1 t Φ`t wt 0 Φ ≤
b0 2ε̃2 f + (ε̃ + | sin θ|)qchd + 0 λchd ,
√
P
2
b0 4ε̃2 f + qchd
2. α1 t Φwt wt 0 Φ ≤
+ 1 λchd ,
3. λmin (A) ≥ (sin2 θ(1 − ε̃2 ) − 2ε̃| sin θ|)(1 − v )λchd − 2ε̃(ε̃ + | sin θ|)v λchd ,
4. λmax (Arest ) ≤ ε̃2 λ+ + ε̃2 v λchd ,
5. kterm11k ≤ ε̃2 f + 2ε̃2 v + ε̃| sin θ|v | λchd .
Using Lemma 7.20 and substituting for 0 , 1 , v , we conclude the following. Conditioned on E0 , with
probability at least 1 − 12n−12 ,
√
√
2
2 b0 2ε̃2 f + (ε̃ + | sin θ|)qchd + b0 4ε̃2 f + qchd
2
+ 2 ε̃ f + 2ε̃2 v + ε̃| sin θ|v + 20 + 1
,
SE(P̂chd , Echd ) ≤
(sin2 θ(1 − ε̃2 ) − 2ε̃| sin θ|)(1 − v ) − 2ε̃(ε̃ + | sin θ|)v − (ε̃2 f + ε̃2 v ) − numer
where numer denotes the numerator expression. The numerator, numer, expression can be simplified to
h p
i
p
2
numer ≤ qchd 2 b0 (ε̃ + | sin θ|) + 0.02| sin θ| + qchd
( b0 + 0.01)
h p
i
+ ε̃ (8 b0 + 2)ε̃f + (2ε̃ + | sin θ|)0.01 + 0.02| sin θ| + 0.01ε̃ .
Further, using ε̃f ≤ 0.01| sin θ|,
√
b0 ≤ 0.1 and qchd ≤ 0.2| sin θ|,
2
numer ≤ | sin θ|(0.222qchd + 0.06ε̃) + 0.11qchd
≤ | sin θ|(0.25qchd + 0.06ε̃)
This can be loosely upper bounded by 0.26 sin2 θ. We use this loose upper bound when this term appears
in the denominator. Following a similar approach for the denominator, denoted denom,
2ε̃
3ε̃2 v
ε̃2 f
numer
2
2
denom ≥ sin θ 1 − ε̃ −
−
−
−
| sin θ| sin2 θ sin2 θ
sin2 θ
h
i
numer
≥ sin2 θ 0.95 −
≥ 0.69 sin2 θ
sin2 θ
24
Thus,
SE(P̂chd , Echd ) ≤
(0.25qchd + 0.06ε̃)| sin θ|
0.36qchd + 0.09ε̃
≤
,
2
| sin θ|
0.69 sin θ
Using (17) and ε̃ ≤ ε̃f ≤ 0.01| sin θ|,
SE(P̂ , Pchd ) ≤ (ε̃ + | sin θ|)
0.36qchd + 0.09ε̃
0.36qchd + 0.08ε̃
≤ 1.01| sin θ|
≤ 0.37qchd + 0.1ε̃.
| sin θ|
| sin θ|
Proof of the last claim: lower bound on λmax (Dobs ). Using Weyl’s inequality,
λmax (Dobs ) ≥ λmax (D) − kHk ≥ λmax (A) − kHk ≥ λmin (A) − kHk.
Using the bounds from Lemmas 7.19 and 7.20 and (14), we get the lower bound.
7.1
Proof of Lemma 7.20: high probability bounds on the sin θ theorem bound terms
Proof of Lemma 7.20. Recall the definition of the event E0 from Remark 6.6. To prove this lemma, we
first bound the probabilities of all the events conditioned on {P̂∗ , Z}, for values of {P̂∗ , Z} ∈ E0 . Then we
use the following simple fact.
Fact 7.21. If Pr(Event|{P̂∗ , Z}) ≥ p0 for all {P̂∗ , Z} ∈ E0 , then,
Pr(Event|E0 ) ≥ p0 .
In the discussion below, we condition on {P̂∗ , Z}, for values of {P̂∗ , Z} in E0 . Conditioned on {P̂∗ , Z},
the matrices Echd , Echd,⊥ , Φ, etc, are constants (not random). All the terms that we bound on this lemma
P
are either of the form t∈J α g1 (P̂∗ , Z)`t `0t g2 (P̂∗ , Z), for some functions g1 (.), g2 (.), or are sub-matrices of
such a term.
Since the pair {P̂∗ , Z} is independent of the `t ’s for t ∈ J α , and these `t ’s are mutually independent,
hence, even conditioned on {P̂∗ , Z}, the same holds: the `t ’s for t ∈ J α are mutually independent. Thus,
once we condition on {P̂∗ , Z}, the summands in the terms we need to bound are mutually independent.
As a result, matrix Bernstein (Theorem A.23) or Vershynin’s sub-Gaussian result (Theorem A.24) are
applicable.
Item 1 : In the proof of this and later items, we condition on {P̂∗ , Z}, for values of {P̂∗ , Z} in E0 .
Since kΦk = 1,
1X
1X
Φ`t wt 0 Φ ≤
Φ`t wt 0 .
α t
α t
To bound the RHS above, we will apply matrix Bernstein (Theorem A.23) with Zt = Φ`t wt 0 . As explained
above, conditioned on {P̂∗ , Z}, the Zt ’s are mutually independent. We first obtain a bound on the
expected value of the time average of the Zt ’s and then compute R and σ 2 needed by Theorem A.23. By
Cauchy-Schwartz,
"
# 2
2
1X
1X
0
0
0
0
E
Φ`t wt
=
ΦP ΛP M1,t M2,t
α t
α t
(a)
1X
1X
M2,t M2,t 0
ΦP ΛP 0 M1,t 0 M1,t P ΛP 0 Φ
α t
α t
h
i
(b)
2
≤ b0 max ΦP ΛP 0 M1,t 0
t
h
2 i
≤ b0 max ΦP∗,fix Λfix P∗,fix 0 M1,t 0 + ΦPchd Λchg Pchd 0 M1,t 0
t
2
≤ b0 ε̃q0 λ+ + (ε̃ + | sin θ|) qchd λchd
≤
25
(20)
where (a) follows by Cauchy-Schwartz (Theorem A.22) with Xt = ΦP ΛP 0 M1,t 0 and Yt = M2,t , (b)
follows from the assumption on M2,t , and the last inequality follows from Lemma 7.17. Using q0 ≤ 2ε̃,
X
p
1
0
E
≤ b0 2ε̃2 λ+ + (ε̃ + | sin θ|) qchd λchd .
Φ`t wt
α
To compute R, using Lemma 7.17 and using q0 ≤ 2ε̃ and qchd < | sin θ|,
p
p
p
p
kZt k ≤ kΦ`t k kwt k ≤ ε̃ ηrλ+ + (ε̃ + | sin θ|) ηλchd q0 ηrλ+ + qchd ηλchd
p
≤ 4ε̃2 ηrλ+ + | sin θ|qchd ηλchd + 2ε̃η rλ+ λchd (qchd + | sin θ|)
≤ c1 ε̃| sin θ|ηrλ+ + c2 | sin θ|qchd ηλchd := R
for numerical constants c1 , c2 . Next we compute σ 2 . Since wt ’s are bounded r.v.’s, we have
1X
1
1X
E[Zt Zt 0 ] =
E Φ`t wt 0 wt `t 0 Φ =
E[kwt k2 Φ`t `t 0 Φ]
α t
α t
α
1X
2
E Φ`t `t 0 Φ
≤ max kwt k
wt
α t
2
ηλchd 2ε̃2 λ+ + sin2 θλchd
≤ 8ε̃2 ηrλ+ + 2qchd
2
sin2 θη(λchd )2 + c2 ε̃2 ηr sin2 θλ+ λchd := σ12
≤ c1 qchd
for P
numerical constants c1 and c2 . The above bounds again used q0 ≤ 2ε̃ and qchd < | sin θ|. For bounding
1
0
t E[Zt Zt ] we get the same expression except for the values of c1 , c2 . Thus, applying Theorem A.23
α
followed by Fact 7.21,
!
p
1X
Pr
Φ`t wt 0 ≤ b0 2ε̃2 λ+ + (ε̃ + | sin θ|)qchd λchd + E0
α t
−α
n 2
o .
≥ 1 − 2n exp
σ1 R
4 max 2 ,
Let = 0 λchd where 0 = 0.01 sin θ(qchd + ε̃). Then, clearly,
R
σ2
≤ cη max{1, f r} = cηf r, and
≤ cη max{1, f r} = cηf r.
2
Hence, for the probability to be of the form 1 − 2n−12 we require that α ≥ α(1) where
α(1) := C · ηf (r log n)
P
Thus, if α ≥ α(1) , conditioned on E0 , the bound on α1 t Φ`t wt 0 Φ given in Lemma 7.20 holds w.p. at
least 1 − 2n−12 .
Item 2 : The proof approach is similar to that of item 1. We provide the proof of this part in Appendix
B.
Item 3 : Expanding the expression for A,
!
1X
0
0
A = Echd ΦP∗,fix
at,fix at,fix P∗,fix 0 ΦEchd
α t
!
1X
0
0
+ Echd ΦPchd
at,chd at,chd Pchd 0 ΦEchd
α t
+ term1 + term10
26
where term1 := Echd 0 ΦP∗,fix
semi-definite,
λmin (A) ≥ λmin
1
α
P
t at,fix at,chd
Echd 0 ΦPchd
0
Pchd 0 ΦEchd . Since the first term on the RHS is positive
1X
at,chd at,chd 0
α t
!
!
Pchd 0 ΦEchd
+ λmin (term1 + term10 )
!
!
1X
0
0
≥ λmin Echd ΦPchd
at,chd at,chd Pchd ΦEchd
α t
!
1X
0
0
− 2 Echd ΦPchd
at,chd at,fix P∗,fix 0 ΦEchd
α t
0
(21)
(22)
P
Under our current assumptions, the at,chd ’s are scalars, so A and α1 t at,chd at,chd 0 are actually scalars.
However, we write things in a general fashion (allowing at,chd ’s to be rchg length vectors), so as to make
our later discussion of the rchg > 1 case easier. Using (22),
!
X
1
λmin (A) ≥ λmin Echd 0 ΦPchd Pchd 0 ΦEchd λmin
at,chd at,chd 0
α t
!
X
1
− 2 Echd 0 ΦPchd P∗,fix 0 ΦEchd
at,chd at,fix 0
α t
!
X
1
≥ (sin2 θ(1 − ε̃2 ) − 2ε̃| sin θ|)λmin
at,chd at,chd 0
(23)
α t
!
1X
0
at,chd at,fix
.
(24)
− 2ε̃(ε̃ + | sin θ|)
α t
The second inequality follows using Echd 0 ΦPchd = Echd 0 Echd Rchd = Rchd and Lemma 7.17. The first
inequality is straightforward if at,chd ’s are scalars (our current setting); it follows using Ostrowski’s theorem
[22] in the general case.
To bound the remaining terms in the above expression, we use Vershynin’s sub-Gaussian result [21,
Theorem 5.39] summarized in√ Theorem A.24. To apply this, recall that (at )i are bounded√ random
variables satisfying |(at )i | ≤ ηλi . Hence they are sub-Gaussian with sub-Gaussian norm ηλi [21].
Using√[21, Lemma
5.24], the vectors at are also sub-Gaussian
p
p with sub-Gaussian norm bounded by
+
maxi ηλi = ηλ . Thus, applying Theorem A.24 with K ≡ ηλ+ , ≡ v λchd , N ≡ α, nw ≡ r, if
α ≥ α(3) :=
C(r log 9 + 10 log n)f 2
,
2v
then,
Pr
1X
at at 0 − Λ ≤ v λchd
α t
!
≥ 1 − 2n−12 .
(25)
P
We could also have used matrix Bernstein to bound k t at at 0 k. However, since the at ’s are r-length
vectors and r n, the Vershynin result requires a smaller lower bound on α.
If B1 is a sub-matrix of a matrix B, then kB1 k ≤ kBk. Thus, we can also use (25) for bounding the
27
norm of various sub-matrices of
Pr λmax
1
α
P
t at at
0
− Λ . Doing this, we get
!
!
1X
at,fix at,fix 0
α t
≥ 1 − 2n−12 ,
≤ λ+ + v λchd
(26)
!
!
1X
0
at,chd at,chd ≤ λchd + v λchd ≥ 1 − 2n−12 ,
Pr λmax
α t
!
!
1X
0
at,chd at,chd ≥ λchd − v λchd ≥ 1 − 2n−12 , and
Pr λmin
α t
!
1X
at,chd at,fix 0 ≤ v λchd ≥ 1 − 2n−12 .
Pr
α t
(27)
(28)
(29)
Combining (23), (28) and (29), if α ≥ α(3) ,
Pr λmin (A) ≥ (sin2 θ(1 − ε̃2 ) − 2ε̃| sin θ|)(1 − v )λchd − 2ε̃(ε̃ + | sin θ|)v λchd ≥ 1 − 4n−12
(30)
Item 4 : Recall that Echd,⊥ 0 ΦPchd = 0. Thus,
λmax (Arest ) ≤ λmax Echd,⊥ 0 ΦP∗,fix P∗,fix 0 ΦEchd,⊥ λmax
1X
at,fix at,fix 0
α t
!
(31)
where the last inequality follows from Ostrowski’s theorem [22]. Using this and (26), if α ≥ α(3) ,
Pr λmax (Arest ) ≤ ε̃2 λ+ + ε̃2 v λchd ≥ 1 − 2n−12
P
Item 5 : Recall that term11 = α1
Echd Echd 0 Φ`t `t 0 ΦEchd,⊥ Echd,⊥ 0 . As in earlier items, we can expand
this into a sum of four terms using `t = P∗,fix at,fix + Pchd at,chd . Then using Echd,⊥ 0 ΦPchd = 0 and
kEchd k = kEchd,⊥ k = 1, we get
!
X
1
kterm11k ≤ kΦP∗,fix k P∗,fix 0 Φ λmax
at,fix at,fix 0
(32)
α t
+ kΦPchd k P∗,fix 0 Φ
1X
at,fix at,chd 0
α t
Using (26) and (29), if α ≥ α(3) , w.p. at least 1−4n−12 , kterm11k ≤ ε̃2 (λ+ +v λchd )+(ε̃(ε̃+| sin θ|))v λchd .
8
Empirical Evaluation
Here we evaluate the s-ReProCS method on real and synthetic data.
Synthetic Data. The purpose of this experiment is to (i) demonstrate the advantage of s-ReProCS for
datasets in which the outlier-frac-row is large, and (ii) to show that s-ReProCS is as good or better than
the other methods when using a random support model. We generated data with n = 5000, d = 8000,
ttrain = 500, r = 5, f = 16, J = 2, t1 = 1000, θ1 = 30◦ , t2 = 4300, and θ2 = 1.01θ1 . The at ’s were zero
mean i.i.d uniform random variables. For the left figure of Fig. 2, we generated Tt using a model to simulate
the support of a person pacing up and down the room while stopping for a while between each move. This
model and other data generation details are in Appendix G. This is one practical example where maximum
outlier fractions per row and per column can be different. Our experiment used outlier-frac-col = s/n = 0.1
28
SE(P̂(t) , P(t) )
101
10
0
ReProCS (37)
GRASTA (1.1)
10−4
10−3
ORPCA (3.0)
Offline ReProCS (85.0)
PCP (89.0)
Alt Proj (130.0)
10−9
10−6
2000
4000
t
6000
8000
RPCA-GD (470.0)
2000
4000
t
6000
8000
Figure 2: Left: SE plot for the moving object model on Tt . Right: SE plot for the Bernoulli model on Tt . Time taken
per frame in milliseconds (ms) is shown in parentheses in the legend. The errors are plotted every kα − 1 time-frames (at
projection-SVD times). Observe the nearly-exponential decay of the subspace error with time.
Original
ReProCS (16.5ms) AltProj(26.0ms) RPCA-GD(29.5ms) GRASTA (2.5ms)
PCP (44.6ms)
Original
ReProCS (85.4ms) AltProj(95.7ms) RPCA-GD(122.5ms)GRASTA (22.6ms) PCP (318.3ms)
Original
ReProCS (72.5ms) AltProj(133.1ms) RPCA-GD(113.6ms)GRASTA (18.9ms) PCP (240.7ms)
Figure 3: Background recovery. For MR and SL sequences (first two rows), only ReProCS background does not contain the
person or even his shadow. All others do. Also ReProCS is faster than all except GRASTA. For LB, ReProCS is as good as
PCP and GRASTA, while others fail. Time taken per frame is shown in parentheses.
and outlier-frac-row ≤ b0 = 0.4 for all t ∈ [ttrain , d]. To ensure that initialization of ReProCS using AltProj
works well we used s/n = 0.05 and b0 = 0.04 for t ∈ [1, ttrain ]. We generated the sparse outlier magnitudes
uniformly at random in the range [xmin , xmax ] with xmin = 10 and xmax = 25. We implemented Algorithm
4 for ReProCS (with initialization using AltProj) with α = Cf 2 (r log n) = 500, K = d−0.8 log(0.9ε̃)e = 5,
ωsupp = xmin /2, and ωevals = 0.0025λ− . Observe that ReProCS outperforms all the compared methods
and offline ReProCS significantly outperforms all of them. The reason is that the outlier fraction per row
are quite large, but ReProCS exploits slow subspace change. The other two online methods - GRASTA
[15] and ORPCA [14] - also have some approach of tracking the subspace, however they still fail because
they do not explicitly exploit the slow subspace change assumption. ReProCS and offline ReProCS are
also faster than all the methods that work to some extent (all batch methods). See time taken in the
figure legend. For the right figure of Fig. 2, the only change was that Tt was generated using an i.i.d
Bernoulli model with parameter ρ = 0.02 for training and ρ = 0.2 for t ∈ [ttrain , d]. In this case, ReProCS
outperforms all the online methods and offline ReProCS outperforms all the methods, while being faster.
The advantage is not as clear as in the left one.
Real Videos. We evaluate ReProCS for foreground-background separation for 3 videos - Meeting
Room (MR), Switch Light (SL), and Lobby (LB). MR is a challenging one because a person comes in,
writes on the board (is mostly static while he writes) and then leaves, this happens a few times with
different people. Also, the color of the person’s shirt and the curtain are hard to differentiate. SL is
challenging because light turns off then turns on again (see Appendix G). We use the first 400 frames to
get an initial subspace estimate using AltProj. We show one recovered background frame for each video in
29
Fig. 3. All algorithms used r = 40 and default parameters in their code. ReProCS used α = 60, K = 3,
√
ξt = kΨ`ˆt−1 k, ωsupp = kyt k / n, ωevals = 0.011λ− . ORPCA failed completely, gave a black background.
Hence it is not shown. Time taken per frame is shown above each image.
The parameter settings in ReProCS is slight deviation from the theory because (i) α is much smaller
than n and this works well because the video sequences tested are not long (typically the number of frames
is much fewer than the dimension n), (ii) the threshold is selected to suggest that the magnitude of outlier
entry is are greater than the average entry, and (iii) ξt measures the “error” incurred in projection step for
the previous frame, which is carried over as the residual to the current frame and so on. All the codes are
available at https://github.com/praneethmurthy/ReProCS
9
Conclusions and Future Work
We obtained the first complete guarantee for any online, streaming or dynamic RPCA algorithm that
holds under weakened versions of standard RPCA assumptions, slow subspace change, and a few simple
assumptions (outlier magnitudes are either large or very small and the `t ’s are mutually independent). Our
guarantee implies that, by exploiting these extra assumptions, one can significantly weaken the required
bound on outlier fractions per row. We analyzed a simple algorithm based on the Recursive Projected
Compressive Sensing (ReProCS) framework introduced in [7]. The algorithm itself is simpler than other
previously studied ReProCS-based methods, it is provably faster, and has near-optimal storage complexity.
Moreover, our guarantee removes all the strong assumptions made by the previous two guarantees for
ReProCS-based methods.
As described earlier, our current result still has limitations, some of which can be removed with a
little more work. For example, it assumes a very simple model on subspace change in which only one
direction can change at any given change time. Of course the changing direction could be different at
different change times, and hence over a long period, the entire subspace could change. In ongoing work,
we are studying an algorithm that does not make any assumption on subspace change, except piecewise
constant-ness and small principal angle between Pj−1 and Pj . It allows all r directions to change. A
corollary of this result also implies that the initial subspace estimate need not be very accurate and it is
still possible to track each subspace change with error below any given within a short delay. Another
issue that we would like to study is whether the lower bound on outlier magnitudes can be removed if we
use the stronger assumption on outlier fractions per row (assume they are of order 1/r). As explained
earlier, it should be possible to do this by borrowing the AltProj [4] or NO-RMC [6] ideas.
A question of practical and theoretical interest is to develop a streaming version of simple-ReProCS.
By streaming we mean that the algorithm makes only one pass through the data and needs storage of
order exactly nr. Simple-ReProCS needs only a little more storage than this, however, it makes multiple
passes through the data in the SVD steps. Algorithmically, streaming ReProCS is easy to develop: one can
replace the projection SVD and SVD steps in the subspace update by their streaming versions, e.g., block
stochastic power method. However, in order to prove that this still works (with maybe an extra factor of
log n in the delay), one would need to analyze the block stochastic power method for the problems of PCA
in data-dependent noise, and for its extension that assumes availability of partial subspace knowledge.
Finally, as explained in [8], any guarantee for dynamic RPCA also provides a guarantee for dynamic
Matrix Completion (MC) as an almost direct corollary. The reason is that MC can be interpreted as
RPCA with outlier supports Tt being known.
References
[1] E. J. Candès, X. Li, Y. Ma, and J. Wright, “Robust principal component analysis?,” J. ACM, vol.
58, no. 3, 2011.
30
[2] V. Chandrasekaran, S. Sanghavi, P. A. Parrilo, and A. S. Willsky, “Rank-sparsity incoherence for
matrix decomposition,” SIAM Journal on Optimization, vol. 21, 2011.
[3] D. Hsu, S.M. Kakade, and T. Zhang, “Robust matrix decomposition with sparse corruptions,” IEEE
Trans. Info. Th., Nov. 2011.
[4] P. Netrapalli, U N Niranjan, S. Sanghavi, A. Anandkumar, and P. Jain, “Non-convex robust pca,” in
Neural Info. Proc. Sys. (NIPS), 2014.
[5] Xinyang Yi, Dohyung Park, Yudong Chen, and Constantine Caramanis, “Fast algorithms for robust
pca via gradient descent,” in Neural Info. Proc. Sys. (NIPS), 2016.
[6] Yeshwanth Cherapanamjeri, Kartik Gupta, and Prateek Jain, “Nearly-optimal robust matrix completion,” arXiv preprint arXiv:1606.07315, 2016.
[7] C. Qiu, N. Vaswani, B. Lois, and L. Hogben, “Recursive robust pca or recursive sparse recovery in
large but structured noise,” IEEE Trans. Info. Th., pp. 5007–5039, August 2014.
[8] B. Lois and N. Vaswani, “Online matrix completion and online robust pca,” in IEEE Intl. Symp.
Info. Th. (ISIT), 2015.
[9] J. Zhan, B. Lois, H. Guo, and N. Vaswani, “Online (and Offline) Robust PCA: Novel Algorithms and
Performance Guarantees,” in Intnl. Conf. Artif. Intell. Stat. (AISTATS), 2016, long version: ArXiv:
1601.07985.
[10] N. Vaswani and H. Guo, “Correlated-pca: Principal components’ analysis when data and noise are
correlated,” in Adv. Neural Info. Proc. Sys. (NIPS), 2016.
[11] N. Vaswani and P. Narayanamurthy, “Finite sample guarantees for pca in non-isotropic and datadependent noise,” arXiv:1709.06255, 2017.
[12] J. Zhan and N. Vaswani, “Robust pca with partial subspace knowledge,” IEEE Trans. Sig. Proc.,
July 2015.
[13] UN Niranjan and Yang Shi, “Streaming robust pca,” 2016.
[14] J. Feng, H. Xu, and S. Yan, “Online robust pca via stochastic optimization,” in Adv. Neural Info.
Proc. Sys. (NIPS), 2013.
[15] J. He, L. Balzano, and A. Szlam, “Incremental gradient on the grassmannian for online foreground
and background separation in subsampled video,” in IEEE Conf. on Comp. Vis. Pat. Rec. (CVPR),
2012.
[16] Y. Li, L. Xu, J. Morphett, and R. Jacobs, “An integrated algorithm of incremental and robust pca,”
in IEEE Intl. Conf. Image Proc. (ICIP), 2003, pp. 245–248.
[17] D. Skocaj and A. Leonardis, “Weighted and robust incremental method for subspace learning,” in
IEEE Intl. Conf. Comp. Vis. (ICCV), Oct 2003, vol. 2, pp. 1494 –1501.
[18] E. Candes, “The restricted isometry property and its implications for compressed sensing,” C. R.
Math. Acad. Sci. Paris. Serie I, pp. 589–592, 2008.
[19] C. Davis and W. M. Kahan, “The rotation of eigenvectors by a perturbation. iii,” SIAM J. Numer.
Anal., vol. 7, pp. 1–46, Mar. 1970.
[20] J. A. Tropp, “User-friendly tail bounds for sums of random matrices,” Found. Comput. Math., vol.
12, no. 4, 2012.
31
[21] R. Vershynin, “Introduction to the non-asymptotic analysis of random matrices,” Compressed sensing,
pp. 210–268, 2012.
[22] R. Horn and C. Johnson, Matrix Analysis, Cambridge Univ. Press, 1985.
[23] Lin Xiao and Tong Zhang, “A proximal-gradient homotopy method for the l1-regularized least-squares
problem,” in ICML, 2012.
A
Preliminaries: Cauchy-Schwartz, matrix Bernstein and Vershynin’s
sub-Gaussian result
Cauchy-Schwartz for sums of matrices says the following [7].
Theorem A.22. For matrices X and Y we have
1X
Xt Yt 0
α t
2
≤
1X
Xt Xt 0
α t
1X
Yt Yt 0
α t
(33)
Matrix Bernstein [20], conditioned on another r.v. X, says the following.
Theorem A.23. Given an α-length sequence of n1 × n2 dimensional random matrices and a r.v. X
Assume the following. For all X ∈ C,
(i)Pconditioned on X, the
Pmatrices Zt are mutually independent, (i)
P(kZt k ≤ R|X) = 1, and (iii) max α1 t E[Zt 0 Zt |X] , α1 t E[Zt Zt 0 |X] ≤ σ 2 . Then, for an > 0,
!
1X
1X
−α2
P
Zt ≤
E[Zt |X] + X ≥ 1 − (n1 + n2 ) exp
for all X ∈ C.
α t
α t
2 (σ 2 + R)
Vershynin’s result for matrices with independent sub-Gaussian rows [21, Theorem 5.39], conditioned
on another r.v. X, says the following.
Theorem A.24. Given an N -length sequence of sub-Gaussian random vectors wi in Rnw , an r.v X, and
a set C. Assume that for all X ∈ C, (i) wi are conditionally independent given X; (ii) the sub-Gaussian
norm of wi is bounded by K for all i. Let W := [w1 , w2 , . . . , wN ]0 . Then for an 0 < < 1 we have
1
1
c2 N
P
W 0 W − E W 0 W |X ≤ X ≥ 1 − 2 exp nw log 9 −
for all X ∈ C.
(34)
N
N
4K 4
B
Proof of simple lemmas from Sec. 6 and Sec. 7
Proof of Lemma 6.7. The proof is straightforward. Let Ct := ITt ITt 0 . Then,
Ct = diag ((ct )1 , (ct )2 , · · · , (ct )n ) ,
where
(
1,
(ct )i =
0,
Since each Ct is diagonal, so is
1
α
if i ∈ Tt ,
= 1{i∈Tt } .
if i ∈
/ Tt
P
Ct . The latter has diagonal entries given by
!
1X
1 X
1 X
Ct
=
(ct )i =
1{i∈Tt }
α t
α
α
α
α
t∈J α
i,i
t∈J
32
t∈J
Thus,
1X
1 X
1X
Ct = max (
Ct )i,i = max
1{i∈Tt } = γ(J α )
i=1,2,...,n α
i=1,2,...,n α
α t
α
t
t∈J
where γ(J α ) defined in (8) is the outlier fraction per row of XJ α . From Theorem 3.1, this is bounded by
b0 .
Proof of Lemma 7.19. Applying the definitions, we get
X
1X
1
0
0
H = D̂ −
Φ`t `t Φ +
Φ`t `t Φ − D
α
α
X
1
1X
0
0
=
Φyt yt Φ −
Φ`t `t Φ
α
α
X
0
0
0
0
0 1
Φ`t `t Φ Echd Echd + Echd,⊥ Echd,⊥ − D
+ Echd Echd + Echd,⊥ Echd,⊥
α
X
1X
1X
1
0
0
0
Φwt `t Φ +
Φ`t wt Φ +
Φwt wt Φ
=
α
α
α
X
X
1
1
0
0
0
0
0
0
Echd Echd Φ`t `t ΦEchd,⊥ Echd,⊥ +
Echd,⊥ Echd,⊥ Φ`t `t ΦEchd Echd
+
α
α
Hence,
kHk ≤ 2
1X
1X
Φwt `t 0 Φ + 2 kterm11k +
Φwt wt 0 Φ .
α
α
Proof of Lemma 6.13. The first four claims were already proved below the lemma statement. Consider
the fifth claim (exact support recovery). Recall that for any t ∈ Jk , bt := Ψ`t satisfies
p
p
p
p
+
kbt k = kΨ`t k ≤ ε̃ rηλ+ + ζchd,k−1
ηλchd ≤ 2ε̃ rηλ+ + | sin θ| ηλchd := bb
From the bound on θ+ in Theorem 3.1, bb < xmin /15. Also, we set ξ = xmin /15. Using these facts, and
δ2s (Ψ) ≤ 0.12 < 0.15 (third claim), [18, Theorem 1.2] implies that
kx̂t,cs − xt k ≤ 7ξ = 7xmin /15
Thus,
|(x̂t,cs − xt )i | ≤ kx̂t,cs − xt k ≤ 7xmin /15 < xmin /2
We have ωsupp = xmin /2. Consider an index i ∈ Tt . Since |(xt )i | ≥ xmin ,
xmin − |(x̂t,cs )i | ≤ |(xt )i | − |(x̂t,cs )i | ≤ |(xt − x̂t,cs )i | <
Thus, |(x̂t,cs )i | >
(xt )j = 0 and so
xmin
2
xmin
2
= ωsupp which means i ∈ T̂t . Hence Tt ⊆ T̂t . Next, consider any j ∈
/ Tt . Then,
|(x̂t,cs )j | = |(x̂t,cs )j )| − |(xt )j | ≤ |(x̂t,cs )j − (xt )j | ≤ bCS <
which implies j ∈
/ T̂t and so T̂t ⊆ Tt . Thus T̂t = Tt .
33
xmin
2
With T̂t = Tt , the sixth claim follows easily. Since Tt is the support of xt , xt = ITt ITt 0 xt , and so
x̂t = ITt ΨTt 0 ΨTt
−1
ΨTt 0 (Ψ`t + Ψxt ) = ITt ΨTt 0 ΨTt
−1
ITt 0 Ψ`t + xt
since ΨTt 0 Ψ = IT0 t Ψ0 Ψ = ITt 0 Ψ. Thus et = x̂t − xt satisfies (11). Using (11) and the earlier claims,
ket k ≤
ΨTt 0 ΨTt
−1
ITt 0 Ψ`t ≤ φ+ ITt 0 Ψ`t
When k = 1, Ψ = I − P̂∗ P̂∗ 0 . Thus, using (15) and kP̂∗ 0 Pnew k ≤ ε̃ (follows from Lemma 6.8),
ITt 0 Ψ`t ≤ kΨP∗,fix k kat,fix k + (kΨP∗,chg cos θk + ITt 0 ΨPnew sin θ ) kat,chd k
p
p
p
≤ ε̃ ηrλ+ + ε̃| cos θ| ηλchd + (0.1 + ε̃)| sin θ| ηλchd
p
p
≤ 2ε̃ ηrλ+ + 0.11| sin θ| ηλchd
p
√
+
ηλchd .
When k > 1, kITt 0 Ψ`t k ≤ kΨ`t k ≤ ε̃ rηλ+ + ζchd,k−1
Proof of Lemma 7.17.
1. This follows because kM1,t P∗ k ≤ q0 ≤ 2ε̃ and kM1,t P∗ k = kM1,t [P∗,fix , P∗,chd ]k ≥ kM1,t P∗,fix k.
Similarly, kM1,t P∗ k ≥ kM1,t P∗,chd k.
2. First two claims follow because kΦP∗ k ≤ ε̃ and the above bound. Third claim uses kΦPnew k ≤
kΦkkPnew k = 1. The fourth claim uses triangle inequality and definition of Pchd .
QR
3. Recall that ΦPchd = Echd Rchd . Thus, σi (Rchd ) = σi (ΦPchd ). Thus kRchd k = kΦPchd k ≤ ε̃+| sin θ|
p
4. From
above,
σ
(R
)
=
σ
(ΦP
).
Moreover,
σ
(ΦP
)
=
λmin (Pchd 0 Φ0 ΦPchd ) =
min
min
min
chd
chd
chd
p
0
0
λmin (Pchd ΦPchd ). We bound this as follows. Recall that Φ = I − P̂∗ P̂∗ .
λmin (Pchd 0 ΦPchd ) ≥ λmin (cos2 θP∗,chd 0 ΦP∗,chd ) + λmin (sin2 θPnew 0 ΦPnew )
− 2| sin θ|| cos θ| kP∗,chd 0 ΦPnew k
≥ 0 + λmin (sin2 θPnew 0 ΦPnew ) − 2ε̃| sin θ|
= sin2 θλmin (I − Pnew 0 P̂∗ P̂∗ 0 Pnew ) − 2ε̃| sin θ|
= sin2 θ(1 − kPnew 0 P̂∗ k2 ) − 2ε̃| sin θ| ≥ sin2 θ(1 − ε̃2 ) − 2ε̃| sin θ|
The last inequality used Lemma 6.8.
5. Using the previous items and the definition of η,
kΦ`t k := kΦ(P∗,fix at,fix + Pchd at,chd )k ≤ kΦP∗,fix at,fix k + kΦPchd at,chd k
p
p
≤ ε̃ ηrλ+ + (ε̃| cos θ| + | sin θ|) ηrchg λchd
p
p
≤ 2ε̃ ηrλ+ + | sin θ| ηλchd
Proof of item 2 of Lemma 7.20. We use Theorem A.23 (matrix Bernstein) with Zt := Φwt wt 0 Φ. The
proof approach is similar to that of the proof of item 1. First we bound the norm of the expectation of the
34
time average of Zt :
X
1
1X
0
E
Φwt wt Φ =
ΦM2,t M1,t P ΛP 0 M1,t 0 M2,t 0 Φ
α
α
1X
M2,t M1,t P ΛP 0 M1,t 0 M2,t 0
≤
α
(a)
h
1X
M2,t M2,t 0 max M2,t M1,t P ΛP 0 M1,t (·)0
≤
t
α t
i
(b) p h
≤ b0 max M1,t P ΛP 0 M1,t 0 M2,t 0
2
!
i 1/2
t
(c)
≤
p
2
b0 q02 λ+ + qchd
λchd
where (a) follows from Cauchy-Schwartz (Theorem A.22) with Xt = M2,t and Yt = M1,t P ΛP 0 M1,t 0 M2,t 0 ,
(b) follows from the assumption on M2,t , and (c) follows from Lemma 7.17. Further, using the bounds
stated in Theorem 6.5,
X
p
1
0
2
λchd
Φwt wt Φ ≤ b0 4ε̃2 λ+ + qchd
E
α
To obtain R,
kZt k = Φwt wt 0 Φ ≤ 2 kΦMt P∗,fix at,fix k2 + kΦMt Pchd at,chd k2
2
2
≤ 2 q02 ηrλ+ + qchd
ηλchd ≤ 8ε̃2 rηλ+ + 2qchd
ηλchd := R
To obtain σ 2 ,
i
1X
1X h
E Φwt (Φwt )0 (Φwt )wt 0 Φ =
E Φwt wt 0 Φ kΦwt k2
α t
α t
2
ΦMt P ΛP 0 Mt 0 Φ
≤ max kΦwt k
wt
2
2
λchd
ηλchd q02 λ+ + qchd
≤ 2 q02 rηλ+ + qchd
4
2
≤ c1 qchd
η(λchd )2 + c2 qchd
ε̃2 ηrλ+ λchd := σ 2
Applying matrix Bernstein (Theorem A.23), we have
Pr
!
p
1X
−α2
0
2 +
2
Φwt wt Φ ≤ b0 4ε̃ λ + qchd λchd + ≥ 1 − n exp
α t
2(σ 2 + R)
2 + ε̃2 ). Then we get
Let = 1 λchd , 1 = 0.01(qchd
R
σ2
≤ cη max{1, rf }, and 2 ≤ cη max{1, rf }.
For the success probability to be of the form 1 − 2n−12 we require α ≥ α(2) where
α(2) := Cη · 13f (r log n)
Thus, if α ≥ α(2) , Pr
1
α
P
t Φwt wt
0Φ
≤
√
2
b0 4ε̃2 f + qchd
+ 1 λchd |E0 ≥ 1 − n−12 .
35
C
Proof of Theorem 3.1: general case
Here we prove Theorem 3.1 without assuming that tj ’s are known. Define
t̂j−1,f in := t̂j−1 + Kα + α − 1, tj,∗ = t̂j−1,f in + d
tj − t̂j−1,f in
eα
α
Thus, t̂j−1,f in is the time at which the (j − 1)-th subspace update is complete; whp, this occurs before tj .
Under this assumption, tj,∗ is such that tj lies in the interval [tj,∗ − α + 1, tj,∗ ]. Recall from the algorithm
that we increment j to j + 1 at t = t̂j + Kα + α := t̂j,f in . Thus, for t ∈ [tj , t̂j,f in ), Φ = I − P̂j−1 P̂j−1 0 ,
while for t ∈ [t̂j,f in , tj+1 ), Φ = I − P̂j P̂j 0 .
Define the events
Ptj,∗
(I − P̂j−1 P̂j−1 0 )`ˆt `ˆ0t (I − P̂j−1 P̂j−1 0 )) > ωevals } and
1. Det0 := {t̂j = tj,∗ } = {λmax ( α1 t=t
j,∗ −α+1
P
t
+α
j,∗
Det1 := {t̂j = tj,∗ + α} = {λmax ( 1
(I − P̂j−1 P̂j−1 0 )`ˆt `ˆ0 (I − P̂j−1 P̂j−1 0 )) > ωevals },
α
t
t=tj,∗ +1
+
2. ProjSVD := ∩K
k=1 ProjSVDk where ProjSVDk := {SE([P̂j−1 , P̂j,chd,k ]) ≤ ζchd,k },
3. Del := {SE(P̂j , Pj ) ≤ ε̃},
4. NoFalseDets := {for all J α ⊆ [t̂j,f in , tj+1 ), λmax ( α1
P
t∈J α (I
− P̂j P̂j 0 )`ˆt `ˆ0t (I − P̂j P̂j 0 )) ≤ ωevals }
5. Γ0,end := {SE(P̂∗ , P∗ ) ≤ ε̃},
∩ (Det0 ∩ ProjSVD ∩ Del ∩ NoFalseDets) ∪ (Det0 ∩ Det1 ∩ ProjSVD ∩ Del ∩
6. Γj,end := Γj−1,end
NoFalseDets) .
Let p0 denote the probability that, conditioned on Γj−1,end , the change got detected at t = tj,∗ , i.e., let
p0 := Pr(Det0|Γj−1,end ).
Thus, Pr(Det0|Γj−1,end ) = 1 − p0 . It is not easy to bound p0 . However, as we will see, this will not be
needed.
Assume that Γj−1,end ∩ Det0 holds. Consider the interval J α := [tj,∗ , tj,∗ + α). This interval starts at or
after tj , so, for all t in this interval, the subspace has changed. For this interval, Ψ = Φ = I − P̂j−1 P̂j−1 0 .
Applying the last item of Theorem 6.5, w.p. at least 1 − 12n−12 ,
!
1 X ˆ ˆ0
Φ`t `t Φ ≥ (0.97 sin2 θ − 0.37qchd | sin θ| − 0.1ε̃| sin θ|)λchd
λmax
α
α
t∈J
where qchd is the bound k(ΨT̂t 0 ΨT̂t )−1 IT̂t 0 ΨPchd k. Theorem 6.5 is applicable for the reasons given in
the proof of Lemma 6.14. Proceeding as in the proof of Lemma 6.14 for k = 1, we get that qchd =
1.2((0.1 + ε̃)| sin θ| + ε̃). Thus, using the bound on ε̃, we can conclude that, w.p. at least 1 − 12n−12 ,
!
1 X ˆ ˆ0
λmax
Φ`t `t Φ ≥ 0.91 sin2 θλchd ≥ 0.91 sin2 θλ− > ωevals
α
α
t∈J
and thus t̂j = tj,∗ + α. This follows since ωevals = 5ε̃2 λ+ = 5ε̃2 f λ− ≤ ε̃f λ− ≤ 0.8∆2min = 0.8 sin2 θ−
(follows from the bound on ε̃ and ∆min = sin θ− ). In other words,
Pr(Det1|Γj−1,end ∩ Det0) ≥ 1 − 12n−12 .
Conditioned on Γj−1,end ∩ Det0 ∩ Det1, the first projection-SVD step is done at t = t̂j + α = tj,∗ + 2α
and so on. We can state and prove Lemma 6.14 with Γk replaced by Γj,end ∩ Det0 ∩ Det1 ∩ ProjSVD1 ∩
36
ProjSVD2 . . . ProjSVDk and with the k-th projection-SVD interval being Jk := [t̂j + (k − 1)α, t̂j + kα).
We can state and prove a similarly changed version of Lemma 6.15 for the simple SVD based deletion step.
Applying Lemma 6.14 for each k, and then apply Lemma 6.15,
Pr(ProjSVD ∩ Del|Γj−1,end ∩ Det0 ∩ Det1) ≥ (1 − 12n−12 )K+1 .
We can also do a similar thing for the case when the change is detected at tj,∗ , i.e. when Det0 holds. In
this case, we replace Γk by Γj,end ∩ Det0 ∩ ProjSVD1 ∩ ProjSVD2 . . . ProjSVDk and conclude that
Pr(ProjSVD ∩ Del|Γj−1,end ∩ Det0) ≥ (1 − 12n−12 )K+1 .
Finally consider the NoFalseDets event. First, assume that Γj−1,end ∩ Det0 ∩ ProjSVD ∩ Del holds.
Consider any interval J α ⊆ [t̂j,f in , tj+1 ). In this interval, Q̂t = P̂j , Ψ = Φ = I − P̂j P̂j 0 and SE(P̂j , Pj ) ≤ ε̃.
Also, using Lemma 6.13, et satisfies (11) for t in this interval. Thus,
X
1 X ˆ ˆ0
1
1 X
Φ`t `t Φ = ΦP (
at at 0 )P 0 Φ +
Φ`t `t 0 Ψ0 ITt (ΨTt 0 ΨTt )−1 IT0 t Φ + (.)0
α
α
α
t∈J α
t∈J α
t∈J α
1 X
+
ΦITt (ΨTt 0 ΨTt )−1 ITt 0 Ψ`t `t 0 Ψ0 ITt (ΨTt 0 ΨTt )−1 IT0 t Φ
α
α
t∈J
We can bound the first term using Vershynin’s sub-Gaussian result (Theorem A.24) and the other terms
using matrix Bernstein (Theorem A.23). The approach is similar to that of the proof of Lemma 7.20. The
derivation is more straightforward in this case, since for the above interval kΨPj k = kΦPj k ≤ ε̃. The
required bounds on α are also the same as those needed for Lemma 7.20 to hold. We conclude that, w.p.
at least 1 − 12n−12 ,
!
p
p
1 X ˆ ˆ0
λmax
Φ`t `t Φ ≤ ε̃2 (λ+ + 0.01λ+ )[1 + b0 (φ+ )2 + 2 b0 φ+ ] ≤ 2.2ε̃2 f λ− < ωevals
α
α
t∈J
This follows since ωevals = 5ε̃2 λ+ = 5ε̃2 f λ− . Since Det0 holds, t̂j = tj,∗ . Thus, we have a tot
−t −Kα−α
t
−t −Kα−α
tal of b j+1 j,∗α
c intervals J α that are subsets of [t̂j,f in , tj+1 ). Moreover, b j+1 j,∗α
c ≤
tj+1 −tj
tj+1 −tj −Kα−α
c ≤ b α c − (K + 1) since α ≤ α. Thus,
b
α
Pr(NoFalseDets|Γj−1,end ∩ Det0 ∩ ProjSVD ∩ Del) ≥ (1 − 12n−12 )b
tj+1 −tj
α
c−(K+1)
On the other hand, if we condition on Γj−1,end ∩ Det0 ∩ Det1 ∩ ProjSVD ∩ Del, then t̂j = tj,∗ + α. Thus,
Pr(NoFalseDets|Γj−1,end ∩ Det0 ∩ Det1 ∩ ProjSVD ∩ Del) ≥ (1 − 12n−12 )b
tj+1 −tj
α
c−(K+2)
We can now combine the above facts to bound Pr(Γj,end |Γj−1,end ). Recall that p0 := Pr(Det0|Γj−1,end ).
Clearly, the events (Det0∩ProjSVD∩Del∩NoFalseDets) and (Det0∩Det1∩ProjSVD∩Del∩NoFalseDets)
are disjoint. Thus,
Pr(Γj,end |Γj−1,end )
= p0 Pr(ProjSVD ∩ Del ∩ NoFalseDets|Γj−1,end ∩ Det0)
+ (1 − p0 ) Pr(Det1|Γj−1,end ∩ Det0) Pr(ProjSVD ∩ Del ∩ NoFalseDets|Γj−1,end ∩ Det0 ∩ Det1)
≥ p0 (1 − 12n−12 )K+1 (1 − 12n−12 )b
tj+1 −tj
α
c−(K+1)
+ (1 − p0 )(1 − 12n−12 )(1 − 12n−12 )K+1 (1 − 12n−12 )b
= (1 − 12n−12 )b
tj+1 −tj
α
c
≥ (1 − 12n−12 )tj+1 −tj .
37
tj+1 −tj c
−(K+2)
α
Since the events Γj,end are nested, the above implies that
Y
Y
Pr(ΓJ,end |Γ0,end ) =
Pr(Γj,end |Γj−1,end ) ≥
(1 − 12n−12 )tj+1 −tj = (1 − 12n−12 )d
j
j
≥ 1 − 12dn−12 .
D
Proof of Corollary 3.2
Proof. Like Theorem 3.1, this corollary is also a direct consequence of Lemmas 6.13, 6.14, and 6.15.
Consider Lemma 6.13. The first four parts remain exactly the same (assuming same bounds from Fact
6.12). The exact support recovery proof given above in Appendix B changes and the expression and
bounds for et change. We explain the changes here. With the extra noise term vt , the bound on ||bt || will
change. It will now contain an extra term equal to bv . With this change, a stronger assumption will be
needed for exact support recovery. The new bound on θ+ (or equivalently on ∆) given in the corollary,
will ensure that exact support recovery holds even with extra noise vt being present. With this, there will
be a simple change to the expression for et .
This change will imply that the proof of Lemma 6.14 (which relies on Theorem 6.5) will need bounds
on two extra terms. We bound these below. The bounds on these extra terms depend on λ+
v . The value of
λ+
assumed
in
the
corollary
ensures
that
Lemma
6.14
holds
with
no
change
and
so
no
change
is needed to
v
+
the definitions of ζchd,k or to Fact 6.12.
Lemma 6.15 will need similar changes which follow by directly using the noisy version of Theorem 6.16
from [11].
P
1. Consider α1 t vt vt 0 . We will apply matrix Bernstein (Theorem A.23). We have kvt vt 0 k = kvt k2 ≤
P
P
2
b2v = R and α1 t E[vt vt 0 vt vt 0 ] ≤ maxt kvt k2 α1 t E[vt vt 0 ] ≤ b2v λ+
v = σ . Now we can apply
Theorem A.23 to get
!
1X
−α2
0
+
Pr
.
vt vt ≤ λv + ≥ 1 − 2n exp
α t
2(σ 2 + R)
2 + −
+
with λ+
v = 0.03 sin θ λ . Let = 0,v λv and 0,v = 0.01. Then,
2
σ R
b2v
1
b2v
max
,
=
max
,
1
=
C
2
0,v
0,v λ+
λ+
v
v
2
Define α(v,0) := C(log n) λb+v .
v
P
1 P
0
2. Consider α t `t vt . We will apply matrix Bernstein (Theorem A.23). We have E α1 t `t vt 0 =
p
P
P
0, k`t vt 0 k = k`t k kvt k ≤ ηrλ+ bv = R and max{ α1 t E[`t vt 0 vt `t 0 ] , α1 t E[vt `t 0 `t vt 0 ] } ≤
2
max{b2v λ+ , ηrλ+ λ+
v } = σ . Applying Theorem A.23,
!
1X
−α2
0
Pr
`t vt ≤ ≥ 1 − 2n exp
,
α t
2(σ 2 + R)
Let = 1,v λchd , and 1,v = 0.01. Then,
(
)
2
√
ηrbv λ+
σ R
b2v f
ηrf λ+
v
max
,
= max 2
,
,
2
1,v λchd 21,v λchd 1,v λchd
√
ηrbv λ+
b2v f
ηrf λ+
v
Define α(v,1) := C log n max 2 λ , 2 λ , 1,v λchd .
1,v chd
1,v chd
The above bounds hold w.p. at least 1 − 4n−10 if α ≥ max{α(v,0) , α(v,1) } = αv,1 . Thus, we replace αchd
√
2
ηrbv λ+
b2v f
ηrf λ+
v
by max(Cηf r(log n), C(log n) max 2 λ , 2 λ , 1,v λchd ) = max(Cηf r(log n), Cf λb+v (log n)).
1,v chd
1,v chd
38
v
E
More than one changed direction
Consider the following model for subspace change with rchg > 1. This is called the double rotation model.
Let Pj,new be a n × rchg basis matrix with directions that belong to span(Pj−1,⊥ ), let Pj−1,chg be an
n × rchg basis matrix for the “changing” subspace of span(Pj−1 ) and let Pj−1,fix be the basis for its fixed
part. Then, we have Pj = [Pj−1,fix , Pj,chd ] where (Pj,chd )i = (Pj,chg )i cos θj,i − (Pj,new )i sin θj,i , for all
i = 1, 2, . . . rchg , and Pj−1,fix is n × (r − rchg ). Thus there are rchg new directions and each gets rotated in
with angle θj,i .
Corollary E.25. With the above model on subspace change, Theorem 3.1 holds with the following simple
changes.
• Redefine θ+ := maxj=1,2,...,J maxi=1,2,...,rchg θj,i and θ− := minj=1,2,...,J mini=1,2,...,rchg θj,i .
q
q
µrchg
0P
• Redefine µ as the smallest real number so that maxi kIi 0 Pj k ≤ µr
and
max
kI
k
≤
i
i j,new
n
n .
• Let fchg =
λ+
chd
λ−
chd
−
with λ+
chd := λmax (Λchd ) and λchd := λmin (Λchd ) being the maximum and minimum
2
eigenvalues along the changing directions. Assume that b0 = 0.01/fchg
and redefine αchd as its
previous value times fchg .
• Estimate P̂j,chd,k as the top rchg singular vectors of the matrix B in Algorithm 4.
• If rchg is known, just the above changes suffices. If it is unknown, we can use the approach suggested
in [8] to estimate it. We estimate rchg as the number of singular values of the matrix B that are
√
larger than ωevals α. For this to work, we will also need to assume that sin2 θ− /fchg > 0.0011.
The above claim follows by an easy modification of our proof for rchg = 1. The reason is that, in most
of the proof, we have written claims so that they hold even when rchg > 1. The main change will be in
proving Theorem 6.5. We set rchg = 2 for explaining the ideas below but nothing changes even if rchg is
larger.
We have the following expression for Pchd . Using this we will re-derive some of the claims of Lemma
7.17.
cos θ1 0
sin θ1 0
Pchd = P∗,chd
− Pnew
0 cos θ2
0 sin θ2
cos θ1 0
sin θ1 0
Let θmax = max{θ1 , θ2 } and θmin = min{θ1 , θ2 }, DC =
and DS =
. Then,
0 cos θ2
0 sin θ2
kΦPchd k ≤ kΦP∗,chd DC − ΦPnew DS k
≤ | cos θmin | kΦP∗,chd k + | sin θmax | kΦPnew k ≤ ε̃ + | sin θmax |.
p
Next, consider σmin (ΦPchd ) = λmin (Pchd 0 ΦPchd ). Using the definition of Pchd and expanding the above
matrix we get
Pchd 0 ΦPchd = (P∗,chd DC − Pnew DS ) 0 Φ (P∗,chd DC − Pnew DS )
= DC 0 P∗,chd 0 ΦP∗,chd DC + DS 0 Pnew 0 ΦPnew DC
− DC 0 P∗,chd 0 ΦPnew DS + DS 0 Pnew 0 ΦP∗,chd DC
Using Weyl’s inequality,
λmin Pchd 0 ΦPchd ≥ λmin DC 0 P∗,chd 0 ΦP∗,chd DC + λmin DS 0 Pnew 0 ΦPnew DS
− λmax DC 0 P∗,chd 0 ΦPnew DS + DS 0 Pnew 0 ΦP∗,chd DC
39
The first term is non-negative since the argument is positive semi-definite. Thus, using Ostrowsky’s
theorem,
λmin Pchd 0 ΦPchd ≥ λmin DS 0 DS λmin Pnew 0 ΦPnew − 2 kDS k kDC k kΦP∗,chd k kΦPnew k
≥ sin2 θmin (1 − ε̃2 ) − 2| cos θmin || sin θmax |ε̃ ≥ sin2 θmin (1 − ε̃2 ) − 2ε̃| sin θmax |
We use these bounds to get the high probability bounds on the terms in Lemma 7.20. Here we only
mention the changes to Lemma 7.20 rather than rewriting the complete proof.
P
1. For α1 t Φ`t wt 0 Φ , we have that
#
"
p
1X
0
Φ`t wt Φ ≤ b0 [2ε̃2 λ+ (ε̃ + | sin θmax |)qchd λ+
E
chd ,
α t
+
R = c1 | sin θmax |qchd ηrchg λ+
chd + c2 | sin θmax |ε̃ηrλ
2
2
2
2
+ +
σ12 = c1 qchd
sin2 θmax η(λ+
chd ) rchg + c2 ε̃ sin θmax ηrλ λchd
2 2
2
Choosing = 0 λ−
chd where 0 = 0.01| sin θmax |(qchd + ε̃) gives σ / ≤ C max{ηfchg rchg , ηrf fchg }
and R/ ≤ C max{ηrchg fchg , ηrf } so that the sample complexity for this term is of the form
α(1) := Cηfchg max{rf, rchg fchg }(log n) = Cηfchg f (r log n).
2. Similarly the wt wt0 term will be bounded.
3. For λmin (A) term, we have the following changes:
+
λmin (A) ≥ sin2 θmin (1 − ε̃2 ) − 2ε̃| sin θmax | λ+
chd − 2ε̃(ε̃ + | sin θmax |)λchd ,
4. The term λmax (Arest ) does not change.
5. The term11 term will have the following change
+
kterm11k ≤ ε̃2 (λ+ + 2 λ+
chd ) + ε̃(ε̃ + | sin θmax |)2 λchd
6. One common change in all the terms is that in all expressions, we need to replace λchd by λ−
chd .
Observe from above that the only changes in our final conclusion of Theorem 6.5 are
1. α(1) gets replaced by its old value times fchg , and hence αchd has the same change.
2. θ gets replaced by θmax everywhere except in the first term in the lower bound on λmin (A) where we
replace it θmin .
The effect of the second item above in our main result, Theorem 3.1, is that θ+ and θ− get redefined.
We can argue that rchg is correctly estimated if we can show that λrchg (Dobs ) > ωevals > λrchg +1 (Dobs ).
For this, by Weyl’s inequality, if λmin (A) > λmax (Arest ), then
λrchg (Dobs ) ≥ λrchg (D) − kHk = λmin (A) − kHk,
λrchg +1 (Dobs ) ≤ λrchg +1 (D) + kHk = λmax (Arest ) + kHk
Applying the modified Lemma 7.20, we get high probability bounds on the above terms. Using those
bounds in the proof of Theorem 3.1 or 3.1, and setting ωevals appropriately, we can argue that rchg is
correctly estimated.
40
F
Time complexity of s-ReProCS
The time-consuming steps of s-ReProCS are either l1 minimization or the subspace update steps. Support
estimation and LS steps are much faster and hence can be ignored for this discussion. The computational
complexity of l1 minimization (if the best solver were used) [23] is the cost of multiplying the CS matrix
or its transpose with a vector times (− log ) if is the bound on the error w.r.t. the true minimizer of
the program. In ReProCS, the CS matrix is of the form I − P̂ P̂ 0 where P̂ is of size n × r or n × (r + 1),
thus multiplying a vector with it takes time O(nr). Thus, the l1 minimization complexity per frame is
O(nr(− log )), and thus the total cost for d − ttrain frames is O(nr(− log )(d − ttrain )).
The subspace update step consists of (d − ttrain − Jα)/α rank one SVD’s on an n × α matrix (for either
detecting subspace change or for projection-SVD) and J rank r SVD’s on an n × α matrix (for subspace
re-estimation). Thus the subspace update complexity is at most O(n(d − ttrain )r(− log )) and the total
ReProCS complexity (without the initialization step) is O(n(d − ttrain )r(− log )).
If we assume that the initialization uses AltProj, AltProj is applied to a matrix of size n × ttrain with
rank r. Thus the initialization complexity is O(nttrain r2 (− log )). If instead GD [5] or NO-RMC [6] is
used, then the time complexity is reduced to O(nttrain rf (− log )) or even lower with NO-RMC. Treating
f as a constant (our discussion treats condition numbers as constants), the final complexity of ReProCS is
O(ndr(− log )).
G
G.1
Detailed Experimental Results
Simulation experiments
Our first simulation experiment is done to illustrate the advantage of ReProCS over existing batch and
online RPCA techniques. As explained earlier, because ReProCS exploits dynamics (slow subspace change),
it is provably able to tolerate a much larger fraction of outliers per row than all the existing techniques
without needing uniformly randomly generated support sets. When the number of subspace changes,
J, is large, it also tolerates a significantly larger fraction of outliers per column. The latter is hard to
demonstrate via simulations (making J large will require a very long sequence). Thus we demonstrate
only the former. Our second experiment shows results with using an i.i.d. Bernoulli model on support
change (which is the model assumed in the other works).
One practical instance where outlier fractions per row can be larger than those per column is in the
case of video moving objects that are either occasionally static or slow moving. The outlier support model
for our first and second experiments is inspired by this example. It models a 1D video consisting of a
person/object of length s pacing up and down in a room with frequent stops. The object is static for β
frames at a time and then moves down. It keeps moving down for a period of τ frames, after which it
turns back up and does the same thing in the other direction. We let β = dc0 τ e for a c0 < 1. With this
model, for any interval of the form [(k1 − 1)τ + 1, k2 τ ] for k1 , k2 integers, the outlier fraction per row is
bounded by c0 . For any general interval of length α ≥ τ , this fraction is still bounded by 2c0 . The fraction
per column is s/n.
Model G.26. Let β = dc0 τ e. Assume that the Tt satisfies the following. For the first τ frames (downward
motion),
[1, s],
t ∈ [1, β]
[s + 1, 2s],
t ∈ [β + 1, 2β]
Tt = .
..
[(1/c0 − 1)s + 1, s/c0 ],
t ∈ [τ − β + 1, τ ]
41
for the next τ frames (upward motion),
[(1/c0 − 1)s + 1, s/c0 ],
[(1/c0 − 2)s + 1, (1/c0 − 1)s],
Tt = .
..
[1, s],
t ∈ [τ + 1, τ + β]
t ∈ [τ + β + 1, τ + 2β]
t ∈ [2τ − β + 1, 2τ ].
Starting at t = 2τ + 1, the above pattern is repeated every 2τ frames until the end, t = d.
This model is similar to the model assumed for the guarantees in older works [8, 9]. The above model
is one practically motivated way to simulate data that is not not generated uniformly at random (or as
i.i.d. Bernoulli, which is approximately the same as the uniform model for large n). It also provides a
way to generate data with a different bounds on outlier fractions per row and per column. The maximum
outlier fraction per column is s/n. For any time interval of length α ≥ τ , the outlier fraction per row is
bounded by 2c0 . Thus, for Theorem 3.1, with this model, b0 = 2c0 . By picking 2c0 larger than s/n we can
ensure larger outlier fractions per row than per column.
We compare Algorithm 4 and its offline counterpart with three of the batch methods with provably
guarantees discussed in Sec. 4 - PCP [1], AltProj [4] and RPCA-GD [5] - and with two recently proposed
online algorithms known to have good experimental performance and for which code was available ORPCA [14] and GRASTA [15]. The code for all these techniques are cloned from the Low-Rank and
Sparse library (https://github.com/andrewssobral/lrslibrary).
For generating data we used d = 8000, ttrain = 500, J = 2, r = 5, f = 16, J = 2 with t1 = 1000,
θ1 = 30◦ , t2 = 4300, θ2 = 1.01θ1 and varying n. The at ’s are zero mean i.i.d uniform random variables
generated exactly as described before and so is Qt . We generated a basis matrix U by ortho-normalizing
the first r + 1 columns of a n × n i.i.d. standard Gaussian matrix. For t ∈ [1, t1 ), we set Qt = P0
with P0 being the first r columns of U . We let P1,new be last column of U , and rotated it in using
(6) with Rj,0 = I and with angle θ1 to get P1 . We set Qt = P1 for t ∈ [t1 , d]. At all times√t, we√let
`t = Qt at with at being zero mean, i.i.d uniform random variables such that (at )i ∼ unif − f , f
for i = 1, · · · , r − 2 and (at )[r−1,r] ∼ unif (−1, 1). With this the condition number is f , the covariance
matrix, Λ = diag(f, f, · · · , f, 1, 1)/3, λ+ = f /3, λchg = λ− = 1/3, and η = 3. We generate Tt using Model
G.26 as follows. For t ∈ [ttrain , d], we used s = 0.1n, c0 = 0.2 and τ = 100. Thus b0 = 0.4. For t ∈ [1, ttrain ],
we used s = 0.05n and c0 = 0.02. This was done to ensure that AltProj (or any other batch technique
works well for this period and provides a good initialization). The magnitudes of the nonzero entries of xt
(outliers) were generated s i.i.d uniform r.v.’s between xmin = 10 and xmax = 25. We show an example
sparse matrix generated as above in Fig. 4. We let yt = `t + xt for each t = 1, 2, . . . , d.
200
400
600
800
50
100
150
200
250
300
350
400
450
500
Figure 4: Illustrating the sparse matrix under the moving-object model (Model G.26). For this figure,
ttrain = 200. The first ttrain frames have fewer outliers than the rest.
42
Automatic ReProCS (Algorithm 4) and its offline counterpart were implemented as described earlier.
We initialized using AltProj applied to Y[1,ttrain ] . For the batch methods used in the comparisons – PCP,
AltProj and RPCA-GD, we implement the algorithms on Y[1,t] every t = ttrain + kα − 1 frames. Further,
√
we set the regularization parameter for PCP 1/ n in accordance with [1]. The other known parameters,
r for Alt-Proj, outlier-fraction for RPCA-GD, are set using the true values. For online methods we
implement the algorithms without
modifications. The regularization parameter for ORPCA was set as
√
√
with λ1 = 1/ n and λ2 = 1/ d according to [14]. We plot the subspace error and the normalized error of
`t over time in Fig. 5(a) and 5(b) for n = 5000. We display the time-averaged error for other values of
n in Table 2. This table also contains the time comparisons. All time comparisons are performed on a
R
Desktop Computer with Intel Xeon E3-1240 8-core CPU @ 3.50GHz and 32GB RAM.
Table 2: Average subspace error SE(Q̂t , Qt ) and time comparison for different values of signal size n. The values
in brackets denote average time taken per frame (– indicates that the algorithm does not work).
n = 500 (in 10−4 s)
n = 500, Bern. (in 10−4 s)
n = 5000 (in 10−2 s)
n = 10, 000 (in 10−2 s)
ReProCS
GRASTA
ORPCA
Offline ReProCS
PCP
AltProj
RPCA-GD
0.066
0.044
0.048
0.090
0.996
0.747
0.999
0.999
0.320 (10)
0.078 (1.8)
0.322 (0.30)
0.3235 (0.68)
8.25 × 10−5 (6.3)
3.9 × 10−7 (9.2)
6.05 × 10−5 (8.5)
0.0006 (36.8)
1.00 (51)
1.2 × 10−4 (395)
0.999 (8.9)
–
0.176 (104)
0.0001 (32)
0.354 (13.0)
–
0.215 (454)
0.303 (329)
0.223 (47.0)
–
(3.1)
(4.8)
(3.7)
(15.6)
(2.8)
(1.9)
(0.11)
(0.25)
As can be seen, ReProCS outperforms all the other methods and offline ReProCS significantly
outperforms all the other methods for this experiment. The reason is that the outlier fraction per row are
quite large, but ReProCS exploits slow subspace change. In principle, even GRASTA exploits slow subspace
change, however, it uses approximate methods for computing the SVD and does not use projection-SVD
and hence it fails. ReProCS and offline ReProCS are faster than all the batch methods especially for large
n. In fact when n = 10000, the batch methods are out of memory and cannot work, while ReProCS still
can. But ReProCS is slower than GRASTA and ORPCA.
Comparison with other algorithms - random outlier support using the i.i.d. Bernoulli model. We
generated data exactly as described above with the following change: Tt was now generated as i.i.d.
Bernoulli with probability of any index i being in Tt being ρs = 0.02 for the first ttrain frames and ρs = 0.2
for the subsequent data. We used n = 500. We show the results for n = 500 in Fig. 5(c) and 5(d). For this
experiment, the batch methods PCP and AltProj have good performance, that is better than ReProCS at
most time instants. Offline ReProCS still outperforms all the other methods.
G.2
Real Video Experiments
In this section we provide simulation results for on real video, specifically the Meeting Room (MR) sequence.
The meeting room sequence is set of 1964 images of resolution 64 × 80. The first 1755 frames consists of
outlier-free data. Henceforth, we consider only the last 1209 frames. For the ReProCS algorithm, we used
the first 400 frames as the training data. In the first 400 frames, a person wearing a black shirt walks in,
writes something on the board and goes back. In the subsequent frames, the person walks in with a white
shirt. This is a challenging video sequence because the color of the person and the color of the curtain are
hard to distinguish. ReProCS algorithm is able to perform the separation at around 43 frames per second.
We obtained an estimate using the Alt Proj algorithm. For the Alt Proj algorithm we set r = 40.
The remaining parameters were used with default setting. For the ReProCS algorithm, we set α = 60,
K = 3, ξt = kΨ`ˆt−1 k2 , θ− = 20◦ . We found that these parameters work for most videos that we verified
our algorithm on. For RPCA-GD we set the “corruption fraction” α = 0.2 as described in the paper.
Lobby dataset: This dataset contains 1555 images of resolution 128 × 160. The first 341 frames are
outlier free. Here we use the first 400 “noisy” frames as training data. The Alt Proj algorithm is used to
obtain an initial estimate with rank, r = 40. The parameters used in all algorithms are exactly the same
as above. ReProCS achieves a “test” processing rate of 16 frames-per-second.
Switch Light dataset: This dataset contains 2100 images of resolution 120 × 160. The first 770 frames are
43
(b)
(a) SE(Q̂t , Qt )
10
k`ˆt −`t k22
k`t k22
3
ReProCS (37)
GRASTA (1.1)
10−2
ORPCA (3.0)
Offline ReProCS (85.0)
PCP (89.0)
Alt Proj (130.0)
10−7
RPCA-GD (470.0)
(c)
(d)
101
10−5
10−11
2000
4000
6000
8000
0
2000
t
4000
6000
t
Figure 5: First row ((a), (b)): Illustrate the subspace error and the normalized `t error for n = 5000 and
outlier supports generated using Model G.26. Both the metrics are plotted every kα − 1 time-frames.
The results are averaged over 100 iterations. Second row ((c), (d)) illustrate the subspace error and the
normalized `t error for n = 500 and Bernoulli outlier support model. They are plotted every kα − 1
time-frames. The plots clearly corroborates the nearly-exponential decay of the subspace error as well as
the error in `t .
outlier free. Here we use the first “noisy” 400 frames as training data. The Alt Proj algorithm is used to
obtain an initial estimate with rank, r = 40. The parameters used in all algorithms are exactly the same
as in the MR and LB dataset. ReProCS achieves a “test” processing rate of 12 frames-per-second.
44
original
ReProCS
AltProj
RPCA-GD
GRASTA
PCP
Figure 6: Comparison of visual performance in Foreground Background separation for the MR dataset.
The images are shown at t = ttrain + 10, 140, 630, 760.
original
ReProCS
AltProj
RPCA-GD
GRASTA
PCP
Figure 7: Comparison of visual performance in Foreground Background separation for the Lobby (LB)
dataset. The recovered background images are shown at t = ttrain + 260, 545, 610. The time taken for each
of the techniques is shown in parentheses. For ReProCS the time denotes training-time + test-time.
45
original
ReProCS
AltProj
RPCA-GD
GRASTA
PCP
Figure 8: Comparison of visual performance in Foreground Background separation for the SwitchLight
(SL) dataset. The recovered background images are shown at t = ttrain + 60, 200, 688, 999. Note that
because of the drastic change in environment, none of the techniques are able to detect the object properly.
46
| 7cs.IT
|
arXiv:1705.03439v1 [stat.ML] 9 May 2017
Frequentist Consistency
of Variational Bayes
Yixin Wang
Department of Statistics
Columbia University
David M. Blei
Department of Computer Science and Statistics
Columbia University
May 10, 2017
Abstract
A key challenge for modern Bayesian statistics is how to perform scalable inference of posterior distributions. To address this challenge, variational Bayes (vb) methods have emerged
as a popular alternative to the classical Markov chain Monte Carlo (mcmc) methods. vb
methods tend to be faster while achieving comparable predictive performance. However, there
are few theoretical results around vb. In this paper, we establish frequentist consistency and
asymptotic normality of vb methods. Specifically, we connect vb methods to point estimates
based on variational approximations, called frequentist variational approximations, and
we use the connection to prove a variational Bernstein-von-Mises theorem. The theorem
leverages the theoretical characterizations of frequentist variational approximations to understand asymptotic properties of vb. In summary, we prove that (1) the vb posterior converges
to the Kullback-Leibler (kl) minimizer of a normal distribution, centered at the truth and (2)
the corresponding variational expectation of the parameter is consistent and asymptotically
normal. As applications of the theorem, we derive asymptotic properties of vb posteriors in
Bayesian mixture models, Bayesian generalized linear mixed models, and Bayesian stochastic
block models. We conduct a simulation study to illustrate these theoretical results.
Keywords: Bernstein–Von Mises, Bayesian inference, variational methods, consistency, asymptotic normality, statistical computing
1
Introduction
Bayesian modeling is a powerful approach for discovering hidden patterns in data. We begin by
setting up a probability model of latent variables and observations. We incorporate prior knowledge by setting priors on latent variables and a functional form of the likelihood. Finally we infer
the posterior, the conditional distribution of the latent variables given the observations.
For many modern Bayesian models, exact computation of the posterior is intractable and
statisticians must resort to approximate posterior inference. For decades, Markov chain Monte
1
Carlo (mcmc) sampling (Hastings, 1970; Gelfand and Smith, 1990; Robert and Casella, 2004)
has maintained its status as the dominant approach to this problem. mcmc algorithms are easy
to use and theoretically sound. In recent years, however, data sizes have soared. This challenges
mcmc methods, for which convergence can be slow, and calls upon scalable alternatives. One
popular class of alternatives is variational Bayes (vb) methods.
To describe vb, we introduce notation for the posterior inference problem. Consider observations
x = x1:n . We posit local latent variables z = z1:n , one per observation, and global latent variables
θ = θ1:d . This gives a joint,
p(θ , z, x) = p(θ )
n
Y
p( z i | θ ) p( x i | z i , θ ).
(1)
i =1
The posterior inference problem is to calculate the posterior p(θ , z | x).
This division of latent variables is common in modern Bayesian statistics.1 In the Bayesian
Gaussian mixture model (gmm) (Roberts et al., 1998), the component means, covariances, and
mixture proportions are global latent variables; the mixture assignments of each observation
are local latent variables. In the Bayesian generalized linear mixed model (glmm) (Breslow and
Clayton, 1993), the intercept and slope are global latent variables; the group-specific random
effects are local latent variables. In the Bayesian stochastic block model (sbm) (Hofman and
Wiggins, 2008), the cluster assignment probabilities and edge probabilities matrix are two sets
of global latent variables; the node-specific cluster assignments are local latent variables. In the
latent Dirichlet allocation (lda) model (Blei et al., 2003), the topic-specific word distributions
are global latent variables; the document-specific topic distributions are local latent variables.
We will study all these examples below.
vb methods formulate posterior inference as an optimization (Jordan et al., 1999; Wainwright
and Jordan, 2008; Blei et al., 2016). We consider a family of distributions of the latent variables
and then find the member of that family that is closest to the posterior.
Here we focus on mean-field variational inference (though our results apply more widely). First,
we posit a family of factorizable probability distributions on latent variables
n
o
Q
Q
Q n+d = q : q(θ , z) = di=1 q θ i (θ i ) nj=1 q z j ( z j ) .
This family is called the mean-field family. It represents a joint of the latent variables with n + d
(parametric) marginal distributions, { q θ1 , . . . , q θd , q z1 , . . . , q zn }.
vb finds the member of the family closest to the exact posterior p(θ, z | x), where closeness is
measured by kl divergence. Thus vb seeks to solve the optimization,
q∗ (θ , z) = arg min
q(θ ,z)∈Q n+d
kl( q(θ, z) || p(θ, z | x)).
(2)
In practice, vb finds q∗ (θ , z) by optimizing an alternative objective, the evidence lower bound
( elbo),
Z
q(θ , z)
elbo( q(θ, z)) = − q(θ, z) log
dθ d z.
(3)
p(θ , z, x)
1 In particular, our results are applicable to general models with local and global latent variables (Hoffman et al.,
2013). The number of local variables z increases with the sample size n; the number of global variables θ does not. We
also note that the conditional independence of Equation (1) is not necessary for our results. But we use this common
setup to simplify the presentation.
2
This objective is called the elbo because it is a lower bound on the evidence log p( x). More
importantly, the elbo is equal to the negative KL plus log p( x), which does not depend on q(·).
Maximizing the elbo minimizes the kl (Jordan et al., 1999).
The optimum q∗ (θ , z) = q∗ (θ ) q∗ ( z) approximates the posterior, and we call it the vb posterior.2
Though it cannot capture posterior dependence across latent variables, it has hope to capture
each of their marginals. In particular, this paper is about the theoretical properties of the vb
posterior q∗ (θ ), the vb posterior of θ . We will also focus on the corresponding expectation of the
global variable, i.e., an estimate of the parameter. It is
Z
∗
θ̂n := θ · q∗ (θ )dθ .
We call θ ∗ the variational Bayes estimate ( vbe).
vb methods are fast and yield good predictive performance in empirical experiments (Blei et al.,
2016). However, there are few rigorous theoretical results. In this paper, we prove that (1) the vb
posterior converges in total variation (tv) distance to the kl minimizer of a normal distribution
centered at the truth and (2) the vbe is consistent and asymptotically normal.
These theorems are frequentist in the sense that we assume the data come from p( z, x ; θ0 )
with a true (nonrandom) θ0 . We then study properties of the corresponding posterior distribution p(θ | x), when approximating it with variational inference. What this work shows
is that the vb posterior is consistent even though the mean field approximating family can
be a brutal approximation. In this sense, vb is a theoretically sound approximate inference
procedure.
1.1
Main ideas
We describe the results of the paper. Along the way, we will need to define some terms: the
variational frequentist estimate (vfe), the variational log likelihood, the vb posterior, the vbe,
and the vb ideal. Our results center around the vb posterior and the vbe. (Table 1 contains a
glossary of terms.)
The variational frequentist estimate (vfe) and the variational log likelihood. The
first idea that we define is the variational frequentist estimate (vfe). It is a point estimate of θ
that maximizes a local variational objective with respect to an optimal variational distribution
of the local variables. (The vfe treats the variable θ as a parameter rather than a random
variable.) We call the objective the variational log likelihood,
M n (θ ; x) = max E q( z) [log p( x, z | θ ) − log q( z)] .
q( z)
(4)
In this objective, the optimal variational distribution q† ( z) solves the local variational inference
problem,
q† ( z) = arg min KL( q( z) || p( z | x, θ )).
q
(5)
Note that q† ( z) implicitly depends on both the data x and the parameter θ .
2 For simplicity we will write q(θ , z) = Q d q(θ ) Q n q( z ), omitting the subscript on the factors q(·). The underi
j
i =1
j =1
standing is that the factor is indicated by its argument.
3
With the objective defined, the vfe is
θ̂n = arg max M n (θ ; x).
θ
(6)
It is usually calculated with variational expectation maximization (em) (Wainwright and Jordan,
2008; Ormerod and Wand, 2010), which iterates between the E step of Equation (5) and the
M step of Equation (6). Recent research has explored the theoretical properties of the vfe for
stochastic block models (Bickel et al., 2013), generalized linear mixed models (Hall et al., 2011b),
and Gaussian mixture models (Westling and McCormick, 2015).
We make two remarks. First, the maximizing variational distribution q† ( z) of Equation (5)
is different from q∗ ( z) in the vb posterior: q† ( z) is implicitly a function of individual values
of θ , while q∗ ( z) is implicitly a function of the variational distributions q(θ ). Second, the
variational log likelihood in Equation (4) is similar to the original objective function for the em
algorithm (Dempster et al., 1977). The difference is that the em objective is an expectation with
respect to the exact conditional p( z | x), whereas the variational log likelihood uses a variational
distribution q( z).
Variational Bayes and ideal variational Bayes. While earlier applications of variational
inference appealed to variational em and the vfe, most modern applications do not. Rather
they use vb, as we described above, where there is a prior on θ and we approximate its posterior
with a global variational distribution q(θ ). One advantage of vb is that it provides regularization through the prior. Another is that it requires only one type of optimization: the same
considerations around updating the local variational factors q( z) are also at play when updating
the global factor q(θ ).
To develop theoretical properties of vb, we connect the vb posterior to the variational log
likelihood; this is a stepping stone to the final analysis. In particular, we define the vb ideal
posterior π∗ (θ | x),
π∗ (θ | x) = R
p(θ ) exp{ M n (θ ; x)}
.
p(θ ) exp{ M n (θ ; x)}dθ
(7)
Here the local latent variables z are constrained under the variational family but the global
latent variables θ are not. Note that because it depends on the variational log likelihood M n (θ ; x),
this distribution implicitly contains an optimal variational distribution q† ( z) for each value of θ ;
see Equations (4) and (5).
Loosely, the vb ideal lies between the exact posterior p(θ | x) and a variational approximation
q(θ ). It recovers the exact posterior when p( z | θ , x) degenerates to a point mass and q† ( z) is
always equal to p( z | θ , x); in that case the variational likelihood is equal to the log likelihood and
Equation (7) is the posterior. But q† ( z) is usually an approximation to the conditional. Thus the
vb ideal usually falls short of the exact posterior.
That said, the vb ideal is more complex that a simple parametric variational factor q(θ ). The
reason is that its value for each θ is defined by the optimization within M n (θ ; x). Such a
distribution will usually lie outside the distributions attainable with a simple family.
In this work, we first establish the theoretical properties of the vb ideal. We then connect it to
the vb posterior.
Variational Bernstein–Von Mises. We have set up the main concepts. We now describe the
main results.
4
Name
Definition
Variational log likelihood
M n ( x ; θ ) := sup q( z)∈Q n
Variational frequentist estimate (vfe)
θ̂n := arg maxθ M n ( x ; θ )
vb ideal
π∗ (θ | x) :=
Evidence Lower Bound (elbo)
elbo( q(θ, z)) :=
vb posterior
q∗ (θ ) := arg max q(θ)∈Q d sup q( z)∈Q n elbo( q(θ , z))
vb estimate (vbe)
R
θ̂n∗ := θ · q∗ (θ )dθ
R
q( z) log
p( z,x | θ )
q( z) d z
R p(θ ) exp{ M n ( x ; θ )}
p(θ ) exp{ M n ( x ; θ )}dθ
RR
p( z,x,θ )
q(θ ) q( z) log q(θ) q( z) dθ d z
Table 1: Glossary of terms
Suppose the data come from a true (finite-dimensional) parameter θ0 . The classical Bernstein–
Von Mises theorem says that, under certain conditions, the exact posterior p(θ | x) approaches a
normal distribution, independent of the prior, as the number of observations tends to infinity.
In this paper, we extend the theory around Bernstein–Von Mises to the variational posterior.
Here we summarize our results.
• Lemma 1 shows that the vb ideal π∗ (θ | x) is consistent and converges to a normal distribution
around the vfe. If the vfe is consistent, the vb ideal π∗ (θ | x) converges to a normal
distribution whose mean parameter is a random vector centered at the true parameter. (Note
the randomness in the mean parameter is due to the randomness in the observations x.)
• We next consider the point in the variational family that is closest to the vb ideal π∗ (θ | x)
in kl divergence. Lemma 2 and Lemma 3 show that this kl minimizer is consistent and
converges to the kl minimizer of a normal distribution around the vfe. If the vfe is
consistent (Bickel et al., 2013; Hall et al., 2011b) then the kl minimizer converges to the kl
minimizer of a normal distribution with a random mean centered at the true parameter.
• Lemma 4 shows that the vb posterior q∗ (θ ) enjoys the same asymptotic properties as the kl
minimizers of the vb ideal π∗ (θ | x).
• Theorem 5 is the variational Bernstein–Von Mises theorem. It shows that the vb posterior
q∗ (θ ) is asymptotically normal around the vfe. Again, if the vfe is consistent then the vb
posterior converges to a normal with a random mean centered at the true parameter. Further,
Theorem 6 shows that the vbe θ̂n∗ is consistent with the true parameter and asymptotically
normal.
• Finally, we prove two corollaries. First, if we use a full rank Gaussian variational family
then the corresponding vb posterior recovers the true mean and covariance. Second, if we
use a mean-field Gaussian variational family then the vb posterior recovers the true mean
and the marginal variance, but not the off-diagonal terms. The mean-field vb posterior is
underdispersed.
5
Related work. This work draws on two themes. The first is the body of work on asymptotic
properties of variational inference. You et al. (2014) and Ormerod et al. (2014) study variational
Bayes for a classical Bayesian linear model. They use normal priors and spike-and-slab priors
on the coefficients, respectively. Wang and Titterington (2005) and Wang et al. (2006) analyze
variational Bayes in Bayesian mixture models with conjugate priors.
On the frequentist side, Hall et al. (2011a,b) establishes consistency of Gaussian variational
em estimates in a Poisson mixed-effects model with a single predictor and a grouped random
intercept. Celisse et al. (2012) and Bickel et al. (2013) proved asymptotic normality of parameter
estimates in the sbm under a mean field variational approximation. Westling and McCormick
(2015) study the consistency of variational em estimates in mixture models through a connection
to M-estimation.
However, all these treatments of variational methods—either in a Bayesian or frequentist
setting—are constrained to specific models and priors. Our work broadens this work by considering more general models. Moreover, the frequentist work focuses on estimation procedures under
a variational approximation. We expand on this work by proving a variational Bernstein–Von
Mises theorem, leveraging the frequentist results to analyze vb posteriors.
The second theme is the Bernstein–Von Mises theorem. The classical (parametric) Bernstein–
p
Von Mises theorem roughly says that the posterior distribution of n(θ − θ0 ) “converges”, under
the true parameter value θ0 , to N ( X , 1/ I (θ0 )), where X ∼ N (0, 1/ I (θ0 )) and I (θ0 ) is the Fisher
information (Ghosh and Ramamoorthi, 2003; Van der Vaart, 2000; Le Cam, 1953; Le Cam
and Yang, 2012). Early forms of this theorem date back to Laplace, Bernstein, and Von Mises
(Laplace, 1809; Bernstein, 1917; Von Mises, 1931). A version also appears in Lehmann and
Casella (2006). Kleijn et al. (2012) established the Bernstein–Von Mises theorem under model
misspecification. Recent advances include extensions to semiparametric and nonparametric
cases (Bickel et al., 2012; Castillo et al., 2014). In particular, Lu et al. (2016) proved a BernsteinVon Mises type result for Bayesian inverse problems, characterizing Gaussian approximations
of probability measures with respect to the kl divergence. Below, we borrow proof ideas from
Lu et al. (2016). But we move beyond the Gaussian approximation to establish frequentist
consistency of variational Bayes.
This paper. The rest of the paper is organized as follows. Section 2 characterizes theoretical
properties of the vb ideal. Section 3 contains the central results of the paper. It first connects
the vb ideal and the vb posterior. It then proves the variational Bernstein–Von Mises theorem,
which characterizes the asymptotic properties of the vb posterior and vb estimate. Section 4
studies three models under this theoretical lens, illustrating how to establish consistency and
asymptotic normality of specific vb estimates. Finally, Section 5 reports simulation studies to
illustrate these theoretical results.
2
The vb ideal
To study the vb posterior q∗ (θ ), we first study the vb ideal of Equation (7). In the next section
we connect it to the vb posterior.
Recall the vb ideal is
π∗ (θ | x) = R
p(θ ) exp( M n (θ ; x))
,
p(θ ) exp( M n (θ ; x)) dθ
6
where M n (θ ; x) is the variational log likelihood of Equation (4). If we embed the variational log
likelihood M n (θ ; x) in a statistical model of x, this model has likelihood
`(θ ; x) ∝ exp( M n (θ ; x)).
We call it the frequentist variational model. The vb ideal π∗ (θ | x) is thus the classical posterior
under the frequentist variational model `(θ ; x); the vfe is the classical maximum likelihood
estimate (mle).
Consider the results around frequentist estimation of θ under variational approximations of the
local variables z (Bickel et al., 2013; Hall et al., 2011b; Westling and McCormick, 2015). These
works consider asymptotic properties of estimators that maximize M n (θ ; x) with respect to θ .
We will first leverage these results to prove properties of the vb ideal and their kl minimizers
in the mean field variational family Q d . Then we will use these properties to study the vb
posterior, which is what is estimated in practice.
This section relies on the consistent testability and the local asymptotic normality (lan) of
M n (θ ; x) (defined later) to show the vb ideal is consistent and asymptotically normal. We will
then show that its kl minimizer in the mean field family is also consistent and converges to the
kl minimizer of a normal distribution in tv distance.
These results are not surprising. Suppose the variational log likelihood behaves similarly to the
true log likelihood, i.e., they produce consistent parameter estimates. Then, in the spirit of the
classical Bernstein–Von Mises theorem under model misspecification (Kleijn et al., 2012), we
expect the vb ideal to be consistent as well. Moreover, the approximation through a factorizable
variational family should not ruin this consistency— point masses are factorizable and thus the
limiting distribution lies in the approximating family.
2.1
The vb ideal
The lemma statements and proofs adapt ideas from Ghosh and Ramamoorthi (2003); Van der
Vaart (2000); Bickel and Yahav (1967); Kleijn et al. (2012); Lu et al. (2016) to the variational
log likelihood. Let Θ be an open subset of RRd . Suppose the observations x = x1:n are a random
sample from the measure Pθ0 with density p( x, z | θ = θ0 ) d z for some fixed, nonrandom value
θ0 ∈ Θ. z = z1:n are local latent variables,
and θ = θ1:d ∈ Θ are global latent variables. We
R
assume that the density maps (θ , x) 7→ p( x, z | θ ) d z of the true model and (θ , x) 7→ `(θ ; x) of the
variational frequentist models are measurable. For simplicity, we also assume that for each n
there exists a single measure that dominates all measures with densities `(θ ; x), θ ∈ Θ as well
as the true measure Pθ0 .
Assumption 1. We assume the following conditions for the rest of the paper:
1. (Prior mass) The prior measure with Lebesgue-density p(θ ) on Θ is continuous and positive
2
on a neighborhood of θ0 . There exists a constant M p > 0 such that |(log p(θ ))00 | ≤ M p e|θ| .
2. (Consistent testability) For every ² > 0 there exists a sequence of tests φn such that
Z
φn ( x) p( x, z | θ0 )d zd x → 0
7
and
Z
sup
θ :||θ −θ0 ||≥²
(1 − φn ( x))
` (θ ; x )
` (θ 0 ; x )
p( x, z | θ0 )d zd x → 0,
3. (Local asymptotic normality (lan)) For every compact set K ⊂ Rd , there exist random
vectors ∆n,θ0 bounded in probability and nonsingular matrices Vθ0 such that
P θ0
1
sup | M n (θ + δn h ; x) − M n (θ ; x) − h> Vθ0 ∆n,θ0 + h> Vθ0 h| → 0,
2
h∈ K
where δn is a d × d diagonal matrix. We have δn → 0 as n → ∞. For d = 1, we commonly
p
have δn = 1/ n.
These three assumptions are standard for Bernstein–Von Mises theorem. The first assumption
is a prior mass assumption. It says the prior on θ puts enough mass to sufficiently small balls
around θ0 . This allows for optimal rates of convergence of the posterior. The first assumption
further bounds the second derivative of the log prior density. This is a mild technical assumption
satisfied by most non-heavy-tailed distributions.
The second assumption is a consistent testability assumption. It says there exists a sequence of
uniformly consistent (under Pθ0 ) tests for testing H0 : θ = θ0 against H1 : ||θ − θ0 || ≥ ² for every
² > 0 based on the frequentist variational model. This is a weak assumption. For example, it
suffices to have a compact Θ and continuous and identifiable M n (θ ; x). It is also true when there
exists a consistent estimator T n of θ . In this case, we can set φn := 1{T n − θ ≥ ²/2}.
The last assumption is a local asymptotic normality assumption on M n (θ ; x) around the true
value θ0 . It says the frequentist variational model can be asymptotically approximated by a
1
normal location model centered at θ0 after a rescaling of δ−
δn
n . This normalizing sequence
p
determines the optimal rates of convergence of the posterior. For example, if δn = 1/ n, then we
p
commonly have θ − θ0 = O p (1/ n).
In the spirit of the last assumption, we perform a change-of-variable step:
1
θ̃ = δ−
n (θ − θ0 ).
(8)
1
We center θ at the true value θ0 and rescale it by the reciprocal of the rate of convergence δ−
n .
This ensures that the asymptotic distribution of θ̃ is not degenerate, i.e., it does not converge to
a point mass. We define π∗ (· | x) as the density of θ̃ when θ has density π∗ (· | x):
θ̃
π∗θ̃ (θ̃ | x) = π∗ (θ0 + δn θ̃ | x) · | det(δn )|.
Now we characterize the asymptotic properties of the vb ideal.
Lemma 1. The vb ideal converges in total variation to a sequence of normal distributions,
P θ0
||π∗θ̃ (· | x) − N (·; ∆n,θ0 , Vθ−01 )||tv → 0.
Proof sketch of lemma 1. This is a consequence of the classical finite-dimensional Bernstein–Von
Mises theorem under model misspecification (Kleijn et al., 2012). Theorem 2.1 of Kleijn et al.
8
(2012) roughly says that the posterior is consistent if the model is locally asymptotically normal
around
the true parameter value θ0 . Here the true data generating measure is Pθ0 with density
R
p( x, z | θ = θ0 ) d z, while the frequentist variational model has densities `(θ ; x), θ ∈ Θ.
What we need to show is that the consistent testability assumption in Assumption 1 implies
assumption (2.3) in Kleijn et al. (2012):
P θ0
Z
|θ̃ |> M n
π∗θ̃ (θ̃ | x) dθ̃ → 0
for every sequence of constants M n → ∞.. To show this, we mimic the argument of Theorem
3.1 of Kleijn et al. (2012), where they show this implication for the iid case with a common
convergence rate for all dimensions of θ . See Appendix A for details.
1
This lemma says the vb ideal of the rescaled θ , θ̃ = δ−
n (θ − θ0 ), is asymptotically normal with
mean ∆n,θ0 . The mean, ∆n,θ0 , as assumed in Assumption 1, is a random vector bounded in probability. The asymptotic distribution N (·; ∆n,θ0 , Vθ−1 ) is thus also random, where randomness is
0
due to the data x being random draws from the true data generating measure Pθ0 . We notice that
1
if the vfe, θ̂n , is consistent and asymptotically normal, we commonly have ∆n,θ0 = δ−
n (θ̂ n − θ0 )
with E(∆n,θ0 ) = 0. Hence, the vb ideal will converge to a normal distribution with a random
mean centered at the true value θ0 .
2.2
The KL minimizer of the vb ideal
Next we study the kl minimizer of the vb ideal in the mean field variational family. We show
its consistency and asymptotic normality. To be clear, the asymptotic normality is in the sense
that the kl minimizer of the vb ideal converges to the kl minimizer of a normal distribution in
tv distance.
Lemma 2. The kl minimizer of the vb ideal over the mean field family is consistent: almost
surely under Pθ0 , it converges to a point mass,
arg min kl( q(θ )||π∗ (θ | x)) → δθ0 .
d
q(θ )∈Q d
Proof sketch of lemma 2. The key insight here is that point masses are factorizable. Lemma 1
above suggests that the vb ideal converges in distribution to a point mass. We thus have its
kl minimizer also converging to a point mass, because point masses reside within the mean
field family. In other words, there is no loss, in the limit, incurred by positing a factorizable
variational family for approximation.
To prove this lemma, we bound the mass of B c (θ0 , η n ) under q(θ ), where B c (θ0 , η n ) is the complement of an η n -sized ball centered at θ0 with η n → 0 as n → ∞. In this step, we borrow ideas
from the proof of Lemma 3.6 and Lemma 3.7 in Lu et al. (2016). See Appendix B for details.
Lemma 3. The kl minimizer of the vb ideal of θ̃ converges to that of N (· ; ∆n,θ0 , Vθ−1 ) in total
0
variation: under mild technical conditions on the tail behavior of Q d (see Assumption 2 in
9
Appendix C),
°
°
°
°
°
−1 °
∗
°arg min kl( q(·)||πθ̃ (· | x)) − arg min kl( q(·)||N (· ; ∆n,θ0 , Vθ0 ))°
°
° q∈Q d
q∈Q d
P θ0
→ 0.
tv
Proof sketch of lemma 3. The intuition here is that, if the two distribution are close in the
limit, their kl minimizers should also be close in the limit. Lemma 1 says that the vb ideal
of θ̃ converges to N (·; ∆n,θ0 , Vθ−1 ) in total variation. We would expect their kl minimizer also
0
converges in some metric. This result is also true for the (full-rank) Gaussian variational family
if rescaled appropriately.
Here we show their convergence in total variation. This is achieved by showing the Γ-convergence
of the functionals of q: kl( q(·)||π∗ (· | x)) to kl( q(·)||N (· ; ∆n,θ0 , Vθ−1 )), for parametric q’s. Γ0
θ̃
convergence is a classical tool for characterizing variational problems; Γ-convergence of functionals ensures convergence of their minimizers (Dal Maso, 2012; Braides, 2006). See Appendix C
for proof details and a review of Γ-convergence.
We characterized the limiting properties of the vb ideal and their kl minimizers. We will next
show that the vb posterior is close to the kl divergence minimizer of the vb ideal. Section 3
culminates in the main theorem of this paper – the variational Bernstein–Von Mises theorem –
showing the vb posterior share consistency and asymptotic normality with the kl divergence
minimizer of vb ideal.
3
Frequentist consistency of variational Bayes
We now study the vb posterior. In the previous section, we proved theoretical properties for the
vb ideal and its kl minimizer in the variational family. Here we first connect the vb ideal to the
vb posterior, the quantity that is used in practice. We then use this connection to understand
the theoretical properties of the vb posterior.
We begin by characterizing the optimal variational distribution in a useful way. Decompose the
variational family as
q(θ , z) = q(θ ) q( z),
Q
Q
where q(θ ) = di=1 q(θ i ) and q( z) = ni=1 q( z i ). Denote the prior p(θ ). Note d does not grow with
the size of the data. We will develop a theory around vb that considers asymptotic properties of
the vb posterior q∗ (θ ).
We decompose the elbo of Equation (3) into the portion associated with the global variable and
the portion associated with the local variables,
Z Z
p(θ , z, x)
elbo( q(θ) q( z)) =
q(θ ) q( z) log
dθ d z
q(θ ) q( z)
Z Z
p(θ ) p( z, x | θ )
=
q(θ ) q( z) log
dθ d z
q (θ ) q ( z )
Z
Z
Z
p(θ )
p( z, x | θ )
= q(θ ) log
dθ + q(θ ) q( z) log
dθ d z.
q(θ )
q( z)
10
The optimal variational factor for the global variables, i.e., the vb posterior, maximizes the
elbo. From the decomposition, we can write it as a function of the optimized local variational
factor,
µ ·
½Z
¾¸
¶
Z
p( z, x | θ )
q∗ (θ ) = arg max sup q(θ ) log p(θ ) exp
q( z) log
d z − log q(θ ) dθ .
(9)
q( z)
q (θ )
q( z)
One way to see the objective for the vb posterior is as the elbo profiled over q( z), i.e., where
the optimal q( z) is a function of q(θ ) (Hoffman et al., 2013). With this perspective, the elbo
becomes a function of q(θ ) only. We denote it as a functional elbo p (·):
µ ·
½Z
¾¸
¶
Z
p( z, x | θ )
elbo p ( q(θ)) := sup q(θ) log p(θ) exp q( z) log
d z − log q(θ ) dθ .
(10)
q( z)
q( z)
We then rewrite Equation (9) as q∗ (θ ) = arg max q(θ) elbo p ( q(θ )). This expression for the vb
posterior is key to our results.
3.1
kl minimizers of the vb ideal
Recall that the kl minimization objective to the ideal vb posterior is the functional kl(·||π∗ (θ | x)).
We first show that the two optimization objectives kl(·||π∗ (θ | x)) and elbo p (·) are close in the
limit. Given the continuity of both kl(·||π∗ (θ | x)) and elbo p (·), this implies the asymptotic
properties of optimizers of kl(·||π∗ (θ | x)) will be shared by the optimizers of elbo p (·).
Lemma 4. The negative kl divergence to the vb ideal is equivalent to the profiled elbo in the
limit: under mild technical conditions on the tail behavior of Q d (see for example Assumption 3
in Appendix D), for q(θ ) ∈ Q d ,
elbo p ( q(θ)) = −kl( q(θ)||π∗ (θ | x)) + o P (1).
Proof sketch of Lemma 4. We first notice that
− kl( q(θ )||π∗ (θ | x))
Z
p(θ ) exp( M n (θ ; x))
= q(θ ) log
dθ
q (θ )
Ã
"
(
Z
Z
=
q(θ ) log p(θ ) exp sup
q( z) log
q( z)
(11)
(12)
p( z, x | θ )
dz
q( z)
)#
!
− log q(θ ) dθ .
(13)
Comparing Equation (13) with Equation (10), we can see that the only difference between
−kl(·||π∗ (θ | x)) and elbo p (·) is in the position of sup q( z) . elbo p (·) allows for a single choice of
optimal q( z) given q(θ ), while −kl(·||π∗ (θ | x)) allows for a different optimal q( z) for each value
of θ . In this sense, if we restrict the variational family of q(θ ) to be point masses, then elbo p (·)
and −kl(·||π∗ (θ | x)) will be the same.
The only members of the variational family of q(θ ) that admit finite limiting −kl( q(θ )||π∗ (θ | x))
are the ones that converge to point masses at rate δn , so we expect elbo p (·) and −kl(·||π∗ (θ | x))
to be close as n → ∞. We prove this by bounding the remainder in the Taylor expansion of M n (θ ; x)
by a sequence converging to zero in probability. See Appendix D for details.
11
3.2
The vb posterior
Section 2 characterizes the asymptotic behavior of the vb ideal π∗ (θ | x) and their kl minimizers.
Lemma 4 establishes the connection between the vb posterior q∗ (θ ) and the kl minimizers
of the vb ideal π∗ (θ | x). Recall arg min q(θ)∈Q d kl( q(θ )||π∗ (θ | x)) is consistent and converges
to the kl minimizer of a normal distribution. We now build on these results to study the vb
posterior q∗ (θ ).
Now we are ready to state the main theorem. It establishes the asymptotic behavior of the vb
posterior q∗ (θ ).
Theorem 5. (Variational Bernstein-von-Mises Theorem)
1. The vb posterior is consistent: almost surely under Pθ0 ,
d
q∗ (θ ) → δθ0 .
2. The vb posterior is asymptotically normal in the sense that it converges to the kl minimizer
of a normal distribution:
°
°
°
° Pθ
° ∗
°
0
(14)
° q θ̃ (·) − arg min kl( q(·)||N (· ; ∆n,θ0 , Vθ−01 ))° → 0.
°
°
d
q∈Q
tv
Here we transform q∗ (θ ) to q θ̃ (θ̃ ), which is centered around the true θ0 and scaled by the
convergence rate; see Equation (8).
Proof sketch of Theorem 5. This theorem is a direct consequence of Lemma 2, Lemma 3, Lemma 4.
We need the same mild technical conditions on Q d as in Lemma 3 and Lemma 4. See Appendix E
for details.
Given the convergence of the vb posterior, we can now establish the asymptotic properties of
the vbe.
Theorem 6. (Asymptotics of the vbe)
R
R
Assume |θ |2 π(θ ) dθ < ∞. Let θ̂n∗ = θ · q∗1 (θ ) dθ denote the vbe.
1. The vbe is consistent: under Pθ0 ,
a.s.
θ̂n∗ → θ0 .
2. The vbe is asymptotically normal in the sense that it converges in distribution to the mean
of the kl minimizer:3 if ∆n,θ0 → X for some X ,
d
d
1 ∗
δ−
n (θ̂ n − θ0 ) →
3 The randomness in the mean of the
Z
θ̃ · arg min kl( q(θ̃ )||N (θ̃ ; X , Vθ−01 )) dθ̃ .
q∈Q d
kl minimizer comes from ∆n,θ0 .
12
Proof sketch of Theorem 6. As the posterior mean is a continuous function of the posterior
distribution, we would expect the vbe is consistent given the vb posterior is. We also know
that the posterior mean is the Bayes estimator under squared loss. Thus we would expect
the vbe to converge in distribution to squared loss minimizer of the kl minimizer of the vb
ideal. The result follows from a very similar argument from Theorem 2.3 of Kleijn et al. (2012),
which shows that the posterior mean estimate is consistent and asymptotically normal under
model misspecification as a consequence of the Bernsterin–Von Mises theorem and the argmax
theorem. See Appendix E for details.
We remark that ∆n,θ0 , as in Assumption 1, is a random vector bounded in Pθ0 probability. The
randomness is due to x being a random sample generated from Pθ0 .
In cases where vfe is consistent, like in all the examples we will see in Section 4, ∆n,θ0 is a
zero mean random vector with finite variance. For particular realizations of x the value of ∆n,θ0
1
might not be zero; however, because we scale by δ−
n , this does not destroy the consistency of vb
posterior or the vbe.
3.3
Gaussian vb posteriors
We illustrate the implications of Theorem 5 and Theorem 6 on two choices of variational families:
a full rank Gaussian variational family and a factorizable Gaussian variational family. In both
cases, the vb posterior and the vbe are consistent and asymptotically normal with different
covariance matrices. The vb posterior under the factorizable family is underdispersed.
Corollary 7. Posit a full rank Gaussian variational family, that is
Q d = { q : q(θ ) = N ( m, Σ)},
(15)
with Σ positive definite. Then
d
1. q∗ (θ ) → δθ0 , almost surely under Pθ0 .
P θ0
2. || q∗ (·) − N (· ; ∆n,θ0 , Vθ−1 )||tv → 0.
θ̃
0
a.s.
3. θ̂n∗ → θ0 .
1 ∗
4. δ−
n (θ̂ n − θ0 ) − ∆ n,θ0 = o Pθ0 (1).
Proof sketch of corollary 7. This is a direct consequence of Theorem 5 and Theorem 6. We only
need to show that Lemma 3 is also true for the full rank Gaussian variational family. The last
d
d
1 ∗
conclusion implies δ−
n (θ̂ n − θ0 ) → X if ∆ n,θ0 → X for some random variable X . We defer the
proof to Appendix F.
This corollary says that under a full rank Gaussian variational family, vb is consistent and
asymptotically normal in the classical sense. It accurately recovers the asymptotic normal
distribution implied by the local asymptotic normality of M n (θ ; x).
13
Before stating the corollary for the factorizable Gaussian variational family, we first present
a lemma on the kl minimizer of a Gaussian distribution over the factorizable Gaussian family. We show that the minimizer keeps the mean but has a diagonal covariance matrix that
matches the precision. We also show the minimizer has a smaller entropy than the original distribution. This echoes the well-known phenomenon of vb algorithms underestimating the variance.
Lemma 8. The factorizable
matches the precision:
kl minimizer of a Gaussian distribution keeps the mean and
kl(N (·; µ0 , Σ0 )||N (·; µ1 , Σ1 )) = µ1 , Σ∗1 ,
arg min
µ0
∈Rd ,Σ
0 ∈diag( d × d )
1
−1
where Σ∗1 is diagonal with Σ∗1,ii = ((Σ−
for i = 1, 2, ..., d . Hence, the entropy of the factorizable
1 ) ii )
kl minimizer is smaller than or equal to that of the original distribution:
H(N (·; µ0 , Σ∗1 )) ≤ H(N (·; µ0 , Σ1 )).
Proof sketch of Lemma 8. The first statement is consequence of a technical calculation of the kl
divergence between two normal distributions. We differentiate the kl divergence over µ0 and
the diagonal terms of Σ0 and obtain the result. The second statement is due to the inequality of
the determinant of a positive matrix being always smaller than or equal to the product of its
diagonal terms (Amir-Moez and Johnston, 1969; Beckenbach and Bellman, 2012). In this sense,
mean field variational inference underestimates posterior variance. See Appendix G for details.
The next corollary studies the vb posterior and the vbe under a factorizable Gaussian variational family.
Corollary 9. Posit a factorizable Gaussian variational family,
Q d = { q : q(θ ) = N ( m, Σ, )}
(16)
where Σ positive definite and diagonal. Then
d
1. q∗ (θ ) → δθ0 , almost surely under Pθ0 .
P θ0
2. || q∗ (·) − N (· ; ∆n,θ0 , Vθ0−1 )||tv → 0,
θ̃
0
where V 0 is diagonal and has the same diagonal entries as Vθ0 .
a.s.
3. θ̂n∗ → θ0 .
1 ∗
4. δ−
n (θ̂ n − θ0 ) − ∆ n,θ0 = o Pθ0 (1).
Proof of corollary 9. This is a direct consequence of Lemma 8, Theorem 5, and Theorem 6.
This corollary says that under the factorizable Gaussian variational family, vb is consistent and
asymptotically normal in the classical sense. The rescaled asymptotic distribution for θ̃ recovers
the mean but underestimates the covariance. This underdispersion is a common phenomenon
we see in mean field variational Bayes.
14
4
Applications
We proved consistency and asymptotic normality of the variational Bayes (vb) posterior (in total
variation (tv) distance) and the variational Bayes estimate (vbe). We mainly relied on the
prior mass condition, the local asymptotic normality of the variational log likelihood M n ( x ; θ )
and the consistent testability assumption of the data generating parameter.
We now apply this argument to three types of Bayesian models: Bayesian mixture models (Bishop,
2006; Murphy, 2012), Bayesian generalized linear mixed models (McCulloch and Neuhaus, 2001;
Jiang, 2007), and Bayesian stochastic block models (Wang and Wong, 1987; Snijders and Nowicki,
1997; Mossel et al., 2012; Abbe and Sandon, 2015; Hofman and Wiggins, 2008). For each model
class, we illustrate how to leverage the known asymptotic results for frequentist variational
approximations to prove asymptotic results for vb. We assume the prior mass condition for
the rest of this section: the prior measure of a parameter θ with Lebesgue density p(θ ) on Θ is
continuous and positive on a neighborhood of the true data generating value θ0 . For simplicity,
we posit a mean field family for the local latent variables and a factorizable Gaussian variational
family for the global latent variables.
4.1
Bayesian Mixture models
The Bayesian mixture model is a versatile class of models for density estimation and clustering
(Bishop, 2006; Murphy, 2012).
Consider a Bayesian mixture of K unit-variance univariate Gaussians with means µ = {µ1 , ..., µK }.
For each observation x i , i = 1, ..., n, we first randomly draw a cluster assignment c i from a
categorical distribution over {1, ..., K }; we then draw x i randomly from a unit-variance Gaussian
with mean µ c i . The model is
µk ∼ p µ ,
k = 1, ..., K,
c i ∼ Categorical(1/K, ..., 1/K ),
x i | c i , µ ∼ N ( c>
i µ, 1).
i = 1, ..., n,
i = 1, ..., n.
For a sample of size n, the joint distribution is
p(µ, c, x) =
K
Y
i =1
p µ (µ i )
n
Y
p ( c i ) p ( x i | c i , µ ).
i =1
Here µ is a K -dimensional global latent vector and c 1:n are local latent variables. We are
interested inferring the posterior of the µ vector.
We now establish asymptotic properties of vb for Bayesian Gaussian mixture model (gmm).
R
Corollary 10. Assume the data generating measure Pµ0 has density p(µ0 , c, x) d c. Let q∗ (µ)
and µ∗ denote the vb posterior and the vbe. Under regularity conditions (A1-A5) and (B1,2,4)
of Westling and McCormick (2015), we have
°
µ
¶°
° ∗
° P µ0
Y 1
° q (µ ) − N µ 0 + p
° → 0,
,
V
(
µ
)
0 0 °
°
n n
tv
15
and
p
d
n(µ∗ − µ0 ) → N (Y , V0 (µ0 )),
where µ0 is the true value of µ that generates the data. We have
Y ∼ N (0, V (µ0 )),
V (µ0 ) = A (µ0 )−1 B(µ0 ) A (µ0 )−1 ,
A (µ) = EPµ0 [D µ2 m(µ ; x)],
B(µ) = EPµ0 [D µ m(µ ; x)D µ m(µ ; x)> ],
Z
p( x, c | µ)
m(µ ; x) = sup
q( c) log
d c.
n
q( c)
q( c)∈Q
The diagonal matrix V0 (µ0 ) satisfies (V0 (µ0 )−1 ) ii = ( A (µ0 )) ii . The specification of Gaussian
mixture model is invariant to permutation among K components; this corollary is true up to
permutations among the K components.
Proof sketch for Corollary 10. The consistent testability condition is satisfied by the existence of
a consistent estimate due to Theorem 1 of Westling and McCormick (2015). The local asymptotic
normality is proved by a Taylor expansion of m(µ ; x) at µ0 . This result then follows directly from
our Theorem 5 and Theorem 6 in Section 3. The technical conditions inherited from Westling
and McCormick (2015) allow us to use their Theorems 1 and 2 for properties around variational
frequentist estimate (vfe). See Appendix H for proof details.
4.2
Bayesian Generalized linear mixed models
Bayesian generalized linear mixed models (glmms) are a powerful class of models for analyzing
grouped data or longitudinal data (McCulloch and Neuhaus, 2001; Jiang, 2007).
Consider a Poisson mixed model with a simple linear relationship and group-specific random
intercepts. Each observation reads ( X i j , Yi j ), 1 ≤ i ≤ m, 1 ≤ j ≤ n, where the Yi j ’s are nonnegative integers and the X i j ’s are unrestricted real numbers. For each group of observations
( X i j , Yi j ), 1 ≤ j ≤ n, we first draw the random effect U i independently from N (0, σ2 ). We follow
by drawing Yi j from a Poisson distribution with mean exp(β0 + β1 X i j + U i ). The probability
model is
β0 ∼ p β0 ,
β1 ∼ p β1 ,
σ 2 ∼ p σ2 ,
iid
U i ∼ N (0, σ2 ),
Yi j | X i j ,U i ∼ Poi(exp(β0 + β1 X i j + U i )).
16
The joint distribution is
p(β0 , β1 ,σ2 ,U1:m , Y1:m,1:n | X 1:m,1:n ) =
n
m Y
m
Y
Y
Poi(Yi j ; exp(β0 + β1 X i j + U i )).
p β0 (β0 ) p β1 (β1 ) p σ2 (σ2 ) N (U i ; 0, σ2 ) ×
i =1 j =1
i =1
We establish asymptotic properties of vb in Bayesian Poisson linear mixed models.
Corollary 11. Consider the true data generating distribution Pβ0 ,β0 ,(σ2 )0 with the global latent
0
1
variables taking the true values {β00 , β01 , (σ2 )0 }. Let q∗β (β0 ), q∗β (β1 ), q∗σ2 (σ2 ) denote the
0
1
vb
posterior of β0 , β1 , σ2 . Similarly, let β∗0 , β∗1 , (σ2 )∗ be the vbes accordingly. Consider m = O ( n2 ).
Under regularity conditions (A1-A5) of Hall et al. (2011b), we have
°
µ
¶°
° ∗
°
Z1 Z2 Z3
° q (β0 ) q∗ (β1 ) q∗ 2 (σ2 ) − N (β0 , β0 , (σ2 )0 ) + ( p
°
,
,
)
,
diag
(
V
,
V
,
V
)
p
p
1
2
3
° β0
°
0 1
β1
σ
n
mn
n
tv
where
Pβ0 ,β0 ,(σ2 )0
0
1
→
0,
Z1 ∼ N (0, (σ2 )0 ), Z2 ∼ N (0, τ2 ), Z3 ∼ N (0, 2{(σ2 )0 }2 ),
1
V1 = exp(−β0 + (σ2 )0 )/φ(β01 ),
2
1
V2 = exp(−β00 + σ2 )/φ00 (β01 ),
2
V3 = 2{(σ2 )0 }2 ,
τ2 =
exp{−(σ2 )0 /2 − β00 }φ(β01 )
φ00 (β01 )φ(β01 ) − φ0 (β01 )2
.
Here φ(·) is the moment generating function of X .
Also,
p
p
p
d
( m(β∗0 − β00 ), mn(β∗1 − β01 ), m((σ2 )∗ − (σ2 )0 )) → N (( Z1 , Z2 , Z3 ), diag(V1 , V2 , V3 )).
Proof sketch for Corollary 11. The consistent testability assumption is satisfied by the existence
of consistent estimates of the global latent variables shown in Theorem 3.1 of Hall et al. (2011b).
The local asymptotic normality is proved by a Taylor expansion of the variational log likelihood
based on estimates of the variational parameters based on equations (5.18) and (5.22) of Hall
et al. (2011b). The technical conditions inherited from Hall et al. (2011b) allow us to leverage
their Theorem 3.1 for properties of the vfe. The result then follows directly from Theorem 5
and Theorem 6 in Section 3. See Appendix I for proof details.
17
4.3
Bayesian stochastic block models
Stochastic block models are an important methodology for community detection in network
data (Wang and Wong, 1987; Snijders and Nowicki, 1997; Mossel et al., 2012; Abbe and Sandon,
2015).
Consider n vertices in a graph. We observe pairwise linkage between nodes A i j ∈ {0, 1}, 1 ≤ i, j ≤
n. In a stochastic block model, this adjacency matrix is driven by the following process: first
assign each node i to one of the K latent classes by a categorical distribution with parameter
π. Denote the class membership as Z i ∈ {1, ..., K }. Then draw A i j ∼ Bernoulli( H Z i ,Z j ). The
parameter H is a symmetric matrix in [0, 1]K ×K that specifies the edge probabilities between
two latent classes; the parameter π are the proportions of the latent classes. The Bayesian
stochastic block model is
π ∼ p(π),
H ∼ p( H ),
iid
Z i | π ∼ Categorical(π),
iid
A i j | Z i , Z j , H ∼ Bernoulli( H Z i Z j ).
The dependence in stochastic block model is more complicated than the Bayesian gmm or the
Bayesian glmm.
Before establishing the result, we reparameterize (π, H ) by θ = (ω, ν), where ω ∈ RK −1 is the
log odds ratio of belonging to classes 1, ..., K − 1, and ν ∈ RK ×K is the log odds ratio of an edge
existing between all pairs of the K classes. The reparameterization is
ω(a) = log
ν(a, b) = log
1−
π( a )
PK −1
b=1
π( b )
a = 1, ..., K − 1
,
H (a, b)
,
1 − H (a, b)
a, b = 1, ..., K.
The joint distribution is
p(θ ,Z, A ) =
KY
−1
KX
−1
a=1
a=1
[ eω(a)n a (1 +
eω(a) )−n ] ×
K Y
K
Y
[ eν(a,b)O ab (1 + eν(a,b) )n ab ]1/2 ,
a=1 b=1
where
na (Z) =
n ab ( Z ) =
n
X
1{ Z i = a},
i =1
n X
n
X
1{ Z i = a, Z j = b},
i =1 j 6= i
O ab ( A, Z ) =
n X
n
X
1{ Z i = a, Z j = b} A i j .
i =1 j 6= i
We now establish the asymptotic properties of vb for stochastic block models.
18
Corollary 12. Consider ν0 , ω0 as true data generating parameters. Let q∗ν (ν), q∗ω (ω) denote the
vb posterior of ν and ω. Similarly, let ν∗ , ω∗ be the vbe. Then
°
Ã
!°
1
−1
°
° P ν ,ω
Σ−
° ∗
°
0 0
1 Y1 Σ2 Y2
∗
, p ), Vn (ν0 , ω0 ) °
→ 0
° q ν (ν) q ω (ω) − N (ν, ω); (ν0 , ω0 ) + ( p
°
°
n
nλ0
tv
where λ0 = EPν0 ,ω0 (degree of each node), (log n)−1 λ0 → ∞. Y1 and Y2 are two zero mean random
vectors with covariance matrices Σ1 and Σ2 , where Σ1 , Σ2 are known functions of ν0 , ω0 . The
diagonal matrix V (ν0 , ω0 ) satisfies V −1 (ν0 , ω0 ) ii = diag(Σ1 , Σ2 ) ii . Also,
p
p
d
1
−1
( nλ0 (ν∗ − ν0 ), n(ω∗ − ω0 )) → N ((Σ−
1 Y1 , Σ2 Y2 ), V (ν0 , ω0 )),
The specification of classes in stochastic block model ( sbm) is permutation invariant. So the
convergence above is true up to permutation with the K classes. We follow Bickel et al. (2013) to
consider the quotient space of (ν, ω) over permutations.
Proof of Corollary 12. The consistent testability assumption is satisfied by the existence of
consistent estimates by Lemma 1 of Bickel et al. (2013). The local asymptotic normality,
p( A, z | ν0 + p
Z
sup
q( z) log
q( z)
q( z)∈Q K
Z
= sup
q( z)∈Q K
t
, ω + psn )
n2 ρ n
q( z) log
dz
1
1
p( A, z | ν0 , ω0 )
d z + s> Y1 + t> Y2 − s> Σ1 s − t> Σ2 t + o P (1),
q( z)
2
2
(17)
for (ν0 , ω0 ) ∈ T for compact T with ρ n = n1 E(degree of each node), is established by Lemma 2
of Bickel et al. (2013). The result then follows directly from our Theorem 5 and Theorem 6 in
Section 3.
5
Simulation studies
We illustrate the implication of Theorem 5 and Theorem 6 by simulation studies on Bayesian
glmm (McCullagh, 1984). We also study the vb posteriors of latent Dirichlet allocation (lda)
(Blei et al., 2003). This is a model that shares similar structural properties with sbm but has
no consistency results established for its vfe.
We use two automated inference algorithms offered in Stan, a probabilistic programming system
(Carpenter et al., 2015): vb through automatic differentiation variational inference (advi)
(Kucukelbir et al., 2016) and Hamiltonian Monte Carlo (hmc) simulation through No-U-Turn
sampler (nuts) (Hoffman and Gelman, 2014). We note that optimization algorithms used for
vb in practice only find local optima.
In both cases, we observe the vb posteriors get closer to the truth as the sample size increases;
when the sample size is large enough, they coincide with the truth. They are underdispersed,
however, compared with hmc methods.
19
5.1
Bayesian Generalized Linear Mixed Models
We consider the Poisson linear mixed model studied in Section 4. Fix the group size as n = 10. We
simulate data sets of size N = (50, 100, 200, 500, 1000, 2000, 5000, 10000, 20000). As the size of
the data set grows, the number of groups also grows; so does the number of local latent variables
U i , 1 ≤ i ≤ m. We generate a four-dimensional covariate vector for each X i j , 1 ≤ i ≤ m, 1 ≤ j ≤ n,
where the first dimension follows i.i.d N (0, 1), the second dimension follows i.i.d N (0, 25), the
third dimension follows i.i.d Bernoulli(0.4), and the fourth dimension follows i.i.d. Bernoulli(0.8).
We wish to study the behaviors of coefficient efficients for underdispersed/overdispersed continuous covariates and balanced/imbalanced binary covariates. We set the true parameters as
β0 = 5, β1 = (0.2, −0.2, 2, −2), and σ2 = 2.
Figure 1 shows the boxplots of vb posteriors for β0 , β1 , and σ2 . All vb posteriors converge to
their corresponding true values as the size of the data set increases. The box plots present
rather few outliers; the lower fence, the box, and the upper fence are about the same size. This
suggests normal vb posteriors. This echoes the consistency and asymptotic normality concluded
from Theorem 5. The vb posteriors are underdispersed, compared to the posteriors via hmc.
This also echoes our conclusion of underdispersion in Theorem 5 and Lemma 8.
Regarding the convergence rate, vb posteriors of all dimensions of β1 quickly converge to
their true value; the vb posteriors center around their true values as long as N ≥ 1000. The
convergence of vb posteriors of slopes for continuous variables (β11 , β12 ) are generally faster
than those for binary ones (β13 , β14 ). The vb posterior of σ2 shares a similarly fast convergence
rate. The vb posterior of the intercept β0 , however, struggles; it is away from the true value until
the data set size hits N = 20000. This aligns with the convergence rate inferred in Corollary 11,
p
p
mn for β1 and m for β0 and σ2 .
Computation wise, vb takes orders of magnitude less time than hmc. The performance of vb
posteriors is comparable with that from hmc when the sample size is sufficiently large; in this
case, we need N = 20000.
5.2
Latent Dirichlet Allocation
Latent Dirichlet Allocation (lda) is a generative statistical model commonly adopted to describe
word distributions in documents by latent topics.
Given M documents, each with N m , m = 1, ..., M words, composing a vocabulary of V words, we
assume K latent topics. Consider two sets of latent variables: topic distributions for document
m, (θm )K ×1 , m = 1, ..., M and word distributions for topic k, (φk )V ×1 , k = 1, ..., K . The generative
process is
θm ∼ pθ ,
m = 1, ..., M,
φk ∼ p φ ,
k = 1, ..., K,
z m, j ∼Mult(θm ),
j = 1, ..., Nm , m = 1, ..., M,
wm, j ∼Mult(φ zm, j ),
j = 1, ..., Nm , m = 1, ..., M.
The first two rows are assigning priors assigned to the latent variables. wm, j denotes word j of
document m and z m, j denotes its assigned topic.
20
algo
algo
MFVB
HMC
3
MFVB
HMC
6
2
5
1
4
0
2
0
1
3
2
2
3
4
1
50
100
200
500
1000
2000
5000
10000
50
20000
100
200
500
2000
5000
10000
20000
(b) Posterior of σ2
(a) Posterior of β0
0.6
0.12
algo
MFVB
HMC
0.5
algo
MFVB
HMC
0.14
0.4
0.16
0.3
0.18
0.2
12
11
1000
N
N
0.1
0.20
0.22
0.0
0.24
0.1
0.26
0.2
50
100
200
500
1000
2000
5000
10000
20000
50
100
200
500
N
1000
2000
5000
10000
20000
N
(c) Posterior of β11
(d) Posterior of β12
2.8
algo
algo
MFVB
HMC
2.6
MFVB
HMC
1.4
2.2
1.8
14
1.6
13
2.4
2.0
2.0
1.8
2.2
1.6
2.4
1.4
50
100
200
500
1000
2000
5000
10000
20000
50
N
100
200
500
1000
2000
5000
10000
20000
N
(e) Posterior of β13
(f) Posterior of β14
Figure 1: vb posteriors and hmc posteriors of Poisson Generalized Linear Mixed Model versus
size of datasets. vb posteriors are consistent and asymptotically normal but underdispersed
than hmc posteriors. β0 and σ2 converge to the truth slower than β1 does. They echo our
conclusions in Theorem 5 and Corollary 11.
21
1.0
1.6
1.4
0.8
KL
KL
1.2
0.6
1.0
0.8
0.4
0.6
0.4
0.2
algo
0.2
10
20
50
100
200
500
1000
M
HMC
MFVB
10
20
50
100
200
500
1000
M
(a) Posterior mean kl divergence of the
K = 10 topics
(b) Boxplots of posterior kl divergence of Topic 2
Figure 2: Mean of kl divergence between the true topics and the fitted vb and hmc posterior
topics versus size of datasets. (a) vb posteriors (dark blue) converge to the truth; they are
very close to the truth as we hit M = 1000 documents. (b) vb posteriors are consistent but
underdispersed compared to hmc posteriors (light blue). These align with our conclusions in
Theorem 5.
We simulate a data set with V = 100 sized vocabulary and K = 10 latent topics in M = (10, 20,
iid
50, 100, 200, 500, 100) documents. Each document has N m words where N m ∼ Poi(100). As
the number of documents M grows, the number of document-specific topic vectors θm grows
while the number of topic-specific word vectors φk stays the same. In this sense, we consider
θm , m = 1, ..., M as local latent variables and φk , k = 1, ..., K as global latent variables. We are
interested in the vb posteriors of global latent variables φk , k = 1, ..., K here. We generate the
iid
data sets with true values of θ and φ, where they are random draws from θm ∼ Dir((1/K )K ×1 )
iid
and φk ∼ Dir((1/V )V ×1 ).
Figure 2 presents the Kullback-Leibler (kl) divergence between the K = 10 topic-specific word
distributions induced by the true φk ’s and the fitted φk ’s by vb and hmc. This kl divergence
P
equals to kl(Mult(φ0k )||Mult(φ̂k )) = Vi=1 φ0ki (log φ0ki − log φ̂ki ), where φ0ki is the i th entry of the
true kth topic and φ̂ki is the i th entry of the fitted kth topic.
Figure 2a shows that vb posterior (dark blue) mean kl divergences of all K = 10 topics get closer
to 0 as the number of documents M increase, faster than hmc (light blue). We become very
close to the truth as the number of documents M hits 1000. Figure 2b4 shows that the boxplots
of vb posterior mean kl divergences get closer to 0 as M increases. They are underdispersed
compared to hmc posteriors. These align with our understanding of how vb posterior behaves
in Theorem 5.
Computation wise, again vb is orders of magnitude faster than hmc. In particular, optimization
in vb in our simulation studies converges within 10,000 steps.
4 We only show boxplots for Topic 2 here. The boxplots of other topics look very similar.
22
6
Discussion
Variational Bayes (vb) methods are a fast alternative to Markov chain Monte Carlo (mcmc)
for posterior inference in Bayesian modeling. However, few theoretical guarantees have been
established. This work proves consistency and asymptotic normality for variational Bayes (vb)
posteriors. The convergence is in the sense of total variation (tv) distance converging to zero
in probability. In addition, we establish consistency and asymptotic normality of variational
Bayes estimate (vbe). The result is frequentist in the sense that we assume a data generating
distribution driven by some fixed nonrandom true value for global latent variables.
These results rest on ideal variational Bayes and its connection to frequentist variational approximations. Thus this work bridges the gap in asymptotic theory between the frequentist
variational approximation, in particular the variational frequentist estimate (vfe), and variational Bayes. It also assures us that variational Bayes as a popular approximate inference
algorithm bears some theoretical soundness.
We present our results in the classical vb framework but the results and proof techniques are
more generally applicable. Our results can be easily generalized to more recent developments
of vb beyond Kullback-Leibler (kl) divergence, f -divergence or α-divergence for example. We
could also allow for model misspecification, as long as the variational loglikelihood M n (θ ; x)
under the misspecified model still enjoys local asymptotic normality.
There are several interesting avenues for future work. The variational Bernstein–Von Mises
theorem developed in this work is parametric; its parameters are of finite dimension. One
direction is to develop a semiparametric or nonparametric counterpart. A second direction is
to characterize the posterior contraction rates of vb posteriors. Finally, we characterized the
asymptotics of an optimization problem, assuming that we obtain the global optimum. Though
our simulations corroborated the theory, vb optimization typically finds a local optimum.
Theoretically characterizing these local optima requires further study of the loss surface.
Acknowledgements. We would like to thank Adji Dieng, Christian Naesseth, and Dustin
Tran for their valuable feedback on our manuscript.
23
References
Abbe, E. and Sandon, C. (2015). Community detection in general stochastic block models:
fundamental limits and efficient recovery algorithms. arXiv preprint arXiv:1503.00609.
Amir-Moez, A. and Johnston, G. (1969). On the product of diagonal elements of a positive matrix.
Mathematics Magazine, 42(1):24–26.
Beckenbach, E. F. and Bellman, R. (2012). Inequalities, volume 30. Springer Science & Business
Media.
Bernstein, S. N. (1917). Theory of Probability.
Bickel, P., Choi, D., Chang, X., Zhang, H., et al. (2013). Asymptotic normality of maximum
likelihood and its variational approximation for stochastic blockmodels. The Annals of
Statistics, 41(4):1922–1943.
Bickel, P., Kleijn, B., et al. (2012). The semiparametric Bernstein–von Mises theorem. The
Annals of Statistics, 40(1):206–237.
Bickel, P. J. and Yahav, J. A. (1967). Asymptotically pointwise optimal procedures in sequential
analysis.
Bishop, C. M. (2006). Pattern recognition. Machine Learning, 128.
Blei, D., Kucukelbir, A., and McAuliffe, J. (2016). Variational inference: A review for statisticians.
Journal of American Statistical Association.
Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent dirichlet allocation. Journal of machine
Learning research, 3(Jan):993–1022.
Braides, A. (2006). A handbook of Γ-convergence. Handbook of Differential Equations: stationary
partial differential equations, 3:101–213.
Breslow, N. E. and Clayton, D. G. (1993). Approximate inference in generalized linear mixed
models. Journal of the American statistical Association, 88(421):9–25.
Carpenter, B., Gelman, A., Hoffman, M., Lee, D., Goodrich, B., Betancourt, M., Brubaker, M. A.,
Guo, J., Li, P., and Riddell, A. (2015). Stan: a probabilistic programming language. Journal
of Statistical Software.
Castillo, I., Nickl, R., et al. (2014). On the Bernstein–von Mises phenomenon for nonparametric
Bayes procedures. The Annals of Statistics, 42(5):1941–1969.
Celisse, A., Daudin, J.-J., Pierre, L., et al. (2012). Consistency of maximum-likelihood and
variational estimators in the stochastic block model. Electronic Journal of Statistics, 6:1847–
1899.
Dal Maso, G. (2012). An introduction to Γ-convergence, volume 8. Springer Science & Business
Media.
Dempster, A., Laird, N., and Rubin, D. (1977). Maximum likelihood from incomplete data via
the EM algorithm. Journal of the Royal Statistical Society, Series B, 39:1–38.
Gelfand, A. E. and Smith, A. F. (1990). Sampling-based approaches to calculating marginal
densities. Journal of the American statistical association, 85(410):398–409.
24
Ghosh, J. and Ramamoorthi, R. (2003). Bayesian Nonparametrics. Springer Series in Statistics.
Springer.
Hall, P., Ormerod, J. T., and Wand, M. (2011a). Theory of gaussian variational approximation
for a Poisson mixed model. Statistica Sinica, pages 369–389.
Hall, P., Pham, T., Wand, M. P., Wang, S. S., et al. (2011b). Asymptotic normality and valid
inference for Gaussian variational approximation. The Annals of Statistics, 39(5):2502–2532.
Hastings, W. (1970). Monte Carlo sampling methods using Markov chains and their applications.
Biometrika, 57:97–109.
Hoffman, M., Blei, D., Wang, C., and Paisley, J. (2013). Stochastic variational inference. Journal
of Machine Learning Research, 14:1303–1347.
Hoffman, M. D. and Gelman, A. (2014). The No-U-Turn sampler. JMLR, 15(1):1593–1623.
Hofman, J. and Wiggins, C. (2008). Bayesian approach to network modularity. Physical Review
Letters, 100(25).
Jiang, J. (2007). Linear and generalized linear mixed models and their applications. Springer
Science & Business Media.
Jordan, M. I., Ghahramani, Z., Jaakkola, T. S., and Saul, L. K. (1999). An introduction to
variational methods for graphical models. Machine learning, 37(2):183–233.
Kleijn, B., Van der Vaart, A., et al. (2012). The Bernstein-von-Mises theorem under misspecification. Electronic Journal of Statistics, 6:354–381.
Kucukelbir, A., Tran, D., Ranganath, R., Gelman, A., and Blei, D. M. (2016). Automatic
differentiation variational inference. arXiv preprint arXiv:1603.00788.
Laplace, P. (1809). Memoire sur les integrales definies et leur application aux probabilites, et
specialement a la recherche du milieu qu’il faut choisir entre les resultats des observations.
Memoires presentes a l’Academie des Sciences, Paris.
Le Cam, L. (1953). On some asymptotic properties of maximum likelihood estimates and related
bayes’ estimates. Univ. Calif. Publ. in Statist., 1:277–330.
Le Cam, L. and Yang, G. L. (2012). Asymptotics in statistics: some basic concepts. Springer
Science & Business Media.
Lehmann, E. L. and Casella, G. (2006). Theory of point estimation. Springer Science & Business
Media.
Lu, Y., Stuart, A. M., and Weber, H. (2016). Gaussian approximations for probability measures
on Rd . arXiv preprint arXiv:1611.08642.
McCullagh, P. (1984). Generalized linear models. European Journal of Operational Research,
16(3):285–292.
McCulloch, C. E. and Neuhaus, J. M. (2001). Generalized linear mixed models. Wiley Online
Library.
Mossel, E., Neeman, J., and Sly, A. (2012). Stochastic block models and reconstruction. arXiv
preprint arXiv:1202.1499.
25
Murphy, K. P. (2012). Machine learning: a probabilistic perspective. MIT press.
Ormerod, J. T. and Wand, M. P. (2010). Explaining variational approximations. The American
Statistician, 64(2):140–153.
Ormerod, J. T., You, C., and Muller, S. (2014). A variational Bayes approach to variable selection.
Technical report, Citeseer.
Robert, C. and Casella, G. (2004). Monte Carlo Statistical Methods. Springer Texts in Statistics.
Springer-Verlag, New York, NY.
Roberts, S. J., Husmeier, D., Rezek, I., and Penny, W. (1998). Bayesian approaches to Gaussian mixture modeling. IEEE Transactions on Pattern Analysis and Machine Intelligence,
20(11):1133–1142.
Snijders, T. A. and Nowicki, K. (1997). Estimation and prediction for stochastic blockmodels for
graphs with latent block structure. Journal of classification, 14(1):75–100.
Van der Vaart, A. W. (2000). Asymptotic statistics, volume 3. Cambridge university press.
Von Mises, R. (1931). Wahrscheinlichkeitsrechnung.
Wainwright, M. J. and Jordan, M. I. (2008). Graphical models, exponential families, and
variational inference. Foundations and Trends® in Machine Learning, 1(1-2):1–305.
Wang, B. and Titterington, D. (2005). Inadequacy of interval estimates corresponding to
variational bayesian approximations. In AISTATS.
Wang, B., Titterington, D., et al. (2006). Convergence properties of a general algorithm for
calculating variational bayesian estimates for a normal mixture model. Bayesian Analysis,
1(3):625–650.
Wang, Y. J. and Wong, G. Y. (1987). Stochastic blockmodels for directed graphs. Journal of the
American Statistical Association, 82(397):8–19.
Westling, T. and McCormick, T. H. (2015). Establishing consistency and improving uncertainty
estimates of variational inference through M-estimation. arXiv preprint arXiv:1510.08151.
You, C., Ormerod, J. T., and Müller, S. (2014). On variational Bayes estimation and variational
information criteria for linear regression models. Australian & New Zealand Journal of
Statistics, 56(1):73–87.
26
Appendix
A
Proof of Lemma 1
What we need to show here is that our consistent testability assumption implies assumption
(2.3) in Kleijn et al. (2012):
Z
P θ0
θ̃ > M n
π∗θ̃ (θ̃ | x) dθ̃ → 0
1
for every sequence of constants M n → ∞, where θ̃ = δ−
n (θ − θ0 ).
This is a consequence of a slight generalization of Theorem 3.1 of Kleijn et al. (2012). That
p
theorem shows this implication for the iid case with a common n-convergence rate for all
dimensions of θ . Specifically, they rely on a suitable test sequence under misspecification of
uniform exponential power around the true value θ0 to split the posterior measure.
p
1
To show this implication in our case, we replace all n by δ−
n in the proofs of Theorem 3.1,
Theorem 3.3, Lemma 3.3, Lemma 3.4 of Kleijn et al. (2012). We refer the readers to Kleijn et al.
(2012) and omit the proof here.
B
Proof of Lemma 2
We first perform a change of variable step regarding the mean field variational family. In light
1
of Lemma 1, we know that the vb ideal degenerates to a point mass at the rate of δ−
n . We need
to assume a variational family that degenerates to points masses at the same rate as the ideal
vb posterior. This is because if the variational distribution converges to a point mass faster
than π(θ | x), then the kl divergence between them will converge to +∞. This makes the kl
minimization meaningless as n increases. 5
To avoid this pathology, we assume a variational family for the rescaled and re-centered θ ,
1
θ̌ := δ−
n (θ − µ), for some µ ∈ Θ. This is a centered and scaled transformation of θ , centered to an
arbitrary µ. (In contrast, the previous transformation θ̃ was centering θ at the true θ0 .) With
this transformation, the variational family is
q θ̌ (θ̌ ) = q(µ + δn θ̌ )|det(δn )|,
(18)
where q(·) is the original mean field variational family. We will overload the notation in Section 1
and write this transformed family as q(θ ) and the corresponding family Q d .
In this family, for each fixed µ, θ degenerates to a point mass at µ as n goes to infinity. µ is
not necessarily equal to the true value θ0 . We allow µ to vary throughout the parameter space
Θ so that Q d does not restrict what the distributions degenerate to. Q d only constrains that
the variational distribution degenerates to some point mass at the rate of δn . This step also
does not restrict the applicability of the theoretical results. In practice, we always have finite
samples with fixed n, so assuming a fixed variational family for θ̌ as opposed to θ amounts to a
change of variable θ = µ + δn θ̌ .
5 Equation (36) and Equation (106) in the proof exemplify this claim.
27
Next we show consistency of the kl minimizer of the vb ideal.
To show consistency, we need to show that the mass of the kl minimizer
q‡ := arg min kl( q(θ )||π∗ (θ | x))
q(θ )∈Q d
concentrates near θ0 as n → ∞. That is,
Z
P θ0
B (θ 0 , ξ n )
q‡ (θ ) dθ → 1,
(19)
for some ξn → 0 as n → ∞. This implies
d
q‡ (θ ) → δθ0
by the Slutsky’s theorem.
To begin with, we first claim that
lim sup min kl( q(θ )||π∗ (θ | x)) ≤ M,
n→∞
(20)
for some constant M > 0, and
Z
Rd \K
q‡ (θ ) dθ → 0,
(21)
where K is the compact set assumed in the local asymptotic normality condition.
The first claim says that the limiting minimum kl divergence is upper bounded. The intuition
is that a choice of q(θ ) with µ = θ0 would have a finite kl divergence in the limit. This is because
(rougly) π∗ (θ | x) converges to a normal distribution centered at θ0 with rate δn , so it suffices to
have a q(θ ) that shares the same center and the same rate of convergence.
The second claim says that the restriction of q‡ (θ ) to the compact set K , due to the set compactness
needed in the local asymptotic normality (lan) condition, will not affect our conclusion in the
limit. This is because the family of Q d we assume has a shrinking-to-zero scale. In this way, as
long as µ resides within K , q‡ (θ ) will eventually be very close to its renormalized restriction to
the compact set K , q‡,K (θ ), where
q‡,K (θ ) = R
q ‡ ( θ ) · Iθ ( K )
q‡ (θ ) · Iθ (K ) dθ
.
We will prove these claims at the end.
P θ0
R
To show B(θ0 ,ξn ) q‡,K (θ ) dθ → 1, we both upper bound and lower bound this integral. This step
mimicks the Step 2 in the proof of Lemma 3.6 along with Lemma 3.7 in Lu et al. (2016).
We first upper bound the integral using the lan condition,
Z
q‡,K (θ ) M n (θ ; x) dθ
Z
= q‡,K (θ )
28
(22)
¸
·
1 −1
>
−1
1
>
[
δ
(
θ
−
θ
)]
V
[
δ
(
θ
−
θ
)]
+
o
(1)
dθ
× M n (θ0 ; x) + δ−
(
θ
−
θ
)
V
∆
−
0
0
0
θ
P
θ
n,
θ
n
n
0
0
0
2 n
d η2 Z
X
≤ M n (θ0 ; x) − C 1
q‡,K (θ ) dθ + o P (1),
2
c
i =1 δ n,ii B(θ0 ,η)
(23)
(24)
for large enough n and η << 1 and some constant C 1 > 0. The first equality is due to the lan
condition. The second inequality is due to the domination of quadratic term for large n.
Then we lower bound the integral using our first claim. By
lim sup kl( q‡,K (θ )||π∗ (θ | x)) ≤ M,
n→∞
we can have
Z
q‡,K (θ ) M n (θ ; x) dθ ≥ M n (θ0 ; x) − M0 ,
(25)
for some large constant M0 > M .
This step is due to a couple of steps of technical calculation and the lan condition. To show this
implication, we first rewrite the kl divergence as follows.
kl( q‡,K (θ)||π∗ (θ | x))
Z
=
=
(26)
q‡,K (θ ) log q‡,K (θ ) dθ −
d Z
X
i =1
Z
‡,K
q‡,K (θ ) log π∗ (θ | x) dθ
‡,K
1
−1
[δ−
n,ii q h,i ( h)] log[δ n,ii q h,i ( h)]δ n,ii d h −
= log | det(δn )|−1 +
d
X
i =1
‡,K
H( q h,i ( h)) −
Z
Z
q‡,K (θ ) log π∗ (θ | x) dθ
q‡,K (θ ) log π∗ (θ | x) dθ ,
(27)
(28)
(29)
where this calculation is due to the form of the Q d family we assume and a change of variable
1
of h = δ−
n (θ − µ); h is in the same spirit as θ̌ above. Notation wise, µ is the location parameter
specific to q‡,K (θ ) and H( q h ( h)) denotes the entropy of distribution q h .
We further approximate the last term by the lan condition.
Z
q‡,K (θ ) log π∗ (θ | x) dθ
(30)
Z
p(θ ) exp( M n (θ ; x))
= q‡,K (θ ) log R
dθ
(31)
p(θ ) exp( M n (θ ; x)) dθ
Z
Z
Z
= q‡,K (θ ) log p(θ ) dθ + q(θ ) M n (θ ; x) dθ − log p(θ ) exp( M n (θ ; x)) dθ
(32)
Z
Z
= q‡,K (θ ) log p(θ ) dθ + q‡,K (θ ) M n (θ ; x) dθ
·
¸
d
1
−
log(2π) − log det Vθ0 + log det(δn ) + M n (θ0 ; x) + log p(θ0 ) + o P (1) .
(33)
2
2
R
This first equality is due to the definition of π∗ (θ | x). The second equality is due to q‡,K (θ ) dθ = 1.
The third equality is due to Laplace approximation and the lan condition.
29
Going back to the kl divergence, this approximation gives
kl( q‡,K (θ)||π∗ (θ | x))
= log | det(δn )|−1 +
d
X
i =1
= log | det(δn )|
−1
+
d
X
i =1
(34)
‡,K
H( q h,i ( h)) −
‡,K
H( q h,i ( h)) −
Z
q‡,K (θ ) log π∗ (θ | x) dθ
Z
q
‡,K
(θ ) log p(θ ) dθ −
Z
(35)
q‡,K (θ ) M n (θ ; x) dθ
¸
d
1
+
log(2π) − log det Vθ0 + log det(δn ) + M n (θ0 ; x) + log p(θ0 ) + o P (1)
2
2
Z
Z
d
X
‡,K
H( q h,i ( h)) − q‡,K (θ ) log p(θ ) dθ − q‡,K (θ ) M n (θ ; x) dθ
=
·
(36)
i =1
+
d
1
log(2π) − log det Vθ0 + M n (θ0 ; x) + log p(θ0 ) + o P (1).
2
2
(37)
The first equality is exactly Equation (29). The second equality is due to Equation (33). The
third equality is due to the cancellation of the two log det(δn ) terms. This exemplifies why
we assumed the convergence rate of the Q d family in the first place; we need to avoid the kl
divergence going to infinity.
By
lim sup kl( q‡,K (θ )||π∗ (θ | x)) ≤ M,
n→∞
we have
Z
q‡,K (θ ) M n (θ ; x) dθ
≥− M +
d
X
i =1
‡,K
H( q h,i ( h)) −
(38)
Z
q‡,K (θ ) log p(θ ) dθ
d
1
+ log(2π) − log det Vθ0 + M n (θ0 ; x) + log p(θ0 ) + o P (1)
2
2
≥ − M0 + M n (θ0 ; x) + o P (1)
(39)
(40)
for some constant M0 > 0. This can be achieved by choosing a large enough M0 to make
the
last inequality true. This is doable because
R ‡,K
R all the terms does not change with n except
q (θ ) log p(θ ) dθ . And we have lim supn→∞ q‡,K (θ ) log p(θ ) dθ < ∞ due to our prior mass
condition.
Now combining Equation (40) and Equation (24), we have
M n (θ0 ; x) − C 1
d η2 Z
X
2
i =1 δ n,ii
B(θ0 ,η) c
q‡,K (θ ) dθ + o P (1) ≥ − M0 + M n (θ0 ; x).
This gives
Z
B(θ0 ,η) c
q‡,K (θ ) dθ + o P (1) ≤
M0 · (min i δn,ii )2
C 2 η2
,
for some constant
C 2 > 0. The right side of the inequality will go to zero as n goes to infinity if we
p
choose η = M0 (min i δn,ii )/C 2 → 0. That is, we just showed Equation (19) with ξn = η.
30
We are now left to show the two claims we made at the beginning.
To show Equation (20), it suffices to show that there exists a choice of q(θ ) such that
lim sup kl( q(θ )||π∗ (θ | x)) < ∞.
n→∞
We choose q̃(θ ) =
Qd
2
i =1 N (θ i ; θ0,i , δ n,ii v i )
for v i > 0, i = 1, ..., d . We thus have
kl( q̃(θ)||π∗ (θ | x))
(41)
Z
Z
d 1
X
d
log(v i ) + + d log(2π) − q̃(θ ) log p(θ ) dθ − q̃(θ ) M n (θ ; x) dθ
=
2
i =1 2
1
log det Vθ0 + M n (θ0 ; x) + log p(θ0 ) + o P (1)
2
d 1
X
d
=
log(v i ) + + d log(2π) − log p(θ0 ) − M n (θ0 ; x)
2
2
i =1
−
1
log det Vθ0 + M n (θ0 ; x) + log p(θ0 ) + o P (1)
2
d 1
X
d
1
log(v i ) + + d log(2π) − log det Vθ0 + C 6 + o P (1),
=
2
2
i =1 2
−
(42)
(43)
(44)
for some constant C 6 > 0. The finiteness of limsup is due to the last term being bounded in the
limit. The first equality is due to the same calculation as in Equation (43). The third equality is
due to the cancellation of the two M n (θ0 ; x) terms and the two p(θ0 ) terms; this renders the
whole term independent of n. The second equality is due to the limit of q̃(θ ) concentrating
around θ0 . Specifically, we expand log p(θ ) to the second order around θ0 ,
Z
q̃(θ ) log p(θ ) dθ
(45)
·
¸
Z
Z
(θ − θ0 )2 1
= log p(θ0 ) + q̃(θ ) (θ − θ0 )(log p(θ0 ))0 +
(log p(ξθ + (1 − ξ)θ0 ))00 (1 − ξ)2 dξ dθ (46)
2
0
½Z
¾
1
≤ log p(θ0 ) +
max
q̃(θ )(θ − θ0 )2 (log p(ξθ + (1 − ξ)θ0 ))00 dθ
(47)
2! ξ∈[0,1]
Z
Mp
1 >
2
−1
|θ |2 e(|θ|+|θ0 |) · e− 2 θ (δn V δn ) θ dθ
(48)
≤ log p(θ0 ) + q
Q
d
(2π)d det(δ2n ) i v i R
Z
Mp
1 >
−1
θ02
≤ log p(θ0 ) + q
e
|θ |2 e− 2 θ [(δn V δn ) −2 I d ]θ
(49)
Q
d
R
(2π)d det(δ2n ) i v i
2
≤ log p(θ0 ) + C 3 M p eθ0 max(δ2n,ii ) det(V −1 − 2δ2n )−1
(50)
≤ log p(θ0 ) + C 4 max(δ2n,ii )
(51)
d
d
where maxd (δ2n,ii ) → 0 and C 3 , C 4 > 0. The first equality is due to Taylor expansion with
integral form residuals. The second inequality is due to the first order derivative terms equal
to zero and taking the maximum of the second order derivative. The third inequality is due
to the prior mass condition where we assume the second derivative of log p(θ ) is bounded by
2
2
M p e|θ| for some constant M p > 0. The fourth inequality is pulling eθ0 out of the integral. The
31
fifth inequality is due to rescaling θ by its covariance matrix and appealing to the mean of a
Chi-squared distribution with d degrees of freedom. The sixth (and last) inequality is due to
det(V −1 − 2δ2n )−1 > 0 for large enough n.
R
We apply the same Taylor expansion argument to the q̃(θ ) M n (θ ; x) dθ .
Z
q̃(θ ) M n (θ ; x) dθ
(52)
Kn
·
¸
Z
1 −1
1
>
>
−1
q̃(θ ) δ−
= M n (θ0 ; x) +
n (θ − θ0 ) Vθ0 ∆ n,θ0 + (δ n (θ − θ0 )) Vθ0 δ n (θ − θ0 ) + o P (1) dθ (53)
2
Kn
1
(54)
≤ M n (θ0 ; x) + T r (Vθ0 V ) + o P (1)
2
≤ M n (θ0 ; x) + C 6 + o P (1)
(55)
where K n is a compact set. The first equality is due to the lan condition. The second inequality
is due to q̃(θ ) centered at θ0 with covariance δn V δn . The third inequalities are true for C 6 >
0.
For the set outside of this compact set K n , we consider for a general choice of q distribution,
Q
q̃(θ ) = N (θ ; θ0 + ∆n,θ0 , δn Vθ0 δn ), of which q̃(θ ) = di=1 N (θ i ; θ0,i , δ2n,ii v i ) we work with is a special
case.
Z
q̃(θ )(log p(θ ) + M n (θ ; x)) dθ
(56)
Rd \K n
Z
≤C 7
N (θ ; θ0 + ∆n,θ0 , δn Vθ0 δn )(log p(θ ) + M n (θ ; x)) dθ
(57)
Rd \K n
Z
≤C 8 [det(δn )−1 log(det(δn )−1 )]
N (θ̃ ; ∆n,θ0 , Vθ0 ) log π∗ (θ̃ | x) det(δn ) dθ̃
(58)
Rd \K n
Z
≤C 9 log(det(δn )−1 )]
[π∗ (θ̃ | x) + o P (1)] log π∗ (θ̃ | x), Vθ0 ) dθ̃
(59)
Rd \K n
Z
≤C 10 log(det(δn )−1 )]
[N (θ̃ ; ∆n,θ0 , Vθ0 ) + o P (1)] log N (θ̃ ; ∆n,θ0 , Vθ0 ) dθ̃
(60)
Rd \K n
(61)
≤ o P (1)
for some C 7 , C 8 , C 9 , C 10 > 0. The first inequality is due to q̃(θ ) centered at θ0 and with rate of
1
convergence δn . The second inequality is due to a change of variable θ̃ = δ−
n (θ − θ0 ). The third
inequality is due to Lemma 1. The fourth inequality is due to Lemma 1 and Theorem 2 in Piera
and Parada (2009). The fifth inequality is due to a choice of fast enough increasing sequence of
compact sets K n .
R
The lower bound of q̃(θ )(log p(θ ) + M n (θ ; x)) dθ can be derived with exactly the same argument.
Our first claim Equation (20) is thus proved.
To show our second claim Equation (21), we first denote B(µ, M ) as the largest ball centered
at µ and contained in the compact set K . We know by the construct of Q d — Q d has a
shrinking-to-zero
scale — that for each ² > 0, there exists an N such that for all n > N we have
R
||θ −µ||> M q(θ ) dθ < ². Therefore, we have
Z
‡
Rd \K
q (θ ) dθ ≤
Z
Rd \B(µ,M )
32
q ‡ (θ ) dθ ≤ ² .
C
Proof of Lemma 3
To show the convergence of optimizers from two minimization problems, we invoke Γ-convergence.
It is a classical technique in characterizing variational problems. A major reason is that if two
functionals Γ−converge, then their minimizer also converge.
We recall the definition of Γ-convergence (Dal Maso, 2012; Braides, 2006).
Definition 13. Let X be a metric space and F² : X → R a family of functionals indexed by
² > 0. Then the existence of a limiting functional F0 , the Γ−limit of F² , as ² → 0, relies on two
conditions:
1. (liminf inequality) for every x ∈ X and for every x² → x, we have
F0 ( x) ≤ lim inf F² ( x² ),
² →0
2. (limsup inequality / existence of a recovery sequence) for every x ∈ X we can find a sequence
x̄² → x such that
F0 ( x) ≥ lim sup F² ( x̄² ).
² →0
The first condition says that F0 is a lower bound for the sequence F² , in the sense that F0 ( x) ≤
F² ( x² ) + o(1) whenever x² → x. Together with the first condition, the second condition implies
that F0 ( x) = lim²→0 F² ( x̄² ), so that the lower bound is sharp.
Γ-convergence is particularly useful for variational problems due to the following fundamental
theorem. Before stating the theorem, we first define equi-coerciveness.
Definition 14. (Equi-coerciveness of functionals) A sequence F² : X → R̄ is equi-coercise if for all
² j → 0 and x j such that F² j ( x j ) ≤ t there exist a subsequence of j (not relabeled) and a converging
sequence x0j such that F² j ( x0j ) ≤ F² j ( x j ) + o(1).
Equi-coerciveness of functionals ensures that we can find a precompact minimizing sequence of
F² such that the convergence x² → x can take place. Now we are ready to state the fundamental
theorem.
Theorem 15. (Fundamental theorem of Γ-convergence) Let X be a metric space. Let (F² ) be an
equi- coercise sequence of functions on X . Let F = Γ − lim²→0 F² , then
arg min F = lim arg min F² .
X
² →0
X
The above theorem implies that if all functions F² admit a minimizer x² then, up to subsequences,
x² converge to a minimum point of F . We remark that the converse is not true; we may have
minimizers of F which are not limits of minimizers of F² , e.g. F² ( t) = ² t2 (Braides, 2006).
In this way, Γ-convergence is convenient to use when we would like to study the asymptotic
behavior of a family of problem F² through defining a limiting problem F0 which is a ‘good
approximation’ such that the minimizers converge: x² → x0 , where x0 is a minimizer of F0 .
33
Conversely, we can characterize solutions of a difficult F0 by finding easier approximating F²
(Braides, 2006).
We now prove Lemma 3 for the general mean field family. The family is parametric as in
Section 1, so we assume it is indexed by some finite dimensional parameter m. We want to show
that the functionals
F n ( m) := kl( q(θ ; m)||π∗ (θ | x))
Γ-converge to
F0 ( m) := kl( q(θ ; m)||N (θ ; θ0 + δn ∆n,θ0 , δn Vθ−01 δn )) − ∆>
n,θ0 Vθ0 ∆ n,θ0
in probability as n → 0. Recall that the mean field family has density
q(θ ) =
d
Y
1
δ−
n,ii q h,i ( h),
i =1
1
where h = δ−
n (θ − θ0 ).
We need the following mild technical conditions on Q d .
Following the change-of-variable step detailed in the beginning of Appendix B, we consider the
Q
1
1
mean field variational family with densities q(θ ) = di=1 δ−
q ( h), where h = δ−
n (θ − µ) for
n,ii h,i
some µ ∈ Θ.
Assumption 2. We assume the following conditions on q h,i :
1. q h,i , i = 1, ..., d have continuous densities.
2. q h,i , i = 1, ..., d have positive and finite entropies.
R
3. q0h,i ( h) d h < ∞, i = 1, ..., d.
The last condition ensures that convergence in finite dimensional parameters imlied convergence
in tv distance. This is due to a Taylor expansion argument:
kl( q(θ; m)|| q(θ; m + δ))
(62)
Z
q(θ ; m) · (log q(θ ; m) − log q(θ ; m + δ)) dθ
(63)
q(θ ; m) · (δ · (log q(θ ; m))0 ) dθ + o(1)
(64)
=
Z
=
=δ
Z
( q(θ ; m))0 dθ
(65)
(66)
<²
The last step is true if Assumption 2 is true for q. We also notice that convergence in kl
divergence implies convergence in tv distance. Therefore, Assumption 2 implies that convergence
in finite dimensional parameter implies convergence in tv distance.
P θ0
Together with Theorem 15, the Γ-convergence of the two functionals implies m n → m 0 where
m n is the minimizer of F n for each n and m 0 is the minimizer of F0 . This is due to the last
34
term of F0 – ∆>
V ∆
is a constant bounded in Pθ0 probability and independent of m. The
n,θ0 θ0 n,θ0
convergence in total variation then follows from Assumption 2 and our argument above.
Lastly, we prove the Γ-convergence of the two functionals for the mean field family.
We first rewrite F n ( m, µ).
F n ( m, µ) := kl( q(θ ; m, µ)||π∗ (θ | x))
Z
d
X
= log | det(δn )|−1 +
H( q h,i ( h; m)) − q(θ ; m, µ) log π∗ (θ | x)) dθ
(67)
(68)
i =1
= log | det(δn )|−1 +
d
X
H( q h,i ( h; m)) −
Z
q(θ ; m, µ) log p(θ ) dθ −
Z
q(θ ; m, µ) M n (θ ; x) dθ
i =1
Z
+ log
(69)
p(θ ) exp( M n (θ ; x)) dθ
= log | det(δn )|−1 +
d
X
H( q h,i ( h; m)) −
Z
q(θ ; m, µ) log p(θ ) dθ −
Z
q(θ ; m, µ) M n (θ ; x) dθ
i =1
¸
d
1
log(2π) − log det Vθ0 + log det(δn ) + M n (θ0 ; x) + log p(θ0 ) + o P (1)
2
2
Z
Z
d
X
=
H( q h,i ( h; m)) − q(θ ; m, µ) log p(θ ) dθ − q(θ ; m, µ) M n (θ ; x) dθ
·
+
(70)
i =1
¸
1
d
log(2π) − log det Vθ0 + M n (θ0 ; x) + log p(θ0 ) + o P (1)
(71)
2
2
·
¸
Z
d
X
d
1
=
H( q h,i ( h; m)) − q(θ ; m, µ) M n (θ ; x) dθ +
log(2π) − log det Vθ0 + M n (θ0 ; x) + o P (1)
2
2
i =1
(72)
Z
d
X
1
>
=
H( q h,i ( h; m)) − q(θ ; m, µ)[ M n (θ0 ; x) + δ−
n (θ − θ0 ) Vθ0 ∆ n,θ0
·
+
i =1
·
¸
1
1
d
1
>
−1
− (δ−
(
θ
−
θ
))
V
δ
(
θ
−
θ
)
+
o
(1)]
d
θ
−
log(2
π
)
−
log
det
V
+
M
(
θ
;
x
)
+
o
(1)
0
0
n 0
θ0 n
P
θ0
P
2 n
2
2
(73)
Z
d
X
1
>
=
H( q h,i ( h; m)) − δ−
n (θ − θ0 ) Vθ0 ∆ n,θ0 · q(θ ; m, µ) dθ
i =1
Z
+
·
¸
d
1
1 −1
>
−1
(δ (θ − θ0 )) Vθ0 δn (θ − θ0 ) · q(θ ; m, µ) dθ −
log(2π) − log det Vθ0 + o P (1)
2 n
2
2
(74)
The first equality is by the definition of kl divergence. The second equality is by the definition
of the vb ideal. The third equality is due to the Laplace approximation of the normalizer like
we did in Equation (33). The fourth equality is due to the cancellation of the two log det(δn )
terms. This again exemplifies why we assume a fixed variational family on the rescale variable
θ̌ . The fifth equality is due to a similar argument as in Equation (43). The sixth equality is due
to the lan condition of M n (θ ; x). The seventh equality is due to the computation of each term
in the integral as an expectation under the distribution q(θ ). To extend the restriction to some
compact set K to the whole space Rd in the sixth equality, we employ the same argument as in
Equation (44).
35
We notice that when µ 6= θ0 , we will have F n ( m) → ∞. On the other hand, we have lim sup F n < ∞.
This echoes our consistency result in Lemma 2.
Now we rewrite F0 ( m, µ).
kl( q(θ; m, µ)||N (θ; θ0 + δn ∆n,θ0 , δn Vθ−01 δn ))
= log | det(δn )|−1 +
d
X
H( q h,i ( h; m)) +
i =1
= log | det(δn )|−1 +
d
X
H( q h,i ( h; m)) +
i =1
Z
+
=
d
X
Z
(75)
q(θ ; m, µ) log N (θ ; θ0 + δn ∆n,θ0 , δn Vθ−01 δn ) dθ
(76)
d
1
log(2π) − log det Vθ0 + log det(δn )
2
2
1
−1
q(θ ; m, µ) · (θ − θ0 − δn ∆n,θ0 )> δ−
n Vθ0 δ n (θ − θ0 − δ n ∆ n,θ0 ) dθ
(77)
1
d
log(2π) − log det Vθ0 + ∆>
n,θ0 Vθ0 ∆ n,θ0
2
2
i =1
Z
Z
1 −1
1
>
1
− δ−
(δ (θ − θ0 ))> Vθ0 δ−
n (θ − θ0 ) Vθ0 ∆ n,θ0 · q(θ ; m, µ) dθ +
n (θ − θ0 ) · q(θ ; m, µ) dθ . (78)
2 n
H( q h,i ( h; m)) +
This gives
F0 ( m, µ) − ∆>
n,θ0 Vθ0 ∆ n,θ0
=
d
X
i =1
Z
−
H( q h,i ( h; m)) −
(79)
d
1
log(2π) + log det Vθ0
2
2
1
>
δ−
n (θ − θ0 ) Vθ0 ∆ n,θ0 · q(θ ; m, µ) dθ
1 −1
1
(δ (θ − θ0 ))> Vθ0 δ−
n (θ − θ0 ) · q(θ ; m, µ) dθ
2 n
d
X
1
d
= + ∞ · (1 − Iµ (θ0 )) + [ H( q h,i ( h; m)) − log(2π) + log det Vθ0
2
2
i =1
Z
1
>
− δ−
n (θ − θ0 ) Vθ0 ∆ n,θ0 · q(θ ; m, µ) dθ
Z
1 −1
1
+
(δ (θ − θ0 ))> Vθ0 δ−
n (θ − θ0 ) · q(θ ; m, µ) dθ ] · Iµ (θ0 ).
2 n
Z
+
The last step is due to our definition of our variational family q(θ ; m, µ) =
(80)
(81)
Qd
−1
i =1 δ n,ii q h,i ( h; m),
where
for some µ ∈ Θ. The last step is true as long as the q h,i distributions are
not point masses at zero. This is ensured by positive entropy in Assumption 2.
1
h = δ−
n (θ − µ )
Comparing Equation (74) and Equation (81), we can prove the Γ convergence. Let m n → m.
When µ 6= θ0 , lim infn→∞ F n ( m n , µ) = +∞. The limsup inequality is automatically satisfied.
When µ = θ0 , we have F n ( m, µ) = F0 ( m, µ) − ∆>
V ∆
+ o P (1). This implies F0 ( m, µ) ≤
n,θ0 θ0 n,θ0
limn→∞ F n ( m n , µ) in Pθ0 probability by the continuity of F n ensured by Assumption 2.
We then show the existence of a recovery sequence. When µ 6= θ0 , F0 ( m, µ) = +∞. The limsup
inequality is automatically satisfied. When µ = θ0 , we can simply choose m n = θ0 . The limsup
inequality is again ensured by F0 ( m, µ) ≤ limn→∞ F n ( m n , µ) in Pθ0 probability and the continuity
of F n . The Γ-convergence of the F functionals is shown.
36
We notice that ∆>
V ∆
does not depend on m or µ so that arg min F0 = arg min F0 −
n,θ θ0 n,θ0
0
∆>
V ∆
. The convergence of the kl minimizers is thus proved.
n,θ0 θ0 n,θ0
D
Proof of Lemma 4
Notice that the mean field variational families Q d = { q : q(θ ) =
1
δ−
n (θ − µ)
form of
Qd
−1
i =1 δ n,ii q h,i ( h),
where h =
for some µ ∈ Θ}, or the Gaussian family { q : q(θ ) = N ( m, δn Σδn )} can be written in the
1
q(θ ) = | det δn |−1 q h (δ−
n (θ − µ))
for some µ ∈ Rd , and q h ( h) d h = 0. This form is due to a change-of-variable step we detailed in
the beginning of Appendix B.
R
We first specify the mild technical conditions on Q d .
Assumption 3. We assume the following conditions on q h .
R
1. If q h is has zero mean, we assume h2 · q h ( h) d h < ∞ and sup z,x |(log p( z, x | θ ))00 | ≤ C 11 ·
q h (θ )−C12 for some C 11 , C 12 > 0; | M n (θ ; x)00 | ≤ C 13 · q h (θ )−C14 for some C 13 , C 14 > 0.
R
2. If q h has nonzero mean, we assume h · q h ( h) d h < ∞ and sup z,x |(log p( z, x | θ ))0 | ≤
C 11 · q h (θ )−C12 for some C 11 , C 12 > 0; | M n (θ ; x)0 ≤ C 13 | · q h (θ )−C14 for some C 13 , C 14 > 0.
The assumption first assumes finite moments for q h so that we can properly apply a Taylor expansion argument. The second part of this assumption makes sure the derivative of
log p( z, x | θ ) does not increase faster than the tail decrease of q h (·). For example, if q h (·) is
normal, then the second part writes sup z,x |(log p( z, x | θ ))00 | ≤ C 15 exp(θ 2 ) for some C 15 > 0, and
| M n (θ ; x)00 | ≤ C 13 exp(θ 2 ) for some C 13 > 0. The latter is satisfied by the lan condition. This is
in general a rather weak condition. We usually would not expect the derivative of log p( z, x | θ )
and M n (θ ; x) to increase this fast as θ increases.
Now we are ready to prove the lemma.
We first approximate the profiled evidence lower bound (elbo), elbo p ( q(θ )).
elbo p ( q(θ))
(82)
µ ·
½Z
¾¸
¶
p( z, x|θ )
:= sup q(θ ) log p(θ ) exp
q( z) log
d z − log q(θ ) dθ ,
q( z)
q( z)
Z
Z
Z
Z
p( z, x|θ )
= q(θ ) log p(θ ) dθ − q(θ ) log q(θ ) dθ + sup q(θ ) q( z) log
d z dθ
q( z)
q( z)
Z
Z
= q(θ ) log p(θ ) dθ − q(θ ) log q(θ ) dθ
µZ
¶0
Z
Z
p( z, x|µ)
p( z, x|µ)
+ sup q(θ )[ q( z) log
d z + (θ − µ)
q( z) log
dz
q( z)
q( z)
q( z)
Z
Z
Z
p( z, x|µ)
dz
= q(θ ) log p(θ ) dθ − q(θ ) log q(θ ) dθ + sup q( z) log
q( z)
q( z)
Z
37
(83)
(84)
¶00
p( z, x|θ̃ † )
q( z) log
d z ] dθ
+
q( z)
Z
Z
Z
p( z, x|µ)
≥ q(θ ) log p(θ ) dθ − q(θ ) log q(θ ) dθ + sup q( z) log
dz
q( z)
q( z)
Z
1
− C 15 q(θ ) |θ − µ|2 q h (θ )−C12 dθ
2
Z
Z
Z
p( z, x|µ)
= q(θ ) log p(θ ) dθ − q(θ ) log q(θ ) dθ + sup q( z) log
dz
q( z)
q( z)
Z
1
− C 15 q h ( h) |δn h|2 q h (µ + δn h)−C12 d h
2
Z
Z
Z
p( z, x|µ)
dz
≥ q(θ ) log p(θ ) dθ − q(θ ) log q(θ ) dθ + sup q( z) log
q( z)
q( z)
Z
1
q(θ ) |θ − µ|2
2
µZ
− C 16 min(δ2n,ii )
i
Z
Z
Z
p( z, x|µ)
= q(θ ) log p(θ ) dθ − q(θ ) log q(θ ) dθ + sup q( z) log
d z + o(1),
q( z)
q( z)
(85)
(86)
(87)
(88)
(89)
for some constant C 16 > 0. The first equality is by the definition of elbo p ( q(θ )). The second
equality is rewriting the integrand. The third equality is due to mean value theorem where θ̃ † is
some value between µ and θ . (A very similar argument for q h with nonzero means can be made
starting from here, that is expanding only to the first order term.) The fourth equality is due to
q h (·) having zero mean. The fifth inequality is due to the second part of Assumption 3. The sixth
inequality is due to a change of variable h = δn (θ − µ). The seventh inequality is due to q h (·)
residing within exponential family with finite second moment (the first part of Assumption 3).
The eighth equality is due to δn → 0 as n → ∞.
We now approximate −kl( q(θ )||π∗ (θ | x)) in a similar way.
− kl( q(θ )||π∗ (θ | x))
Z
p(θ ) exp( M n (θ ; x))
= q(θ ) log
dθ
q (θ )
Ã
"
(
Z
Z
(90)
(91)
)#
!
p( z, x|θ )
− log q(θ ) d θ
q( z)
q( z)
Z
Z
Z
Z
p( z, x|µ)
= q(θ ) log p(θ ) dθ − q(θ ) log q(θ ) dθ + q(θ )[sup q( z) log
dz
q( z)
q( z)
Ã
!0
Ã
!00
Z
Z
p( z, x|µ)
1
p( z, x|θ̃˜ † )
2
+ (θ − µ) sup q( z) log
d z + |θ − µ| sup q( z) log
d z ] dθ
q( z)
2
q( z)
q( z)
q( z)
Z
Z
Z
p( z, x|µ)
dz
= q(θ ) log p(θ ) dθ − q(θ ) log q(θ ) dθ + sup q( z) log
q( z)
q( z)
Ã
!00
Z
Z
1
p( z, x|θ̃˜ † )
2
+ q(θ ) |θ − µ| sup q( z) log
d z ] dθ
2
q( z)
q( z)
Z
Z
Z
p( z, x|µ)
≤ q(θ ) log p(θ ) dθ − q(θ ) log q(θ ) dθ + sup q( z) log
dz
q( z)
q( z)
Z
1
+ C 17 q(θ ) |θ − µ|2 q h (θ )−C14 dθ
2
=
q(θ ) log p(θ ) exp sup
q( z) log
38
(92)
(93)
(94)
(95)
Z
=
q(θ ) log p(θ ) dθ −
Z
q(θ ) log q(θ ) dθ + sup
Z
q( z) log
q( z)
p( z, x|µ)
dz
q( z)
1
+ C 17 q h ( h) |δn h|2 q h (µ + δn h)−C14 d h
2
Z
Z
Z
p( z, x|µ)
d z + C 17 max(δ2n,ii )
≤ q(θ ) log p(θ ) dθ − q(θ ) log q(θ ) dθ + sup q( z) log
i
q( z)
q( z)
Z
Z
Z
p( z, x|µ)
= q(θ ) log p(θ ) dθ − q(θ ) log q(θ ) dθ + sup q( z) log
d z + o(1)
q( z)
q( z)
Z
(96)
(97)
(98)
for some constant C 17 > 0. The first equality is by the definition of kl divergence. The second
equality is rewriting the integrand. The third equality is due to mean value theorem where θ̃˜ † is
some value between µ and θ . (A very similar argument for q h with nonzero means can be made
starting from here, that is expanding only to the first order term.) The fourth equality is due to
q h (·) having zero mean. The fifth inequality is due to the third part of Assumption 3. The sixth
inequality is due to a change of variable h = δn (θ − µ). The seventh inequality is due to q h (·)
residing within exponential family with finite second moment (the first part of Assumption 3).
The eighth equality is due to δn → 0 as n → ∞.
Combining the above two approximation, we have −kl( q(θ )||π∗ (θ | x)) ≤ elbo p ( q(θ )) + o(1). On
the other hand, we know that elbo p ( q(θ )) ≤ −kl( q(θ )||π∗ (θ | x)) by definition. We thus conclude
elbo p ( q(θ)) = −kl( q(θ)||π∗ (θ| x)) + o P (1).
E
Proof of Theorem 5 and Theorem 6
Theorem 5 is a direct consequence of Lemma 2, Lemma 3, and Lemma 4. Lemma 2 and Lemma 3
characterizes the consistency and asymptotic normality of the kl minimizer of the vb ideal.
Lemma 4 says the vb posterior shares the same asymptotic properties as the vb ideal. All of
them together give the consistency and asymptotic normality of vb posteriors.
Theorem 6 is a consequence of a slight generalization of Theorem 2.3 of Kleijn et al. (2012). The
theorem characterizes the consistency and asymptotic normality of the posterior mean estimate
p
p
under model specification with a common n-convergence rate. We only need to replace all n
1
by δ−
n in their proof to obtain the generalization.
Specifically, we first show that, for any M n → ∞,
P θ0
Z
||θ̃ ||> M n
||θ̃ ||2 q∗ (θ̃ ) dθ̃ → 0.
This is ensured by Assumption 3.1.
We then consider three stochastic processes: fix some compact set K and for given M > 0,
Z
t 7→ Z n,M ( t) =
( t − θ̃ )2 · q∗θ̃ (θ̃ ) dθ̃ ,
(99)
||θ̃ ||≤ M
Z
t 7→ Wn,M ( t) =
( t − θ̃ )2 · N (θ̃ ; ∆n,θ0 , Vθ−01 ) dθ̃ ,
(100)
||θ̃ ||≤ M
Z
t 7→ WM ( t) =
( t − θ̃ )2 · N (θ̃ ; X , Vθ−01 ) dθ̃ .
(101)
||θ̃ ||≤ M
39
We note that θn∗ is the minimizer of t 7→ Z n,∞ ( t) and
is the minimizer of t 7→ W∞ ( t).
R
θ̃ · arg min q∈Q d kl( q(θ̃ )||N (θ̃ ; X , Vθ−1 )) dθ̃
0
By sup t∈K,||h||≤ M ( t − h)2 < ∞, we have Z n,M − Wn,M = o Pθ0 (1) in `∞ (K ) by Theorem 5. Since
d
∆n,θ0 → X , the continuous mapping theorem implies that Wn,M − WM = o Pθ0 (1) in `∞ (K ). By
R
θ · q∗ (θ ) < ∞, we have WM − W∞ = o Pθ (1) as M → ∞. We conclude that there exists a sequence
0
M n → ∞ such that Z n,M n − W∞ = o Pθ0 (1). We also have from above that Z n,M n − Z n,∞ = o Pθ0 (1)
in `∞ (K ). We conclude that Z n,∞ − W∞ = o Pθ0 (1) in `∞ (K ). By the continuity and convexity
of
theorem and conclude that θn∗ converges weakly to
R the squared loss, we invoke the argmax
−1
θ̃ · arg min q∈Q d kl( q(θ̃ )||N (θ̃ ; X , Vθ )) dθ̃ .
0
F
Proof of Corollary 7
We prove Lemma 3 for the Gaussian family. We want to show that the functionals
F n ( m, Σ) := kl(N (θ ; m, δn Σδn )||π∗ (θ | x))
Γ-converge to
F0 ( m, Σ) := kl(N (θ ; m, δn Σδn )||N (θ ; θ0 + δn ∆n,θ0 , δn Vθ−01 δn )) − ∆>
n,θ0 Vθ0 ∆ n,θ0
in probability as n → 0. We note that this is equivalent to the Γ-convergence of the functionals
of q: kl( q(·)||π∗ (· | x)) to kl( q(·)||N (· ; ∆n,θ0 , Vθ−1 )). This is because the second statement is the
θ̃
0
same as the first up to a change of variable step from θ to θ̃ .
P θ0
Together with Theorem 15, this implies m n , Σn → m 0 , Σ0 where ( m n , Σn ) is the minimizer of F n
for each n and ( m 0 , Σ0 ) is the minimizer of F0 . This is due to the last term of F0 – ∆>
V ∆
n,θ0 θ0 n,θ0
is a constant bounded in Pθ0 probability and independent of m, Σ. The convergence in total
variation then follows from Lemma 4.9 of Klartag (2007), which gives an upper bound on tv
distance between two Gaussian distributions.
Now we prove the Γ-convergence.
We first rewrite F n ( m, Σ).
F n ( m, Σ)
(102)
:=kl(N (θ ; m, δn Σδn )||π (θ | x))
Z
Z
= N (θ ; m, δn Σδn ) log N (θ ; m, δn Σδn ) dθ − N (θ ; m, δn Σδn ) log π∗ (θ | x)) dθ
Z
d d
1
= − − log(2π) − log det(δn ) − log det(Σ) − N (θ ; m, δn Σδn ) log p(θ ) dθ
2 2
2
Z
Z
− N (θ ; m, δn Σδn ) M n (θ ; x) dθ + log p(θ ) exp( M n (θ ; x)) dθ
Z
d d
1
= − − log(2π) − log det(δn ) − log det(Σ) − N (θ ; m, δn Σδn ) log p(θ ) dθ
2 2
2
Z
− N (θ ; m, δn Σδn ) M n (θ ; x) dθ
∗
40
(103)
(104)
(105)
¸
1
d
log(2π) − log det Vθ0 + log det(δn ) + M n (θ0 ; x) + log p(θ0 ) + o P (1)
2
2
Z
d 1
= − − log det(Σ) − N (θ ; m, δn Σδn ) log p(θ ) dθ
2 2
¸
·
Z
1
− N (θ ; m, δn Σδn ) M n (θ ; x) dθ + − log det Vθ0 + M n (θ0 ; x) + log p(θ0 ) + o P (1)
2
Z
d 1
= − − log det(Σ) − N (θ ; m, δn Σδn ) M n (θ ; x) dθ
2 2
·
¸
1
+ − log det Vθ0 + M n (θ0 ; x) + o P (1)
2
Z
d 1
1
>
= − − log det(Σ) − N (θ ; m, δn Σδn )[ M n (θ0 ; x) + δ−
n (θ − θ0 ) Vθ0 ∆ n,θ0
2 2
·
¸
1
1
1
>
−1
(
θ
−
θ
))
V
δ
(
θ
−
θ
)
+
o
(1)]
d
θ
−
log
det
V
−
M
(
θ
;
x
)
+
o
(1)
− (δ−
0
0
n 0
θ0 n
P
θ0
P
2 n
2
d 1
1 −1
1
>
>
−1
= − − log det(Σ) − δ−
n ( m − θ0 ) Vθ0 ∆ n,θ0 + (δ n ( m − θ0 )) Vθ0 δ n ( m − θ0 )
2 2
2
1
1
+ T r (Vθ0 · Σ) − log det Vθ0 + o P (1).
2
2
·
+
(106)
(107)
(108)
(109)
(110)
The first equality is by the definition of kl divergence. The second equality is calculating
the entropy of multivariate Gaussian distribution. The third equality is due to the Laplace
approximation of the normalizer like we did in Equation (33). The fourth equality is due to the
cancellation of the two log det(δn ) terms. This again exemplifies why we assume the family to
have variance δn Σδn . The fifth equality is due to a similar argument as in Equation (43). The
intuition is that N ( m, δn Σδn ) converges to a point mass as n → ∞. The sixth equality is due to
the lan condition of M n (θ ; x). The seventh equality is due to the computation of each term in
the integral as an expectation under the Gaussian distribution N ( m, δn Σδn ). To extend the
restriction to some compact set K to the whole space Rd in the sixth equality, we employ the
same argument as in Equation (44).
We notice that when m 6= θ0 , we will have F n ( m, Σ) → ∞. On the other hand, we have lim sup F n <
∞. This echoes our consistency result in Lemma 2.
Now we rewrite F0 ( m, Σ).
kl(N (θ; m, δn Σδn )||N (θ; θ0 + δn ∆n,θ0 , δn Vθ−01 δn ))
d 1
1
−1
= − + T r (Vθ0 · Σ) + ( m − θ0 − δn ∆n,θ0 )> δ−
n Vθ0 δ n ( m − θ0 − δ n ∆ n,θ0 )
2 2
1
1
− log det(Σ) − log det Vθ0
2
2
d 1
1 −1
>
>
−1
1
= − − log det(Σ) − δ−
n ( m − θ0 ) Vθ0 ∆ n,θ0 + (δ n ( m − θ0 )) Vθ0 δ n ( m − θ0 )
2 2
2
1
1
+ T r (Vθ0 · Σ) − log det Vθ0 + ∆>
n,θ0 Vθ0 ∆ n,θ0 .
2
2
(111)
(112)
(113)
This gives
F0 ( m, Σ) − ∆>
n,θ0 Vθ0 ∆ n,θ0
(114)
d 1
1 −1
1
>
>
−1
= − − log det(Σ) − δ−
n ( m − θ0 ) Vθ0 ∆ n,θ0 + (δ n ( m − θ0 )) Vθ0 δ n ( m − θ0 )
2 2
2
41
1
1
+ T r (Vθ0 · Σ) − log det Vθ0
2
2
d 1
1
1
= + ∞ · (1 − Im (θ0 )) + [− − log det(Σ) + T r (Vθ0 · Σ) − log det Vθ0 ] · Im (θ0 ).
2 2
2
2
(115)
(116)
This equality is due to the kl divergence between two multivariate Gaussian distributions.
Comparing Equation (110) and Equation (116), we can prove the Γ convergence. Let m n → m and
Σn → Σ. When m 6= θ0 , lim infn→∞ F n ( m n , Σn ) = +∞. The limsup inequality is automatically
satisfied. When m = θ0 , we have F n ( m, Σ) = F0 ( m, Σ) − ∆>
V ∆
+ o P (1). This implies
n,θ0 θ0 n,θ0
F0 ( m, Σ) ≤ limn→∞ F n ( m n , Σn ) in Pθ0 probability by the continuity of F n .
We then show the existence of a recovery sequence. When m 6= θ0 , F0 ( m, Σ) = +∞. The limsup
inequality is automatically satified. When m = θ0 , we can simply choose Σm = Σ and m n = θ0 .
The limsup inequality is again ensured by F0 ( m, Σ) ≤ limn→∞ F n ( m n , Σn ) in Pθ0 probability and
the continuity of F n . The Γ-convergence of the F functionals is shown.
We notice that ∆>
V ∆
does not depend on any of m, Σ so that arg min F0 = arg min F0 −
n,θ θ0 n,θ0
0
∆>
V ∆
. The convergence of the kl minimizers is thus proved.
n,θ0 θ0 n,θ0
G
Proof of Lemma 8
We first note that
kl(N (·; µ0 , Σ0 ||N (·; µ1 , Σ1 ))
1
det(Σ1 )
1
> −1
= [T r (Σ−
].
1 Σ0 ) + (µ1 − µ0 ) Σ1 (µ1 − µ0 ) − d + log
2
det(Σ0 )
(117)
(118)
Clearly, the optimal choice of µ0 is µ̂0 = µ1 . Next, we write Σ0 = diag(λ1 , ..., λd ). The
divergence minimization objective thus becomes
kl
d
d
X
1 X
1
[ (Σ−
log λ i ].
1 ) ii λ i + log det(Σ1 ) −
2 i=1
i =1
Taking its derivative with respect to each λ i and setting it to zero, we have
1
−1
(Σ −
1 ) ii = λ i .
1
−1
The optimal Σ0 thus should be diagonal with Σ̂0,ii = ((Σ−
for i = 1, 2, ..., d . In this sense,
1 ) ii )
mean field (factorizable) approximation matches the precision matrix at the mode.
Moreover, by the inequality (Amir-Moez and Johnston, 1969; Beckenbach and Bellman, 2012)
Y −1
1
1
det(Σ−
(Σ1 ) ii = det(Σ̂−
1 )≤
0 ),
i
we have
1
H(Σ̂−
0 )=
1
1
log((2π e)d · det(Σ̂0 )) ≤ log((2π e)d · det(Σ1 )) = H(Σ1 ).
2
2
42
H
Proof of Corollary 10
We only need to verify the local asymptotic normality of L n (µ ; x) here. By Equation (2) of Westling
P
and McCormick (2015), we know the variational log likelihood writes L n (µ ; x) = i m(µ ; x i ). We
Taylor-expand it around the true value µ0 :
s
L n (µ 0 + p ; x )
n
X
s
= m(µ0 + p ; x i )
n
i
X
X
s X
1
D µ m(µ0 ; x i ) + s> [ D µ2 m(µ0 ; x i )] s.
= m(µ0 ; x i ) + p
n
n i
i
i
(119)
(120)
(121)
Due to X i ’s being independent and identically distributed, we have
P
p
1 X
i D µ m(µ0 ; x i ) d
D µ m(µ0 ; x i ) = n ·
→ N (0, B(µ))
p
n
n i
under Pµ0 , where B(µ) = EPµ0 [D µ m(θ ; x)D µ m(µ ; x)> ]. The convergence in distribution is due to
central limit theorem. The mean zero here is due to conditions B2, B4, and B5 of Westling and
McCormick (2015) (See point 4 in the first paragraph of Proof of Theorem 2 in Westling and
McCormick (2015) for details.)
By strong law of large numbers, we also have
P θ0
1X 2
D µ m(µ0 ; x i ) → EPθ0 [D µ2 m(µ0 ; x i )].
n i
This gives the local asymptotic normality, for s in a compact set,
s
1
L n (µ0 + p ; x) = L n (µ0 ; x) + s> ΣY − s> Σ s + o P (1),
2
n
where
Y ∼ N (0, A (µ0 )−1 B(µ0 ) A (µ0 )−1 ),
and
Σ = A (µ0 ) = EPθ0 [D µ2 m(µ0 ; x i )]
is a positive definite matrix.
The consistent testability assumption is satisfied by the existence of consistent estimators. This
is due to Theorem 1 of Westling and McCormick (2015).
The corollary then follows from Theorem 5 and Theorem 6 in Section 3.
I
Proof of Corollary 11
We first verify the local asymptotic normality of the variational log likelihood:
`(β, σ2 ) := sup `(β, σ2 , µ, λ)
(122)
µ, λ
43
= sup
n
m X
X
{Yi j(β0 +β1 X i j +µ i )−exp(β0 +β1 X i j +µ i +λ i /2) − log(Yi j !)}
µ,λ i =1 j =1
−
n
m
m
1X
m
1 X
(µ2i + λ i ) +
log(λ i ).
log(σ2 ) + − 2
2
2 2σ i=1
2 i=1
(123)
We take the Taylor expansion of the variational log likelihood at the true parameter values
β00 , β01 , (σ2 )0 :
s
t
r
`(β00 + p , β01 + p
, (σ2 )0 + p )
m
mn
m
∂
∂
s
t
=`(β00 , β01 , (σ2 )0 ) + p
`(β00 , β01 , (σ2 )0 ) + p
`(β00 , β01 , (σ2 )0 )
m ∂β0
mn ∂β1
(124)
s2 ∂2
t2 ∂2
r ∂
0 0
2 0
0 0
2 0
`
(
β
,
β
,
(
σ
)
)
+
`
(
β
,
β
,
(
σ
)
)
+
`(β00 , β01 , (σ2 )0 )
+p
0 1
0 1
m ∂β20
mn ∂β21
m ∂σ2
+
r 2 ∂2
`(β00 , β01 , (σ2 )0 ) + o P (1).
m ∂(σ2 )2
(125)
Denote µ̂ i and λ̂ i as the optimal µ and λ at the true values (β00 , β01 , (σ2 )0 ). Also write Yi· =
Pn
Pn
exp(β0 + β1 X i j ). Let β̂1 , β̂0 , σ̂2 be the maximizers of the `. Hence,
j =1 Yi j and B i =
Pn j=1
we write B̂ i = j=1 exp(β̂0 + β̂1 X i j ). Finally, the moment generating function of X writes
φ( t) = E{exp( tX )}.
The cross terms are zero due to Equation (5.21), Equation (5.29), Equation (5.37), and Equation
(5.50) of Hall et al. (2011).
We next compute each of the six derivatives terms.
For the first term
∂
`(β00 , β01 , (σ2 )0 ),
∂β0
we have
1 ∂
`(β00 , β01 , (σ2 )0 )
p
m ∂β0
m
1 X
1
=p
[Yi· − B i exp(µ̂ i + λ̂ i )]
2
m i=1
m
m
X
1 X
1
1
1
[Yi· − B̂ i exp(µ̂ i + λ̂ i )] − p (β0 − β̂0 )
B̂ i exp(µ̂ i + λ̂ i )
=p
2
2
m i=1
m
i =1
m
X
1
1
− p (β1 − β̂1 )
B̂ i exp(µ̂ i + λ̂ i ) + o P (1)
2
m
i =1
m X
n
p 1 X
1
= m
((β0 − β̂0 ) + (β1 − β̂1 ) X i j ) exp(β̂0 + β̂1 X i j ) exp(µ̂ i + λ̂ i ) + o P (1)
m i=1 j=1
2
(126)
(127)
(128)
(129)
m X
n
p 1 X
1
= m
((β1 − β̂1 )( X i j − γ(β01 )) + Ū ) exp(β̂0 + β̂1 X i j ) exp(µ̂ i + λ̂ i ) + o P (1)
m i=1 j=1
2
(130)
m
p
1 X
1
= mŪ
B̂ i exp(µ̂ i + λ̂ i ) + o P (1)
m i=1
2
(131)
m
p
1 X
1
= mŪ
B̂ i exp(µ̂ i + λ̂ i ) + o P (1)
m i=1
2
(132)
44
d
→N (0, (σ2 )0 )[exp(β00 )E exp(β1 X i j + U i )] + o P (1)
(133)
1
=N (0, (σ2 )0 )[exp(β00 − (σ2 )0 )φ(β01 )].
2
(134)
The first equality is due to differentiation with respect to β0 . The second equality is due to
Taylor expansion around the vfe. The third equality is due to Equation (3.5) of Hall et al.
(2011). The fourth equation is due to Equation (5.21) of Hall et al. (2011). The fifth equation
is due to Equation (5.1) of Hall et al. (2011). The sixth equation is due to the weak law of law
numbers and Slutsky’s theorem together with Equation (3.4), Equation (5.16), and Equation
(5.18) of Hall et al. (2011). The seventh equality is due to the equation below Equation (5.80) of
Hall et al. (2011).
We then compute the fourth term.
=
1 ∂2
`(β00 , β01 , (σ2 )0 )
m ∂β20
(135)
m
1 X
1
B̂ i exp(µ̂ i + λ̂ i ) + o P (1)
m i=1
2
(136)
1
P
→ exp(β0 − (σ2 )0 )φ(β01 ).
2
(137)
This step is due to a similar computation to above. The last step is due to weak law of large
numbers and the equation below Equation (5.80) of Hall et al. (2011).
We now compute the third term.
1 ∂
`(β00 , β01 , (σ2 )0 )
p
m ∂σ2
m
X
m
1
1
= p [−
+
(µ̂2 + λ̂ i )]
2
0
2
0
2
2{(σ ) } i=1 i
m 2(σ )
1
σ̂2
)(1
−
)
2(σ2 )0
(σ2 )0
p
1
=
m(σ̂2 − (σ2 )0 )
2
0
2
2{(σ ) }
1
d
→N (0, 2{(σ2 )0 }2 )
.
2{(σ2 )0 }2
p
= − m(
(138)
(139)
(140)
(141)
(142)
The first equality is due to differentiation with respect to σ2 . The second equality is due to
Equation (5.3) of Hall et al. (2011). The third equality is rearranging the terms. The fourth
equation is due to Equation (3.6) of Hall et al. (2011).
We then compute the sixth term.
1 ∂2
`(β00 , β01 , (σ2 )0 )
m ∂(σ2 )2
m
1
1
1
1 X
=[−
(µ̂2 + λ̂ i )]
+
2
0
2
2
0
3
2 {(σ ) }
{(σ ) } m i=1 i
=[−
1
1
σ̂2
+
] + o P (1)
2 {(σ2 )0 }2 {(σ2 )0 }3
45
(143)
(144)
(145)
P
→
1
.
2{(σ2 )0 }2
(146)
This is due to a similar computation to above. The last step is due to Equation (3.6) of Hall et al.
(2011) and the weak law of large numbers.
We next compute the second term.
∂
1
`(β00 , β01 , (σ2 )0 )
mn ∂β1
n
m X
1
1 X
X i j (Yi j − exp(β00 + β01 X i j + µ̂ i + λ̂ i ))
=p
2
mn i=1 j=1
p
=p
(147)
(148)
n
m X
X
1
X i j (Yi j − exp(β̂0 + β̂1 X i j + µ̂ i + λ̂ i ))
2
mn i=1 j=1
1
+ (β00 − β̂0 ) p
+ (β01 − β̂1 ) p
m X
n
X
1
mn i=1 j=1
1
X i j (− exp(β̂0 + β̂1 X i j + µ̂ i + λ̂ i ))
2
m X
n
X
1
1
X i2j (− exp(β̂0 + β̂1 X i j + µ̂ i + λ̂ i )) + o P (1)
2
mn i=1 j=1
(149)
m X
n
p
1 X
1
= mn
((β0 − β̂0 ) + (β01 − β̂1 ) X i j ) X i j (− exp(β̂0 + β̂1 X i j + µ̂ i + λ̂ i ))
mn i=1 j=1 0
2
+ o P (1)
m X
n
p
1 X
1
= mn
((β01 − β̂1 )( X i j − γ(β01 )) + Ū ) X i j (− exp(β̂0 + β̂1 X i j + µ̂ i + λ̂ i ))
mn i=1 j=1
2
(150)
(151)
+ o P (1)
1
d
→N (0, τ2 ) exp(β00 − σ2 )φ00 (β01 ).
2
(152)
The first equality is due to differentiation with respect to β1 . The second equality is due to
Taylor expansion around vfe. The third equality is due to Equation (3.4) of Hall et al. (2011).
The fourth equality is due to Equation (5.16), Equation (5.18), and Equation (5.21) of Hall et al.
(2011). The fifth equation is due to the weak law of law numbers and Slutsky’s theorem together
with Equation (3.5) and the equation below Equation (5.80) of Hall et al. (2011).
We lastly compute the fifth term.
1 ∂2
`(β00 , β01 , (σ2 )0 )
mn ∂β21
(153)
=
m X
n
1 X
1
X i2j (exp(β00 + β01 X i j + µ̂ i + λ̂ i ))
mn i=1 j=1
2
(154)
=
m X
n
1 X
X 2 (exp(β00 + β01 X i j + U i ))
mn i=1 j=1 i j
(155)
1
P
→ exp(β00 − σ2 )φ00 (β01 ).
2
(156)
This is due to a similar computation to above. The last step is due to the weak law of large
numbers.
46
The calculation above gives the full local asymptotic expansion of `(β0 , β1 , σ2 ).
The consistent testability assumption is satisfied by the existence of consistent estimators. This
is due to Theorem 3.1 of Hall et al. (2011). The corollary then follows directly from Theorem 5
and Theorem 6 in Section 3.
47
References
Amir-Moez, A. and Johnston, G. (1969). On the product of diagonal elements of a positive matrix.
Mathematics Magazine, 42(1):24–26.
Beckenbach, E. F. and Bellman, R. (2012). Inequalities, volume 30. Springer Science & Business
Media.
Braides, A. (2006). A handbook of Γ-convergence. Handbook of Differential Equations: stationary
partial differential equations, 3:101–213.
Dal Maso, G. (2012). An introduction to Γ-convergence, volume 8. Springer Science & Business
Media.
Hall, P., Pham, T., Wand, M. P., Wang, S. S., et al. (2011). Asymptotic normality and valid
inference for Gaussian variational approximation. The Annals of Statistics, 39(5):2502–2532.
Klartag, B. (2007). A central limit theorem for convex sets. Inventiones mathematicae, 168(1):91–
131.
Kleijn, B., Van der Vaart, A., et al. (2012). The Bernstein-von-Mises theorem under misspecification. Electronic Journal of Statistics, 6:354–381.
Lu, Y., Stuart, A. M., and Weber, H. (2016). Gaussian approximations for probability measures
on Rd . arXiv preprint arXiv:1611.08642.
Piera, F. J. and Parada, P. (2009). On convergence properties of Shannon entropy. Problems of
Information Transmission, 45(2):75–94.
Westling, T. and McCormick, T. H. (2015). Establishing consistency and improving uncertainty
estimates of variational inference through M-estimation. arXiv preprint arXiv:1510.08151.
48
| 10math.ST
|
SPECTRUM HOLE PREDICTION BASED ON HISTORICAL
DATA: A NEURAL NETWORK APPROACH
Bara’u Gafai Najashi1*, Feng Wenjiang2 Mohammed Dikko Almustapha3
1,2
College of Communication Engineering, Chongqing University,
Chongqing, China, Postal Code 400044
174 Shazheng Street, Shapingba District, Chongqing
3
Department of Electrical and Computer Engineering, Ahmadu Bello University Zaria, Nigeria
ABSTRACT
The concept of cognitive radio pioneered by Mitola
promises to change the future of wireless communication
especially in the area of spectrum management. Currently,
the command and control strategy employed in spectrum
assignment is too rigid and needs to be reviewed. Recent
studies have shown that assigned spectrum is underutilized
spectrally and temporally. Cognitive radio provides a viable
solution whereby licensed users can share the spectrum
with unlicensed users opportunistically without causing
interference. Unlicensed users must be able to sense
weather the channel is busy or idle, failure to do so will
lead to interference to the licensed user. In this paper, a
neural network based prediction model for predicting the
channel status using historical data obtained during a
spectrum occupancy measurement is presented. Genetic
algorithm is combined with LM BP for increasing the
probability of obtaining the best weights thus optimizing
the network. The results obtained indicate high prediction
accuracy over all bands considered.
Keywords: Cognitive radio, spectrum sensing, prediction
model, neural network, genetic algorithm
1. INTRODUCTION
The field of wireless communication has witnessed
tremendous evolution over the years. The demand for
bandwidth has grown with the introduction of several
wireless standards. This advancement has brought
about a perceived spectrum scarcity. Recent studies
have shown that there is ample spectrum available
that is not being utilized. So the problem can be said
to be spectrum under utilization and not spectrum
scarcity. Cognitive radio is seen as a great contender
for better spectrum utilization because of its ability to
make spectrum sharing possible. A situation whereby
a licensed user is allowed to share the spectrum with
an unlicensed user also called a secondary user
opportunistically without causing interference will
surely increase the spectrum utilization. In order to
achieve this, the unlicensed user must be able to
sense when a channel is idle so that it can use, or
busy so that it can vacate it. It has therefore become
imperative to develop models capable of predicting
the channel state accurately.
Figure 1: Waterfall figure showing sparsely used spectrum with
low utilization level
Several spectrum prediction models have been
proposed. Most of these works employ Markov
chains for the prediction problem and assume that the
primary user traffic follows a Poisson process
[1][2][3]. In [4], a neural network based spectrum
prediction using Multilayer Perceptron MLP was
proposed. An hour long data was divided into 60
slots and converted into a time series. This was used
to train and test the network. A practical spectrum
behavior learning method based on MLP artificial
neural network (ANN) was introduced in [5].
Performances were evaluated with an existing 7-days
spectrum data set from a previous measurement,
which was conducted in a metro city located in south
China. In this paper, a neural network based
prediction model is presented. The training and test
data were obtained from [10]. Unlike other neural
network based prediction methods, a 12-hour long
data was used for this work. The problem of weight
selection which is common in neural networks was
tackled using GA.
their simplicity in terms of training. Whereas other
prediction schemes require continuous training,
neural network are trained once in an offline fashion
when the observed process is stationary [6]. The key
element of this paradigm is the novel structure of the
information processing system. It is composed of a
large number of highly interconnected processing
elements (neurons) working in unison to solve
specific problems. An ANN is configured for a
specific application, through a learning process.
Learning in biological system involves adjustment to
the synaptic connections that exist between the
neurons [7]. There are basically 2 types of
architectures namely, the feed forward and the
recurrent networks. The feed forward architecture is
where the connections are feed forward this means
that a neuron will not accept an input from a neuron it
has previously feed. The recurrent networks are those
that allow feedback between neurons. For this work
the feed forward architecture is used.
2.1
Figure 2: The red dot shows a bird’s eye view of Gwarinpa, the
measurement location in [10]
The rest of the paper is organized as follows. Section
2 provides a review of neural network, Multi-layer
Perceptron and application of genetic algorithm in
optimizing the interconnecting weights. In Section 3,
the model used in this work is presented. Section 4
contains results obtained; conclusions are drawn from
section 5.
2.
Multilayered Perceptron (MLP)
An MLP network is a multi layered structure
consisting of an input, output and a few hidden
layers. Each layer (excluding the input layer) consists
of a number of computing units called neurons which
calculate a weighted sum of the inputs and perform a
nonlinear transformation on the sum. . The nonlinear
transform is implemented using sigmoid functions
(e.g. a hyperbolic tangent function). Neurons
belonging to different layers are connected through
adaptive weights. The output of a neuron j in the ith
layer, denoted by yij, can be represented as:
NEURAL NETWORK
=
Hidden layer
1 − exp (− )
1 + exp (− )
(1)
Input layer
Output layer
Figure 3: A basic neural network architecture showing the input,
hidden and output layers
Artificial Neural Networks (ANNs) are non-linear
mapping structures based on the function of the
human brain. They are powerful tools for modeling
and are widely used in prediction problems due to
Where
=∑
is the weighted sum of the
inputs coming from the output of the neurons in the
( − 1)
layer using the adaptive weights (or
parameters)
connecting the neuron in the
layer and neuron in the ( − 1) layer. Due to the
nonlinear transform using the hyperbolic tangent
function on
in (1), lies in the range of [−1, +1].
If the inputs come from the input layer,
is
calculated using the corresponding inputs instead of
.The total number of inputs in the input layer is
referred to as the order of the MLP network and is
denoted by . The number of hidden layers and the
number of neurons in each layer depend on the
application [8].
2.2
MLP Training
Once the architecture of a neural network has been
determined for a particular application, the network is
then ready for training. There are two types of
training namely supervised and unsupervised
training. In supervised learning, the input and output
data is provided while in unsupervised learning the
network has to make its decision on the input without
any provision of the output. Over the years, MLP also
known as a feed forward neural network has been
applied successfully in several combinatorial
problems. It works extremely well as a universal
approximation in which input signal propagates in
forward direction [11].
J
J netk
netk
k
wkj netk wkj
wkj
Where sensitivity of
k
J
J zk
(t k zk ) f netk )
net k
zk net k
wkj k y j (tk zk ) f netk ) y j
J
J y j net j
w ji y j net j w ji
J
y j y j
(9)
zk netk
netk y j
(10)
(tk zk ) f netk ) w jk
(11)
(t
k
zk )
k 1
(2)
c
Where is the desired output and
is the actual
output, t and z are the target network and output
vectors with length c respectively. W represents the
weights in the network. The initial weights are
normally selected randomly and are updated in a way
that reduces the error. This method is based on the
gradient descent algorithm.
J
J
or ( w pq
)
w
w pq
(3)
is the learning rate, which indicates the size of the
change in weights. This iterative algorithm requires
taking a weight vector at iteration m and updating it
as
w(m 1) w(m) w(m)
(7)
(8)
1 c
2
2 (tk zk )
k 1
c
c
w
(6)
Similarly, considering the input-hidden weights
=
1
1
(tk zk )2 (t z )2
2 k 1
2
can be defined to be
Therefore, the weight update (learning rule) for the
hidden-output weights given as
The general approach in learning involves presenting
an input training pattern to an untrained network
which is then used to get the output. The error is
some scalar function of the weights that is minimized
when the network outputs match the desired outputs.
Since the actual output is known beforehand, the
weights are adjusted to minimize the error so that the
desired output could be reached. A Constant number
1/2 is added to do the mathematical derivation
conveniently [5].
J ( w)
(5)
(4)
Where m indexes the particular pattern presentation.
Considering the hidden-output weights
and
applying chain rule differentiation
k 1
Thus the sensitivity of the hidden neuron can be
given as
j
c
J
J zk netk y j
net j
k 1 z k net k y j net j
(12)
c
netk y j
k
y j net j
k 1
c
f netk ) ( k wkj )
(13)
k 1
Finally the learning rule for the input-hidden is given
as
c
wkj xi j xi f net j ) ( wkj k ) (14)
k 1
By substituting equation (7) and (14) into equation
(4), the updated weights could be obtained. The
process will continue until the desired error goal is
achieved.
1 c
J (W ) (tk zk ) 2
2 k 1
2.3
(15)
Levenberg-Marquardt BP Algorithm
The Levenberg-Marquardt
back
propagation
algorithm has been shown to be good at training
moderately sized feed forward neural networks [12].
It is an approximation of Newton’s method while
back propagation with gradient descent technique is a
steepest descent algorithm. It updates the weights
using the expression below
∆
=−
+
( )
( )
∇ ( )
(16)
Where ( ) is the Jacobian matrix of the error
( ) evaluated in , and is the identity
vector
( ) is the error of the
matrix. The error vector
network for pattern , that is
( )=
−
( )
(17)
The parameter is increased or decreased at each
step. If the error is reduced, then is divided by a
factor and multiplied by in other case [13].
2.4
Genetic Algorithm
Genetic algorithms are stochastic search techniques
that guide a population of solutions towards an
optimum using the principles of evolution and natural
genetics [7]. Genetic algorithms are inspired by the
evolution of populations. In a particular environment,
individuals who better fit the environment will be
able to survive and hand down their chromosomes to
their descendants, while less fit individuals will
become extinct. The aim of genetic algorithms is to
use simple representations to encode complex
structures and simple operations to improve these
structures. Genetic algorithms therefore are
characterized by their representation and operators.
In the original genetic algorithm an individual
chromosome is represented by a binary string. The
bits of each string are called genes and their varying
values alleles. A group of individual chromosomes
are called a population. [5]. Genetic algorithms are
especially capable of handling problems in which the
objective function is discontinuous or non
differentiable, non-convex, multimodal or noisy.
Since the algorithms operate on a population instead
of a single point in the search space, they climb many
peaks in parallel and therefore reduce the probability
of finding local minima [7]. Genetic algorithm
involves 3 main operators: these include selection,
cross-over (recombination) and mutation.
Selection is usually the first step in a genetic
algorithm process. It involves selecting chromosomes
from the population to crossover and produce
offspring. The selection is base on Darwin’s
evolution theory “survival of the fittest” that the best
chromosomes should survive and produce offspring.
Extracting a subset of genes from a population
according a definition of quality (fitness function) is
what the selection process is all about. Fitness
function is the measure used to measure the
optimality or closeness of the selected chromosomes
to the desired ones. Many selection methods exist;
they include Roulette wheel selection, Boltzmann
selection, Tournament selection, rank selection, and
steady state selection.
Crossover is usually the second stage. It’s basically a
process whereby two chromosomes combine to
produce an offspring. The concept behind crossover
is that an offspring may possess better qualities than
its parents if the best qualities are transferred to the
offspring. Crossover occurs during evolution
according to a user-definable probability. Crossover
selects genes from parent chromosomes and create a
new offspring. Examples of crossover operators
include one point crossover, two-point crossover,
arithmetic and heuristic crossovers.
Mutation is an important step in genetic algorithm it
helps in ensuring that the population is not stagnated
at local minima. It is widely regarded as the last stage
in the GA process. It helps in ensuring generic
diversity from generation of population to the next.
The process involves altering the gene values of one
chromosome or even more. This can lead to the
addition of totally new genes in the gene pools which
will subsequently lead to a better result than initially
obtained. Commonly used mutation operators include
Flip-bit, boundary, non-uniform, uniform and
Gaussian [14].
2.5
Optimizing BP-ANN with GA
Two factors generally influence modeling a network
during the learning and training session. One is the
initial interconnecting weights of the network, and
another is their modified quantities [9]. Generally the
initial interconnecting weights of BP ANN are often
stochastically and blindly produced, this might cause
the network to run into partial optimization and
therefore decrease the probability to obtain the
optimal solutions. Moreover, because the Delta rule
is always adopted to modify the interconnecting
weights of BP ANN, the convergence velocity is
always slow, or sometimes the network does not
even converge. These shortages of BP ANN are
quite necessary to be optimized and improved [9].
The problem of partial minimum of a BP ANN can
be solved by adjusting the initial interconnecting
weights of the network. This can be achieved through
the application of Genetic algorithm because the
problem is a non linear problem. GA is a nonlinear
optimization method that has very strong ability of
global searching.
weights and thresholds of BP ANN is considered to
be an individual of the population and coded to be a
chromosome. With the errors of BP ANN as the
adaptive function, all the chromosomes are changed
using GA operators (selection, crossover and
mutation). The chromosomes that correspond to the
best initial interconnecting weights and thresholds of
BP ANN are gradually evolved. Finally, the initial
interconnecting weights and thresholds optimized
with GA are used into the learning and training of BP
ANN again, and a better BP ANN model can be
constructed. Because the initial interconnecting
weights and thresholds of BP ANN are globally
optimized by utilizing GA, an appropriate search
space can be located in the complex solution space,
and the partial minimum or non-convergence can be
avoided.
3. MODEL IMPLEMENTATION
Based on the 12 hour spectrum occupancy
measurement conducted in [10], a total of 2,700
samples were obtained since each complete sweep
takes 16 seconds. These raw power level data were
processed and feed to the neural network. The input
and output data were randomly selected from three
different services namely: GSM band, broadcasting
band, and the 3G cellular band. The input is a set of
power level values obtained from the spectrum
analyzer during the measurement for a single
channel. In [10], the channel size was given as 200
KHz. The binary hypothesis defined as
=1
if ≥ , otherwise = 0, where are the power
values from the spectrum analyzer and m is the
decision threshold which was applied to get output
data. The training and testing was achieved using
data from the 875 MHz broadcasting band, GSM
downlink 905 MHz, 3G downlink 1865 MHz and
GSM uplink 890 MHz band. The entire process can
be summarized in the following steps:
Figure 4: The flowchart of BP-MLP weight optimization using
GA
After the structure and parameters of an ANN such as
the number of layers and the number of neurons in
every layer are determined, the approximate ranges
of interconnecting weights and thresholds can be
computed by using the algorithm of back
propagation. Then, an initial population in the field of
solution can be randomly produced by adopting
genetic algorithm. Each group of the interconnecting
SERVICES
Broadcasting
GSM Uplink
GSM Downlink
3G 1800 Downlink
3G 1900 Downlink
FREQUENCY
700-806 MHz
890-895 MHz
905-910 MHz
1865-1880 MHz
1883-1890 MHz
Table 1: A table of different services considered
Step 1: The time series power level data is converted
into a binary time series of 0’s and 1’s using
thresholding.
Step 2: The data from step one is then grouped into
the different services examined. The training and
testing data sets are then obtained.
Step 3: The training and testing data for the service to
be predicted is selected, architecture and training
parameters are also selected.
Step 4: The network is trained using MLP-BP and
weight selection is optimized through the use of GA.
Step 5: The network is tested with the test data to
ascertain the prediction accuracy.
The simulation was conducted using MATLAB
version r2012a. After a trial and error process, the
number of neurons was set at 10 which were found to
be adequate. GA and MLP- BP were used during the
learning process and the prediction accuracy was
determined across the bands considered.
Parameter
Number of Hidden Neurons
Training samples
Testing samples
Transfer Function
Training Algorithm
Value
10
2700
2700
Sigmoid
LM+GA
Table 2: Parameters used in modeling
4. RESULTS AND DISCUSSION
In this section, the performance of the designed
neural network is given. The main attraction of using
neural network based spectrum prediction lies in the
fact that cognitive radios can learn and train itself
from historical information obtained by the cognitive
radio without redesigning the whole system
completely as is the case with other methods. Unlike
other models such as the Markov chain approach to
spectrum prediction, the neural network model need
only to be updated with the most recent data as its
input. This approach saves power, sensing time and
manpower. Machine learning has already been
proposed as an integral part of future cognitive
radios, the high prediction accuracy realized in this
model will greatly reduce the time required to sense
whole bands. In addition, the processing power
required at the base station will be also reduced. The
high prediction accuracy will also drastically reduce
the rate of interference between the primary and
secondary users in a cognitive radio scenario. It has
therefore become paramount to improve machine
learning based spectrum hole prediction accuracy in
order to attain high prediction accuracy.
For the
purpose of this work, five popular services were
considered. The GSM 900 uplink channels licensed
to Etisalat had a mean prediction error of about 0.035
over five channels that were selected randomly. We
observed a mean prediction error of about 0.005 in
the GSM 900 downlink band. 0.0179 prediction
error was recorded in the 3G downlink band. 0.0004
and 0.0007 were observed in the broadcasting band
and 3G downlink band (licensed to Starcomms
Nigeria) respectively. The difference between the
uplink and downlink prediction accuracy can be due
to the fact that the base station is continuously
transmitting information to the mobile users which is
not the case with the uplink band. It has already been
stated in [5] that the sparsely used uplink band can
provide an inspiration for future deployment of some
form of cognitive radio in the future. The similarity
in both GSM and 3G downlink bands can be due to
the similar nature of the bands in terms of their
behavior. The 1900 downlink band had an error of
around 0.0007 which is very low. This value might
be deceiving. This band is currently being used by
Starcomms Nigeria which uses a CDMA based
technology for their services. The spectrum analyzer
might not be able to accurately detect the presence or
absence of a signal during the spectrum occupancy
measurement, therefore the low value experienced
might be misleading because of the low power nature
of the signals in this band. In a worst case scenario
whereby a signal suffers from wireless effects such
shadowing, multipath, and attenuation the power
considerably degrades making detection almost
impossible. Overall, the performance of neural
network based spectrum prediction using GA for
weight optimization is good.
97.00%
96.50%
96.00%
95.50%
Figure 5: Spectrum Prediction Accuracy for 900 GSM Uplink
SPECTRUM BAND
GSM 900 UPLINK
GSM 900 DOWNLINK
3G 1900 DOWNLINK
BROADCASTING
3G 1800 DOWNLINK
100.20%
100.00%
99.80%
99.60%
99.40%
99.20%
PREDICTION ERROR
0.035
0.005
0.0179
0.0004
0.0007
Table 1: Prediction error for the services investigated
ChannelChannel ChannelChannel
1
2
3
4
Figure 6: Spectrum Prediction Accuracy for 900 GSM Downlink
102.00%
100.00%
98.00%
96.00%
94.00%
92.00%
ChannelChannel ChannelChannel
1
2
3
4
Figure 7: Spectrum Prediction Accuracy for 3G 1800 Downlink
99.98%
99.96%
99.94%
99.92%
99.90%
99.88%
99.86%
99.84%
ChannelChannel ChannelChannel
1
2
3
4
Figure 8: Spectrum Prediction Accuracy for broadcasting band
100.20%
100.00%
99.80%
99.60%
99.40%
99.20%
5. CONCLUSION
The importance of spectrum sensing is key to the
development and eventual implementation of
cognitive radios. Due to the random nature of
spectrum which varies from one place to another,
there is need to better understand this scarce resource
so that its behavior can be predicted with little or no
error. This knowledge could be obtained through
extensive measurements that can help in developing
models capable of predicting this behavior.
Ultimately, these models could help in reducing
sensing time and power consumption in cognitive
radios. Unlike other models, neural network based
models need not to be built from scratch once the
model has been designed. Since it has already been
suggested that cognitive radio will utilize a geolocation database which it will use to query
information with respect to its current environment,
machine learning based method could prove
invaluable because of its ability to update itself
whenever new information is available. Neural
network based spectrum prediction has already been
explored as was stated earlier; the aim of this paper
was to explore the possibility of optimizing the initial
weights in a neural network because they tend to run
into partial optimization therefore reducing the
accuracy of the deigned model. In this paper, a
neural network based spectrum prediction model is
presented. Unlike other models, the issue of local
minimum in selecting interconnecting weights was
addressed by using Genetic algorithm. Results
indicate high prediction accuracy across all bands
considered.
Acknowledgement
ChannelChannel ChannelChannel
1
2
3
4
Figure 9: Spectrum Prediction Accuracy for 1900 3G Downlink
Our thanks go to the College of Communication
Engineering, Chongqing University for their support
and understanding during this work.
REFERENCES
1.
2.
3.
4.
5.
6.
7.
8.
9.
I. A. Akbar and W. H. Tranter, “Dynamic
spectrum allocation in cognitive radio using
hidden Markov models: Poisson distributed
case,” in Proc. IEEE Southeast Con, 2007,
pp. 196–201
C.-H. Park, S.-W. Kim, S.-M. Lim, and M.S. Song, “HMM based channel status
predictor for cognitive radio,” in Proc. 2007
Asia-Pacific
Microwave
Conference
(APMC), Dec. 2007, pp. 1–4.
V. K. Tumuluru, P. Wang, and D. Niyato,
“Channel status prediction for cognitive
radio networks,” Wireless Commun. and
Mobile Comput, Aug. 2010.
Zhao Jianli, Wang Mingwei, and Yuan
Jinsha, “Based on neural network spectrum
prediction of Cognitive Radio” Electronics,
Communications and Control International
Conference(IECC),
2011
Liang Yin, SiXing Yin, Weijun Hong,
ShuFang Li, “Spectrum Behavior Learning
in Cognitive Radio Based on Artificial
Neural Network” Military Communication
Conference, Track 1, Waveforms and Signal
Processing, 2011.
Vamsi Krishna Tumuluru, Ping Wang and
Dusit Niyato, “A Neural Network Based
Spectrum Prediction Scheme for Cognitive
Radio” IEEE International Conference on
Communication (ICC), 2010.
Ms. Dharmistha, D. Vishwakarma, “Genetic
Algorithm based Weights Optimization of
Artificial Neural Network”, International
Journal of Advanced Research in Electrical,
Electronics
and
Instrumentation
Engineering, August 2012.
Vamsi Krishna Tumuluru, Ping Wang and
Dusit Niyato,” A Neural Network Based
Spectrum Prediction Scheme for Cognitive
Radio”,
Qiuwen Zhang, Cheng Wang,” Using
Genetic Algorithm to Optimize Artificial
Neural Network: A Case Study on
Earthquake
Prediction”,
International
Conference on Genetic and Evolutionary
Computing
10. Bara’u Gafai Najashi,”Insight into Spectrum
Occupancy in Nigeria”, International
Journal of Computer Science Issues”, Vol
10, January 2013.
11. Shah, Habib, Rozaida Ghazali, and Nazri
Mohd Nawi. "Using artificial bee colony
algorithm for MLP training on earthquake
time series data prediction. "arXiv preprint
arXiv:1112.4628 (2011).
12. Kuryati Kipli, Mohd Saufee Muhammad,
Sh. Masniah Wan Masra, Nurdiani Zamhari,
Kasumawati Lias, Dayang Azra Awang
Mat,” Performance of Levenberg-Marquardt
Backpropagation for Full Reference Hybrid
Image Quality Metrics”, Proceedings of
International
Conference
of
MutiConference of Engineers and Computer
Scientists”, March 2012.
13. Alba, Enrique, and J. Chicano. "Training
neural networks with GA hybrid
algorithms." Genetic and Evolutionary
Computation–GECCO
2004.
Springer
Berlin/Heidelberg, 2004.
14. Goldberg, D. E.: Genetic Algorithms in
Search, Optimization, and Machine Learning
[M]. Addison- Wesley Pub. Co., 1989.
| 9cs.NE
|
1
arXiv:1711.03272v1 [cs.LO] 9 Nov 2017
Go with the Flow: Compositional Abstractions for
Concurrent Data Structures (Extended Version)∗
SIDDHARTH KRISHNA, New York University, USA
DENNIS SHASHA, New York University, USA
THOMAS WIES, New York University, USA
Concurrent separation logics have helped to significantly simplify correctness proofs for concurrent data
structures. However, a recurring problem in such proofs is that data structure abstractions that work well in
the sequential setting are much harder to reason about in a concurrent setting due to complex sharing and
overlays. To solve this problem, we propose a novel approach to abstracting regions in the heap by encoding
the data structure invariant into a local condition on each individual node. This condition may depend on a
quantity associated with the node that is computed as a fixpoint over the entire heap graph. We refer to this
quantity as a flow. Flows can encode both structural properties of the heap (e.g. the reachable nodes from the
root form a tree) as well as data invariants (e.g. sortedness). We then introduce the notion of a flow interface,
which expresses the relies and guarantees that a heap region imposes on its context to maintain the local
flow invariant with respect to the global heap. Our main technical result is that this notion leads to a new
semantic model of separation logic. In this model, flow interfaces provide a general abstraction mechanism for
describing complex data structures. This abstraction mechanism admits proof rules that generalize over a
wide variety of data structures. To demonstrate the versatility of our approach, we show how to extend the
logic RGSep with flow interfaces. We have used this new logic to prove linearizability and memory safety of
nontrivial concurrent data structures. In particular, we obtain parametric linearizability proofs for concurrent
dictionary algorithms that abstract from the details of the underlying data structure representation. These
proofs cannot be easily expressed using the abstraction mechanisms provided by existing separation logics.
CCS Concepts: • Theory of computation → Logic and verification; Separation logic; Concurrent
algorithms; Programming logic;
Additional Key Words and Phrases: memory safety, linearizability, flow interfaces, separation algebra
1
INTRODUCTION
With the advent of concurrent separation logics (CSLs), we have witnessed substantial inroads
into solving the difficult problem of concurrent data structure verification [Bornat et al. 2005;
da Rocha Pinto et al. 2014, 2016; Dinsdale-Young et al. 2013, 2010; Dodds et al. 2016; Gu et al. 2015;
Jung et al. 2015; Nanevski et al. 2014; O’Hearn 2004; Vafeiadis 2008; Vafeiadis and Parkinson 2007].
CSLs provide compositional proof rules that untangle the complex interference between concurrent
threads operating on shared memory resources. At the heart of these logics lies separation logic
(SL) [O’Hearn et al. 2001; Reynolds 2002]. The key ingredient of SL is the separating conjunction
operator, which allows the global heap memory to be split into disjoint subheaps. An operation
working on one of the subheaps can then be reasoned about compositionally in isolation using
so-called frame rules that preserve the invariant of the rest of the heap. To support reasoning about
complex data structures (lists, trees, etc.), SL is typically extended with predicates that are defined
∗ This work is funded in parts by NYU WIRELESS and by the National Science Foundation under grants MCB-1158273, IOS1339362, MCB-1412232, and CCF-1618059.
Authors’ addresses: Siddharth Krishna, New York University, USA, [email protected]; Dennis Shasha, New York
University, USA, [email protected]; Thomas Wies, New York University, USA, [email protected].
2017. XXXX-XXXX/2017/11-ART1 $15.00
https://doi.org/10.1145/nnnnnnn.nnnnnnn
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:2
Siddharth Krishna, Dennis Shasha, and Thomas Wies
inductively in terms of separating conjunction to express invariants of unbounded heap regions.
However, inductive predicates are not a panacea when reasoning about concurrent data structures.
One problem with inductive predicates is that the recursion scheme must follow a traversal of the
data structure in the heap that visits every node exactly once. Such definitions are not well-suited
for describing data structures with unbounded sharing and overlays of multiple data structures
with separate roots whose invariants have mutual dependencies [Hobor and Villard 2013]. Both of
these features are prevalent in concurrent algorithms (examples include B-link trees [Lehman and
Yao 1981] and non-blocking lists with explicit memory management [Harris 2001]).
Another challenge is that proofs involving inductive predicates rely on lemmas that show how the
predicates compose, decompose, and interact. For example, SL proofs of algorithms that manipulate
linked lists often use the inductive predicate lseg(x, y), which denotes subheaps containing a list
segment from x to y. A common lemma about lseg used in such proofs is that two disjoint list
segments that share an end and a start point compose to a larger list segment (under certain side
conditions). Unfortunately, these lemmas do not easily generalize from one data structure to another
since the predicate definitions may follow different traversal patterns and generally depend on
the data structure’s implementation details. Hence, there is a vast literature describing techniques
to derive such lemmas automatically, either by expressing the predicates in decidable fragments
of SL [Berdine et al. 2004; Bouajjani et al. 2012; Cook et al. 2011; Enea et al. 2017; Iosif et al. 2014;
Pérez and Rybalchenko 2011; Piskac et al. 2013; Reynolds et al. 2017; Tatsuta et al. 2016] or by using
heuristics [Brotherston et al. 2011; Chlipala 2011; Enea et al. 2015; Nguyen and Chin 2008; Pek
et al. 2014]. However, these techniques can be brittle, in particular, when the predicate definitions
involve constraints on data.
What appears to be missing in existing SL variants is an abstraction mechanism for heap regions
that is agnostic to specific traversal patterns, yet can still express inductive properties of the
represented data structure. To address this shortcoming, we here take a radically different approach
to abstraction in separation logic. Instead of relying on data-structure-specific inductive predicates,
we introduce a new abstraction mechanism that specifies inductive structural and data properties
uniformly but independently of each other, without fixing a specific traversal strategy.
As a first step, we describe data structure invariants in a uniform way by expressing them in
terms of a condition local to each node of the heap graph. However, this condition is allowed to
depend on a quantity associated with the node that is computed inductively over the entire graph.
We refer to this inductively defined quantity as a flow. An example of a flow is the function pc that
maps each node n to the number of paths from the root of the data structure to n. The property that
a given heap graph is a tree can then be expressed by the condition that for all nodes n, pc(n) = 1.
Flows may also depend on the data stored in the heap graph. For instance, in a search data structure
that implements a dictionary of key/value pairs, we can define the inset flow. The inset of a node n
is a set of keys. Intuitively, a key k is in the inset of n if and only if a search for k that starts from
the root of the data structure will have to traverse the node n. The inset flow can be used to express
the invariants of search data structure algorithms in a way that abstracts from the concrete data
structure implementation (i.e., whether the concrete data structure is a list, a tree, or some more
complicated structure with unbounded sharing) [Shasha and Goodman 1988].
We would like to reason compositionally about flows in SL using separating conjunctions. A
separating conjunction is defined in terms of a composition operator on the semantic models of SL
that yields a so-called separation algebra [Calcagno et al. 2007]. For the standard heap graph model
of SL, composition is disjoint graph union: G = G 1 ⊎ G 2 . To enable local reasoning about flows, the
composition operator needs to ensure that the flow on the resulting graph G maintains the local
conditions with respect to the flows on its constituents G 1 and G 2 . However, this is typically not the
case for disjoint union. For instance, consider again our example of the path-counting flow where
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended)
1:3
the flow condition bounds the path count of each node. Even if G 1 and G 2 individually satisfy the
bound on the path count, G 1 ⊎ G 2 may not since the union of the graphs can create cycles.
We therefore define a disjoint union operator on graphs that additionally asserts that the two
subgraphs respect their mutual constraints on the flow of the composite graph. Our key technical
contribution is to identify a class of flows that give rise to a separation algebra with this new
composition operator. This class is closed under product (which enables reasoning about data
structure overlays) and subsumes many natural examples of flows, including the ones given above.
To express abstract SL predicates that describe unbounded graph regions and their flows, we
introduce the notion of a flow interface. A flow interface of a graph G expresses the constraints
on G’s contexts that G relies upon in order to satisfy its internal flow conditions, as well as the
guarantees that G provides its contexts so that they can satisfy their flow conditions. In the example
of path-counting flows used to express “treeness”, the rely of the flow interface specifies how many
paths G expects to exist from the global root in the composite graph to each of the roots in G (there
can be many roots in G since it may consist of disconnected subtrees), and the guarantee specifies
how many paths there are between each pair of root and sink node in G.
The algebraic properties of flows that guarantee that flow composition is well-defined also give
rise to generic proof rules for reasoning about flow interfaces. These include rules that allow a
flow interface to be split into arbitrary chunks which can be modified and recomposed, enabling
reasoning about data structure algorithms that do not follow a fixed traversal strategy.
To demonstrate the usefulness of our new abstraction mechanism, we have instantiated rely/guarantee separation logic (RGSep) [Vafeiadis 2008; Vafeiadis and Parkinson 2007] with flow interfaces
(other CSL flavors can be extended in a similar fashion). We have used the new logic to obtain
simple correctness proofs of intricate concurrent data structure algorithms such as the Harris list
with explicit memory management [Harris 2001], which cannot be easily verified with existing
logics. Moreover, we show how the new logic can be used to formalize the edgeset framework
for verifying concurrent dictionary data structures [Shasha and Goodman 1988]. We apply this
formalization to proving linearizability [Herlihy and Wing 1990] of an algorithm template for
concurrent dictionary implementations. The template abstracts from the specifics of the underlying
data structure representation allowing it to be refined to diverse concrete implementations (e.g.
using linked lists, B+ trees, etc.). By using flow interfaces, the correctness proof of the template can
also abstract from the concrete implementation details, enabling proof reuse across the different
refinements. We are not aware of any other logic that provides an abstraction mechanism to support
this style of proof modularization at the level of data structure algorithms.
This is an extended version of a conference paper [Krishna et al. 2018], containing the following
additional material: §5 contains formal definitions of the syntactic shorthands we use in our
logic, the formal semantics of our new ghost commands, and some additional lemmas for proving
entailments; §A contains proof sketches of the full insert procedure and the delete procedure on the
Harris list; §B contains an instantiation of our dictionary framework to the B+ tree data structure;
and §C contains an example stability proof, needed to prove safety under interference of other
threads, in our logic.
2
MOTIVATING EXAMPLE
Our aim is to prove memory safety and functional properties of concurrent data structures using
local reasoning. We begin with a quick reminder of the standard way of specifying data structures
in separation logic using inductive predicates. Using the example of the Harris list [Harris 2001],
we see the limitations of the standard approach when dealing with concurrent data structures.
Separation logic (SL) is an extension of Hoare logic built for reasoning about programs that access
and mutate data stored on the heap. Assertions in SL describe partial heaps (usually represented as
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:4
Siddharth Krishna, Dennis Shasha, and Thomas Wies
mh
−∞
3
5
9
10
fh
2
6
1
7
ft
12
∞
Fig. 1. A potential state of the Harris list with explicit memory management. fnext pointers are shown with
dashed edges, marked nodes are shaded gray, and null pointers are omitted for clarity.
a partial function from program locations to values), and can be interpreted as giving the program
the permission to access the described heap region. Assertions are composed using the separating
conjunction P ∗ Q which states that P and Q hold for disjoint regions of the heap. This enables
local reasoning by means of a frame rule that allows one to remove regions of the heap that are not
modified by a program fragment while reasoning about it.
The standard way of describing dynamic data structures (i.e. structures covering an unbounded
and statically-unknown set of locations) in SL is using inductive predicates. For example, one can
describe a singly-linked null-terminated acyclic list using the predicate
ls(x) B x = null ∧ emp ∨ ∃y. x 7→ y ∗ ls(y).
A heap h consists of a list from x if either x = null and h is empty (emp) or h contains a location x
whose value is y (x 7→ y) and a disjoint null-terminated list beginning at y. The fact that the list is
acyclic is enforced by the separating conjunction, which implies that x is not a node in ls(y).
One can write similar inductive predicates to describe any other data structure that can be
unrolled and decomposed into subpredicates that do not share nodes. Sequential data structures
such as sorted lists, doubly-linked lists, and binary search trees have been successfully verified in
this manner. However, concurrent data structures often use intricate sharing and overlays.
For instance, consider the concurrent non-blocking linked list algorithm due to Harris [2001].
This algorithm implements a set data structure as a sorted list, and uses atomic compare-and-swap
(CAS) operations to allow a high degree of parallelism. As with the sequential linked list, Harris’
algorithm inserts a new key k into the list by finding nodes k 1 , k 2 such that k 1 < k < k 2 , setting k
to point to k 2 , and using a CAS to change k 1 to point to k only if it was still pointing to k 2 . However,
a similar approach fails for the delete operation. If we had consecutive nodes k 1 , k 2 , k 3 and we
wanted to delete k 2 from the list (say by setting k 1 to point to k 3 ), there is no way to ensure with a
single CAS that k 2 and k 3 are also still adjacent (another thread could have inserted or deleted in
between them).
Harris’ solution is a two step deletion: first atomically mark k 2 as deleted (by setting a mark bit
on its successor field) and then later remove it from the list using a single CAS. After a node is
marked, no thread can insert or delete to its right, hence a thread that wanted to insert k ′ to the
right of k 2 would first remove k 2 from the list and then insert k ′ as the successor of k 1 .
As described so far, this data structure can still be specified and verified using inductive predicates
as the underlying shape is still a standard linked list. Things get complicated when we want to
free deleted nodes. Since marked nodes can still be traversed by other threads, there may still be
suspended threads accessing a marked node even after we remove it from the list. Thus, we cannot
free the node immediately. A common solution is the so-called drain technique which maintains a
second “free list” to which marked nodes are added before they are unlinked from the main list.
These nodes are then labeled with a timestamp indicating when they were unlinked. When the
starting timestamp of all active threads is greater than the timestamp of a node in the free list, the
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended)
1:5
node can be safely freed. This leads to the kind of data structure shown in Fig. 1, where each node
has two pointer fields: a next field for the main list and an fnext field for the free list (shown as
dashed edges). Threads that have been suspended while holding a reference to a node that was
added to the free list can simply continue traversing the next pointers to find their way back to
the unmarked nodes of the main list.
Even if our goal is to verify seemingly simple properties such as that the Harris list is memory
safe and not leaking memory, the proof will rely on the following nontrivial invariants
(a) The data structure consists of two (potentially overlapping) lists: a list on next edges beginning at mh and one on fnext edges beginning at fh.
(b) The two lists are null terminated and ft is an element in the free list.
(c) The next edges from nodes in the free list point to nodes in the free list or main list.
(d) All nodes in the free list are marked.
One may be tempted to describe the Harris list simply as two lists (using the inductive predicate
ls defined above), each using a different successor field. However, the two lists may overlap and so
cannot be described by a separating conjunction of the two predicates. For the same reason, we
can also not describe the entire structure with a single inductive predicate that uses separating
conjunction because the predicate must describe each node exactly once1 . Here, we could have
arbitrarily many nodes in the free list with next edges to the same node in the main list (imagine a
series of successive deletions of the predecessor of a node, such as 6, 7, 5 in Fig. 1). Because of this,
we cannot describe the list nodes using separating conjunction as there is no way of telling if we
are “double-counting” them. Also, using an inductive list predicate to describe the main list would
work well only for reasoning about traversals starting from mh, since that is the only possible
unrolling of the predicate. By contrast, threads may enter the main list at arbitrary points in the
list by following next edges from the free list.
To further complicate matters, the next edges of free list nodes can even point backward to
predecessors in the free list (for e.g. 1 points back to 2 in Fig. 1, creating a cycle). This means that
even ramification-style approaches that use the overlapping conjunction [Hobor and Villard 2013]
would not be able to specify this structure without switching to a coarse abstraction of an arbitrary
potentially-cyclic graph (akin to indexed separating conjunctions [Reynolds 2002; Yang 2001]).
However, a coarse graph abstraction could not easily capture key invariants such as that every
node must be reachable from mh or fh, which are critical for proving the absence of memory leaks.
In the rest of this paper, we show how to solve these problems and obtain a compositional
SL-based framework for reasoning about complex data structures like the Harris list. In §4 we
define flows, show how they can be used to encode data structure invariants (both shape and
data properties), build a flow-based semantic model for separation logic, and define an abstraction,
flow interfaces, to reason compositionally about flow-based properties. §4.2 demonstrates the
expressivity of our flow-based approach by showing how to describe many intricate data structures.
We then extend a concurrent separation logic with flow interfaces in §5, show some lemmas to
prove entailments in §5.5, and use them in §6 to verify memory safety and absence of memory
leaks of a skeleton of the Harris list insert operation that abstracts from some of the algorithm’s
details but whose proof still relies on the above invariants (a) to (d). Our lemmas for reasoning
about flows are data-structure-agnostic, i.e., they generalize across many different data structures.
To demonstrate this feature, we show how to formalize the edgeset framework for concurrent
dictionaries using flows in §7 and obtain a template algorithm with an implementation-agnostic
proof of memory safety and linearizability.
1 This
is essentially because the disjointness of ∗ makes x 7→ _ ∗ x 7→ _ unsatisfiable
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:6
3
Siddharth Krishna, Dennis Shasha, and Thomas Wies
PRELIMINARIES
Notation. The term ITE(b, t 1 , t 2 ) denotes t 1 if condition b holds and t 2 otherwise. If f is a function
from A to B, we write f [x ↣ y] to denote the function from A ∪ {x } defined by f [x ↣ y](z) B
ITE(z = x, y, f (z)). We use {x 1 ↣ y1 , . . . , x n ↣ yn } for pairwise different x i to denote the function
ϵ[x 1 ↣ y1 ] · · · [x n ↣ yn ], where ϵ is the function on an empty domain. If f : A × B → C,
then we also write dom1 (f ), dom2 (f ) for A, B respectively. For f : A → B and C ⊆ A we write
f |C : C → B for the function obtained from f by restricting its domain to C. For f 1 : A → B
and f 2 : C → B if A ∩ C = ∅ then we write f 1 ⊎ f 2 for the function f : A ∪ C → B given by
f (x) B ITE(x ∈ A, f 1 (x), f 2 (x)). Moreover, we denote by f (C) the set { f (c) | c ∈ C}.
Semirings, ω-cpos, and continuous functions. A semiring (D, +, ·, 0, 1) is a set D equipped with
binary operators +, · : D × D → D. The operation + is called addition, and the operation · multiplication. The two operators must satisfy the following properties: (1) (D, +, 0) is a commutative
monoid with identity 0; (2) (D, ·, 1) is a monoid with identity 1; (3) multiplication left and right
distributes over addition; and (4) multiplication with 0 annihilates D, i.e., 0 · d = d · 0 = 0 for all
d ∈ D.
An ω-complete partial order (ω-cpo) is a set D equipped with a partial order ⊑ on D such that all
increasing chains in D have suprema
, ⊑1 )
à in D.ÃA function f : D 1 → D 2 between two ω-cpos (D 1Ã
and (D 2 , ⊑2 ) is continuous if f ( 1 X ) = 2 f (X ) for any increasing chain X ⊆ D 1 . Here, i X
denotes the supremum of a chain X in D i .
4
THE FLOW FRAMEWORK
This section presents the formal treatment of our flow framework.
Our theory is parameterized by the domain over which flows range. This domain is equipped
with operations for calculating flows, which must satisfy certain algebraic properties.
Definition 4.1 (Flow Domain). A flow domain (D, ⊑, ⊔, +, ·, 0, 1) is a positive partially ordered
semiring that is ω-complete. That is, (1) (D, ⊑) is an ω-cpo and ⊔ its join; (2) (D, +, ·, 0, 1) is a
semiring; (3) + and · are continuous with respect to ⊑; and (4) 0 is the smallest element of D. We
identify a flow domain with its support set D.
Example 4.2. The natural numbers extended with infinity, N∞ B (N ∪ {∞} , ≤, max, +, ·, 0, 1),
form a flow domain and so does (D, ⊑, ⊔, ⊔, ⊓, ⊥, ⊤) for any completely distributive lattice (D, ⊑
, ⊔, ⊓, ⊥, ⊤).
In our formalization, we consider graphs whose nodes are labeled from some set of node labels
A that encode relevant information contained in each node (e.g. the node’s content or the id of
the thread holding a lock on the node). We will later build abstractions of graphs that compute
summaries of these node labels. For this purpose, we require that A is equipped with a partial order
⊑ that induces a join-semilattice (A, ⊑, ⊔, ae ) with join ⊔ and smallest element ae .
For the remainder of this section we fix a flow domain (D, ⊑, ⊔, +, ·, 0, 1) and node label domain
(A, ⊑, ⊔, ae ). We use the same symbols for the partial order and join operator on the two domains.
However, it will always be clear from the context which one is meant. All definitions are implicitly
parameterized by D and A.
4.1 Flows
Flows are calculated over directed graphs that serve as an intermediate abstraction of heap graphs.
The graphs are partial, i.e., they may have outgoing edges to nodes that are not itself part of the
graph. The edges are labeled by elements of the flow domain and the nodes with node labels.
Formally, given a (potentially infinite) set of nodes N , a (partial) graph G = (N , N o , λ, ε) consists
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended)
1
1
n0 : 1
n1 : 1
n2 : 1
n6 : 1
1
n1 : 1
1
=
n3 : 1
n4 : 1
n5 : 1
•
n4 : 1
1
n0 : 1
n2 : 1
1:7
n6 : 1
1
1
n3 : 1
n5 : 1
Fig. 2. A decomposition of an inflow/graph pair (in, G) into the boxed blue region (in1 , G 1 ) and its context
(in2 , G 2 ). All shown edges have edge label 1, missing edges have label 0, and the path-counting flow is used.
Non-zero inflows are shown as curved arrows at each node, nodes are labeled with names and the resulting
flows, and sink nodes are shown as dotted circles.
of a finite set of nodes N ⊆ N , a finite set of sink nodes N o ⊆ N disjoint from N , a node labeling
function λ : N → A, and an edge function ε : N × (N ∪ N o ) → D. Note that the edge function is
total on N × N o . The absence of an edge between two nodes n, n ′ is indicated by ε(n, n ′) = 0. We
let dom(G) = N and sometimes identify G and dom(G) to ease notational burden. The (unique)
graph defined over the empty set of nodes and sinks is denoted by G e . We define the disjoint union
of two graphs G 1 ⊎ G 2 in the expected way:
(
(N 1 ∪ N 2 , (N 1o \ N 2 ) ∪ (N 2o \ N 1 ), λ 1 ⊎ λ 2 , ε ′) if N 1 ∩ N 2 = ∅
G1 ⊎ G2 B
undefined
otherwise
ε (n , n ) n 1 ∈ N 1 ∧ n 2 ∈ N 1 ∪ N 1o
1 1 2
where ε (n 1 , n 2 ) B ε 2 (n 1 , n 2 ) n 1 ∈ N 2 ∧ n 2 ∈ N 2 ∪ N 2o
0
otherwise.
′
A graph G 1 is a subgraph of another graph G, denoted G 1 ⊆ G, if and only if there exists G 2 such
that G = G 1 ⊎ G 2 .
Example 4.3. The graph G 1 in the middle of Fig. 2 is ({n 1 , n 2 , n 4 } , {n 3 , n 5 , n 6 } , λ 1 , ε 1 ) for some λ 1
and an ε 1 that sets all pairs in {(n 1 , n 3 ), (n 1 , n 4 ), (n 2 , n 6 )} to 1 and all other pairs to 0. Similarly, G 2
on the right is ({n 0 , n 3 , n 5 , n 6 } , {n 1 , n 2 } , λ 2 , ε 2 ) for appropriate λ 2 and ε 2 , and the composed graph
G = G 1 ⊎ G 2 is shown on the left of Fig. 2.
Let G = (N , N o , λ, ε) be a graph. A flow of G is a function flow(in, G) : N → D that is calculated
as a certain fixpoint over G’s edge function starting from a given inflow in : N → D going into G.
To motivate the definition of flow(in, G), let us use the flow domain N∞ from Example 4.2. We can
use this flow domain to define a flow that counts the number of paths between a fixed node and any
other node. First, view the edge function ε as an adjacency matrix E indexed by the nodes N ∪ N o .
If one looks at the matrix product E · E = E 2 (using the addition and multiplication operations of
the flow domain), then the (n, n ′)-th entry corresponds to the sum over the products of edge labels
on all paths of length 2 between n and n ′. For G from Fig. 2, for example, the (n 0 , n 3 )-th entry of
E 2 is 1 as there is exactly 1 non-zero length 2 path, while the (n 0 , n 1 )-th entry is 0 as there is no
non-zero length 2 path. Extending this idea, for any graph where edges are labeled from {0, 1},
the (n, n ′)-th entry of matrix E l is the number of l length paths from n to n ′. Thus, the matrix
C = I + E + E 2 + . . . tells us the number of all paths between two nodes. Given a root node r , we
can look at the r -th row of this matrix to get the number of paths to any node from the root. To
generalize this to multiple roots, we can calculate the number of paths to any node by looking at
® · C where in
® is a row vector with a 1 for every root node and 0 otherwise.
in
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:8
Siddharth Krishna, Dennis Shasha, and Thomas Wies
Formally, first let the capacity cap(G) : N × (N ∪ N o ) → D be the least fixpoint of the following
equation:
Õ
cap(G)(n, n ′) = init(n, n ′) +
ε(n, n ′′) · cap(G)(n ′′, n ′) where init(n, n ′) B ITE(n = n ′, 1, 0).
n ′′ ∈G
Note that the fixpoint exists because D is ω-complete and + and · are continuous. Then define
Õ
flow(in, G)(n) B
in(n ′) · cap(G)(n ′, n).
n ′ ∈G
Example 4.4. In Fig. 2, the capacity of G 2 is equal to its edge function, the capacity of G 1 is its edge
function with (n 1 , n 5 ) additionally 1, and cap(G) is 1 for all pairs (n, n ′) such that n is an ancestor
of n ′ in the tree. For the inflow in(x) = ITE(x = n 0 , 1, 0), flow(in, G) = (λx .1).
Finally, a key property of flows is that they can be composed by considering the product flow on
the product of their flow domains.
Lemma 4.5. Let (D 1 , ⊑1 , ⊔1 , +1 , ·1 , 01 , 11 ) and (D 2 , ⊑2 , ⊔2 , +2 , ·2 , 02 , 12 ) be flow domains. (D 1 ×D 2 , ⊑
, ⊔, +, ·, (01 , 02 ), (11 , 12 )), where ⊑, ⊔, +, and · operate on each component respectively, is a flow domain.
Moreover, the flow on the product domain is the product of the flows on each component.
4.2
Expressivity
We now give a few examples of flows to demonstrate the range of data structures that can be
described using local conditions on flows.
The shape of all common structures that can be described with inductive predicates can be
described using flows, e.g. lists (singly and doubly linked, cyclic), trees, and nested combinations
of these. By considering products with flows for data properties, we can also describe structures
such as sorted lists, binary heaps, and search trees. Going beyond inductive predicates, the flow
framework can describe structures with unbounded sharing and irregular traversals (e.g. Harris
list), overlays (e.g. threaded and B-link trees), as well as those with irregular shape (e.g. DAGs and
arbitrary graphs).
In this section, we assume that the local condition on the flow of each node is specified by a good
condition. This is a predicate good(n, in, a, f ) taking as arguments the node n, the inflow function in
on the singleton graph containing n, the node label a of n, and the edge function f of that singleton
graph. The singleton inflow of n is equal to the flow at n in the global graph representing the data
structure of interest containing n. This fact, as well as the reason for this formalization, will be seen
in §4.4. We also assume that we can specify the global inflow of our data structure.
4.2.1 Path-counting Flow. Consider the flow domain N∞ from Example 4.2, with an arbitrary
node label domain.
Lemma 4.6. For any graph G, if the edge label of every edge (n, n ′) is interpreted as the number
of edges between n and n ′, then cap(G)(n, n ′) is the number of paths from n to n ′. If the inflow in
takes values in {0, 1} then the path-counting flow flow(in, G)(n) is the number of paths from nodes in
{n ′ | in(n ′) = 1} to n.
We can use this flow to describe shapes such as lists and trees. For instance, a graph is a tree if
and only if every node has exactly one path from the root. Thus, if a graph G satisfies the good
condition
goodtr (n, in, _, _) B in(n) = 1
at all nodes n, given a global inflow satisfying in(n) = ITE(n = n 0 , 1, 0), then G must be a tree rooted
at n 0 . As a singly-linked list is a tree where every node has exactly one outgoing edge, we can
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended)
1:9
describe it using the good condition
goodls (n, in, _, f ) B ∃n ′ . in(n) = 1 ∧ (f = ϵ ∨ f = {(n, n ′) ↣ 1})
where ϵ stands for the empty function (i.e. the node has no outgoing edges). A graph G satisfying
goodls under a global inflow in with in(n) = ITE(n = n 0 , 1, 0) for all n must be a list beginning at n 0 .
To model a null-terminated list, we must further require that the entire graph G has no outgoing
edges. This can be done using our flow interface abstraction in §4.4. If we want to model a list that
terminates at a specific node, say ft as in the free list shown in Fig. 1, we can replace the last clause
of goodls with f = ITE(n = ft, ϵ, {(n, n ′) ↣ 1}).
Similarly, one can describe cyclic lists by requiring that each node has exactly one outgoing edge
and that the path count of each node is ∞. To describe a tree of a particular arity, like a binary tree,
one can constrain the path count to be 1 and add a further condition restricting the number of
outgoing edges at each node appropriately. For overlaid structures like the threaded tree, which is
a tree where all elements also form a list (in an arbitrary order), we can use the product of two
path-counting flows, one from the root of the tree and one from the head of the list.
4.2.2 Last-edge Flow. We next show a flow domain where the flow along a path is equal to the
label of the last edge on the path.
Lemma 4.7. Given an ω-cpo (D, ⊑, ⊔) with smallest element 0, let 1 be a fresh element. There exists
a flow domain (D ⊎ {1} , ⊑, ⊔, ⊔, ⋊, 0, 1), where multiplication is the projection operator, defined as
d 1 ⋊ d 2 B ITE(d 1 = 0, 0, ITE(d 2 = 1, d 1 , d 2 )) and the ordering is extended with 0 ⊑ 1.
This flow has many applications; for instance, we can start with a lattice of types and use the
last-edge flow to enforce that all reachable nodes are well-typed in the presence of type-unsafe
operations. Another example is doubly linked lists, which can be described by using a product
of two path-counting flows, one counting paths from the head and the other from the tail, and a
last-edge flow to ensure that for any node n, n.next.prev is equal to n.
We can also use this flow to encode nested combinations of structures such as trees of lists. We
start with a path-counting flow to enforce that the entire graph forms a tree and then take a product
with the last-edge flow domain built from the ω-cpo on {0, dT , d L } where 0 is the smallest element
and the others are unordered. We then use the good condition to label all edges to tree nodes with
dT , all edges to list nodes with d L , and enforce that the label on the incoming edge matches the
type of the node and that list nodes have no edges to tree nodes.
4.2.3 Upper and Lower-bound Flows. We next turn to data properties like sortedness. Consider
the lower-bound flow domain formed by the integers extended with positive and negative infinities:
(Z ∪ {−∞, ∞} , ≥, min, min, max, ∞, −∞). Assume further that the node label domain consists of
sets of data values and that each node is labeled with the singleton set of its content. If we label
each edge of a list with the value at the source node and start with an inflow of −∞ at the root h,
then the flow at a node is the maximum value of its predecessors. This value is like a lower bound
for the value of the current node; if the value at every node reachable from h is greater than its
lower-bound flow, then the list beginning at h is sorted in ascending order. To do this, we use the
product of the path-counting and the lower-bound flow domains. We can use the good condition
goodsls (n, in, a, f ) B ∃n ′, l, k. in(n) = (1, l) ∧ a = {k} ∧ f = ITE(n ′ = null, ϵ, {(n, n ′) ↣ k}) ∧ l ≤ k
and a global inflow in(n) = ITE(n = n 0 , (1, −∞), (0, ∞)) to describe a sorted list beginning at n 0 .
Note that, unlike in the definitions of inductive SL predicates, the data constraints are not tied to
the shape constraints – we could just as easily use the lower-bound flow condition of sortedness
and the path-counting flow condition of being a tree to describe a min-heap.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:10
Siddharth Krishna, Dennis Shasha, and Thomas Wies
Analogously, we can use the upper-bound flow domain (Z ∪ {−∞, ∞} , ≤, max, max, min, −∞, ∞)
to describe lists sorted in descending order and max-heaps. We can also use a product of a lowerbound and an upper-bound flow to describe a binary search tree by enforcing that each node
propagates to its child the appropriate bounds on values that can be present in the child’s subtree.
We will see a more general flow, the inset flow, that can be used to specify data properties of
dictionary data structures that subsumes the examples above in §7.
4.3
Flow Graph Algebras
We next show how to define a separation algebra that can be used to give semantics to separation
logic assertions in a way that allows us to maintain flow-dependent invariants. This will enable us
to reason locally about graphs and their flows using the ∗ operator of SL.
Definition 4.8 (Separation Algebra [Calcagno et al. 2007]). A separation algebra is a cancellative,
partial commutative monoid (Σ, •, u). That is, • : Σ × Σ ⇀ Σ and u ∈ Σ such that the following
properties hold:
(1)
(2)
(3)
(4)
Identity: ∀σ ∈ Σ. σ • u = σ .
Commutativity: ∀σ1 , σ2 ∈ Σ. σ1 • σ2 = σ2 • σ1 .
Associativity: ∀σ1 , σ2 , σ3 ∈ Σ. σ1 • (σ2 • σ3 ) = (σ1 • σ2 ) • σ3 .
Cancellativity: ∀σ , σ1 , σ1′, σ2 ∈ Σ. σ = σ1 • σ2 ∧ σ = σ1′ • σ2 ⇒ σ1 = σ1′.
Here, equality means that either both sides are defined and equal, or both sides are undefined.
A simple way of obtaining a separation algebra of graphs is to define the composition • in
terms of disjoint union. Effectively, this yields the standard model of separation logic. However,
suppose we have a graph G satisfying some local condition on the flow at each node that we want
to decompose into two subgraphs G 1 and G 2 , say to verify a program that traverses the Harris
list. To infer that the subgraph G 1 also satisfies the same local condition on the flow, we need to
be able to compute the flow of nodes in G 1 without looking at the entire graph. We thus want a
stricter version of composition that allows us to decompose a flow on G into flows on G 1 and G 2 .
Our solution is to use graphs equipped with inflows as the elements of our separation algebra.
Since we allow our graphs to have cycles, it is hard to define the composed inflow of two inflows
on subgraphs. Instead, we start with an inflow/graph pair (in, G) and show how to decompose it
into an inflow for its constituent subgraphs G i . To do this, we project the inflow in to obtain a new
inflow function ini for G i that results in the same flow at each node of G i as in on G. The projected
inflow in1 (n) of a node in G 1 is equal to in(n) plus the contribution of G 2 to the flow of n (and vice
versa for in2 on G 2 ). We will use projection to define the composition of inflow/graph pairs.
Definition 4.9. The projection of an inflow in on a graph G = G 1 ⊎ G 2 onto G 1 is a function
proj(in, G)(G 1 ) : dom(G 1 ) → D defined as:
Õ
proj(in, G)(G 1 )(n) B in(n) +
flow(in, G)(n ′) · ε(n ′, n).
n ′ ∈G\G 1
The following lemma tells us that the flow induced by the projection of an inflow in onto a
subgraph is the same as the flow induced by in on the larger graph. In other words, we can view
the inflow as containing all information about the context of a graph that is needed for calculation
of the flow of nodes in the graph.
Lemma 4.10. If in is an inflow on G and in1 = proj(in, G)(G 1 ) for some G 1 ⊆ G, then flow(in, G)|G 1 =
flow(in1 , G 1 ).
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:11
Example 4.11. Fig. 2 uses the path-counting flow and shows a split of a graph G into the boxed
blue region G 1 and its context G 2 . The inflow on G is in(n) = ITE(n = n 0 , 1, 0), and its projections
onto G 1 and G 2 are shown as curved arrows entering the nodes in each graph.
We next define a first approximation of a composition operator on inflow/graph pairs that
preserves the flow at each node. The composition of two inflow/graph pairs (in1 , G 1 ) and (in2 , G 2 ),
denoted (in1 , G 1 ) • (in2 , G 2 ), is the set of all pairs (in, G) such that G 1 and G 2 compose to G and the
component inflows in1 and in2 are the projections of in onto the respective subgraphs:
(in, G) ∈ (in1 , G 1 ) • (in2 , G 2 ) :⇔ G = G 1 ⊎ G 2 ∧ ∀i ∈ {1, 2} . proj(in, G)(G i ) = ini .
Lemma 4.10 then ensures that the flow is the same in the composite graph as in the subgraphs. Note
that this set may be empty even if G 1 and G 2 are disjoint since the inflows of the two subgraphs
may not be compatible.
Unfortunately, (in1 , G 1 ) • (in2 , G 2 ) may contain more than one element since there may be many
possible composite inflows in whose projection onto the subgraphs yields in1 and in2 . For example,
let G 1 , G 2 be singleton graphs on nodes n 1 , n 2 respectively that each have an edge with label ∞ ∈ N∞
to the other. If in1 = {n 1 ↣ ∞} and in2 = {n 2 ↣ ∞}, then the composite inflow can be {n 1 ↣ 1}
or {n 2 ↣ 1} (among many others). So we cannot immediately use • to define the partial monoid
for our separation algebra. However, as we shall see, all inflows in satisfying the condition of the
composition induce the same flow on G. Thus, let us define an equivalence relation ∼G on inflows
for a given graph G. The relation ∼G relates all inflows that induce the same flows:
in ∼G in ′ :⇔ flow(in, G) = flow(in ′, G).
Given an inflow in, we denote its equivalence class with respect to ∼G as [in]G and extend flow
from inflows to their equivalence classes in the expected way. We also write dom(In) B dom(in)
for In = [in]G , and use IN to denote the set of all inflow equivalence classes.
Using Lemma 4.10 we can show that the equivalence relation ∼ is a congruence on •:
Lemma 4.12. The following two implications hold:
(1) (in, G) ∈ (in1 , G 1 ) • (in2 , G 2 ) ∧ in1 ∼G 1 in1′ ∧ in2 ∼G 2 in2′ ∧ (in ′, G) ∈ (in1′ , G 1 ) • (in2′ , G 2 ) ⇒
in ∼G in ′.
(2) (in, G) ∈ (in1 , G 1 ) • (in2 , G 2 ) ∧ in ∼G in ′ ⇒
∃in1′ , in2′ . in1 ∼G 1 in1′ ∧ in2 ∼G 2 in2′ ∧ (in ′, G) ∈ (in1′ , G 1 ) • (in2′ , G 2 ).
Lemma 4.12 implies that • yields a partial function on pairs of graphs G and their inflow
equivalence classes. This suggests the following definition for our separation algebra.
Definition 4.13 (Flow Graph Algebra). The flow graph algebra (FG, •, H e ) for flow domain D and
node label domain A is defined by
H ∈ FG B {([in]G , G) | in : dom(G) → D}
(
([in]G , G) G = G 1 ⊎ G 2 ∧ ∀i ∈ {1, 2} . proj(in, G)(G i ) ∈ [ini ]G i
([in1 ]G 1 , G 1 ) • ([in2 ]G 2 , G 2 ) B
undefined otherwise
H e B([ine ]G e , G e )
where G e is the empty graph and ine the inflow on an empty domain. We call the elements H ∈ FG
flow graphs. We again let dom(H ) = dom(G) and write H for dom(H ) when it is clear from the
context.
Theorem 4.14. The flow graph algebra (FG, •, H e ) is a separation algebra.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:12
4.4
Siddharth Krishna, Dennis Shasha, and Thomas Wies
Abstracting Flow Graphs with Flow Interfaces
We can now use flow graphs to give semantics to separation logic assertions. However, we also
want to be able to use predicates in the logic that describe sets of such graphs so that the flow
of each graph in the set satisfies certain local invariants at each of its nodes. In the following,
we introduce semantic objects, which we call flow interfaces, that provide such abstractions. We
will then lift the composition operator • from flow graphs to flow interfaces and prove important
properties about flow interface composition. In §5, we will use these properties to justify general
lemmas for proving entailments in a separation logic with flow interfaces.
Our target abstraction must have certain properties. Recall from §4.2 that to describe data
structure invariants as conditions on the flow we fixed a particular inflow to the global graph.
Thus, our abstraction must fix the inflow of the abstracted graph. Secondly, our abstraction must
be sound with respect to flow graph composition: if we have a graph G = G 1 ⊎ G 2 , and G 1 and G 1′
satisfy the same interface, then G 1′ ⊎ G 2 must be defined and satisfy the same interface as G.
The second property is a challenging one, for the effect of a local modification in a flow graph
are not necessarily local: small changes in G 1 may affect the flow of nodes in G 2 . For example, in
Fig. 2 if we modified G 1 by adding a new outgoing edge from n 1 to n 6 , then this would increase the
path count of n 6 to 2 and we can no longer compose with G 2 as it expects an inflow of 1 at n 6 . We
need to find a characterization of modifications to G 1 that preserve the inflow of every node in G 2 .
Intuitively, the internal structure of G 1 should be irrelevant for computing the flow of G 2 ; from
the perspective of G 2 , G 1 should be replaceable by any other flow graph that provides G 2 with the
same inflow. One may try to define an outflow quantity, analogous to inflow, that specifies the flow
that a graph gives to each of its sink nodes. If the interface abstracting G 1 also fixes its outflow,
then one expects the flow of G 2 to be preserved. However, this is not true in the presence of cycles.
On the other hand, if we preserve the capacity between every pair of node and sink node in the
modification of G 1 to G 1′ , then the flow of all nodes in G 2 stays the same. This is a consequence of
the semiring properties of the flow domain and the fact that + and · are continuous. Moreover, as G 2
can only route flow into nodes in G 1 with a non-zero inflow, we only need to preserve the capacity
from these source nodes. We capture this idea formally by defining the flow map of H = (In, G) as
the restriction of cap(G) onto G’s source-sink pairs as specified by In:
flm((In, G)) B cap(G)| {n ∈N
| ∃in∈In. in(n)>0}×N o .
This gives us sufficient technical machinery to define our abstraction of flow graphs.
Definition 4.15 (Flow Interface). Given a flow graph (In, G) ∈ FG, its flow interface is the tuple
consisting of its inflow equivalence class, the join of all its node labels, and its flow map:
int(H ) B (In, ⊔n ∈H λ(n), flm(H ))
where H = (In, G) and G = (N , N o , λ, ε).
The set of all flow interfaces is FI B {int(H ) | H ∈ FG}. The denotation of a flow interface is a set
of flow graphs defined as JI K B {H ∈ FG | int(H ) = I }.
Example 4.16. The interface of H 1 = ([in1 ]G 1 , G 1 ) from Fig. 2 is I 1 = ({in1 } , ae , f 1 ) where f 1 is
the flow map that sets {(n 1 , n 3 ), (n 1 , n 5 ), (n 2 , n 6 )} to 1 and every other pair to 0 (assuming all nodes
are labeled with ae ). If we now modified H 1 to H 1′ by removing the edge (n 1 , n 4 ) and adding an edge
(n 1 , n 5 ), then H 1′ will also be in the denotation JI 1 K. Note that H 1′ still composes with the flow graph
on G 2 , and the flow at all nodes in G 2 is the same in this composition.
If we look at a flow interface I = (In, a, f ) in the rely-guarantee setting, a flow graph satisfying I
relies on getting some inflow in In from its context, and guarantees the flow map f to its context.
The third component, a, provides a summary of the information contained in the nodes of a graph,
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:13
which will be useful, for example, to reason about the contents of data structures. We identify the
domain of an interface with the domain of the graphs it abstracts, i.e. dom((In, a, f )) = dom(In).
Note that both the inflow class and flow map of a flow interface I fully determine the domains of
the graphs being described by the interface. That is, I does not abstract from the identity of the
nodes in the graphs JI K, which is intentional as it allows us to define separation logic predicates in
terms of flow interfaces that have precise semantics, without encoding specific traversal patterns
of the underlying data structure and without restricting sharing as other methods do.
If we consider the equivalence relation on flow graphs induced by having the same flow interface,
then we have the required property that this relation is a congruence on •.
Lemma 4.17. If H 1 , H 1′ ∈ JI 1 K, H 2 , H 2′ ∈ JI 2 K and H 1 • H 2 ∈ JI K, then H 1′ • H 2′ ∈ JI K.
We can now lift flow graph composition to flow interfaces, denoted I = I 1 ⊕ I 2 , as follows:
I = I 1 ⊕ I 2 :⇔ ∃H ∈ JI K, H 1 ∈ JI 1 K, H 2 ∈ JI 2 K. H = H 1 • H 2 .
Note that ⊕ is also a partial function for the domains of the graphs may overlap, or the inflows
may not be compatible. The identity flow interface Ie is defined by Ie B ({ine } , ae , f e ), where f e
is the flow map on an empty domain. A simple corollary of Lemma 4.17 is that the flow interface
algebra (FI, ⊕, Ie ) is also a separation algebra.
As we saw in §4.2, to express properties of different data structures we need to use abstractions
of flow graphs that additionally satisfy certain conditions on the flow at each node. To formalize
this, we assume the node condition is specified by a predicate good(n, in, a, f ) which takes a node n,
the inflow in of the singleton flow graph containing n, the node label a of n, and the edge function
f of that singleton graph as arguments. The denotation of a flow interface I with respect to a good
condition good is defined as
JI Kgood B {(In, (N , N o , λ, ε)) ∈ JI K | ∀n ∈ N . good(n, inn , λ(n), εn )}
where inn B {n ↣ flow(In, G)(n)} and εn B {(n, n ′) ↣ ε(n, n ′) | n ′ ∈ N ∪ N o , ε(n, n ′) , 0}. Note
that the interface of the singleton flow graph containing n is ({inn } , λ(n), ITE(inn (n) = 0, ϵ, εn )),
but we choose the above formulation for notational convenience.
The following lemma tells us that imposing a good condition on flow interfaces is essentially
reducing the domain of flow graphs to those that satisfy the good condition. In other words, the
equivalence on flow graphs induced by the denotation of interfaces with respect to a good condition
is also a congruence on •.
Lemma 4.18. For all good conditions good and interfaces I = I 1 ⊕ I 2 , if H 1 ∈ JI 1 Kgood and H 2 ∈
JI 2 Kgood then H 1 • H 2 is defined and in JI Kgood .
When reasoning about programs, we will sometimes need to modify a flow graph region in a
way that will increase the set of inflows. For instance, if we were to add a node n to a cyclic list,
then the equivalence class of inflows will now also contain inflows that have n as a source. As
the flow map is defined from all source nodes, this modification will also change the flow map.
However, as long as the flow map from the existing sources is preserved, then the modified flow
graph should still be able to compose with any context. Formally, we say an interface (In, a, f ) is
contextually extended by (In ′, a ′, f ′), written (In, a, f ) ≾ (In ′, a ′, f ′), if and only if In ⊆ In ′ and
∀n ∈ dom1 (f ). ∃in ∈ In. in(n) , 0 ⇒ f (n, ·) = f ′(n, ·).
where f (n, ·) is the function that maps a node n ′ to f (n, n ′). The following theorem states that
contextual extension preserves composability and is itself preserved under interface composition.
Theorem 4.19 (Replacement). If I = I 1 ⊕ I 2 , and I 1 ≾ I 1′ , then there exists I ′ = I 1′ ⊕ I 2 such that
I ≾ I ′.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:14
Siddharth Krishna, Dennis Shasha, and Thomas Wies
1
1
l
1
r
1
r
l
1
n
⇝
1
1
n
flm(H {l } ) = {(l, r ) ↣ 1}
′
flm(H {l
} ) = {(l, n) ↣ 1}
flm(H {l,n} ) = {(l, r ) ↣ 1}
′
flm(H {l,n}
) = {(l, r ) ↣ 1}
Fig. 3. Inserting a new node n into a list H between existing nodes l and r obtaining a new list H ′ . Edges are
labeled with edge labels for path counting, and the nodes are labeled with their names. The flow maps of
certain subgraphs before and after the modification are shown below.
5
A CONCURRENT SEPARATION LOGIC WITH FLOW INTERFACES
We next show how to extend a concurrent separation logic with flow interfaces. We use RGSep for
this purpose as it suffices to prove the correctness of interesting programs, yet is relatively simple
compared to other CSL flavors. However, many concurrent separation logics, including RGSep,
are parametric in the separation algebra over which the semantics of programs is defined. So the
technical development in this section can be easily transferred to these other logics.
RGSep is parametric in the program states (states must form a separation algebra), the language
of the assertions (the original paper used the standard variant of separation logic), and the basic
commands of the programming language (their semantics need to satisfy a locality property to
obtain the frame rule of separation logic). For the most part, we piggyback on the development
in [Vafeiadis 2008]. We denote the resulting logic by RGSep[FI]. Once we have shown that RGSep[FI]
is a valid instantiation of RGSep, the soundness of all RGSep proof rules immediately carries over to
RGSep[FI]. The only extra work that we will need to do before we can apply the logic is to derive a
few generic lemmas for proving entailments between assertions that involve flow interfaces.
5.1
Reasoning about the Actual Heap
We first discuss the insert procedure on a singly-linked list as an example to motivate the design
of our logic. The insert procedure first traverses the list to find two adjacent nodes l and r , with
some appropriate properties, between which to insert the given node n. It then sets n’s next field
to point to r , and finally swings l’s next pointer to n. A pictorial representation of the crucial step
in this procedure is shown in the top portion of Fig. 3.
Using flow graphs directly as our program state, as presented thus far, poses three key challenges. Firstly, a key assumption SL makes on the programming language is that commands are
local [Reynolds 2002]: informally, that if a command runs successfully on a state σ1 and we run it
on a composed state σ1 · σ2 then it leaves σ2 unchanged. However, even a basic operation on a flow
graph, for instance changing the edge (l, r ) to an edge (l, n), can potentially change the flows of all
downstream nodes and is hence not local. We have seen that if we can find a region around the
modification where the flow map is preserved, then the change will not affect any nodes outside
the region. But which region do we choose as the footprint of a given command? Secondly, the
programs that we wish to verify may temporarily violate the good condition during an update
or a maintenance operation. For example, at the state on the left of Fig. 3 the new node n has no
paths from the root, which violates the good condition describing a list. Thus we also need a way
to describe these intermediate states. Finally and most immediately, the programs that we wish to
verify operate on a program state that are more akin to standard heap representations than to flow
graphs. We thus need a way to abstract the heap to a flow graph.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:15
Our solution to all three problems is based on using the product of heaps and flow graphs as
program states. The heap component represents the concrete state, and the standard commands
such as heap allocation or mutation affect only the heap. Thus, these standard commands inherit
locality from the standard semantics. The graph component is a ghost state and serves as an
abstraction of the heap. Since standard commands only modify the heap, this abstraction may not
be up-to-date with the heap. We add a new ghost command to sync a region of the graph with the
current state of the heap, i.e. change the graph component to one that abstracts the heap component
(using some abstraction as described below). We say the resulting region of the state is in sync.
This sync command takes as an argument the new interface of the heap region to sync, and is only
successful if the new interface contextually extends the old interface of that region. For instance, in
Fig. 3, we cannot sync the region {l } as l’s outgoing edges have changed, we would have to sync the
region {l, n} (note that n is not a source node in this subgraph). Recall that Theorem 4.19 guarantees
that this new region will still compose with the old context, making sync a local command and
solving our first problem.
The fact that the graph portion of the state is not up-to-date with the heap also gives us a way
to solve the problem of temporary violations of the good condition. When executing a series of
commands, we delay syncing the graph portion of the state until the point when the good condition
has been re-established. Thus, the graph component can be thought of as an abstraction of the heap
at the last time it was in a good state. In the concurrent setting, one may wonder if we might expose
such dirty regions (where the graph is not in sync with the heap) to other threads; however, if a
thread is breaking some data structure invariants in a region, then it must be locked or marked off
from interference from other threads in some manner. We add a dirty predicate [ϕ]I to our logic to
describe a region where the heap component satisfies ϕ but the flow graph component is potentially
out of sync and satisfies the interface I . So to verify the list insert procedure, we would use a dirty
predicate to describe l and n until we have modified both edges and are ready to use sync.
Finally, to solve the problem of relating the concrete heap to the abstract graph, we make use of
the good condition. We define good by means of an SL predicate γ that not only describes the good
condition on the flow in the graph component, but also describes the relation between a good node
in the graph and its representation on the heap. For instance, here is the γ that we use to describe a
singly-linked list:
γ (n, in, _, f ) B ∃n ′ . n 7→ n ′ ∧ in(n) = 1 ∧ f = ITE(n ′ = null, ϵ, {(n, n ′) ↣ 1}).
Note that this is essentially goodls from §4.2 with the additional constraint that the edge label in
the graph depends on n ′, the next pointer of n on the heap. We will show how to define a good
graph predicate Gr(I ) using such a γ to ensure that the heap and the graph are in sync.
5.2
Program States
We model the heap h as usual in standard separation logic. That is, the heap is a partial map from
addresses Addr to values Val. Values include addresses, as well as the flow and node label domains
and some auxiliary values for representing ghost state related to flow interfaces:
Val ⊇ Addr ∪ D ∪ A ∪ (Addr ⇀ D) ∪ IN ∪ (Addr × Addr ⇀ D) ∪ FI
A state augments the heap with an auxiliary ghost state in the form of a flow graph H that
maintains an abstract view of the heap. Each graph node n ∈ dom(H ) is an abstraction of potentially
multiple heap nodes, one of which is chosen as the representative n. That is, we require dom(H ) ⊆
dom(h). Also, we have a node map r that labels each heap node with the graph node that abstracts
it. This map is partial, i.e., we allow heap nodes that are not described by the flow graph.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:16
Siddharth Krishna, Dennis Shasha, and Thomas Wies
T F x | (x 1 , x 2 ) | Td | Ta | Tin | TIn | Tf | TI
Ta F a | A | Ta ⊔ Ta
Td F d | D | Tin (x) | Tf (x 1 , x 2 ) | Td + Td | Td · Td
TIn F In | {Tin }
v, x, d, in, In, f , a, I ∈ Var
Tin F in | ϵ | {x ↣ Td , . . . } | Tin + Tin | 0
Tf F f | ϵ | {(x 1 , x 2 ) ↣ Td , . . . } | Tf + Tf | 0 TI F I | (TIn ,Ta ,Tf ) | TI ⊕ TI
P F T = T | Td ⊑ Td | Ta ⊑ Ta | x ∈ TI | (x 1 , x 2 ) ∈ Tf | Tin ∈ TIn | TI ≾ TI
ϕ F P | emp | true | x 7→ T | Gr(I ) | [ϕ]I | ϕ ∗ ϕ | ϕ −∗ ϕ | ϕ ∧ ϕ | ∃v. ϕ | ¬ϕ
Fig. 4. Syntax of RGSep[FI] terms and assertions.
Definition 5.1. A state σ = (h, H, r ) ∈ State consists of a heap h : Addr ⇀ Val, a flow interface
graph H ∈ FG such that dom(H ) ⊆ dom(h), and a node map r : dom(h) ⇀ dom(H ) such that
r |dom(r )∩dom(H ) = (λx .x).
The composition operator on states is a partial function σ1 · σ2 , defined as
(
(h 1 · h 2 , H 1 • H 2 , r 1 ⊎ r 2 ) if h 1 · h 2 and H 1 • H 2 defined
(h 1 , H 1 , r 1 ) · (h 2 , H 2 , r 2 ) B
undefined
otherwise.
Here, h 1 · h 2 denotes the standard disjoint union of heaps. The empty state σe is defined as σe B
(he , H e , r e ) where he and r e are the empty heap and empty node map, respectively. The fact that
states form a separation algebra easily follows from Theorem 4.14.
Theorem 5.2. The structure (State, ·, σe ) is a separation algebra.
For simplicity we do not use the variable as resource model [Bornat et al. 2006] for the treatment
of stack variables. Instead, we follow [Vafeiadis 2008] and assume that all stack variables are local
to a thread, and that all global variables are read-only. Stack variables are allocated on the local
heap and global variables on the shared heap. We assume that each program variable x is located at
a fixed address &x. In all assertions occurring in program specifications throughout the rest of the
paper we have an implicit &x 7→ x for every program variable x in scope.
5.3
Assertions
The syntax of assertions is summarized in Fig. 4. We assume an infinite set of (implicitly sorted)
logical variables Var. The meta variable v stands for a variable of an arbitrary sort and the meta
variable x for a variable of a sort denoting addresses. The sorts of other variables are as expected.
Assertions are built from sorted terms – one sort per semantic domain of interest. The terms Td
denote flow domain values, Ta node label values, etc. For inflow terms Tin , the constant ϵ denotes
the empty inflow ine , Tin + Tin denotes the lifting of + to partial functions2 , and the constant 0
denotes the total inflow that maps all addresses to 0. Terms for flow maps are constructed similarly.
The atomic pure predicates P include equalities, domain membership tests for interfaces, flow
maps, and inflow equivalence classes, as well as contextual extension of flow interfaces. The actual
separation logic assertions ϕ include the standard SL assertions as well as the extended assertions
Gr(I ) and [ϕ]I related to flow interfaces.
The good graph predicate Gr(I ) describes a heap region that is abstracted by a good flow graph
satisfying the flow interface I . The semantics of the logic is parametric in what constitutes a good
graph. For simplicity, we consider only one type of good graph predicate, i.e., all instances of good
graph predicates imply the same data structure invariant. The logic can be easily extended to
2 (f
+ д)(x ) B ITE(x ∈ dom(f ), ITE(x ∈ dom(д), f (x ) + д(x ), f (x )), ITE(x ∈ dom(д), д(x ), undefined))
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:17
ϕ[_]
B
∃v. ϕ[v]
ϕ[I In ]
v fresh in ϕ
B
∃In. I = (In, _, _) ∧ ϕ[In]
where In fresh in ϕ, similarly for ϕ[I a ] and ϕ[I f ]
ϕ[I In (x)]
B
∃in. in ∈ I In ∧ ϕ[in(x)]
ϕ 1 −∗∗ ϕ 2
B
(ϕ 1 ∗ true) ∧ ϕ 2
N(x, I )
B
Gr(I ) ∧ {x ↣ _} ∈ I In
I ≈ I′
B
I ≾ I′ ∧ I′ ≾ I
Fig. 5. Some syntactic shorthands for RGSep[FI] assertions.
(h, H, r ), i
|= T ∈ TI, f
⇔
(h, H, r ), i
|= Tin ∈ TIn
⇔
(h, H, r ), i
|= TI ≾ TI′
⇔
(h, H, r ), i
|= emp
⇔
h = he ∧ H = He
(h, H, r ), i
|= x 7→ T
⇔
(h, H, r ), i
|= Gr(I )
⇔
dom(h) = {JxKi } ∧ h(JxKi ) = JT Ki ∧ H = H e
(h, H, r ), i
|= [ϕ]I
⇔
|= ϕ 1 ∗ ϕ 2
⇔
...
σ,i
JT Ki ∈ dom(JTI,f Ki )
JTin Ki ∈ JTIn Ki
JTI Ki ≾ JTI′Ki
dom(h) = dom(r ) ∧ H ∈ JJI Ki Kgoodh,r
dom(h) = dom(r ) ∧ H ∈ JJI Ki K ∧ ∃H ′, r ′ . (h, H ′, r ′), i |= ϕ
∃σ1 , σ2 . (σ = σ1 · σ2 ) ∧ (σ1 , i |= ϕ 1 ) ∧ (σ2 , i |= ϕ 2 )
Fig. 6. Semantics of assertions.
support more than one such predicate or, if we were to extend higher-order separation logic [Dodds
et al. 2016; Jung et al. 2015; Krebbers et al. 2017], the parametrization could be pushed into the
assertion language itself.
We also need to permit certain regions of the heap that are under modification to be not good.
These regions are described using the dirty predicate [ϕ]I . This describes a state where the flow
graph satisfies I but the heap satisfies ϕ.
We use some syntactic shorthands for RGSep[FI] assertions. For instance, N(x, I ) denotes a
singleton good graph Gr(I ) containing x, and the separating inclusion operator ϕ 1 −∗∗ ϕ 2 defined
as (ϕ 1 ∗ true) ∧ ϕ 2 . For an interface I = (In, a, f ) we write I In , aa and I f for In, a and f . These, and
some additional syntactic shorthands for RGSep[FI] assertions, are defined in Fig. 5.
Terms are interpreted with respect to an interpretation i : Var → Val that maps the logical
variables to values. For a variable v ∈ Var, we write JvKi to mean i(v) and also extend this function
to terms T . We omit the definition of this partial function as it is straightforward. Note that some
terms, for instance interface composition TI ⊕ TI , may be undefined; we define the semantics of
any atom containing an undefined term to be false. The semantics of assertions ϕ is given by the
satisfaction relation σ , i |= ϕ and we denote the induced entailment relation by ϕ 1 |= ϕ 2 . The
satisfaction relation is defined in Fig. 6. Again, we omit some of the obvious cases. Most of the
cases are straightforward. In particular, the standard SL assertions have their standard semantics.
We only discuss the cases for Gr(I ) and [ϕ]I .
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:18
Siddharth Krishna, Dennis Shasha, and Thomas Wies
&x 7→ _ ∗ ∃y . ϕ(y) x :| ϕ(x) &x 7→ x ∗ ϕ(x )
x 7→ v ∗ [ϕ]I ∧ y ∈ I mark(x,y) [x 7→ v ∗ ϕ]I
n
o
[x 7→ v ∗ ϕ]I ∧ x < I In unmkark(x) x 7→ v ∗ [ϕ]I
I ≾ I ′ ∧ [Gr(I ′ )]I sync(I’) Gr(I ′ )
x = y ∧ x 7→ v mark(x,y) [x 7→ v]({{x ↣0}}, ae , ϵ )
n
o
[x 7→ v]I ∧ I In = { {x ↣ 0} } unmkark(x) x 7→ v
Fig. 7. Specifications of ghost commands
The predicate Gr(I ) describes a region whose heap is abstracted by a good flow graph. We tie
the concrete representation in the heap h to the flow graph H of the state using the good node
condition, which we assume is defined by an SL predicate γ (x, in, a, f ). This predicate specifies
the heap representation of a node x in the graph as well as any invariants that x must satisfy on
flow(H )(x). We restrict γ to standard SL assertions, i.e., γ is not allowed to include occurrences of
the predicates Gr(I ) and [ϕ]I .
The predicate γ implicitly defines a good node condition goodh,r on flow interfaces:
goodh,r (n, in, a, f ) :⇔ (h|r −1 (n) , H e , r e ), ϵ |= γ (n, in, a, f )
The semantics of Gr(I ) then uses this condition to tie the flow graph H to the heap h and ensures
that each node in the graph satisfies its local invariant on the flow. The semantics of [ϕ]I is slightly
more complicated. It states that the current flow graph H satisfies I and there is some flow interface
graph H ′ that satisfies ϕ in the current heap.
A precise assertion is one that if it holds for any substate, then it holds for a unique substate. To
use our new predicates in actions to describe interference, we require them to be precise.
Lemma 5.3. The graph predicates Gr(I ) and [ϕ]I are precise.
5.4 Programming Language and Semantics
The programming language is very similar to the one used in RGSep and its semantics is also
mostly identical to the one given in Section 3.2 of [Vafeiadis 2008]:
C ∈ Com F
cF
skip | c | C 1 ; C 2 | C 1 + C 2 | C ∗ | ⟨C⟩ | C ∥C
x :| ϕ(x) | sync(I) | mark(x, y) | unmkark(x) | . . .
The standard commands include the empty command (skip), basic commands, sequential composition, non-deterministic choice, looping, atomic commands, and parallel composition. Apart from
the standard basic commands, we add several special ghost commands to the language, which we
describe next. Their axiomatic specifications are given in Fig. 7.
The wishful assignment x :| ϕ(x) is a ghost command that allows us to bring a witness to an
existentially quantified assertion onto the stack. This helps us to get a handle on the updated flow
interface of a modified region so that we can pass it to the sync. The special ghost command sync
is used to bring a flow graph region back to sync with the heap. The mark command3 is used to
formally associate a (potentially new) heap node with the graph node that abstracts it. This is
useful in cases where many heap nodes are abstracted by the same graph node. Similarly, unmkark
removes the association between a heap and a graph node. Each command has two successful cases,
one where the node being (un)marked is the representative of a graph node, and one where it is
represented by another node.
The operational semantics of commands C in RGSep are given by a reduction relation
→
− ⊆ (Com × State) × ((Com × State) ⊎ {abort}).
3 not
to be confused with the notion of marking logically deleted nodes in Harris’ list
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:19
(x :| ϕ(x), (h, H, r )) →
−
(skip, (h[l → v], H, r ))
if h(&x) = l ∧ ϕ(x) is precise ∧
∃i. (h, H , r ), i[x → v] |= ϕ(x) ∗ true
otherwise.
abort
(skip, (h 1 , H 1′, r 1′ ) · σ2 ) if ∃I , I ′, i. I ≾ I ′ ∧ σ2 |= &I’ 7→ I ′ ∗ true ∧
(sync(I’), (h 1 , H 1 , r 1 ) · σ2 ) →
−
(h 1 , H 1 , r 1 ) |= [true]I ∧ (h 1 , H 1′, r 1′ ) |= Gr(I ′)
abort
otherwise.
(skip, (h, H, r [l ↣ n]))
if l = h(&x) ∈ dom(h) \ dom(H ) ∧
n = h(&y) ∈ dom(H )
(mark(x, y), (h, H, r )) →
− (skip, (h, H ′, r [n ↣ n])) if n = h(&x) = h(&y) ∈ dom(h) \ dom(H ) ∧
∃Hn ∈ J({{n ↣ 0}} , ae , ϵ)K. H ′ = H • Hn
abort
otherwise.
(skip, (h, H, r |dom(r )\{l } )) if l = h(&x) ∧ l ∈ dom(r ) \ dom(H )
(skip, (h, H ′, r |dom(r )\{n } )) if n = h(&x) ∧ r −1 (n) = {n} ∧
(unmkark(x), (h, H, r )) →
−
∃Hn ∈ J({{n ↣ 0}} , _, _)K. H = H ′ • Hn
abort
otherwise.
Fig. 8. Semantics of ghost commands.
The semantics of the new ghost commands are shown in Fig. 8. The semantics of all other commands
are inherited from [Vafeiadis 2008]. That is, these commands only affect the heap component of
the state and leave the flow graph and node map components unchanged. If a standard command
C modifies an unmarked heap location, then we can use the standard SL specification to reason
about it. If it modifies a heap location that is marked, then we can lift C’s SL specification to an
RGSep[FI] specification using the following rule
ψ1 C ψ2
[ψ 1 ]I C [ψ 2 ]I
where ψ _ denotes an assertion that does not contain the predicates Gr(I ) and [ϕ]I . We shall see
generic lemmas to convert a good graph node into a dirty region in §5.5.
Each RGSep command must be local. A command C is local if for all C ′, σ , σ ′, σ1 , and σ2 :
(1) If (C, σ1 · σ2 ) →
− (C ′, σ ′), then either (C, σ1 ) →
− abort, or there exists σ1′ such that σ ′ = σ1′ · σ2
′
′
and (C, σ1 ) →
− (C , σ1 ).
(2) If (C, σ ) →
− abort and σ = σ1 · σ2 , then (C, σ1 ) →
− abort.
Theorem 5.4. The commands :|, sync, mark, unmkark are local and their semantics satisfies the
specifications given in Fig. 7.
Our logic RGSep[FI] thus satisfies all the requirements for a correct instantiation of RGSep.
5.5 Proving Entailments
Finally, we show several generic lemmas for proving entailments about SL assertions with flow
interfaces. We only focus on lemmas that involve flow interfaces and omit lemmas about simple pure
assertions involving inflows and flow maps. We also omit lemmas about the algebraic properties of
flow interface composition, such as associativity. The lemmas are shown in Fig.9.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:20
Siddharth Krishna, Dennis Shasha, and Thomas Wies
Gr(I ) ∧ x ∈ I |= ∃I 1 , I 2 . N(x, I 1 ) ∗ Gr(I 2 ) ∧ I = I 1 ⊕ I 2
(Gr(I 1 ) ∗ true) ∧ Gr(I ) |= ∃I 2 . Gr(I 1 ) ∗ Gr(I 2 )
′
Q 1 (I 1 ) I1 ∗ Q 2 (I 2′ ) I2 ∧ I ′ = I 1′ ⊕ I 2′ |= Q 1 (I 1′ ) ∗ Q 2 (I 2′ ) I1 ⊕I2
(Decomp)
(GrDecomp)
(Comp)
Gr(I 1 ) ∗ Gr(I 2 ) |= Gr(I 1 ⊕ I 2 )
(GrComp)
(Gr(I 1 ) ∗ Gr(I 2 )) ∧ (N(x, I x ) ∗ Gr(I 3 )) |= ∃I 4 . Gr(I 1 ) ∗ N(x, I x ) ∗ Gr(I 4 )
(Disj)
∧ I x a ̸⊑ I 1 a
N(x, I ) ≡ ∃in, a, f . [γ (x, in, a, f )]I ∧ in = {x ↣ _} (Conc)
∧ I = ({in} , a, ITE(in(x) = 0, ϵ, f ))
[N(x, I )]I ≡ ∃in, a, f . [γ (x, in, a, f )]I ∧ in = {x ↣ _}
′
(Abs)
∧ I = ({in} , a, ITE(in(x) = 0, ϵ, f ))
′
[true]I ∧ [true]I ′ |= I = I ′
I = I ⊕ ({{x ↣ 0}} , _, _) ∧ in ∈ I |= ∃in ∈ I
′
′
f
I = I ⊕ (_, _, ϵ) ∧ I = ϵ |= I
′
I1 ≾
I 1′
′f
(Uniq)
′ In
. in + 0 = in + 0
′
=ϵ
|= (I 1 ⊕ I 2 ) ≾
(AddF)
(I 1′
⊕ I2 )
I ≾ I ′ ∧ in ∈ I |= in ∈ I ′In
′
f
f
f
I ≾ I ∧ I = ϵ |= I
′f
(AddIn)
=ϵ
I = I 1 ⊕ I 2 ∧ (x, y) ∈ I 1 ∧ I = ϵ |= y ∈ I 2
I = I 1 ⊕ _ ∧ x ∈ I 1 |= x ∈ I
N(x, I x ) −∗∗ Gr(I ) ∧ (x, y) ∈ I x f ∧ I f = ϵ |= ∃Iy . (N(x, I x ) ∗ N(y, Iy )) −∗∗ Gr(I )
(Repl)
(ReplIn)
(ReplF)
(Step)
(Climb)
(Step−∗∗)
Fig. 9. Generic lemmas for proving entailments
For example, the lemma (Decomp) implies that we can pull an arbitrary node N(x, I 1 ) from a
good graph Gr(I ) containing x, obtaining a flow interface constraint I = I 1 ⊕ I 2 that remembers that
the two parts composed to I . Lemma (Comp) uses Q(I ) to denote either Gr(I ) or [ϕ]I for some ϕ and
tells us that we can combine two dirty regions if the interfaces they are expecting (I 1′ and I 2′ ) are
compatible. The lemma (GrComp) states that we can always abstract two good graphs Gr(I 1 )∗Gr(I 2 )
to a larger good graph Gr(I ) that satisfies the composite interface I = I 1 ⊕ I 2 . The validity of this
entailment follows from Lemma 4.18. (Conc) and (Abs) allow us to shift our reasoning from graph
nodes to their heap representation and back. (AddIn) and (AddF) can be used to reason about
adding new nodes to the state.
One advantage of our logic is that a lemma like (GrComp) is very general and can be used to
compose two lists, two sorted lists, or two min-heaps. The good condition will ensure, for instance,
that the composition of sorted lists is itself sorted.
6
APPLICATION 1: THE HARRIS LIST
We now demonstrate our flow framework by describing and verifying the Harris list. For simplicity
of presentation, we consider a data structure without keys and abstract the algorithm to one that
non-deterministically chooses where to insert a node or which node to delete. This is without loss
of generality, as we are only proving memory safety and absence of memory leaks. Our proof can
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:21
1
procedure insert() {
n
Φ
o
// Where: Φ B ∃I . Gr(I ) ∧ φ(I ), φ(I ) B ∃in ∈ I in . in + 0 = {mh ↣ (1, 0)} + {fh ↣ (0, 1)} + 0 ∧ I f = ϵ
n
o
var l := mh;
N(l, Il ) −∗∗ Φ ∧ l = mh
2
3
4
var r := getUnmarked(l.next);
5
while (r != null && nondet()) {
o
(N(l, Il ) ∗ N(r, I r )) −∗∗ Φ
n
o
l := r; r := getUnmarked(l.next);
(N(l, Il )) −∗∗ Φ ∧ r , null ⇒ (N(l, Il ) ∗ N(r, I r )) −∗∗ Φ
6
7
}
8
if (!isMarked(r)) {
o
N(l, Il ) −∗∗ Φ ∧ ¬M (r )
n
o
var n := new Node(r, null);
N(l, Il ) −∗∗ Φ ∗ n 7→ r, null ∧ ¬M (r )
n
o
mark(n, n);
N(l, Il ) −∗∗ Φ ∗ [n 7→ r, null]In ∧ ¬M (r ) // Where: I n := ({n ↣ (0, 0)} , ⊥, ϵ )
9
10
n
atomic { // CAS(l.next, r, n)
if (l.next == r) {
o
n
[l 7→ r, _]Il ∧ N(l, Il ) ∗ [n 7→ r, null]In ∗ Gr(I 2 ) ∧ Il a = ♦ ∧ I ′ = Il ⊕ I n ⊕ I 2 ∧ φ(I ′ )
11
12
13
l.next := n;
n
o
N(l, Il′ ) ∗ N(n, I n′ ) I ∗ Gr(I 2 ) ∧ Il′ a = ♦ ∧ I n′ a = ♦ ∧ I ′ = Il ⊕ I n ⊕ I 2 ∧ φ(I ′ ) ∧ I 1 = Il ⊕ I n ≈ Il′ ⊕ I n′
1
var I′1 :| Gr(I 1′ ) I ∧ {l ↣ _, n ↣ _} ∈ I 1 In ;
1
a
a
sync(I′1 ); N(l, Il′ ) ∗ N(n, I n′ ) ∗ Gr(I 2 ) ∧ Il′ = ♦ ∧ I n′ = ♦ ∧ I ′′ = Il′ ⊕ I n′ ⊕ I 2 ∧ φ(I ′′ )
a
a
var b := true; (N(l, Il′ ) ∗ N(l, Il′ )) −∗∗ Φ ∧ Il′ = ♦ ∧ I n′ = ♦ ∧ b
n
o
} else var b := false;
N(l, Il ) −∗∗ Φ ∗ [n 7→ r, null]In ∧ ¬b
14
15
16
17
18
19
} ... // If CAS failed, unmark and free n, and call insert() again
20
}
21
22
n
}
Fig. 10. A fragment from the insert procedure on the Harris list.
be extended to prove these properties on the full Harris list, and our framework can also prove the
key invariants needed to prove linearizability.
We now describe the flows and good conditions we use for this proof. To describe the structural
properties, we use the product of two path-counting flows, one counting paths from the head of the
main list mh and one from the head of the free list fh. To reason about marking, we use the node
domain N ∪ {♦, ⊤} under the ordering where ♦ is the smallest element, ⊤ is the largest element,
and all other elements are unordered. We label unmarked nodes with ♦, and marked nodes with the
thread ID t ∈ N of the thread that marked the node. This is in order to enforce that only the thread
that marks a node may link it to the free list, a property needed to prove that the free list is acyclic.
Our good condition is specified by the following predicate:
γ (n, in, a, f ) B ∃n ′, n ′′ . n 7→ n ′, n ′′ ∧ a , ⊤ ∧ (M(n ′) ⇔ a , ♦) ∧ (0, 0) < in(n) ≤ (1, 1)
∧ (in(n) ≥ (0, 1) ⇒ a , ♦) ∧ (n = ft ⇒ in(n) ≥ (0, 1)) ∧ (in(n) ≤ (1, 0) ⇒ n ′′ = null)
∧ f = ITE(u(n ′) = null, ϵ, {(n, u(n ′)) ↣ (1, 0)}) + ITE(n ′′ = null, ϵ, {(n, n ′′) ↣ (0, 1)}).
This condition expresses that every node n is a heap cell containing two pointers n ′ (for the next
field) and n ′′ (for the fnext field). n is either unmarked or marked with a thread ID (a , ⊤), but
only if the value n ′ of field next has its mark bit set (encoded using the predicate M). We next use
the path-counting flows to say that n has exactly 1 path on next edges from mh, on fnext edges
from fh, or both. This enforces that all nodes are in at least one of the two lists, and this is how
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:22
Siddharth Krishna, Dennis Shasha, and Thomas Wies
we establish absence of memory leaks. The next few conjuncts say that all nodes in the free list
are marked (in ≥ (0, 1) ⇒ a , ♦), that ft is a node in the free list, and that main list nodes have no
fnext edges. The final line describes the edges: n has a next edge (encoded with the edge label
(1, 0)) to u(n ′) (the unmarked version of n ′, i.e. the actual pointer obtained from n ′ by masking the
mark bit), and an fnext edge (label (0, 1)) to n ′′, but only if they are not null.
We use the global data structure invariant Φ shown in line 2 of Fig. 10, which gives the appropriate
non-zero inflow to fh and mh. It is easy to see that with good condition γ , Φ describes a structure
satisfying properties (a) to (d) of the Harris list from §2.
We now describe the RGSep[FI] proof for the insert procedure, shown in Fig. 10. Variables
that are not program variables in the annotations are implicitly existentially quantified. All the
entailments in this proof sketch can be proved with the help of our library of lemmas from §5.5.
The procedure starts in a state satisfying Φ and sets a variable l to equal the head of the main
list. Since l is in the domain of the inflow I in we use (Decomp) to decompose Φ and get a single
node (N(l, Il )) included in the larger graph satisfying Φ (line 3). We then read l.next, store the
unmarked version in r, and enter a loop. At the beginning of the loop (line 5), since r , null, we
know by γ that l has a next edge to r . But since Φ implies I f = ϵ, we can use (Step−∗∗) to extract
the node corresponding to r and obtain the annotation on line 5.
Inside the loop, we move l to r and then again read l.next, unmark it, and set it to r (line 6).
This dereference is memory safe because we know that we have access permission to l (by N(r , Ir )
on line 5). Finally, to establish the annotation on line 6, we “drop” the node predicate corresponding
to the old value of l and absorb it into Φ. The second conjunct is derived similar to the annotation
on line 5. Note that the annotation at the end of the loop implies the annotation at the beginning of
the loop if the loop check succeeds.
The loop terminates when the non-deterministic loop condition fails – presumably at the correct
position to insert the new node. Since we only want to insert new nodes into the main list, we
ensure l is unmarked (line 8). We then create the new node with next field r (which may equal
null) in line 9. As this allocation only modifies the heap, the resulting node is described using a
standard SL points-to predicate in the local state of the thread. To create a corresponding node in
the graph we use the mark ghost command, which creates a new graph node with a zero inflow, the
bottom node label, and no outgoing edges. The resulting state is described with a dirty predicate
(line 10) as it is neither in sync nor a good node.
The next step is to swing l’s next pointer from r to n using a CAS operation. The CAS is expanded
into an atomic block (lines 11 to 20) in order to show the intermediate proof steps. Note that inside
the atomic block there is no need for separation of shared and local state, as in RGSep one can
reason sequentially about atomic executions. If the compare portion of the CAS succeeds, then we
know that l is unmarked (Il a = ♦) and its next field equals r (we use a dirty predicate around l to
express this). We use (GrDecomp), (Comp), (AddIn), and (AddF) to decompose Φ at this point and
infer that I ′, the global interface extended with n, also satisfies the global conditions φ(I ′) (line 13).
The modification to l.next in the next line is memory safe since we have access permission to l
(by l 7→ r , _).
Before we can bring the graph abstraction back to sync, we must establish that the change to
the interface of some region is a contextual extension. As we saw with the example of inserting in
a singly-linked list in Fig. 3, we must consider the region containing {l, n}. Some pure reasoning
about path-counts in this region can be used to infer that Il ⊕ In ≈ Il′ ⊕ In′ (line 15)4 . We then use a
wishful assignment (line 16) to take a snapshot of the new interface I1′ of the region we wish to
sync.
4I
≈ I ′ is shorthand for I ≾ I ′ ∧ I ′ ≾ I .
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:23
The other fact to prove before syncing the graph is to show that the region under modification
is itself a good state. This is indicated in line 15 by using good node predicates inside the dirty
region. To establish that n is a good node, for instance, we use a new interface In′ which gives it
an inflow of (1, 0), a node abstraction of ♦, and a flow map of one edge to r with label (1, 0). We
then check that the heap representation of n along with this interface satisfies γ . Similarly, we
check that l, with its new edge to n is still a good node. As the entire dirty region is a good state
with a contextually extended interface, we can now use sync (line 17) to update the graph. We can
then use the (Repl), (ReplIn), (ReplF), and (Comp) rules to establish that the new global interface
satisfies the invariant φ(I ′′).
The final state (line 18) is once again a good state, which means that we have shown both
memory safety and absence of memory leaks of this procedure. We have omitted the rest of the
insert procedure due to space constraints, which frees the node n and restarts if the CAS failed,
but note that this can be proved in a similar fashion. We have also omitted the reasoning about
interference by other threads, which is done in RGSep by checking that each intermediate assertion
is stable under the action of other threads. This is a syntactic check that can be done with the
help of the entailment lemmas presented in Fig. 9. The full proof can be found in §A. One can
also prove linearizability – for the Harris list, this requires a technique such as history variables
since linearization points are dynamic – but the invariants of the data structure needed to show
linearizability are expressible using flows.
This example shows that reasoning about programs using flow graphs and interfaces is as natural
as with inductive predicates, and we can use similar unrolling and abstraction lemmas to reason
about traversals of the data structure. Note that the intermediate assertions would have looked
almost identical if we did not have a free list – the only place where we reason about the free list is
in the global interface on line 2 and when we prove the local property Il ⊕ In ≈ Il′ ⊕ In′ on line 15.
7 APPLICATION 2: DICTIONARIES
In this section, we use the flow framework to verify a large class of concurrent dictionary implementations. We base our approach on the edgeset framework of Shasha and Goodman [1988] that
provides invariants, in terms of reachability properties of sets of keys, for proving linearizability
of dictionary operations (search, insert, and delete). Linearizability means, informally, that each
operation appears to happen atomically, i.e. at a single instant in time, and that if operation o 1
finishes before o 2 begins, then o 1 will appear to happen before o 2 .
By encoding the edgeset framework using flows, we obtain a method of proving linearizability
as well as memory safety. More importantly, we can use the power of flows to encode the data
constraints independently of the shape. Thus our encoding will be data-structure-agnostic, meaning
that we can verify any dictionary implementation that falls within the edgeset framework.
We first briefly describe the original edgeset framework, then show how to encode it using
flows. We then give an abstract algorithm template and specifications for the dictionary operations
that can be instantiated to concrete implementations. In §B we describe such an instantiation to a
nontrivial implementation based on B+ trees, and show how we can verify it using this framework.
7.1
The Edgeset Framework
A dictionary is a key-value store that implements three basic operations: search, insert, and delete.
For simplicity of exposition, we ignore the data values and treat the dictionary as containing only
keys. We refer to a thread seeking to search for, insert, or delete a key k as an operation on k, and
to k as the operation’s query key. Let KS be the set of possible keys, e.g., all integers.
The edgeset framework describes certain invariants on the data layout in dictionary implementations and an abstract algorithm that will be correct if the invariants are maintained. These
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:24
Siddharth Krishna, Dennis Shasha, and Thomas Wies
invariants do not specify the shape of the data structure to be say a tree, list, or hash table, and
instead describe properties of an abstract graph representation of the data structure (rather like the
flow graph). The nodes in the graph can represent, for instance, an actual heap node (in the case of
a list), an array cell (in the case of a hash table), or even a collection of fields and arrays (in the case
of a B-tree). Nodes are labeled with the set of keys stored at the node (henceforth, the contents of
the node). Each edge is labeled by a set of keys that we call the edgeset, and is defined as follows.
When a dictionary operation arrives at a node n, the set of query keys for which the operation
traverses an edge (n, n ′) is called the edgeset of (n, n ′). For example, in a BST, an operation on k
moves from a node to its left child nl if k is less than the key contained at the node, hence the
edgeset of (n, nl ) is {k | k < n.key}. Note that k can be in the edgeset of (n, n ′) even if n is not
reachable from any root; the edgeset is the set of query keys for which an operation would traverse
(n, n ′) assuming it somehow found itself at n.
The pathset of a path between nodes n 1 and n 2 is defined as the intersection of edgesets of every
edge on the path, and is thus the set of keys for which operations starting at n 1 would arrive at n 2
assuming neither the path nor the edgesets along that path change. For example, in a sorted list,
the pathset of a path from the head of a list to a node n is equal to the edgeset of the edge leading
into n (i.e. the set {k | n ′ .key < k} where n ′ is n’s predecessor). With this, we define the inset of a
node n as the union of the pathset of all paths from the root node to n. 5 If we take the inset of a
node n, and remove all the keys in the union of edgesets of edges leaving n, we get the keyset of n.
There are three desirable conditions on graphs representing dictionary data structures:
(GS1) The keysets of two distinct nodes are disjoint6 .
(GS2) The contents of every node are a subset of the keyset of that node.
(GS3) The edgesets of two distinct edges leaving a node are disjoint.
Intuitively, (GS1) and (GS2) tell us we can treat the keyset of n as the set of keys that n can potentially
contain. In this case, k is in the inset of n if and only if operations on k pass through n, and k is in
the keyset of n if and only if operations on k end up at n. (GS3) requires that there is a deterministic
path that operations follow, which is a desirable property that is true of all data structures in
common use. A dictionary state that satisfies these conditions is called a good state.
The Keyset Theorem of Shasha and Goodman [1988] states, informally, that if every atomic
operation preserves the good state property and k is in the keyset of a node n at the point when
the operation looks for, inserts, or deletes k at n, then the algorithm is linearizable. Intuitively, if k
is in the keyset of n, then since the keysets are disjoint we know that no other thread is performing
an operation on k at any other node. And once this operation acquires a lock on n (or establishes
exclusive access in another way, e.g. through a compare and swap), we know that operations on n
will be atomic.
The challenge with using this framework for a formal proof in an SL-based program logic is that
the invariants depend on quantities, like the inset, that are not local. The inset of a node n depends
both on the global root as well as all paths in the data structure from the root to n. We show next
how to convert the good state conditions into local properties of nodes using flows.
7.2
Encoding the Edgeset Framework using Flows
Given a potentially-infinite set of keys, KS, the set of subsets of KS forms a flow domain: (2KS , ⊆
, ∪, ∪, ∩, ∅, KS). The node domain must contain a set of keys to keep track of the contents of each
node, but we also wish to reason about locking. The challenge here is that threads may modify a
5 If
there are multiple roots, then each edge has a different edgeset depending on the root, and hence the definitions of
pathset and inset also depend on the particular root. The formalism can be extended to handle multiple roots in this manner.
6 The original paper required the keysets to partition KS, but we note that this weaker condition is sufficient for linearizability.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:25
o
n
procedure dictionaryOp(Key k) {
Φ
n
o
var c := r;
N(c, Ic ) −∗∗ Φ
n
o
while (true) {
N(c, Ic ) −∗∗ Φ
n
o
lock(c);
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t })
1
2
3
4
5
var n;
6
if (inRange(c, k)) {
n
o
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) ∧ k ∈ Ic In (c)
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) ∧ k ∈ Ic In (c)
7
n := findNext(c, k);
8
if (n == null) break;
o
n
(N(c, Ic ) ∗ N(n, I n )) −∗∗ Φ ∧ Ic a = (_, {t })
9
∧ (n , null ∧ k ∈ Ic f (c, n) ∨ n = null ∧ ∀x ∈ Ic f . k < Ic f (c, x ))
} else {
10
o
n
n := r;
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) ∧ n = r
o
n
}
(N(c, Ic ) ∗ N(n, I n )) −∗∗ Φ ∧ Ic a = (_, {t }) ∨ N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) ∧ c = n = r
11
12
unlock(c);
n
o
c := n;
N(c, Ic ) −∗∗ Φ
o
n
}
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) ∧ k ∈ Ic In (c) ∧ ∀x ∈ Ic f . k < Ic f (c, x )
n
o
var res := decisiveOp(c, k);
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t })
o
n
unlock(c);
N(c, Ic ) −∗∗ Φ
n
o
return res;
Φ
13
14
15
16
17
18
}
19
Fig. 11. The give-up template, with proof annotations in our logic.
n
N(c, Ic ) −∗∗ Φ
o
n
N(c, Ic ) −∗∗ Φ ∧ Ic a
n
N(c, Ic ) −∗∗ Φ ∧ Ic a
o
N(c, Ic′ ) −∗∗ Φ ∧ Ic′ a = (_, {t }) ∧ Ic ≈ Ic′
o
n
o
N(c, Ic′ ) −∗∗ Φ ∧ Ic ≈ Ic′
= (_, {t }) unlock(c);
o
n
o
= (_, {t }) res := inRange(c, k);
N(c, Ic ) −∗∗ Φ ∧ (r es ⇒ k ∈ Ic In (c))
lock(c);
n
o
N(c, Ic ) −∗∗ Φ ∧ n , null ∧ k ∈ Ic f (c, n)
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) n := findNext(c, k);
f
f
∨ n = null ∧ ∀x ∈ Ic . k < Ic (c, x )
a
In
n
o
N(c, Ic ) −∗∗ Φ ∧ Ic = (C, {t }) ∧ k ∈ Ic (c)
N(c, Ic′ ) −∗∗ Φ ∧ Ic ≈ Ic′ ∧ Ψ
res := decisiveOp(c, k);
f
f
∧ ∀x ∈ Ic . k < Ic (c, x )
a
′
I
=
(C,
∧
r
es
⇔
k
∈
C
for member
{t
})
c
where Ψ B Ic′ a = (C ∪ {k } , {t }) ∧ r es ⇔ k < C for insert
I ′ a = (C \ {k } , {t }) ∧ r es ⇔ k ∈ C for delete
c
n
Fig. 12. Specifications for helper functions.
locked node using a series of atomic operations such that the node does not satisfy the good state
conditions in between operations, thus making the Keyset Theorem inapplicable. We do not want
to use the dirty predicate to reason about such nodes, as this will complicate the global shared
state invariant. Instead, we label nodes with elements of N ⊎ N, where N B {x | x ∈ N}. Here, 0
denotes an unlocked node, t denotes a node locked by thread t, and t denotes a node locked by
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:26
Siddharth Krishna, Dennis Shasha, and Thomas Wies
thread t whose heap representation is out of sync. Formally, we use a product of sets of keys and
sets of augmented thread IDs as the node domain: (2KS × 2N⊎N , ⊆, ∪, (∅, ∅)), where ⊆ and ∪ are
lifted component-wise. In the following, all uses of t implicitly assert that the label is not x for any
x ∈ N.
Lemma 7.1. For any graph G, cap(G)(n, n ′) is the set of keys k for which there exists a path from n
to n ′, every edge of which contains k in its edge label. In particular, if in(x) = ITE(x = r , KS, ∅) is an
inflow on G, then flow(in, G)(n) is the set of keys k for which such a path exists from r to n.
If we label each graph edge with its edgeset, then by Lemma 7.1, the flow at each node is the
inset of that node. We can encode (GS3) easily using the good condition at each node, and by using
sets of keys as the node domain, we can also encode (GS2). We further observe that for graphs with
one root, edgesets leaving a node being pairwise disjoint implies that the keysets of every pair of
nodes is disjoint7 . We can thus use a global data structure invariant that the inflow of the root is
KS and all other nodes ∅ to obtain (GS1).
This motivates us to use the following good condition
γ (x, in, (C,T ), f ) B ∃t . (γд (x, in, C, t, f ) ∧ T = {t } ∨ γb (x, t) ∧ t , 0 ∧ T = {t })
∧ C ⊆ in(x) ∧ ∀y. (C ∩ f (x, y) = ∅ ∧ ∀z. f (x, y) ∩ f (x, z) = ∅)
where γд and γb are user-specified SL predicates. γд is to be instantiated with the heap implementation of a node in sync, and γb with a description of a node that may not be in sync. The global data
structure invariant enforces that the graph has no outgoing edges and that only the root gets a
non-zero inflow:
Φ B ∃I , in. Gr(I ) ∧ in ∈ I In ∧ in + 0 = {r ↣ KS} + 0 ∧ I f = ϵ
Lemma 7.2. The flow graph component of any state satisfying Φ satisfies the good state conditions
(GS1) to (GS3).
The original edgeset paper [Shasha and Goodman 1988] set out three template algorithms, based
on different locking disciplines, along with invariants that implied their linearizability. Of these,
we formalize the template based on the give-up technique here. Our method can be easily extended
to the lock-coupling template, but we note that most practical algorithms use more fine-grained
locking schemes. For the link technique, the third template, we need to encode the inreach, an
inductive quantity depending on the keyset and edgesets, viz. k is in the inreach of a node n if k is
in the keyset of n or k is in the keyset of n ′ and for every edge e in the path from n to n ′, k is in the
edgeset of e. Being able to express the inreach would also give a simpler way to prove linearizability
of the full Harris list. We would need to extend our framework to support second-order flows to
define the inreach, and we leave this for future work.
We now describe the give-up template algorithm, shown in Fig. 11. This template can be used to
build implementations of all three dictionary operations by defining the function decisiveOp as
described below. The basic idea is that every node stores a range, which is an under-approximation
(i.e. always a subset) of its inset. An operation on key k proceeds by starting at the root and enters
a loop at line 3 where it follows edges that contain k in their edgeset. Between two nodes, the
algorithm has no locks on either node, to allow for more parallelism. Because of this, when it
arrives at some node c, the first thing it does after locking it is to check c’s range field (the call to
inRange in line 6) to ensure that the dictionary operation is at a node whose inset8 contains k. If
the check succeeds, then it calls another helper, findNext at line 7, that checks if there exists a
7 For
graphs with several roots, each edgeset is defined with respect to each root and then this property holds for each root.
is shorthand for in(x ) for some in ∈ I In .
8 I In (x )
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:27
node n such that k is in the edgeset of (c, n). If there is no such node, then we know that k must be
in the keyset of c, and we break from the loop (line 8) while holding the lock on c. If the inRange
check fails, then the algorithm gives up and starts again from the root r (line 13)9 . If the search
continues, the algorithm first unlocks c on line 15 before reassigning c to the next node n.
When the algorithm breaks out of the loop, k must be in the keyset of c (line 15), so it calls
decisiveOp to perform the operation on c. It then unlocks c, and returns the result of decisiveOp.
Proof. The proof annotations for this template (also shown in Fig. 11) are fairly straightforward,
and entailments between them can be derived using the lemmas in Fig. 9, assuming that the userprovided implementations of the helper functions satisfy the specifications in Fig. 12. Note that
these specifications are in terms of the entire shared state, even though the helper functions only
modify the current node c. This is a limitation of RGSep, and one can obtain local specifications
for these functions by switching to a more advanced logic such as [Feng 2009a]. To reason about
interference, we use the following actions to specify the modifications to the shared state allowed
by a set of thread IDs T :
t ∈ T ∧ N(x, (In, (C, {0}), f )) ⇝ N(x, (In, (C,T ′), f )) ∧ T ′ ⊆ {t, t }
(Lock)
t ∈ T ∧ emp ⇝ N(x, ({{x ↣ 0}} , (∅, {t }), ϵ))
a
t ∈ T ∧ Gr(I ) ∧ I ⊑ (_, {t, t }) ⇝ Gr(I ) ∧ I
′
′a
⊑ (_, {0, t, t }) ∧ I ≾ I
(Alloc)
′
(Sync)
(Lock) allows a thread t ∈ T to lock an unlocked node; (Alloc) allows t to add new nodes with
no inflow, contents, or outgoing edges; and (Sync) allows t to modify a locked region arbitrarily,
as long as the new interface contextually extends the old one. The last action also allows it to
unlock nodes it has locked. The guarantee of thread with id t 0 is made up of the above actions with
T = {t 0 }, while the rely constitutes the above actions with T = N \ {t 0 }. To complete the proof that
the template algorithm is memory safe and preserves the global data structure invariant, one must
show the stability of every intermediate assertion in Fig. 11. This can be done syntactically using
our lemmas, and an example of such a proof can be seen in §C.
Proving Linearizability. To prove that the template algorithm is linearizable, we adapt the Keyset
Theorem to the language of flows. To show that all atomic operations maintain the good state
conditions, we require that every intermediate assertion in our proof implies Φ. This is true of
assertions in our template proof, but also needs to be true in the intermediate assertions used to
prove that implementations of helper functions meet their specification. The condition about the
query key k being in the keyset of the node at which a thread performs its operation is captured by
the specifications in Fig. 12. We additionally require that each dictionary operation only modifies
the contents of the global once, and that no other operation modifies the contents. We thus obtain
the following re-statement of the Keyset Theorem:
Theorem 7.3. An implementation of the give-up template from Fig. 11 is memory safe and linearizable if the following conditions hold:
(1) The helper functions satisfy the specification in Fig. 12 under the rely and guarantee specified
above, with all intermediate assertions implying Φ ∗ true.
(2) Every execution of decisiveOp must have at most one call to sync that changes the contents
of the flow graph.
(3) All other calls to sync, including those in maintenance operations, do not change the contents of
the graph region on which they operate.
9 We
could, in theory, jump to any ancestor of that node. This will require a variable to keep track of the ancestor, and
similar reasoning can be used.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:28
Siddharth Krishna, Dennis Shasha, and Thomas Wies
Alternatively, our RGSep[FI] proof has established sufficient invariants to directly prove linearizability. One way to do this [Vafeiadis 2009] is to use auxiliary variables to track the abstract state of
the data structure (we already have this as the set of contents in the global interface) and atomically
execute the specification of each operation on this abstract state at the linearization point. One can
then use a write-once variable to store the result of the abstract operation, and at the end of the
operation prove that the implementation returns the same value as the specification.
This template algorithm can now be instantiated to any concrete implementation by providing
predicates γд and γb to describe the heap layout of the data structure and implementing the helper
functions such that they satisfy the conditions of Theorem 7.3. If there are any maintenance
operations, for example splitting or merging nodes, these must also satisfy the invariants in the
theorem under the given rely and guarantee. An example implementation of this template, the B+
tree, can be seen in §B.
8
RELATED WORK
Abstraction Mechanisms in Separation Logic. The prevalent mechanism for abstracting unbounded
heap regions in separation logic is based on inductive predicates defined by separating conjunctions [Berdine et al. 2004; Brotherston et al. 2011; Cook et al. 2011; Enea et al. 2015; Iosif et al.
2014; Pek et al. 2014; Reynolds 2002]. For simple inductive data structures whose implementation
follows regular traversal patterns, inductive predicates can be easier to work with than flow-based
abstractions. Certain abstractions, for instance abstracting a list as a sequence of values, while
possible to encode in flows, would be more natural with inductive predicates. However, as discussed
in §2, inductive predicates are often ill-suited for abstracting concurrent data structures. The reasons
include the dependence on specific traversal patterns in the inductive definitions of the predicates,
and the implied restrictions on expressing sharing and data structure overlays with separating
conjunction.
Ramifications [Hobor and Villard 2013; Mehnert et al. 2012] offer an alternative approach to
reasoning about overlaid data structures and data structures with unrestricted sharing. These
approaches express inductively defined graph abstractions by using overlapping conjunctions of
subheaps instead of separating conjunction. However, this necessitates complicated ramification
entailments involving magic wand to reason about updates. We do not need to use combinations of
separating conjunction and overlapping conjunction or reason about entailments involving the
notorious magic wand. Instead, we shift the reasoning to the composition and decomposition of
inflows and flow maps. While showing that a flow graph satisfies conditions on the flow in general
requires computing fixpoints over the graph, in proofs we only need to reason about them when
a heap region is modified. These computations are typically easy because concurrent algorithms
modify a bounded number of nodes at a time in order to minimize interference. Finally, we also
obtain a uniform, and decoupled, way to reason about both shape and data properties, yielding
abstractions that generalize over a wide variety of data structures such as in our encoding of the
edgeset framework.
Yet another alternative for abstracting arbitrary graphs in SL is to use iterated separating
conjunction [Müller et al. 2016; Raad et al. 2016; Reynolds 2002; Yang 2001]. Similar to flow
interfaces, such abstractions are not tied to specific traversal patterns and can capture invariants
that are expressible as local conditions on nodes such as that the graph is closed. However, unlike
flow interfaces, iterated separating conjunctions cannot capture inductive properties of a graph
(e.g. that the reachable nodes from a root form a tree). In essence, flow interfaces occupy a sweet
spot between inductive predicates and iterated separating conjunctions.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:29
Nondeterministic Monoidal Models of Separation Logic. Our notion of flow graph composition
naturally yields a nondeterministic monoidal model of SL where separating conjunction ∗ is
interpreted as a ternary relation. However, the conventional meta-theory of SL [Calcagno et al.
2007; Dockins et al. 2009] requires ∗ to be partial-deterministic. We overcome this mismatch here
by defining an appropriate equivalence relation on flow graphs (or, more precisely, their inflows)
to enforce a functional interpretation of ∗. However, this solution leads to a slightly stronger
model than is strictly necessary. It also adds some artificial complexity to the logic as the inflow
equivalence classes must be reasoned about at the syntactic level. We believe that both of these
issues can be avoided by considering a nondeterministic monoidal semantics as the starting point
for the development of the program logic. Such semantics have been studied in the context of
substructural logics, Boolean BI [Galmiche and Larchey-Wendling 2006], and more recently to
obtain a more general proof theory of propositional abstract SL [Hóu et al. 2014]. To our knowledge,
flow graphs constitute the first example of a separation algebra with nondeterministic monoidal
structure that has practical applications in program verification.
Concurrent Separation Logics. CSL was introduced by O’Hearn [2004]. A recent article by Brookes
and O’Hearn [2016] provides a survey of the development of CSLs since then. Among the many
improvements that have been developed are the idea of fractional permissions to reason about
shared reads [Bornat et al. 2005; Heule et al. 2013], combinations of rely/guarantee reasoning [Feng
et al. 2007; Vafeiadis and Parkinson 2007], and abstraction mechanisms for reasoning about different aspects of concurrency such as synchronization protocols for low-level lock implementations [Nanevski et al. 2014], and atomicity abstractions [da Rocha Pinto et al. 2014; Dinsdale-Young
et al. 2010; Xiong et al. 2017]. We only discuss the most closely related works and otherwise refer
the reader to the detailed discussion in [Brookes and O’Hearn 2016].
Vafeiadis and Parkinson introduced RGSep [Vafeiadis and Parkinson 2007] to marry rely/guarantee reasoning with separation logic. RGSep allows the actions of threads to be described by
SL assertions, thereby abstracting from the specifics of the synchronization mechanism between
threads. We chose RGSep as the basis for our logic because it provides a good trade-off between
flexibility and simplicity when reasoning about concurrent data structure implementations. However, the logic also has limitations. One limitation is that it does not allow the shared state to be
split and framed out. This is less of an issue in our work as the global shared state can always be
described by a single Gr predicate and carrying this predicate through all proofs does not add too
much complexity. This limitation of RGSep was addressed in local RGSep [Feng 2009b] and our
results can be easily adapted to that logic.
Another limitation of RGSep is that it lacks abstraction mechanisms that enable compositional
reasoning about data structure clients. A number of solution to this problem have been proposed,
including the views framework [Dinsdale-Young et al. 2013], Concurrent Abstract Predicates
(CAP) [Dinsdale-Young et al. 2010], and TaDA [da Rocha Pinto et al. 2014]. The abstractions
provided by these logics are orthogonal to the ones developed here. These logics have been used,
e.g., to verify specifications of concurrent dictionary implementations based on B-trees and skip
lists that enable compositional client verification [da Rocha Pinto et al. 2011; Xiong et al. 2017]. We
believe that the proofs developed in [Xiong et al. 2017] can be further simplified by introducing
flow interfaces as an intermediate abstraction of the considered data structures.
Higher-order concurrent separation logic [Jung et al. 2015; Krebbers et al. 2017] can express
ghost state within the assertion language of the logic itself. This feature can be used to eliminate
the restriction of RGSep[FI] that the semantics of the flow interface predicates is defined on the
meta level and that it does not support nesting of flow interface abstractions (i.e., cases where a
node of a flow graph should abstract from another flow graph contained in the node). Similarly,
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:30
Siddharth Krishna, Dennis Shasha, and Thomas Wies
higher-order CSL can express complex linearizability proofs directly without relegating a part of
the proof argument to the meta level.
Invariant Inference. There is a large body of work on inferring invariants for heap-manipulating
programs (e.g., [Sagiv et al. 2002]), including techniques based on separation logic [Calcagno et al.
2009; Distefano et al. 2006; Vafeiadis 2010]. Many of these approaches rely on some form of abstract
interpretation [Cousot and Cousot 1977]. We believe that the least fixpoint characterization of
flows in flow interfaces lends itself well to abstract interpretation techniques.
9
CONCLUSION
We have introduced flow interfaces as a novel approach to the abstraction of unbounded data
structures in separation logic. The approach avoids several limitations of common solutions to
such abstraction, allows unrestricted sharing and arbitrary traversals of heap regions, and provides
a uniform treatment of data constraints. We have shown that flow interfaces are particularly
well suited for reasoning about concurrent data structures and that they hold great promise for
developing automated techniques for reasoning about implementation-agnostic abstractions.
REFERENCES
Josh Berdine, Cristiano Calcagno, and Peter O’Hearn. 2004. A Decidable Fragment of Separation Logic. In FSTTCS. Springer.
Richard Bornat, Cristiano Calcagno, Peter O’Hearn, and Matthew Parkinson. 2005. Permission Accounting in Separation
Logic. In Proceedings of the 32nd ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL ’05).
ACM, New York, NY, USA, 259–270. https://doi.org/10.1145/1040305.1040327
Richard Bornat, Cristiano Calcagno, and Hongseok Yang. 2006. Variables As Resource in Separation Logic. Electron. Notes
Theor. Comput. Sci. 155 (May 2006), 247–276. https://doi.org/10.1016/j.entcs.2005.11.059
Ahmed Bouajjani, Cezara Dragoi, Constantin Enea, and Mihaela Sighireanu. 2012. Accurate Invariant Checking for Programs
Manipulating Lists and Arrays with Infinite Data. In ATVA (LNCS), Vol. 7561. Springer, 167–182.
Stephen Brookes and Peter W. O’Hearn. 2016. Concurrent separation logic. SIGLOG News 3, 3 (2016), 47–65. https:
//doi.org/10.1145/2984450.2984457
James Brotherston, Dino Distefano, and Rasmus Lerchedahl Petersen. 2011. Automated Cyclic Entailment Proofs in
Separation Logic. In 23rd International Conference on Automated Deduction, CADE-23 (Lecture Notes in Computer Science),
Vol. 6803. Springer, 131–146. https://doi.org/10.1007/978-3-642-22438-6_12
Cristiano Calcagno, Dino Distefano, Peter W. O’Hearn, and Hongseok Yang. 2009. Compositional shape analysis by means
of bi-abduction. In POPL.
Cristiano Calcagno, Peter W. O’Hearn, and Hongseok Yang. 2007. Local Action and Abstract Separation Logic. In 22nd IEEE
Symposium on Logic in Computer Science (LICS 2007). IEEE Computer Society, 366–378. https://doi.org/10.1109/LICS.2007.
30
Adam Chlipala. 2011. Mostly-automated verification of low-level programs in computational separation logic. In Proceedings
of the 32nd ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI 2011, San Jose, CA,
USA, June 4-8, 2011. ACM, 234–245. https://doi.org/10.1145/1993498.1993526
Byron Cook, Christoph Haase, Joël Ouaknine, Matthew Parkinson, and James Worrell. 2011. Tractable Reasoning in a
Fragment of Separation Logic. In CONCUR. Springer.
Patrick Cousot and Radhia Cousot. 1977. Abstract interpretation: a unified lattice model for static analysis of programs by
construction or approximation of fixpoints. In Proc. 4th POPL.
Pedro da Rocha Pinto, Thomas Dinsdale-Young, Mike Dodds, Philippa Gardner, and Mark J. Wheelhouse. 2011. A simple
abstraction for complex concurrent indexes. In Proceedings of the 26th Annual ACM SIGPLAN Conference on ObjectOriented Programming, Systems, Languages, and Applications, OOPSLA 2011. ACM, 845–864. https://doi.org/10.1145/
2048066.2048131
Pedro da Rocha Pinto, Thomas Dinsdale-Young, and Philippa Gardner. 2014. TaDA: A Logic for Time and Data Abstraction.
In 28th European Conference on Object-Oriented Programming, ECOOP 2014 (Lecture Notes in Computer Science), Vol. 8586.
Springer, 207–231. https://doi.org/10.1007/978-3-662-44202-9_9
Pedro da Rocha Pinto, Thomas Dinsdale-Young, Philippa Gardner, and Julian Sutherland. 2016. Modular Termination
Verification for Non-blocking Concurrency. In 25th European Symposium on Programming, ESOP 2016 (Lecture Notes in
Computer Science), Vol. 9632. Springer, 176–201. https://doi.org/10.1007/978-3-662-49498-1_8
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:31
Thomas Dinsdale-Young, Lars Birkedal, Philippa Gardner, Matthew J. Parkinson, and Hongseok Yang. 2013. Views:
compositional reasoning for concurrent programs. In The 40th Annual ACM SIGPLAN-SIGACT Symposium on Principles
of Programming Languages, POPL ’13. ACM, 287–300. https://doi.org/10.1145/2429069.2429104
Thomas Dinsdale-Young, Mike Dodds, Philippa Gardner, Matthew J. Parkinson, and Viktor Vafeiadis. 2010. Concurrent
Abstract Predicates. In 24th European Conference on Object-Oriented Programming, ECOOP 2010 (Lecture Notes in Computer
Science), Vol. 6183. Springer, 504–528. https://doi.org/10.1007/978-3-642-14107-2_24
Dino Distefano, Peter W. O’Hearn, and Hongseok Yang. 2006. A Local Shape Analysis Based on Separation Logic. In Tools
and Algorithms for the Construction and Analysis of Systems, 12th International Conference, TACAS 2006 (Lecture Notes in
Computer Science), Vol. 3920. Springer, 287–302. https://doi.org/10.1007/11691372_19
Robert Dockins, Aquinas Hobor, and Andrew W. Appel. 2009. A Fresh Look at Separation Algebras and Share Accounting.
In Proceedings of the 7th Asian Symposium on Programming Languages and Systems, APLAS 2009. Springer-Verlag, Berlin,
Heidelberg, 161–177. https://doi.org/10.1007/978-3-642-10672-9_13
Mike Dodds, Suresh Jagannathan, Matthew J. Parkinson, Kasper Svendsen, and Lars Birkedal. 2016. Verifying Custom
Synchronization Constructs Using Higher-Order Separation Logic. ACM Trans. Program. Lang. Syst. 38, 2, Article 4 (Jan.
2016), 72 pages. https://doi.org/10.1145/2818638
Constantin Enea, Ondrej Lengál, Mihaela Sighireanu, and Tomás Vojnar. 2017. SPEN: A Solver for Separation Logic. In
NASA Formal Methods, NFM 2017 (Lecture Notes in Computer Science), Vol. 10227. Springer, 302–309. https://doi.org/10.
1007/978-3-319-57288-8_22
Constantin Enea, Mihaela Sighireanu, and Zhilin Wu. 2015. On Automated Lemma Generation for Separation Logic with
Inductive Definitions. In 13th International Symposium on Automated Technology for Verification and Analysis, ATVA 2015
(Lecture Notes in Computer Science), Vol. 9364. Springer, 80–96. https://doi.org/10.1007/978-3-319-24953-7_7
Xinyu Feng. 2009a. Local rely-guarantee reasoning. In Proceedings of the 36th ACM SIGPLAN-SIGACT Symposium on
Principles of Programming Languages, POPL 2009, Savannah, GA, USA, January 21-23, 2009, Zhong Shao and Benjamin C.
Pierce (Eds.). ACM, 315–327. https://doi.org/10.1145/1480881.1480922
Xinyu Feng. 2009b. Local Rely-guarantee Reasoning. In Proceedings of the 36th Annual ACM SIGPLAN-SIGACT Symposium
on Principles of Programming Languages (POPL ’09). ACM, New York, NY, USA, 315–327. https://doi.org/10.1145/1480881.
1480922
Xinyu Feng, Rodrigo Ferreira, and Zhong Shao. 2007. On the Relationship Between Concurrent Separation Logic and
Assume-Guarantee Reasoning. In 16th European Symposium on Programming, ESOP 2007 (Lecture Notes in Computer
Science), Vol. 4421. Springer, 173–188. https://doi.org/10.1007/978-3-540-71316-6_13
Didier Galmiche and Dominique Larchey-Wendling. 2006. Expressivity Properties of Boolean BI Through Relational Models.
In 26th International Conference on Foundations of Software Technology and Theoretical Computer Science, FSTTCS 2006,
S. Arun-Kumar and Naveen Garg (Eds.). Springer, Berlin, Heidelberg, 357–368. https://doi.org/10.1007/11944836_33
Ronghui Gu, Jérémie Koenig, Tahina Ramananandro, Zhong Shao, Xiongnan (Newman) Wu, Shu-Chun Weng, Haozhong
Zhang, and Yu Guo. 2015. Deep Specifications and Certified Abstraction Layers. In Proceedings of the 42nd Annual ACM
SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL 2015. ACM, 595–608. https://doi.org/10.
1145/2676726.2676975
Timothy L. Harris. 2001. A Pragmatic Implementation of Non-blocking Linked-Lists. In Distributed Computing, 15th
International Conference, DISC 2001, Lisbon, Portugal, October 3-5, 2001, Proceedings (Lecture Notes in Computer Science),
Vol. 2180. Springer, 300–314. https://doi.org/10.1007/3-540-45414-4_21
Maurice Herlihy and Jeannette M. Wing. 1990. Linearizability: A Correctness Condition for Concurrent Objects. ACM Trans.
Program. Lang. Syst. 12, 3 (1990), 463–492. https://doi.org/10.1145/78969.78972
Stefan Heule, K. Rustan M. Leino, Peter Müller, and Alexander J. Summers. 2013. Abstract Read Permissions: Fractional
Permissions without the Fractions. In 14th International Conference on Verification, Model Checking, and Abstract Interpretation, VMCAI 2013 (Lecture Notes in Computer Science), Vol. 7737. Springer, 315–334. https://doi.org/10.1007/
978-3-642-35873-9_20
Aquinas Hobor and Jules Villard. 2013. The Ramifications of Sharing in Data Structures. In Proceedings of the 40th Annual
ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL ’13). ACM, New York, NY, USA,
523–536. https://doi.org/10.1145/2429069.2429131
Zhé Hóu, Ranald Clouston, Rajeev Goré, and Alwen Tiu. 2014. Proof Search for Propositional Abstract Separation Logics via
Labelled Sequents. In Proceedings of the 41st ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages,
POPL 2014. ACM, New York, NY, USA, 465–476. https://doi.org/10.1145/2535838.2535864
Radu Iosif, Adam Rogalewicz, and Tomás Vojnar. 2014. Deciding Entailments in Inductive Separation Logic with Tree
Automata. In 12th International Symposium on Automated Technology for Verification and Analysis, ATVA 2014 (Lecture
Notes in Computer Science), Vol. 8837. Springer, 201–218. https://doi.org/10.1007/978-3-319-11936-6_15
Ralf Jung, David Swasey, Filip Sieczkowski, Kasper Svendsen, Aaron Turon, Lars Birkedal, and Derek Dreyer. 2015. Iris:
Monoids and Invariants As an Orthogonal Basis for Concurrent Reasoning. In Proceedings of the 42Nd Annual ACM
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:32
Siddharth Krishna, Dennis Shasha, and Thomas Wies
SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL ’15). ACM, New York, NY, USA, 637–650.
https://doi.org/10.1145/2676726.2676980
Robbert Krebbers, Ralf Jung, Ales Bizjak, Jacques-Henri Jourdan, Derek Dreyer, and Lars Birkedal. 2017. The Essence of
Higher-Order Concurrent Separation Logic. In 26th European Symposium on Programming, ESOP 2017 (Lecture Notes in
Computer Science), Vol. 10201. Springer, 696–723. https://doi.org/10.1007/978-3-662-54434-1_26
Siddharth Krishna, Dennis Shasha, and Thomas Wies. 2018. Go with the Flow: Compositional Abstractions for Concurrent
Data Structures. Proc. ACM Program. Lang. 2, POPL, Article 37 (Jan. 2018), 31 pages. https://doi.org/10.1145/3158125
Philip L. Lehman and s. Bing Yao. 1981. Efficient Locking for Concurrent Operations on B-trees. ACM Trans. Database Syst.
6, 4 (Dec. 1981), 650–670. https://doi.org/10.1145/319628.319663
Hannes Mehnert, Filip Sieczkowski, Lars Birkedal, and Peter Sestoft. 2012. Formalized Verification of Snapshotable Trees:
Separation and Sharing. In Verified Software: Theories, Tools, Experiments - 4th International Conference, VSTTE 2012,
Philadelphia, PA, USA, January 28-29, 2012. Proceedings (Lecture Notes in Computer Science), Vol. 7152. Springer, 179–195.
https://doi.org/10.1007/978-3-642-27705-4_15
Peter Müller, Malte Schwerhoff, and Alexander J. Summers. 2016. Automatic Verification of Iterated Separating Conjunctions
Using Symbolic Execution. In Computer Aided Verification CAV (Lecture Notes in Computer Science), Vol. 9779. Springer,
405–425. https://doi.org/10.1007/978-3-319-41528-4_22
Aleksandar Nanevski, Ruy Ley-Wild, Ilya Sergey, and Germán Andrés Delbianco. 2014. Communicating State Transition
Systems for Fine-Grained Concurrent Resources. In 23rd European Symposium on Programming, ESOP 2014 (Lecture Notes
in Computer Science), Vol. 8410. Springer, 290–310. https://doi.org/10.1007/978-3-642-54833-8_16
Huu Hai Nguyen and Wei-Ngan Chin. 2008. Enhancing Program Verification with Lemmas. In 20th International Conference
on Computer Aided Verification, CAV 2008 (Lecture Notes in Computer Science), Vol. 5123. Springer, 355–369. https:
//doi.org/10.1007/978-3-540-70545-1_34
Peter W. O’Hearn. 2004. Resources, Concurrency and Local Reasoning. In CONCUR 2004 - Concurrency Theory, 15th
International Conference, London, UK, August 31 - September 3, 2004, Proceedings (Lecture Notes in Computer Science),
Vol. 3170. Springer, 49–67. https://doi.org/10.1007/978-3-540-28644-8_4
Peter W. O’Hearn, John C. Reynolds, and Hongseok Yang. 2001. Local Reasoning about Programs that Alter Data Structures.
In Computer Science Logic, 15th International Workshop, CSL 2001. 10th Annual Conference of the EACSL, Paris, France,
September 10-13, 2001, Proceedings (Lecture Notes in Computer Science), Laurent Fribourg (Ed.), Vol. 2142. Springer, 1–19.
https://doi.org/10.1007/3-540-44802-0_1
Edgar Pek, Xiaokang Qiu, and P. Madhusudan. 2014. Natural proofs for data structure manipulation in C using separation
logic. In ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI ’14, Edinburgh, United
Kingdom - June 09 - 11, 2014. ACM, 440–451. https://doi.org/10.1145/2594291.2594325
Juan Antonio Navarro Pérez and Andrey Rybalchenko. 2011. Separation logic + superposition calculus = heap theorem
prover. In PLDI. ACM, 556–566.
Ruzica Piskac, Thomas Wies, and Damien Zufferey. 2013. Automating Separation Logic Using SMT. In CAV (LNCS), Vol. 8044.
Springer, 773–789.
Azalea Raad, Aquinas Hobor, Jules Villard, and Philippa Gardner. 2016. Verifying Concurrent Graph Algorithms. In Asian
Symposium on Programming Languages and Systems APLAS (Lecture Notes in Computer Science), Vol. 10017. 314–334.
https://doi.org/10.1007/978-3-319-47958-3_17
Andrew Reynolds, Radu Iosif, and Cristina Serban. 2017. Reasoning in the Bernays-Schönfinkel-Ramsey Fragment of
Separation Logic. In Verification, Model Checking, and Abstract Interpretation, VMCAI 2017 (Lecture Notes in Computer
Science), Vol. 10145. Springer, 462–482. https://doi.org/10.1007/978-3-319-52234-0_25
John C. Reynolds. 2002. Separation Logic: A Logic for Shared Mutable Data Structures. In 17th IEEE Symposium on Logic
in Computer Science (LICS 2002), 22-25 July 2002, Copenhagen, Denmark, Proceedings. IEEE Computer Society, 55–74.
https://doi.org/10.1109/LICS.2002.1029817
Shmuel Sagiv, Thomas W. Reps, and Reinhard Wilhelm. 2002. Parametric shape analysis via 3-valued logic. ACM Trans.
Program. Lang. Syst. 24, 3 (2002), 217–298. https://doi.org/10.1145/514188.514190
Dennis E. Shasha and Nathan Goodman. 1988. Concurrent Search Structure Algorithms. ACM Trans. Database Syst. 13, 1
(1988), 53–90. https://doi.org/10.1145/42201.42204
Makoto Tatsuta, Quang Loc Le, and Wei-Ngan Chin. 2016. Decision Procedure for Separation Logic with Inductive Definitions
and Presburger Arithmetic. In 14th Asian Symposium on Programming Languages and Systems, APLAS 2016 (Lecture Notes
in Computer Science), Vol. 10017. Springer, 423–443. https://doi.org/10.1007/978-3-319-47958-3_22
Viktor Vafeiadis. 2008. Modular fine-grained concurrency verification. Ph.D. Dissertation. University of Cambridge, UK.
http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.612221
Viktor Vafeiadis. 2009. Shape-Value Abstraction for Verifying Linearizability. In Verification, Model Checking, and Abstract
Interpretation, 10th International Conference, VMCAI 2009, Savannah, GA, USA, January 18-20, 2009. Proceedings (Lecture
Notes in Computer Science), Neil D. Jones and Markus Müller-Olm (Eds.), Vol. 5403. Springer, 335–348. https://doi.org/10.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:33
1007/978-3-540-93900-9_27
Viktor Vafeiadis. 2010. RGSep Action Inference. In Verification, Model Checking, and Abstract Interpretation, 11th International
Conference, VMCAI 2010, Madrid, Spain, January 17-19, 2010. Proceedings (Lecture Notes in Computer Science), Gilles Barthe
and Manuel V. Hermenegildo (Eds.), Vol. 5944. Springer, 345–361. https://doi.org/10.1007/978-3-642-11319-2_25
Viktor Vafeiadis and Matthew J. Parkinson. 2007. A Marriage of Rely/Guarantee and Separation Logic. In 18th International
Conference on Concurrency Theory, CONCUR 2007 (Lecture Notes in Computer Science), Vol. 4703. Springer, 256–271.
https://doi.org/10.1007/978-3-540-74407-8_18
Shale Xiong, Pedro da Rocha Pinto, Gian Ntzik, and Philippa Gardner. 2017. Abstract Specifications for Concurrent Maps. In
26th European Symposium on Programming, ESOP 2017 (Lecture Notes in Computer Science), Vol. 10201. Springer, 964–990.
https://doi.org/10.1007/978-3-662-54434-1_36
Hongseok Yang. 2001. An example of local reasoning in BI pointer logic: the Schorr-Waite graph marking algorithm. In
Proceedings of the SPACE Workshop.
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:34
A
Siddharth Krishna, Dennis Shasha, and Thomas Wies
THE HARRIS LIST: COMPLETE PROOF ANNOTATIONS
In the following proof sketches, t is the thread ID of the current thread. We use an augmented
global shared state invariant that also enforces that ft is a node in the shared state:
Φ B ∃I . Gr(I )∧φ(I ),
1
2
3
φ(I ) B ∃in ∈ I in . in+0 = {mh ↣ (1, 0)}+{fh ↣ (0, 1)}+0∧ft ∈ I ∧I f = ϵ
procedure insert() {
o
n
Φ
n
o
var l := mh;
N(l, Il ) −∗∗ Φ ∧ l = mh
4
var r := getUnmarked(l.next);
5
while (r != null && nondet()) {
o
(N(l, Il ) ∗ N(r, I r )) −∗∗ Φ
n
o
l := r; r := getUnmarked(l.next);
(N(l, Il )) −∗∗ Φ ∧ r , null ⇒ (N(l, Il ) ∗ N(r, I r )) −∗∗ Φ
6
7
}
8
if (!isMarked(r)) {
10
11
13
14
15
16
17
18
19
n
o
N(l, Il ) −∗∗ Φ ∧ ¬M (r )
n
o
var n := new Node(r, null);
N(l, Il ) −∗∗ Φ ∗ n 7→ r, null ∧ ¬M (r )
n
o
mark(n, n);
N(l, Il ) −∗∗ Φ ∗ [n 7→ r, null]In ∧ ¬M (r ) // Where: I n := ({n ↣ (0, 0)} , ⊥, ϵ )
9
12
n
atomic { // CAS(l.next, r, n)
if (l.next == r) {
o
n
[l 7→ r, _]Il ∧ N(l, Il ) −∗∗ Φ ∗ [n 7→ r, null]In ∧ Il a = ♦ // (Conc)
n
o
[l 7→ r, _]Il ∧ N(l, Il ) ∗ Gr(I 2 ) ∗ [n 7→ r, null]In ∧ Il a = ♦ ∧ I = Il ⊕ I 2 ∧ φ(I ) // (GrDecomp)
n
o
[l 7→ r, _]Il ∧ N(l, Il ) ∗ [n 7→ r, null]In ∗ Gr(I 2 ) ∧ Il a = ♦ ∧ I ′ = Il ⊕ I n ⊕ I 2 ∧ φ(I ′ ) // (Comp), (AddIn),
(AddF)
l.next := n;
n
o
N(l, Il′ ) I ∗ [n 7→ r, null]In ∗ Gr(I 2 ) ∧ Il a = ♦ ∧ I ′ = Il ⊕ I n ⊕ I 2 ∧ φ(I ′ )
l
n
o
N(l, Il′ ) ∗ N(n, I n′ ) I ∗ Gr(I 2 ) ∧ Il′ a = ♦ ∧ I n′ a = ♦ ∧ I ′ = Il ⊕ I n ⊕ I 2 ∧ φ(I ′ ) ∧ I 1 = Il ⊕ I n ≈ Il′ ⊕ I n′
1
var I′1 :| Gr(I 1′ ) I ∧ {l ↣ _, n ↣ _} ∈ I 1 In ;
1
sync(I′1 );
N(l, Il′ ) ∗ N(n, I n′ ) ∗ Gr(I 2 ) ∧ Il′ a = ♦ ∧ I n′ a = ♦ ∧ I ′ = Il ⊕ I n ⊕ I 2 ∧ φ(I ′ ) ∧ Il ⊕ I n ≈ Il′ ⊕ I n′
N(l, Il′ ) ∗ N(n, I n′ ) ∗ Gr(I 2 ) ∧ Il′ a = ♦ ∧ I n′ a = ♦ ∧ I ′′ = Il′ ⊕ I n′ ⊕ I 2 ∧ φ(I ′′ )
a
a
var b := true; (N(l, Il′ ) ∗ N(l, Il′ )) −∗∗ Φ ∧ Il′ = ♦ ∧ I n′ = ♦ ∧ b
o
n
} else var b := false;
N(l, Il ) −∗∗ Φ ∗ [n 7→ r, null]In ∧ ¬b
20
21
22
23
24
25
}
26
if (!b) {
N(l, Il ) −∗∗ Φ ∗ [n 7→ r, _]In ∧ ¬b
n
o
unmark(n);
Φ ∗ n 7→ r, _
n
o
free(n);
Φ
n
o
insert();
Φ
27
28
29
}
30
}
31
32
n
}
33
34
35
36
37
procedure delete() {
n
o
Φ
n
o
var l := mh;
N(l, Il ) −∗∗ Φ ∧ l = mh
var r := getUnmarked(l.next);
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
o
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:35
o
(N(l, Il ) ∗ N(r, I r )) −∗∗ Φ
n
o
l := r; r := getUnmarked(l.next);
N(l, Il ) −∗∗ Φ ∧ (r , null ⇒ (N(l, Il ) ∗ N(r, I r )) −∗∗ Φ )
while (r != null && nondet()) {
38
39
}
if (r == null) return;
var x := r.next;
if (isMarked(x)) delete();
n
o
(N(l, Il ) ∗ N(r, I r )) −∗∗ Φ ∧ ¬M (x )
40
41
42
43
44
atomic { // CAS(r.next, x, getMarked(x))
n
o
N(l, Il ) ∗ [r 7→ x, _]Ir ∧ N(r, I r ) −∗∗ Φ ∧ I r a = ♦ ∧ I r In = (1, 0)
if (r.next == x) {
o
n
r.next := getMarked(x);
N(l, Il ) ∗ r 7→ x, _ ∧ N(r, I r′ ) I −∗∗ Φ ∧ I r ≈ I r′ ∧ I r′ a = t ∧ I r′ In = (1, 0)
r
var I′r :| N(r, I r′ ) I ∧ {r ↣ _} ∈ I r In ;
r
n
o
sync(I′r );
N(l, Il ) ∗ [r 7→ x, _]_ ∧ N(r, I r′ ) −∗∗ Φ ∧ I r′ a = t ∧ I r′ In (r ) = (1, 0)
45
46
47
48
49
var b := true;
} else var b := false;
50
51
52
}
53
if (b) {
n
o
[r 7→ x, _]_ ∧ N(r, I r ) −∗∗ Φ ∧ l ∈ I ∧ I r a = t ∧ I r In (r ) = (1, 0) ∧ r , ft
while (true) {
N(ft, I f ) ∗ [r 7→ x, _]_ ∧ N(r, I r ) −∗∗ Φ ∧ l ∈ I ∧ I r a = t ∧ I r In (r ) = (1, 0)
54
atomic { // CAS(ft.fnext, null, r)
if (ft.fnext == null) {
n
o
[ft 7→ _, null]_ ∧ N(ft, I f ) ∗ [r 7→ x, _]_ ∧ N(r, I r ) −∗∗ Φ ∧ l ∈ I ∧ I r a = t ∧ I r In (r ) = (1, 0)
nh
o
i
N(ft, I f′ ) ∗ [r 7→ x, _]_ ∧ N(r, I r′ )
−∗∗ Φ ∧ l ∈ I ∧ I 1 = I f ⊕ I r ≈ I f′ ⊕ I r′ ∧ I r′ In = (1, 1)
ft.fnext := r;
I1
var I′1 :| Gr(I 1′ ) I ∧ {ft ↣ _, r ↣ _} ∈ I 1 In ;
1
o
n
sync(I′1 );
N(ft, I f′ ) ∗ [r 7→ x, _]_ ∧ N(r, I r′ ) −∗∗ Φ ∧ l ∈ I ∧ I r′ In = (1, 1)
55
56
57
58
59
60
var c := true;
} else var c := false;
61
62
63
}
64
if (c) {
ft := r;
65
n
−∗∗ Φ ∧ l ∈ I ∧ I r In = (1, 1)
o
7 x, _]_ ∧ N(r, I r ) −∗∗ Φ ∧ l ∈ I
[r →
N(ft, I f ) ∗ [r 7→ x, _]_ ∧ N(r, I r )
break;
66
67
}
68
}
n
o
[r 7→ x, _]_ ∧ N(r, I r ) −∗∗ Φ ∧ l ∈ I
atomic { // CAS(l.next, r, x)
[l 7→ r, _]_ ∧ N(l, Il ) ∗ [r 7→ x, _]_ ∧ N(r, I r ) −∗∗ Φ
if (l.next == r) {
n
o
′
′
′
′
N(l, Il ) ∗ N(r, I r ) I −∗∗ Φ ∧ I 1 = Il ⊕ I r ≈ Il ⊕ I r
l.next := x;
1
′
′
var I1 :| Gr(I 1 ) I ∧ {l ↣ _, r ↣ _} ∈ I 1 In ;
1
N(l, Il′ ) ∗ N(r, I r′ ) −∗∗ Φ
sync(I′1 );
}
n
o
}
Φ
n
o
} else {
Φ
n
o
delete();
Φ
69
70
71
72
73
74
75
76
77
}
78
79
n
}
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:36
Siddharth Krishna, Dennis Shasha, and Thomas Wies
(−∞, ∞)
3
1
...
5
(3, ∞)
7
2
3
4
(3, 5)
(5, 7)
5
7
1
2
8
(7, ∞)
2
(a)
{}
{3 ≤ k < ∞}
...
{}
{3 ≤ k < 5}
{5 ≤ k < 7}
{3, 4}
{5}
{7 ≤ k < ∞}
{7, 8}
(b)
Fig. 13. A B+ tree (a) and its representation as a flow graph (b). The nodes in the B+ tree contain an array of
keys on the top row, an array of pointers in the bottom row, the range in the top right, and l (number of keys)
in the bottom right. The flow graph edges are labelled with the edgesets and nodes with their contents. Lock
information is omitted from this figure.
B
THE B+ TREE
In this section we describe the B+ tree data structure and show how it is an instantiation of the
give-up template, hence ensuring its memory safety and linearizability.
A B+ tree is a generalization of a binary search tree that implements a dictionary data structure.
In a binary search tree (BST), each node contains a value k and up to two pointers yl and yr . An
operation takes the left branch if its query key is less than k and the right branch otherwise. A
B+ tree generalizes this by having l values k 0 , . . . , kl −1 and l + 1 pointers y0 , . . . , yl at each node,
such that l + 1 is between B and 2B for some constant B. At internal nodes, an operation takes the
branch yi if its query key is between ki−1 and ki . From the perspective of the edgeset framework,
the edgeset of an edge to yi is {k | ki−1 ≤ k < ki }. We store the inset of each node in a range field,
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:37
shown on the top right of each node as two values (r 0 , r 1 ) that denote the set {k | r 0 ≤ k < r 1 }. The
values stored at a leaf node are keys and make up the contents of the B+ tree. Nodes also contain l,
the number of values in a node, and a lock field to store the thread ID of the thread that holds a
lock on the node. Fig. 13(a) shows an example of a B+ tree with B = 2, and the corresponding flow
graph is shown in Fig. 13(b).
In our code, we assume that a B+ tree node is implemented by the following struct definition:
1
2
3
4
5
6
struct Node {
Int lock, len;
Key range[2];
Key keys[2B];
Node ptrs[2B];
}
This node structure is described in our logic by the following instantiations of the γд and γb
predicates:
® y.
γд (x, in, C, t, f ) B ∃l, r 0 , r 1 , k,
® x 7→ t, l, r 0 , r 1 , k 0 , . . . , k 2B−1 , y0 , . . . , y2B−1
∧ 0 ≤ l < 2B ∧ in(x) = {k | r 0 ≤ k < r 1 } ∧ r 0 ≤ k 0 ∧ kl −1 < r 1
∧ C = ITE(y0 , null, ∅, {ki | 0 ≤ i < l })
∧ f = {(x, yi ) ↣ {k | ITE(i ≤ 0, −∞, ki−1 ) ≤ k < ITE(i ≥ l, ∞, ki )} | yi , null}
∧ ∀0 < i < 2B. (i < l ⇒ ki−1 < ki ) ∧ (i ≥ l ⇒ ki = null)
∧ ∃l ′ . (l ′ = 0 ∨ l ′ = l + 1) ∧ ∀i < 2B. (l ′ ≤ i ⇔ yi = null)
γb (x, t) B x 7→ t, _, . . . , _
We also re-write the global data structure invariant as follows, to simplify notation in our proofs:
Φ B ∃I . Gr(I ) ∧ φ(I )
φ(I ) B ∃in ∈ I In . in + 0 = {r ↣ KS} + 0 ∧ I f = ϵ
We now show the instantations of the helper functions for the give-up template algorithm, along
with the implementation of the delete operation (the insert and member operations are similar).
The code is annotated with intermediate assertions that all imply the global data structure invariant
Φ. We assume that t is the thread ID of the current thread. We have omitted the stability proofs,
but these are syntactic and are similar to the example stability proof shown in §C.
1
2
n
o
procedure lock(Node c) {
N(c, Ic ) −∗∗ Φ
n
o
while (true) {
N(c, Ic ) −∗∗ Φ
atomic { // Action: (Lock)
[c 7→ 0, _, . . . , _]_ ∧ N(c, Ic ) ∗ Gr(I 2 ) ∧ I = Ic ⊕ I 2 ∧ φ(I )
if (c.lock == 0) {
n
o
N(c, Ic′ ) I ∗ Gr(I 2 ) ∧ I = Ic ⊕ I 2 ∧ φ(I ) ∧ Ic ≈ Ic′ ∧ Ic a = (C, 0) ∧ Ic′ a = (C, t )
c.lock := t;
c
var I′c :| Gr(Ic′ ) I ∧ {c ↣ _} ∈ Ic In ;
c
a
sync(I′c ); N(c, Ic′ ) ∗ Gr(I 2 ) ∧ I ′ = Ic′ ⊕ I 2 ∧ φ(I ′ ) ∧ Ic′ = (C, t )
break;
}
}
n
o
N(c, Ic′ ) −∗∗ Φ ∧ Ic′ a = (_, {t })
}
3
4
5
6
7
8
9
10
11
12
}
13
14
procedure unlock(Node c) {
n
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t })
o
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:38
Siddharth Krishna, Dennis Shasha, and Thomas Wies
atomic { // Action: (Sync)
[c 7→ t, _, . . . , _]_ ∧ N(c, Ic ) ∗ Gr(I 2 ) ∧ I = Ic ⊕ I 2 ∧ φ(I )
o
n
N(c, Ic′ ) I ∗ Gr(I 2 ) ∧ I = Ic ⊕ I 2 ∧ φ(I ) ∧ Ic ≈ Ic′ ∧ Ic a = (C, t ) ∧ Ic′ a = (C, 0)
c.lock := 0;
c
var I′c :| Gr(Ic′ ) I ∧ {c ↣ _} ∈ Ic In ;
c
a
sync(I′c ); N(c, Ic′ ) ∗ Gr(I 2 ) ∧ I ′ = Ic′ ⊕ I 2 ∧ φ(I ′ ) ∧ Ic′ = (C, 0)
n
o
N(c, Ic′ ) −∗∗ Φ ∧ Ic′ a ̸ ⊑ (KS, {t })
}
15
16
17
18
19
20
21
}
22
23
24
n
o
procedure inRange(Node c, Key k) {
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t })
γд (c, _, _, t, _) I −∗∗ Φ ∧ Ic a = (_, {t })
c
26
if (c.range[0] <= k && k < c.range[1]) {
n
o
return true;
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) ∧ k ∈ Ic In (c)
27
}
28
return false;
25
29
n
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) ∧ k < Ic In (c)
o
}
30
31
32
n
o
procedure findNext(Node c, Key k) {
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t })
(
)
h
i
® y®
c 7→ t, l, _, _, k,
−∗∗ Φ ∧ Ic a = (_, {t })
Ic
var i := 0;
while (i < c.len && k >= c.keys[i]) {
i := i + 1;
(
)
h
i
® y®
c 7→ t, l, _, _, k,
−∗∗ Φ ∧ Ic a = (_, {t }) ∧ ITE(i ≤ 0, −∞, k i −1 ) ≤ k < ITE(i ≥ l, ∞, k i )
}
33
34
35
36
Ic
38
if (i == c.len) {
n
o
return null;
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) ∧ n = null ∧ ∀x ∈ Ic f . k < Ic f (c, x )
39
} else {
37
return c.ptrs[i];
40
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) ∧ n , null ∧ k ∈ Ic f (c, n)
o
}
41
42
n
}
43
44
45
46
47
48
49
50
51
52
53
54
55
n
o
procedure delete(Node c, Key k) {
N(c, Ic ) −∗∗ Φ ∧ Ic a = (C, {t }) ∧ k ∈ Ic In (c) ∧ ∀x ∈ Ic f . k < Ic f (c, x )
n
o
N(c, Ic ) −∗∗ Φ ∧ k ∈ Ic In (c) ∧ Ic f = ϵ // Because union of outgoing edgesets is KS
var i := 0;
while (i < c.len && k >= c.keys[i]) {
i := i + 1;
(
)
h
n
oi
® _
c 7→ t, l, _, _, k,
−∗∗ Φ ∧ k ∈ Ic In (c) ∧ Ic f = ϵ ∧ 0 ≤ i ≤ l ∧ (i = l ∨ k < k i )
}
Ic
if (i == c.len || c.keys[i] != k) {
o
n
return false;
N(c, Ic ) −∗∗ Φ ∧ Ic a = (C, {t }) ∧ k < C
(
)
h
n
oi
®
c 7→ t, l, _, _, k, _
−∗∗ Φ ∧ 0 ≤ i < l ∧ k = k i
} else {
Ic
// First change node label to t to allow breaking good condition
var I′c :| Gr(Ic ) ∧ {c ↣ _} ∈ Ic In ∧ Ic ≈ Ic′ ∧ Ic a = (C, {t }) ∧ Ic′ a = (C, {t });
sync(I′c ); // Action: (Sync)
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
Go with the Flow: Compositional Abstractions for Concurrent Data Structures (Extended) 1:39
( h
n
oi
® _
c 7→ t, l, _, _, k,
56
−∗∗ Φ ∧ 0 ≤ i < l ∧ k = k i ∧ Ic′ a = (_, {t })
Ic′
)
while (i < c.len - 1) { // All these actions are also (Sync)s, with unchanged interfaces
c.keys[i] := c.keys[i + 1];
i := i + 1;
}
c.keys[i] := null;
N(c, Ic′′ ) I ′ ∧ N(c, Ic′ ) −∗∗ Φ ∧ Ic′ ≈ Ic′′ ∧ Ic′ a = (C, {t }) ∧ Ic′′ a = (C ′, {t }) ∧ C ′ = C \ {k }
c
var I′′c :| Gr(Ic′′ ) I ∧ {c ↣ _} ∈ Ic In ;
57
58
59
60
61
62
63
c
sync(I′′c ); // Action: (Sync)
n
o
N(c, Ic′′ ) −∗∗ Φ ∧ Ic′′ a = (C ′, {t }) ∧ k < C ′
64
65
return true;
66
}
67
68
}
C
EXAMPLE STABILITY PROOF
In RGSep, to show that an assertion S is stable under an action P ⇝ Q, we need to show that
(P −⊛ S) ∗ Q ⇒ S, where −⊛ is the septraction operator. Septraction is the dual operator to magic
wand (P −⊛ Q ⇔ ¬(P −∗ ¬Q)), and intuitively (P −⊛ S) is the state obtained when we remove a
state satisfying P from a state satisfying S.
As an example, we now prove that the intermediate assertion at line 6 of Fig. 11 is stable under
the action (Sync). This shows that if a thread has a reference to a node c that it has locked and
it knows that k is in the inset of c, then this knowledge is preserved after any combination of
interference by other threads. Moreover, the global invariant of the data structure described by
the good node condition is maintained. Note that this proof does not depend on the concrete
good condition predicates, and so holds for any concrete implementation. Proving stability of all
intermediate assertions is similarly easy, and furthermore the proofs are purely syntactic, and use
generic lemmas that are not specific to the inset flow.
Lemma C.1. The formula N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) ∧ k ∈ Ic In (c) is stable under action (Sync).
Proof.
(1) We start with
Gr(I 1 ) ∧ I 1 a ⊑ (_, {t ′, t ′ }) −⊛ N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) ∧ k ∈ Ic In (c)
∗ Gr(I 1′ ) ∧ I 1′a ⊑ (_, {0, t ′ }) ∧ I 1 ≾ I 1′
(2) By expanding Φ, the spatial part of this formula is
(Gr(I 1 ) −⊛ (N(c, Ic ) −∗∗ Gr(I ))) ∗ Gr(I 1′ )
(3) N(c, Ic ) −∗∗ Gr(I ) ≡ (N(c, Ic ) ∗ true) ∧ Gr(I ), so we use (GrDecomp) to get
(Gr(I 1 ) −⊛ ((N(c, Ic ) ∗ Gr(I 2 )) ∧ Gr(I ))) ∗ Gr(I 1′ )
(4) We use the SL identity P −⊛ Q ≡ P −⊛ (Q ∧ (P ∗ true)) to get
(Gr(I 1 ) −⊛ ((Gr(I 1 ) ∗ true) ∧ (N(c, Ic ) ∗ Gr(I 2 )) ∧ Gr(I ))) ∗ Gr(I 1′ )
(5) By (GrDecomp)
(Gr(I 1 ) −⊛ ((Gr(I 1 ) ∗ Gr(_)) ∧ (N(c, Ic ) ∗ Gr(I 2 )) ∧ Gr(I ))) ∗ Gr(I 1′ )
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
1:40
Siddharth Krishna, Dennis Shasha, and Thomas Wies
(6) As Ic a = (_, {t }) ∧ I 1 a ⊑ (_, {t ′, t ′ }) ∧ t , t ′ |= Ic a ̸⊑ I 1 a , use (Disj)
(Gr(I 1 ) −⊛ ((Gr(I 1 ) ∗ N(c, Ic ) ∗ Gr(I 3 )) ∧ Gr(I ))) ∗ Gr(I 1′ )
(7) Using (GrComp) and (Uniq) we get I = I 1 ⊕ Ic ⊕ I 3 .
(8) By P −⊛ (P ∗ Q) ⇒ Q for precise P,
N(c, Ic ) ∗ Gr(I 3 ) ∗ Gr(I 1′ )
(9) As I 1 ≾ I 1′ , we use (Repl) to get I = I 1 ⊕ Ic ⊕ I 3 ≾ I 1′ ⊕ Ic ⊕ I 3 = I ′.
(10) The pure parts of Φ were: ∃in ∈ I In . in + 0 = {r ↣ KS} + 0 ∧ I f = ϵ. By (ReplIn) and (ReplF)
on I ≾ I ′, we get in ∈ I ′In ∧ I ′ f = ϵ.
(11) Now use (GrComp) on (8), and by (7) we have
N(c, Ic ) ∗ Gr(I 3 ) ∗ Gr(I 1′ ) ∧ Gr(I ′)
(12) This implies
N(c, Ic ) −∗∗ Φ ∧ Ic a = (_, {t }) ∧ k ∈ Ic In (c)
□
, Vol. 1, No. 1, Article 1. Publication date: November 2017.
| 8cs.DS
|
Robust Budget Allocation via Continuous Submodular Functions
Matthew Staib and Stefanie Jegelka
arXiv:1702.08791v2 [cs.LG] 13 Jun 2017
Computer Science and Artificial Intelligence Laboratory
Massachusetts Institute of Technology
{mstaib, stefje}@mit.edu
Abstract
The optimal allocation of resources for maximizing influence, spread of information or coverage, has gained attention in the past years, in particular in machine learning and data mining.
But in applications, the parameters of the problem are rarely known exactly, and using wrong
parameters can lead to undesirable outcomes. We hence revisit a continuous version of the Budget Allocation or Bipartite Influence Maximization problem introduced by Alon et al. [2012]
from a robust optimization perspective, where an adversary may choose the least favorable parameters within a confidence set. The resulting problem is a nonconvex-concave saddle point
problem (or game). We show that this nonconvex problem can be solved exactly by leveraging
connections to continuous submodular functions, and by solving a constrained submodular minimization problem. Although constrained submodular minimization is hard in general, here, we
establish conditions under which such a problem can be solved to arbitrary precision .
1
Introduction
The optimal allocation of resources for maximizing influence, spread of information or coverage, has
gained attention in the past few years, in particular in machine learning and data mining [Domingos
& Richardson, 2001; Kempe et al., 2003; Chen et al., 2009; Gomez Rodriguez & Schölkopf, 2012;
Borgs et al., 2014].
In the Budget Allocation Problem, one is given a bipartite influence graph between channels
S and people T , and the task is to assign a budget y(s) to each channel s in S with the goal of
maximizing the expected number of influenced people I(y). Each edge (s, t) ∈ E between channel
s and person t is weighted with a probability pst that, e.g., an advertisement on radio station s will
influence person t to buy some product. The budget y(s) controls how many independent attempts
are made via the channel s to influence the people in T . The probability that a customer t is
influenced when the advertising budget is y is
Y
It (y) = 1 −
[1 − pst ]y(s) ,
(1)
(s,t)∈E
P
and hence the expected number of influenced people is I(y) = t∈T It (y). We write I(y; p) = I(y)
to make the dependence on the probabilities pst explicit. The totalPbudget y must remain within
some feasible set Y which may encode e.g. a total budget limit s∈S y(s) ≤ C. We allow the
budgets y to be continuous, as in [Bian et al., 2017].
1
Since its introduction by Alon et al. [2012], several works have extended the formulation of
Budget Allocation and provided algorithms Bian et al. [2017]; Hatano et al. [2015]; Maehara et al.
[2015]; Soma et al. [2014]; Soma & Yoshida [2015]. Budget Allocation may also be viewed as
influence maximization on a bipartite graph, where information spreads as in the Independent
Cascade model. For integer y, Budget Allocation and Influence Maximization are NP-hard. Yet,
constant-factor approximations are possible, and build on the fact that the influence function is
submodular in the binary case, and DR-submodular in the integer case [Soma et al., 2014; Hatano
et al., 2015]. If y is continuous, the problem is a concave maximization problem.
The formulation of Budget Allocation assumes that the transmission probabilities are known
exactly. But this is rarely true in practice. Typically, the probabilities pst , and possibly the
graph itself, must be inferred from observations [Gomez Rodriguez et al., 2010; Du et al., 2013;
Narasimhan et al., 2015; Du et al., 2014; Netrapalli & Sanghavi, 2012]. In Section 4 we will see that
a misspecification or point estimate of parameters pst can lead to much reduced outcomes. A more
realistic assumption is to know confidence intervals for the pst . Realizing this severe deficiency,
recent work studied robust versions of Influence Maximization, where a budget y must be chosen
that maximizes the worst-case approximation ratio over a set of possible influence functions [He &
Kempe, 2016; Chen et al., 2016; Lowalekar et al., 2016]. The resulting optimization problem is
hard but admits bicriteria approximations.
In this work, we revisit Budget Allocation under uncertainty from the perspective of robust
optimization [Bertsimas et al., 2011; Ben-Tal et al., 2009]. We maximize the worst-case influence –
not approximation ratio – for p in a confidence set centered around the “best guess” (e.g., posterior
mean). This avoids pitfalls of the approximation ratio formulation (which can be misled to return
poor worst-case budgets, as demonstrated in Appendix A), while also allowing us to formulate the
problem as a max-min game:
max min I(y; p),
(2)
y∈Y
p∈P
where an “adversary” can arbitrarily manipulate p within the confidence set P. With p fixed, I(y; p)
is concave in y. However, the influence function I(y; p) is not convex, and not even quasiconvex,
in the adversary’s variables pst .
The new, key insight we exploit in this work is that I(y; p) has the property of continuous
submodularity in p – in contrast to previously exploited submodular maximization in y – and can
hence be minimized by generalizing techniques from discrete submodular optimization Bach [2015].
The techniques in [Bach, 2015], however, are restricted to box constraints, and do not directly apply
to our confidence sets. In fact, general constrained submodular minimization is hard [Svitkina &
Fleischer, 2011; Goel et al., 2009; Iwata & Nagano, 2009]. We make the following contributions:
1. We present an algorithm with optimality bounds for Robust Budget Allocation in the nonconvex adversarial scenario (2).
2. We provide the first results for continuous submodular minimization with box constraints
and one more “nice” constraint, and conditions under which the algorithm is guaranteed to
return a global optimum.
1.1
Background and Related Work
We begin with some background material and, along the way, discuss related work.
2
1.1.1
Submodularity over the integer lattice and continuous domains
Submodularity is perhaps best known as a property of set functions. A function F : 2V → R
defined on subsets S ⊆ V of a ground set V is submodular if for all sets S, T ⊆ V , it holds that
F (S) + F (T ) ≥ F (S ∩ T ) + F (S ∪ T ). A similar definition extends to functions defined over a
distributive lattice L, e.g. the integer lattice. Such a function f is submodular if for all x, y ∈ L,
it holds that
f (x) + f (y) ≥ f (x ∨ y) + f (x ∧ y).
(3)
For the integer lattice and vectors x, y, x ∨ y denotes the coordinate-wise maximum and x ∧ y the
coordinate-wise minimum. Submodularity has also been considered on continuous domains X ⊂ Rd ,
where, if f is also twice-differentiable, the property of submodularity means that all off-diagonal
∂f (x)
entries of the the Hessian are nonpositive, i.e., ∂x
≤ 0 for all i 6= j [Topkis, 1978, Theorem 3.2].
i ∂xj
These functions may be convex, concave, or neither.
Submodular functions on lattices can be minimized by a reduction to set functions, more precisely, ring families Birkhoff [1937]. Combinatorial algorithms for submodular optimization on lattices are discussed in [Khachaturov et al., 2012]. More recently, Bach [2015] extended results based
on the convex Lovász extension, by building on connections to optimal transport. The subclass of
L\ -convex functions admits strongly polynomial time minimization [Murota, 2003; Kolmogorov &
Shioura, 2009; Murota & Shioura, 2014], but does not apply in our setting.
Similarly, results for submodular maximization extend to integer lattices, e.g. [Gottschalk &
Peis, 2015]. Stronger results are possible if the submodular function also satisfies diminishing
returns: for all x ≤ y (coordinate-wise) and i such that y + ei ∈ X , it holds that f (x + ei ) − f (x) ≥
f (y + ei ) − f (y). For such DR-submodular functions, many approximation results for the set
function case extend [Bian et al., 2017; Soma & Yoshida, 2015; Soma et al., 2014]. In particular,
Ene & Nguyen [2016] show a generic reduction to set function optimization that they apply to
maximization. In fact, it also applies to minimization:
Q
Proposition 1.1. A DR-submodular function f defined on ni=1 [ki ] can be minimized in strongly
polynomial time O(n4 log4 k · log2 (n log k) · EO + n4 log4 k · logO(1) (n log k)), where k = maxi ki and
EO is the time complexity of evaluating f . Here, [ki ] = {0, 1, . . . , ki − 1}.
Proof. The function f can be reduced to a submodular set function g : 2V → R via [Ene & Nguyen,
2016], where |V | = O(n log k). The function g can be evaluated via mapping from 2V to the domain
of f , and then evaluating f , in time O(n log k · EO). We can directly substitute these complexities
into the runtime bound from [Lee et al., 2015].
In particular, the time complexity is logarithmic in k. For general lattice submodular functions,
this is not possible without further assumptions.
1.1.2
Related Problems
A sister problem of Budget Allocation is Influence Maximization on general graphs, where a set
of seed nodes is selected to start a propagation process. The influence function is still monotone
submodular and amenable to the greedy algorithm Kempe et al. [2003], but it cannot be evaluated
explicitly and requires approximation Chen et al. [2010]. Stochastic Coverage [Goemans & Vondrák,
2006] is a version of Set Cover where the covering sets Si ⊂ V are random. A variant of Budget
Allocation can be written as stochastic coverage with multiplicity. Stochastic Coverage has mainly
3
been studied in the online or adaptive setting, where logarithmic approximation factors can be
achieved [Golovin & Krause, 2011; Deshpande et al., 2016; Adamczyk et al., 2016].
Our
Q objective function (2) is a signomial in p, i.e., a linear combination of monomials of the
form i xci i . General signomial optimization is NP-hard [Chiang, 2005], but certain subclasses are
tractable: posynomials with all nonnegative coefficients can be minimized via Geometric Programming [Boyd et al., 2007], and signomials with a single negative coefficient admit sum of squares-like
relaxations [Chandrasekaran & Shah, 2016]. Our problem, a constrained posynomial maximization,
is not in general a geometric program. Some work addresses this setting via monomial approximation [Pascual & Ben-Israel, 1970; Ecker, 1980], but, to our knowledge, our algorithm is the first
that solves this problem to arbitrary accuracy.
1.1.3
Robust Optimization
Two prominent strategies of addressing uncertainty in parameters of optimization problems are
stochastic and robust optimization. If the distribution of the parameters is known (stochastic optimization), formulations such as value-at-risk (VaR) and conditional value-at-risk (CVaR) Rockafellar & Uryasev [2000, 2002] apply. In contrast, robust optimization [Ben-Tal et al., 2009; Bertsimas
et al., 2011] assumes that the parameters (of the cost function and constraints) can vary arbitrarily
within a known confidence set U , and the aim is to optimize the worst-case setting, i.e.,
min sup {g(y; u) s.t. Ay ≤ b}.
y
(4)
u,A,b∈U
Here, we will only have uncertainty in the cost function.
In this paper we are principally concerned with robust maximization of the continuous influence
function I(y), but mention some results for the discrete case. While there exist results for robust
and CVaR optimization of modular (linear) functions [Nikolova, 2010; Bertsimas & Sim, 2003],
submodular objectives do not in general admit such optimization Maehara [2015], but variants
admit approximations [Zhang et al., 2014]. The brittleness of submodular optimization under noise
has been studied in [Balkanski et al., 2016, 2017; Hassidim & Singer, 2016].
Approximations for robust submodular and influence optimization have been studied in [Krause
et al., 2008; He & Kempe, 2016; Chen et al., 2016; Lowalekar et al., 2016], where an adversary can
pick among a finite set of objective functions or remove selected elements Orlin et al. [2016].
2
Robust and Stochastic Budget Allocation
The unknown parameters in Budget Allocation are the transmission probabilities pst or edge weights
in a graph. If these are estimated from data, we may have posterior distributions or, a weaker
assumption, confidence sets for the parameters. For ease of notation, we will work with the failure
probabilities xst = 1 − pst instead of the pst directly, and write I(y; x) instead of I(y; p).
2.1
Stochastic Optimization
If a (posterior) distribution of the parameters is known, a simple strategy is to use expectations.
We place a uniform prior on xst , and observe nst independent observations drawn from Ber(xst ).
If we observe αst failures and and βst successes, the resulting posterior distribution on the variable
4
Xst is Beta(1 + αst , 1 + βst ). Given such a posterior, we may optimize
max I(y; E[X]), or
(5)
max E[I(y; X)].
(6)
y∈Y
y∈Y
Proposition 2.1. Problems (5) and (6) are concave maximization problems over the (convex) set
Y and can be solved exactly.
Concavity of (6) follows since it is an expectation over concave functions, and the problem can
be solved by stochastic gradient ascent or by explicitly computing gradients.
Merely maximizing expectation does not explicitly account for volatility and hence risk. One
option is to include variance Ben-Tal & Nemirovski [2000]; Bertsimas et al. [2011]; Atamtürk &
Narayanan [2008]:
p
min −E[I(y; X)] + ε Var(I(y; X)),
(7)
y∈Y
but in our case this CVaR formulation seems difficult:
p
Fact 2.1. For y in the nonnegative orthant, the term Var(I(y; X)) need not be convex or concave,
and need not be submodular or supermodular.
This observation does not rule out a solution, but the apparent difficulties further motivate a
robust formulation that, as we will see, is amenable to optimization.
2.2
Robust Optimization
The focus of this work is the robust version of Budget Allocation, where we allow an adversary to
arbitrarily set the parameters x within an uncertainty set X . This uncertainty set may result, for
instance, from a known distribution, or simply assumed bounds. Formally, we solve
max min I(y; x),
y∈Y x∈X
(8)
where Y ⊂ RS+ is a convex set with an efficient projection oracle, and X is an uncertainty set
containing an estimate x̂. In the sequel, we use uncertainty sets X = {x ∈ Box(l,Q
u) : R(x) ≤ B},
where R is a distance (or divergence) from the estimate x̂, and Box(l, u) is the box (s,t)∈E [lst , ust ].
The intervals [lst , ust ] can be thought of as either confidence intervals around x̂, or, if [lst , ust ] = [0, 1],
enforce that each xst is a valid probability.
Common examples of uncertainty sets used in Robust Optimization are Ellipsoidal and D-norm
uncertainty sets Bertsimas et al. [2011]. Our algorithm in Section 3.1 applies to both.
Ellipsoidal uncertainty. The ellipsoidal or quadratic uncertainty set is defined by
X Q (γ) = {x ∈ Box(0, 1) : (x − x̂)T Σ−1 (x − x̂) ≤ γ},
where Σ is the covariance of the random vector X of probabilities distributed according to our
Beta posteriors. In our case, since the distributions on each xst are independent, Σ−1 is actually
diagonal. Writing Σ = diag(σ 2 ), we have
n
o
X
X Q (γ) = x ∈ Box(0, 1) :
Rst (xst ) ≤ γ ,
(s,t)∈E
5
−2
where Rst (x) = (xst − x̂st )2 σst
.
D-norm uncertainty. The D-norm uncertainty set is similar to an `1 -ball around x̂, and is
defined as
n
X D (γ) = x : ∃c ∈ Box(0, 1) s.t.
xst = x̂st + (ust − x̂st )cst ,
X
o
cst ≤ γ .
(s,t)∈E
Essentially, we allow an adversary to increase x̂st up to some upper bound ust , subject to some
total budget γ across all terms xst . The set X D (γ) can be rewritten as
o
n
X
Rst (xst ) ≤ γ ,
X D (γ) = x ∈ Box(x̂, u) :
(s,t)∈E
where Rst (xst ) = (xst − x̂st )/(ust − x̂st ) is the fraction of the interval [x̂st , ust ] we have used up in
increasing xst .
The min-max formulation maxy∈Y minx∈X I(y; x) has several merits: the model is not tied to
a specific learning algorithm for the probabilities x as long as we can choose a suitable confidence
set. Moreover, this formulation allows to fully hedge against a worst-case scenario.
3
Optimization Algorithm
As noted above, the function I(y; x) is concave as a function of y for fixed x. As a pointwise
minimum of concave functions, F (y) := minx∈X I(y; x) is concave. Hence, if we can compute
subgradients of F (y), we can solve our max-min-problem via the subgradient method, as outlined
in Algorithm 1.
A subgradient gy ∈ ∂F (y) at y is given by the gradient of I(y; x∗ ) for the minimizing x∗ ∈
arg minx∈X I(y; x), i.e., gy = ∇y I(y; x∗ ). Hence, we must be able to compute x∗ for any y. We
also obtain a duality gap: for any x0 , y 0 we have
min I(y 0 ; x) ≤ max min I(y; x) ≤ max I(y; x0 ).
x∈X
y∈Y x∈X
y∈Y
(9)
This means we can estimate the optimal value I ∗ and use it in Polyak’s stepsize rule for the
subgradient method Polyak [1987].
But I(y; x) is not convex in x, and not even quasiconvex. For example, standard methods
√
[Wainwright & Chiang, 2004, Chapter 12] imply that f (x1 , x2 , x3 ) = 1 − x1 x2 − x3 is not quasiconvex on R3+ . Moreover, the above-mentioned signomial optimization techniques do not apply for
an exact solution either. So, it is not immediately clear that we can solve the inner optimization
problem.
The key insight we will be using is that I(y; x) has a different beneficial property: while not
convex, I(y; x) as a function of x is continuous submodular.
Lemma 3.1. Suppose we have n ≥ 1 differentiable functions
fi : R → R+ , for i = 1, . . . , n, either
Q
all nonincreasing or all nondecreasing. Then, f (x) = ni=1 fi (xi ) is a continuous supermodular
function from Rn to R+ .
6
Algorithm 1 Subgradient Ascent
Input: suboptimality tolerance ε > 0, initial feasible budget y (0) ∈ Y
Output: ε-optimal budget y for Problem (8)
repeat
x(k) ← arg minx∈X I(y (k) ; x)
g (k) ← ∇y I(y (k) ; x(k) )
L(k) ← I(y (k) ; x(k) )
U (k) ← maxy∈Y I(y; x(k) )
γ (k) ← (U (k) − L(k) )/kg (k) k22
y (k+1) ← projY (y (k) + γ (k) g (k) )
k ←k+1
until U (k) − L(k) ≤ ε
Proof. For n = 1, the resulting function is modular and therefore supermodular. In the case n ≥ 2,
we simply need to compute derivatives. The mixed derivatives are
Y
∂f
= fi0 (xi )fj0 (xj ) ·
fk (xk ).
∂xi ∂xj
(10)
k6=i,j
By monotonicity, fi0 and fj0 have the same sign, so their product is nonnegative, and since each fk
is nonnegative, the entire expression is nonnegative. Hence, f (x) is continuous supermodular by
Theorem 3.2 of [Topkis, 1978].
Corollary 3.1. The influence function I(y; x) defined in Section 2 is continuous submodular in x
over the nonnegative orthant, for each y ≥ 0.
Proof. Since submodularity is preserved under summation, it suffices to show that each function
It (y) is continuous submodular. By Lemma 3.1, since fs (z) = z y(s) is nonnegative and monotone
Q
y(s)
nondecreasing for y(s) ≥ 0, the product (s,t)∈E xst is continuous supermodular in x. Flipping
the sign and adding a constant term yields It (y), which is hence continuous submodular.
Conjecture 3.1. Strong duality holds, i.e.
max min I(y; x) = min max I(y; x).
y∈Y x∈X
x∈X y∈Y
(11)
If strong duality holds, then the duality gap maxy∈Y I(y; x∗ ) − minx∈X I(y ∗ ; x) in Equation (9)
is zero at optimality. If I(y; x) were quasiconvex in x, strong duality would hold by Sion’s min-max
theorem, but this is not the case. In practice, we observe that the duality gap always converges to
zero.
Bach [2015] demonstrates how to minimize a continuous submodular function H(x) subject to
box constraints x ∈ Box(l, u), up to an arbitrary suboptimality gap ε > 0. The constraint set X in
our Robust Budget Allocation problem, however, has box constraints with an additional constraint
R(x) ≤ B. This case is not addressed in any previous work. Fortunately, for a large class of
functions R, there is still an efficient algorithm for continuous submodular minimization, which we
present in the next section.
7
3.1
Constrained Continuous Submodular Function Minimization
We next address an algorithm for minimizing a monotone continuous submodular function H(x)
subject to box constraints x ∈ Box(l, u) and a constraint R(x) ≤ B:
minimize H(x)
s.t.
R(x) ≤ B
x ∈ Box(l, u).
(12)
If H and R were convex, the constrained problem would be equivalent to solving, with the right
Lagrange multipler λ∗ ≥ 0:
minimize H(x) + λ∗ R(x)
(13)
s.t.
x ∈ Box(l, u).
Although H and R are not necessarily convex here, it turns out that a similar approach indeed
applies. The main idea of our approach bears similarity with [Nagano et al., 2011] for the set
function case, but our setting with continuous functions and various uncertainty sets is more general, and requires more argumentation. We outline our theoretical results here, and defer further
implementation details and proofs to the appendix.
Following Bach [2015], we discretize the problem; for a sufficiently fine discretization,Qwe will
achieve arbitraryQ
accuracy. Let A be an interpolation mapping that maps the discrete set ni=1 [ki ]
into Box(l, u) = ni=1 [li , ui ] via the componentwise interpolation functions Ai : [ki ] → [li , ui ]. We
say Ai is δ-fine if Ai (xi + 1) − Ai (xi ) ≤ δ for all xi ∈ {0, 1, . . . , ki − 2}, and we say the full
interpolation function A is δ-fine if eachQAi is δ-fine.
Q
This mapping yields functions H δ : ni=1 [ki ] → R and Rδ : ni=1 [ki ] → R via H δ (x) = H(A(x))
and Rδ (x) = R(A(x)). H δ is lattice submodular (on the integer lattice). This construction leads
to a reduction of Problem (12) to a submodular minimization problem over the integer lattice:
δ
minimize H δ (x)
Qn+ λR (x)
s.t.
x ∈ i=1 [ki ].
(14)
Ideally, there should then exist a λ such that the associated minimizer x(λ) yields a close to optimal
solution for the constrained problem. Theorem 3.1 below states that this is indeed the case.
Moreover, a second benefit of submodularity is that we can find the entire solution path for
Problem (14) by solving a single optimization problem.
Lemma 3.2. Suppose H is continuous
submodular, and suppose the regularizer R is strictly inP
creasing and separable: R(x) = ni=1 Ri (xi ). Then we can recover a minimizer x(λ) for the induced
discrete Problem (14) for any λ ∈ R by solving a single convex optimization problem.
The problem in question arises from a relaxation h↓ that extends H δ in each coordinate i to
a function on distributions over the domain [ki ]. These distributions are represented via their
inverse cumulative distribution functions ρi , which take the coordinate xi as input, and output the
probability of exceeding xi . The function h↓ is an analogue of the Lovász extension of set functions
to continuous submodular functions [Bach, 2015], it is convex and coincides with H δ on lattice
points.
Formally, this resulting single optimization problem is:
P P −1
minimize h↓ (ρ) + ni=1 kjii=1
aixi (ρi (xi ))
Qn
(15)
ki −1
s.t.
ρ ∈ i=1 R↓
8
where Rk↓ refers to the set of ordered vectors z ∈ Rk that satisfy z1 ≥ z2 ≥ · · · ≥ zk , the notation
ρi (xi ) denotes the xi -th coordinate of the vector ρi , and the aixi are strictly convex functions given
by
1
aixi (t) = t2 · [Riδ (xi ) − Riδ (xi − 1)].
(16)
2
Problem (15) can be solved by Frank-Wolfe methods [Frank & Wolfe, 1956; Dunn & Harshbarger,
1978; Lacoste-Julien, 2016; Jaggi, 2013]. This is because the greedy algorithm for computing
subgradients of the Lovász extension can be generalized, and yields a linear optimization oracle for
the dual of Problem (15). We detail the relationship between Problems (14) and (15), as well as
how to implement the Frank-Wolfe methods, in Appendix C.
Let ρ∗ be the optimal solution for Problem (15). For any λ, we obtain a rounded solution x(λ)
for Problem (14) by thresholding: we set x(λ)i = max{j | 1 ≤ j ≤ ki − 1, ρ∗i (j) ≥ λ}, or zero if
ρ∗i (j) < λ for all j. Each x(λ0 ) is the optimal solution for Problem (14) with λ = λ0 . We use the
largest parameterized solution x(λ) that is still feasible, i.e. the solution x(λ∗ ) where λ∗ solves
min H δ (x(λ))
s.t. λ ≥ 0
Rδ (x(λ)) ≤ B.
(17)
This λ∗ can be found efficiently via binary search or a linear scan.
Theorem 3.1. Let H be continuous submodular and monotone decreasing, with `∞ -Lipschitz constant G, and let R be strictly increasing and separable. Assume all entries ρ∗i (j) of the optimal
solution ρ∗ of Problem (15) are distinct. Let x0 = A(x(λ∗ )) be the thresholding corresponding to the
optimal solution λ∗ of Problem (17), mapped back into the original continuous domain X . Then x0
is feasible for the continuous Problem (12), and is a 2Gδ-approximate solution:
H(x0 ) ≤ 2Gδ +
min
H(x).
x∈Box(l,u), R(x)≤B
Theorem 3.1 implies an algorithm for solving Problem (12) to ε-optimality: (1) set δ = ε/G,
(2) compute ρ∗ which solves Problem (15), (3) find the optimal thresholding of ρ∗ by determining
the smallest λ∗ for which Rδ (x(λ∗ )) ≤ B, and (4) map x(λ∗ ) back into continuous space via the
interpolation mapping A.
Optimality Bounds.
the discretized mesh
Theorem 3.1 is proved by comparing x0 and x∗ to the optimal solution on
x∗d ∈
argmin
x∈
Qn
i=1 [ki ]:R
H δ (x).
δ (x)≤B
Beyond the theoretical guarantee of Theorem 3.1, for any problem instance and candidate solution
x0 , we can compute a bound on the gap between H(x0 ) and H δ (x∗d ). The following two bounds are
proved in the appendix:
1. We can generate a discrete point x(λ+ ) satisfying
H(x0 ) ≤ [H(x0 ) − H δ (x(λ+ ))] + H δ (x∗d ).
2. The Lagrangian yields the bound
H(x0 ) ≤ λ∗ (B − R(x0 )) + H δ (x∗d ).
9
Improvements. The requirement in Theorem 3.1 that the elements of ρ∗ be distinct may seem
somewhat restrictive, but as long as ρ∗ has distinct elements in the neighborhood of our particular
λ∗ , this bound still holds. We see in Section 4.1.1 that in practice, ρ∗ almost always has distinct
elements in the regime we care about, and the bounds of Remark 3.1 are very good.
If H is DR-submodular and R is affine in each coordinate, then Problem (14) can be represented
more compactly via the reduction of Ene & Nguyen [2016], and hence problem (12) can be solved
more efficiently. In particular, the influence function I(y; x) is DR-submodular in x when for each
s, y(s) = 0 or y(s) ≥ 1.
3.2
Application to Robust Budget Allocation
The above algorithm directly applies to Robust Allocation with the uncertainty
sets in Section 2.2.
P
Q
The ellipsoidal uncertainty set X corresponds to the constraint that (s,t)∈E Rst (xst ) ≤ γ with
−2
Rst (x) = (xst − x̂st )2 σst
, and x ∈ Box(0, 1). By the monotonicity of I(x, y), there is never incentive
to reduce any xst below x̂st , so we can replace Box(0, 1) with Box(x̂, 1). On this interval, each Rst
is strictly increasing, and Theorem 3.1 applies.
For D-norm sets, we have Rst (xst ) = (xst − x̂st )/(ust − x̂st ). Since each Rst is monotone,
Theorem 3.1 applies.
Runtime and Alternatives. Since the core algorithm is Frank-Wolfe, it is straightforward to
show that Problem (15) can be solved to ε-suboptimality in time O(ε−1 n2 δ −3 α−1 |T |2 log nδ −1 ),
where α is the minimum derivative of the functions Ri . If ρ∗ has distinct elements separated by η,
then choosing ε = η 2 αδ/8 results in an exact solution to (14) in time O(η −2 n2 δ −4 α−2 |T |2 log nδ −1 ).
Noting that H δ + λRδ is submodular for all λ, one could instead perform binary search over
λ, each time converting the objective into a submodular set function via Birkhoff’s theorem and
solving submodular minimization e.g. via one of the recent fast methods [Chakrabarty et al., 2017;
Lee et al., 2015]. However, we are not aware of a practical implementation of the algorithm in [Lee
et al., 2015]. The algorithm in [Chakrabarty et al., 2017] yields a solution in expectation. This
approach also requires care in the precision of the search over λ, whereas our approach searches
directly over the O(nδ −1 ) elements of ρ∗ .
4
Experiments
We evaluate our Robust Budget Allocation algorithm on both synthetic test data and a real-world
bidding dataset from Yahoo! Webscope yah to demonstrate that our method yields real improvements. For all experiments, we used Algorithm 1 as the outer loop. For the inner submodular
minimization step, we implemented the pairwise Frank-Wolfe algorithm
of Lacoste-Julien & Jaggi
P
S
[2015]. In all cases, the feasible set of budgets Y is {y ∈ R+ : s∈S y(s) ≤ C} where the specific
budget C depends on the experiment. Our code is available at git.io/vHXkO.
4.1
Synthetic
On the synthetic data, we probe two questions: (1) how often does the distinctness condition of
Theorem 3.1 hold, so that we are guaranteed an optimal solution; and (2) what is the gain of using
a robust versus non-robust solution in an adversarial setting? For both settings, we set |S| = 6 and
10
D-norm, C = 0.4
0.028
Ellipsoidal, C = 0.4
0.01
0.005
0.014
0
0
0
0.026
50
D-norm, C = 4
100
0
50
Ellipsoidal, C = 4
100
0
50
100
0.01
0.013
0.005
0
0
0
50
100
Figure 1: Visualization of the sorted values of ρ∗i (j) (blue dots) with comparison to the particular Lagrange multiplier λ∗ (orange line). In most regimes there are no duplicate values, so that
Theorem 3.1 applies. The theorem only needs distinctness at λ∗ .
|T | = 2 and discretize with δ = 0.001. We generated true probabilties pst , created Beta posteriors,
and built both Ellipsoidal uncertainty sets X Q (γ) and D-norm sets X D (γ).
4.1.1
Optimality
Theorem 3.1 and Remark 3.1 demand that the values ρ∗i (j) be distinct at our chosen Lagrange
multiplier λ∗ and, under this condition, guarantee optimality. We illustrate this in four examples:
for Ellipsoidal or a D-norm uncertainty set, and a total influence budget C ∈ {0.4, 4}. Figure 3
shows all elements of ρ∗ in sorted order, as well as a horizontal line indicating our Lagrange
multiplier λ∗ which serves as a threshold. Despite some plateaus, the entries ρ∗i (j) are distinct
in most regimes, in particular around λ∗ , the regime that is needed for our results. Moreover, in
practice (on the Yahoo data) we observe later in Figure 3 that both solution-dependent bounds
from Remark 3.1 are very good, and all solutions are optimal within a very small gap.
4.1.2
Robustness and Quality
Next, we probe the effect of a robust versus non-robust solution for different uncertainty sets and
budgets γ of the adversary. We compare our robust solution with using a point estimate for x,
11
D-norm, C = 0.400
0.3
Ellipsoidal, C = 0.400
0.3
0.2
0.2
0.1
0.1
0
0
0
5
10
D-norm, C = 3.740
15
1.5
1.5
1
1
0.5
0.5
0
0
400
Ellipsoidal, C = 3.740
800
0
400
800
0
0
5
10
15
Figure 2: Comparison of worst-case expected influences for D-norm uncertainty sets X D (γ) (left)
and ellipsoidal uncertainty sets X Q (γ) (right), for different total budget bounds C. For any particular adversary budget γ, we compare minx∈X (γ) I(y; x) for each candidate allocation y.
i.e., ynom ∈ argmaxy∈Y I(y; x̂), treating estimates as ground truth, and the stochastic solution
yexpect ∈ argmaxy∈Y E[I(y; X)] as per Section 2.1. These two optimization problems were solved
via standard first-order methods using TFOCS Becker et al. [2011].
Figure 2 demonstrates that indeed, the alternative budgets are sensitive to the adversary and
the robustly-chosen budget yrobust performs better, even in cases where the other budgets achieve
zero influence. When the total budget C is large, yexpect performs nearly as well as yrobust , but
when resources are scarce (C is small) and the actual choice seems to matter more, yrobust performs
far better.
4.2
Yahoo! data
To evaluate our method on real-world data, we formulate a Budget Allocation instance on advertiser
bidding data from Yahoo! Webscope [yah]. This dataset logs bids on 1000 different phrases by
advertising accounts. We map the phrases to channels S and the accounts to customers T , with an
edge between s and t if a corresponding bid was made. For each pair (s, t), we draw the associated
transmission probability pst uniformly from [0, 0.4]. We bias these towards zero because we expect
people not to be easily influenced by advertising in the real world. We then generate an estimate
12
Figure 3: Convergence properties of our algorithm on real data. In the first plot, ‘p’ and ‘d’ refer to
primal and dual values, with dual gap shown on the second plot. The third plot demonstrates that
the problem-dependent suboptimality bounds of Remark 3.1 (x for x(λ+ ) and L for Lagrangian)
are very small (good) for all inner iterations of this run.
p̂st and build up a posterior by generating nst samples from Ber(pst ), where nst is the number of
bids between s and t in the dataset.
This transformation yields a bipartite graph with |S| = 1000, |T | = 10475, and more than
50,000 edges that we use for Budget Allocation. In our experiments, the typical gap between the
naive ynom and robust yrobust was 100-500 expected influenced people. We plot convergence of the
outer loop in Figure 3, where we observe fast convergence of both primal influence value and the
dual bound.
4.3
Comparison to first-order methods
Given the success of first-order methods on nonconvex problems in practice, it is natural to compare
these to our method for finding the worst-case vector x. On one of our Yahoo problem instances with
D-norm uncertainty set, we compared our submodular minimization scheme to Frank-Wolfe with
fixed stepsize as in [Lacoste-Julien, 2016], implementing the linear oracle using MOSEK [MOSEK
ApS, 2015]. Interestingly, from various initializations, Frank-Wolfe finds an optimal solution, as
verified by comparing to the guaranteed solution of our algorithm. Note that, due to non-convexity,
there are no formal guarantees for Frank-Wolfe to be optimal here, motivating the question of global
convergence properties of Frank-Wolfe in the presence of submodularity.
It is important to note that there are many cases where first-order methods are inefficient or
do not apply to our setup. These methods require either a projection oracle (PO) onto or linear
optimization oracle (LO) over the feasible set X defined by `, u and R(x). The D-norm set admits
a LO via linear programming, but we are not aware of any efficient LO for Ellipsoidal uncertainty,
nor PO for either set, that does not require quadratic programming. Even more, our algorithm
applies for nonconvex functions R(x) which induce nonconvex feasible sets X . Such nonconvex sets
may not even admit a unique projection, while our algorithm achieves provable solutions.
13
10500
9500
8500
7500
0
40
80
Figure 4: Convergence properties of Frank-Wolfe (FW), versus the optimal value attained with our
scheme (SFM).
5
Conclusion
We address the issue of uncertain parameters (or, model mis-specification) in Budget Allocation or
Bipartite Influence Maximization [Alon et al., 2012] from a robust optimization perspective. The
resulting Robust Budget Allocation is a nonconvex-concave saddle point problem. Although the
inner optimization problem is nonconvex, we show how continuous submodularity can be leveraged
to solve the problem to arbitrary accuracy ε, as can be verified with the proposed bounds on the
duality gap. In particular, our approach extends continuous submodular minimization methods
[Bach, 2015] to more general constraint sets, introducing a mechanism to solve a new class of constrained nonconvex optimization problems. We confirm on synthetic and real data that our method
finds high-quality solutions that are robust to parameters varying arbitrarily in an uncertainty set,
and scales up to graphs with over 50,000 edges.
There are many compelling directions for further study. The uncertainty sets we use are standard in the robust optimization literature, but have not been applied to e.g. Robust Influence
Maximization; it would be interesting to generalize our ideas to general graphs. Finally, despite
the inherent nonconvexity of our problem, first-order methods are often able to find a globally optimal solution. Explaining this phenomenon requires further study of the geometry of constrained
monotone submodular minimization.
Acknowledgements
We thank the anonymous reviewers for their helpful suggestions. We also thank MIT Supercloud
and the Lincoln Laboratory Supercomputing Center for providing computational resources. This
research was conducted with Government support under and awarded by DoD, Air Force Office of
Scientific Research, National Defense Science and Engineering Graduate (NDSEG) Fellowship, 32
CFR 168a, and also supported by NSF CAREER award 1553284.
14
References
Yahoo! Webscope dataset ydata-ysm-advertiser-bids-v1 0. URL http://research.yahoo.com/
Academic_Relations.
Adamczyk, Marek, Sviridenko, Maxim, and Ward, Justin. Submodular Stochastic Probing on
Matroids. Mathematics of Operations Research, 41(3):1022–1038, 2016.
Alon, Noga, Gamzu, Iftah, and Tennenholtz, Moshe. Optimizing Budget Allocation Among Channels and Influencers. In WWW. 2012.
Atamtürk, Alper and Narayanan, Vishnu. Polymatroids and mean-risk minimization in discrete
optimization. Operations Research Letters, 36(5):618–622, 2008.
Bach, Francis. Submodular Functions: From Discrete to Continous Domains. arXiv:1511.00394,
2015.
Balkanski, Eric, Rubinstein, Aviad, and Singer, Yaron. The power of optimization from samples.
In NIPS, 2016.
Balkanski, Eric, Rubinstein, Aviad, and Singer, Yaron. The limitations of optimization from samples. In STOC, 2017.
Becker, Stephen R., Candès, Emmanuel J., and Grant, Michael C. Templates for convex cone
problems with applications to sparse signal recovery. Mathematical programming computation, 3
(3):165–218, 2011.
Ben-Tal, Aharon and Nemirovski, Arkadi. Robust solutions of Linear Programming problems
contaminated with uncertain data. Mathematical Programming, 88(3):411–424, 2000.
Ben-Tal, Aharon, El Ghaoui, Laurent, and Nemirovski, Arkadi. Robust Optimization. Princeton
University Press, 2009.
Bertsimas, Dimitris and Sim, Melvyn. Robust discrete optimization and network flows. Mathematical programming, 98(1):49–71, 2003.
Bertsimas, Dimitris, Brown, David B., and Caramanis, Constantine. Theory and Applications of
Robust Optimization. SIAM Review, 53(3):464–501, 2011.
Best, Michael J. and Chakravarti, Nilotpal. Active set algorithms for isotonic regression; A unifying
framework. Mathematical Programming, 47(1-3):425–439, 1990.
Bian, Andrew An, Mirzasoleiman, Baharan, Buhmann, Joachim M., and Krause, Andreas. Guaranteed Non-convex Optimization: Submodular Maximization over Continuous Domains. In AISTATS, 2017.
Birkhoff, Garrett. Rings of sets. Duke Mathematical Journal, 3(3):443–454, 1937.
Borgs, Christian, Brautbar, Michael, Chayes, Jennifer, and Lucier, Brendan. Maximizing Social
Influence in Nearly Optimal Time. In SODA, 2014.
15
Boyd, Stephen, Kim, Seung-Jean, Vandenberghe, Lieven, and Hassibi, Arash. A tutorial on geometric programming. Optimization and engineering, 8(1):67–127, 2007.
Chakrabarty, Deeparnab, Lee, Yin Tat, Sidford, Aaron, and Wong, Sam Chiu-wai. Subquadratic
submodular function minimization. In STOC, 2017.
Chandrasekaran, Venkat and Shah, Parikshit. Relative Entropy Relaxations for Signomial Optimization. SIAM Journal on Optimization, 26(2):1147–1173, 2016.
Chen, Wei, Wang, Yajun, and Yang, Siyu. Efficient influence maximization in social networks. In
KDD, 2009.
Chen, Wei, Wang, Chi, and Wang, Yajun. Scalable Influence Maximization for Prevalent Viral
Marketing in Large-scale Social Networks. In KDD, 2010.
Chen, Wei, Lin, Tian, Tan, Zihan, Zhao, Mingfei, and Zhou, Xuren. Robust influence maximization.
In KDD. 2016.
Chiang, Mung. Geometric Programming for Communication Systems. Commun. Inf. Theory, 2
(1/2):1–154, 2005.
Deshpande, Amol, Hellerstein, Lisa, and Kletenik, Devorah. Approximation Algorithms for
Stochastic Submodular Set Cover with Applications to Boolean Function Evaluation and MinKnapsack. ACM Trans. Algorithms, 12(3):42:1–42:28, 2016.
Domingos, Pedro and Richardson, Matt. Mining the network value of customers. In KDD, 2001.
Du, Nan, Song, Le, Gomez Rodriguez, Manuel, and Zha, Hongyuan. Scalable influence estimation
in continuous-time diffusion networks. In NIPS. 2013.
Du, Nan, Liang, Yingyu, Balcan, Maria-Florina, and Song, Le. Influence function learning in
information diffusion networks. In ICML, 2014.
Dunn, Joseph C. and Harshbarger, S. Conditional gradient algorithms with open loop step size
rules. Journal of Mathematical Analysis and Applications, 62(2):432–444, 1978.
Ecker, Joseph. Geometric Programming: Methods, Computations and Applications. SIAM Review,
22(3):338–362, 1980.
Ene, Alina and Nguyen, Huy L. A Reduction for Optimizing Lattice Submodular Functions with
Diminishing Returns. arXiv:1606.08362, 2016.
Frank, Marguerite and Wolfe, Philip. An algorithm for quadratic programming. Naval Research
Logistics Quarterly, 3(1-2):95–110, 1956.
Goel, Gagan, Karande, Chinmay, Tripathi, Pushkar, and Wang, Lei. Approximability of combinatorial problems with multi-agent submodular cost functions. In FOCS, 2009.
Goemans, Michel and Vondrák, Jan. Stochastic Covering and Adaptivity. In LATIN 2006: Theoretical Informatics. Springer Berlin Heidelberg, 2006.
16
Golovin, Daniel and Krause, Andreas. Adaptive Submodularity: Theory and Applications in Active
Learning and Stochastic Optimization. Journal of Artificial Intelligence, 42:427–486, 2011.
Gomez Rodriguez, Manuel and Schölkopf, Bernhard. Influence maximization in continuous time
diffusion networks. In ICML, 2012.
Gomez Rodriguez, Manuel, Leskovec, Jure, and Krause, Andreas. Inferring networks of diffusion
and influence. In KDD, 2010.
Gottschalk, Corinna and Peis, Britta. Submodular function maximization on the bounded integer
lattice. In Approximation and Online Algorithms: 13th International Workshop (WAOA), 2015.
Hassidim, Avinatan and Singer, Yaron. Submodular optimization under noise. arXiv preprint
arXiv:1601.03095, 2016.
Hatano, Daisuke, Fukunaga, Takuro, Maehara, Takanori, and Kawarabayashi, Ken-ichi. Lagrangian
Decomposition Algorithm for Allocating Marketing Channels. In AAAI, 2015.
He, Xinran and Kempe, David. Robust influence maximization. In KDD. 2016.
Iwata, Satoru and Nagano, Kiyohito. Submodular function minimization under covering constraints. In FOCS, 2009.
Jaggi, Martin. Revisiting Frank-Wolfe: Projection-Free Sparse Convex Optimization. In ICML,
2013.
Kempe, David, Kleinberg, Jon, and Tardos, Éva. Maximizing the Spread of Influence Through a
Social Network. In KDD, 2003.
Khachaturov, Vladimir R., Khachaturov, Roman V., and Khachaturov, Ruben V. Supermodular
programming on finite lattices. Computational Mathematics and Mathematical Physics, 52(6):
855–878, 2012.
Kolmogorov, Vladimir and Shioura, Akiyoshi. New algorithms for convex cost tension problem
with application to computer vision. Discrete Optimization, 6:378–393, 2009.
Krause, Andreas, McMahan, H Brendan, Guestrin, Carlos, and Gupta, Anupam. Robust submodular observation selection. Journal of Machine Learning Research, 9(Dec):2761–2801, 2008.
Lacoste-Julien, Simon.
arXiv:1607.00345, 2016.
Convergence Rate of Frank-Wolfe for Non-Convex Objectives.
Lacoste-Julien, Simon and Jaggi, Martin. On the global linear convergence of Frank-Wolfe optimization variants. In NIPS, 2015.
Lee, Yin Tat, Sidford, Aaron, and Wong, Sam Chiu-wai. A faster cutting plane method and its
implications for combinatorial and convex optimization. In FOCS, 2015.
Lowalekar, Meghna, Varakantham, Pradeep, and Kumar, Akshat. Robust Influence Maximization:
(Extended Abstract). In AAMAS, 2016.
17
Maehara, Takanori. Risk averse submodular utility maximization. Operations Research Letters, 43
(5):526–529, 2015.
Maehara, Takanori, Yabe, Akihiro, and Kawarabayashi, Ken-ichi. Budget Allocation Problem with
Multiple Advertisers: A Game Theoretic View. In ICML, 2015.
MOSEK ApS. MOSEK MATLAB Toolbox 8.0.0.57, 2015. URL http://docs.mosek.com/8.0/
toolbox/index.html.
Murota, Kazuo. Discrete convex analysis. SIAM, 2003.
Murota, Kazuo and Shioura, Akiyoshi. Exact bounds for steepest descent algorithms of l-convex
function minimization. Operations Research Letters, 42:361–366, 2014.
Nagano, Kiyohito, Kawahara, Yoshinobu, and Aihara, Kazuyuki. Size-constrained submodular
minimization through minimum norm base. In ICML, 2011.
Narasimhan, Harikrishna, Parkes, David C, and Singer, Yaron. Learnability of influence in networks. In NIPS. 2015.
Netrapalli, Praneeth and Sanghavi, Sujay. Learning the graph of epidemic cascades. In SIGMETRICS. 2012.
Nikolova, Evdokia. Approximation algorithms for reliable stochastic combinatorial optimization.
In APPROX. 2010.
Orlin, James B., Schulz, Andreas, and Udwani, Rajan. Robust monotone submodular function
maximization. In IPCO, 2016.
Pascual, Luis D. and Ben-Israel, Adi. Constrained maximization of posynomials by geometric
programming. Journal of Optimization Theory and Applications, 5(2):73–80, 1970.
Polyak, Boris T. Introduction to Optimization. Number 04; QA402. 5, P6. 1987.
Rockafellar, R Tyrrell and Uryasev, Stanislav. Optimization of conditional value-at-risk. Journal
of risk, 2:21–42, 2000.
Rockafellar, R Tyrrell and Uryasev, Stanislav. Conditional value-at-risk for general loss distributions. Journal of banking & finance, 26(7):1443–1471, 2002.
Soma, Tasuku and Yoshida, Yuichi. A Generalization of Submodular Cover via the Diminishing
Return Property on the Integer Lattice. In NIPS, 2015.
Soma, Tasuku, Kakimura, Naonori, Inaba, Kazuhiro, and Kawarabayashi, Ken-ichi. Optimal Budget Allocation: Theoretical Guarantee and Efficient Algorithm. In ICML, 2014.
Svitkina, Zoya and Fleischer, Lisa. Submodular approximation: Sampling-based algorithms and
lower bounds. SIAM Journal on Computing, 40(6):1715–1737, 2011.
Topkis, Donald M. Minimizing a submodular function on a lattice. Operations research, 26(2):
305–321, 1978.
18
Wainwright, Kevin and Chiang, Alpha.
McGraw-Hill Education, 2004.
Fundamental Methods of Mathematical Economics.
Zhang, Peng, Chen, Wei, Sun, Xiaoming, Wang, Yajun, and Zhang, Jialin. Minimizing seed set
selection with probabilistic coverage guarantee in a social network. In KDD, 2014.
A
Worst-Case Approximation Ratio versus True Worst-Case
Consider the function f (x; θ) defined on {0, 1} × {0, 1}, with values given by:
(
(
1
x=0
1 x=0
f (x; 0) =
f (x; 1) =
0.6 x = 1,
2 x = 1.
(18)
We wish to choose x to maximize f (x; θ) robustly with respect to adversarial choices of θ. If θ were
fixed, we could directly choose x∗θ to maximize f (x; θ). In particular, x∗0 = 0 and x∗1 = 1. Of course,
we want to deal with worst-case θ. One option is to maximize the worst-case approximation ratio:
max min
x
θ
f (x; θ)
.
f (x∗θ ; θ)
(19)
One can verify that the best x according to this criterion is x = 1, with worst-case approximation
ratio 0.6 and worst-case function value 0.6. In this paper, we optimize the worst-case of the actual
function value:
max min f (x; θ).
(20)
x
θ
This criterion will select x = 0, which has a worse worst-case approximation ratio of 0.5, but
actually guarantees a function value of 1, significantly better than the 0.6 achieved by the other
formulation of robustness.
B
DR-submodularity and L\ -convexity
A function is L\ -convex if it satisfies a discrete version of midpoint convexity, i.e. for all x, y it
holds that
x+y
x+y
f (x) + f (y) ≥ f
+f
.
(21)
2
2
Remark B.1. An L\ -convex function need not be DR-submodular, and vice-versa. Hence algorithms for optimizing one type may not apply for the other.
Proof. Consider f1 (x1 , x2 ) = −x21 − 2x1 x2 and f2 (x1 , x2 ) = x21 + x22 , both defined on {0, 1, 2} ×
{0, 1, 2}. The function f1 is DR-submodular but violates discrete midpoint convexity for the pair
of points (0, 0) and (2, 2), while f2 is L\ -convex but does not have diminishing returns in either
dimension.
Intuitively-speaking, L\ -convex functions look like discretizations of convex functions. The
continuous objective function I(x, y) we consider need not be convex, hence its discretization need
not be L\ -convex, and we cannot use those tools. However, in some regimes (namely if each
y(s) ∈ {0} ∪ [1, ∞)), it happens that I(x, y) is DR-submodular in x.
19
C
Constrained Continuous Submodular Function Minimization
Define Rn↓ to be the set of vectors ρ in Rn which are monotone nonincreasing, i.e. ρ(1) ≥ ρ(2) ≥
· · · ≥ ρ(n). As in the main text, define [k] = {0, 1, . . . , k − 1}. One
Qof the key results from Bach
[2015] is that an arbitrary submodular function H(x) defined on ni=1 [ki ] can be extended to a
particular convex function h↓ (ρ) so that
minimize h↓ (ρ)
minimize H(x)Q
Q
⇔
n
s.t.
x ∈ i=1 [ki ]
s.t.
ρ ∈ ni=1 R↓ki −1 .
(22)
Moreover, Theorem 4 from Bach [2015] states that, if aiyi are strictly convex functions for all
i = 1, . . . , n and each yi ∈ [ki ], then the two problems
P P
minimize H(x) + ni=1 xyii=1 a0iyi (λ)
Q
(23)
s.t.
x ∈ ni=1 [ki ].
and
P P
minimize h↓ (ρ) + ni=1 kxii−1
=1 aixi [ρi (xi )]
Qn
ki −1
s.t.
ρ ∈ i=1 R↓
(24)
are equivalent. In particular, one recovers a solution to Problem (23) for any λ just as alluded to
in Lemma 3.2: find ρ∗ which solves Problem (24) and, for each component i, choose xi to be the
maximal value for which ρ∗i (xi ) ≥ λ.
C.1
Proof of Lemma 3.2
Proof.
The discretized form of the regularizer Rδ is also separable and can be written Rδ (x) =
Pn
1 2
δ
δ
i=1 Ri (x). For each i = 1, . . . , n and each yi ∈ [ki ] with yi ≥ 1, define aiyi (t) = 2 t · [Ri (yi ) −
Riδ (yi − 1)], so that a0iyi (t) = t · [Riδ (yi ) − Riδ (yi − 1)]. Since we assumed R(x) is strictly increasing,
the coefficient of t2 in each aiyi (t) is strictly positive, so that each aiyi (t) is strictly convex. Then,
xi
X
Riδ (yi ) − Riδ (yi − 1)
λRiδ (xi ) = λ · Riδ (0) +
(25)
yi =1
= λRiδ (0) +
xi
X
a0iyi (λ),
(26)
yi =1
so that the discretized version of the minimization problem can be written as
P P
minimize H δ (x) + λRδ (0) + ni=1 xyii=1 a0iyi (λ)
Q
s.t.
x ∈ ni=1 [ki ].
Since the term Rδ (0) does not depend on the variable x, this minimization is equivalent to
P P
minimize H δ (x) + ni=1 xyii=1 a0iyi (λ)
Q
s.t.
x ∈ ni=1 [ki ].
(27)
(28)
This problem is in the precise form where we can apply the preceding equivalence result between
Problems (23) and (24), so we are done.
20
C.2
Proof of Theorem 3.1
Proof. The general idea of this proof is to first show that the integer-valued point x∗d which solves
x∗d ∈
H δ (x)
argmin
x∈
Qn
i=1 [ki ]:R
δ (x)≤B
is also nearly a minimizer of the continuous version of the problem, due to the fineness of the
discretization. Then, we show that the solutions traced out by x(λ) get very close to x∗d . These
two results are simply combined via the triangle inequality.
C.2.1
Continuous and Discrete Problems
We begin by proving that
H δ (x∗d ) ≤ Gδ +
min
H(x).
(29)
x∈X :R(x)≤B
Consider x∗ ∈ arg minx∈X :R(x)≤B H(x). If x∗ corresponds to an integral point in the discretized
coordinate.
domain, then H(x∗ ) = H δ (x∗d ) and we are done. Else, x∗ has at least one non-integral
Q
By rounding coordinatewise, we can construct a set X = {x1 , . . . , xm } ⊆ ni=1 [ki ] so that x∗ ∈
conv({A(x1 ), . . . , A(xm )}. By monotonicity, there must be some xi ∈ X with Rδ (xi ) ≤ B, i.e.
A(xi ) is feasible for the original continuous problem. By construction, since the discretization
given by A is δ-fine, we must have kx∗ − A(xi )k∞ ≤ δ. Applying the Lipschitz property of H and
the optimality of x∗ , we have
Gδ ≥ H(A(xi )) − H(x∗ ) = H δ (xi ) − H(x∗ ) ≥ H δ (x∗d ) − H(x∗ ),
from which (29) follows.
C.2.2
Discrete and Parameterized Discrete Problems
Define λ− and λ+ by
λ− ∈
H δ (x(λ))
argmin
and
λ≥0:Rδ (x(λ))≤B
λ+ ∈
H δ (x(λ)).
argmax
λ≥0:Rδ (x(λ))≥B
The next step in proving our suboptimality bound is to prove that
H δ (x(λ+ )) ≤ H δ (x∗d ) ≤ H δ (x(λ− )),
from which it will follow that
H δ (x(λ− )) ≤ Gδ + H δ (x∗d ).
21
(30)
We begin by stating the min-max inequality, i.e. weak duality:
n
o
δ
δ
δ
max
H
(x)
+
λ(R
(x)
−
B)
min
H
(x)
=
min
Q
Q
δ
λ≥0
x∈ n
x∈ n
i=1 [ki ]
i=1 [ki ]:R (x)≤B
n
o
δ
δ
≥ max Q
min
H
(x)
+
λ(R
(x)
−
B)
λ≥0 x∈ n
i=1 [ki ]
n
o
= max H δ (x(λ)) + λ(Rδ (x(λ)) − B)
λ≥0
n
o
≥
max
H δ (x(λ)) + λ(Rδ (x(λ)) − B)
λ≥0:Rδ (x(λ))≥B
≥
max
λ≥0:Rδ (x(λ))≥B
H δ (x(λ))
(31)
(32)
(33)
(34)
(35)
= H δ (x(λ+ )).
(36)
We can also bound the optimal value of H δ (x∗d ) from the other side:
H δ (x∗d ) =
x∈
Qn
min
δ
i=1 [ki ]:R (x)≤B
H δ (x) ≤
min
λ≥0:Rδ (x(λ))≤B
H δ (x) = H δ (x(λ− ))
(37)
Q
because the set of x(λ) parameterized by λ is a subset of the full set {x ∈ ni=1 [ki ] : Rδ (x) ≤ B}.
We have now bounded the optimal value of H δ (x∗d ) on either side by optimization problems
where we seek an optimal λ ≥ 0 for the parameterization x(λ):
H δ (x(λ+ )) ≤ H δ (x∗d ) ≤ H δ (x(λ− )).
(38)
Recall that x(λ) comes from thresholding the values of ρ∗ by λ, and that we assume that the
elements of ρ∗ are unique. Hence, as we increase λ, the components of x decrease by 1 each time.
Combining this with the strict monotonicity of R, we see that kx(λ+ ) − x(λ− )k∞ ≤ 1. By the
Lipschitz properties of H δ , it follows that H δ (x(λ+ )) − H δ (x(λ− )) ≤ Gδ. Since H δ (x∗d ) lies in
the interval between H δ (x(λ+ )) and H δ (x(λ− )), it follows that H δ (x∗d ) − H δ (x(λ− )) ≤ Gδ.
C.3
Proof of Remark 3.1
Define λ∗ = λ− as in the previous section, so that x0 = A(x(λ∗ ). The x(λ+ ) bound is a simple
consequence from the above result that
H δ (x(λ+ )) ≤ H δ (x∗d ) ≤ H δ (x(λ− )) = H(x0 ).
As for the Lagrangian bound, since x(λ∗ ) is a minimizer for the regularized function H δ (x) +
λ∗ (Rδ (x) − B), it follows that
H δ (x(λ∗ )) + λ∗ (Rδ (x(λ∗ )) − B) ≤ H δ (x∗d ) + λ∗ (Rδ (x∗d ) − B).
(39)
Rearranging, and observing that Rδ (x∗d ) ≤ B because x∗d is feasible, it holds that
H(x0 ) = H δ (x(λ∗ )) ≤ H δ (x∗d ) + λ∗ (Rδ (x∗d ) − Rδ (x(λ∗ ))) ≤ H δ (x∗d ) + λ∗ (B − R(x0 )).
(40)
One can also combine either of these bounds with the result from the proof of Theorem 3.1 that
H δ (x∗d ) ≤ Gδ + H(x∗ ) yielding e.g.
H(x0 ) ≤ Gδ + λ∗ (B − R(x0 )) + H δ (x∗d ).
22
C.4
Solving the Optimization Problem
Now that we have proven equivalence results between the constrained problem we want to solve
and the convex problem (24), we need to actually solve the convex problem. At the beginning of
Section 5.2 in Bach [2015], it is stated that this surrogate problem can optimized via the FrankWolfe method and its variants, but only the the version of Problem (24) without the extra functions
aixi is elaborated upon. Here we detail how Frank-Wolfe algorithms can be used to solve the more
general parametric regularized problem. Our aim is to spell out very clearly the applicability of
Frank-Wolfe to this problem, for the ease of future practitioners.
Bach [2015] notes that by duality, Problem (24) is equivalent to:
min
Q
ki −1
ρ∈ n
i=1 R↓
h↓ (ρ) − H(0) +
n kX
i −1
X
aixi [ρi (xi )] =
i=1 xi =1
max hρ, wi +
min
Q
ki −1 w∈B(H)
ρ∈ n
i=1 R↓
= max
min
w∈B(H) ρ∈Qn
k −1
i
i=1 R↓
n kX
i −1
X
aixi [ρi (xi )]
i=1 xi =1
hρ, wi +
n kX
i −1
X
i=1 xi =1
aixi [ρi (xi )]
:= max f (w).
w∈B(H)
Here, the base polytope B(H) happens to be the convex hull of all vectors w which could be output
by the greedy algorithm in Bach [2015].
It is the dual problem, where we maximize over w, which is amenable to Frank-Wolfe. For FrankWolfe methods, we need two oracles: an oracle which, given w, returns ∇f (w); and an oracle which,
given ∇f (w), produces a point s which solves the linear optimization problem maxs∈B(H) hs, ∇f (w)i.
Per Bach [2015], an optimizer of the linear problem can be computed directly from the greedy
algorithm. For the gradient oracle, recall that we can find a subgradient of g(x) = miny h(x, y)
at the point x0 by finding y(x0 ) which is optimal for the inner problem, and then computing
∇x h(x, y(x0 )). Moreover, if such y(x0 ) is the unique optimizer, then the resulting vector is indeed
the gradient of g(x) at x0 . Hence, in our case, it suffices to first find ρ(w) which solves the inner
problem, and then ∇f (w) is simply ρ(w) because the inner function is linear in w. Since each
function aixi is strictly convex, the minimizer ρ(w) is unique, confirming that we indeed get a
gradient of f , and that f is differentiable.
Of course, we still need to compute the minimizer ρ(w). For a given w, we want to solve
min
ρ∈
Qn
i=1
k −1
hρ, wi +
R↓i
n kX
i −1
X
aixi [ρi (xi )]
i=1 xi =1
There are no constraints coupling the vectors ρi , and the objective is similarly separable, so we can
independently solve n problems of the form
min hρ, wi +
ρ∈Rk−1
↓
k−1
X
aj (ρj ).
j=1
Recall that each function aiyi (t) takes the form
1 2
2 t riyi
23
for some riyi > 0. Let D = diag(r), the
(k − 1) × (k − 1) matrix with diagonal entries rj . Our problem can then be written as
k−1
min hρ, wi +
ρ∈Rk−1
↓
1X
1
rj ρ2j = min hρ, wi + hDρ, ρi
2
2
ρ∈Rk−1
↓
j=1
1
= min hD1/2 ρ, D−1/2 wi + hD1/2 ρ, D1/2 ρi.
k−1
2
ρ∈R↓
Completing the square, the above problem is equivalent to
min kD
ρ∈Rk−1
↓
1/2
ρ+D
−1/2
wk22
= min
ρ∈Rk−1
↓
= min
k−1
X
1/2
−1/2
(rj ρj + rj
wj )2
j=1
k−1
X
ρ∈Rk−1
↓
j=1
rj (ρj + rj−1 wj )2 .
This last expression is precisely the problem which is called weighted isotonic regression: we are
fitting ρ to diag(r−1 )w, with weights r, subject to a monotonicity constraint. Weighted isotonic
regression is solved efficiently via the Pool Adjacent Violators algorithm of Best & Chakravarti
[1990].
C.5
Runtime
Frank-Wolfe returns an ε-suboptimal solution in O(ε−1 D2 L) iterations, where D is the diameter of
the feasible region, and L is the Lipschitz constant for the gradient of the objective [Jaggi, 2013].
Our optimization problem is maxw∈B(H) f (w) as defined in the previous section. Each w ∈ B(H)
has O(nδ −1 ) coordinates of the form H δ (x + ei ) − H δ (x). Since H δ is an expected influence in
the range [0, T ], we can bound the magnitude of each coordinate of w by T and hence D2 by
O(nδ −1 T 2 ). If α is the minimum derivative of the functions Ri , then the smallest coefficient of the
functions aixi (t) is bounded below by αδ. Hence the objective is the conjugate of an αδ-strongly
convex function, and therefore has α−1 δ −1 -Lipschitz gradient. Combining these, we arrive at the
O(ε−1 nδ −2 α−1 T 2 ) iteration bound. The most expensive step in each iteration is computing the
subgradient, which requires sorting the O(nδ −1 ) elements of ρ in time O(nδ −1 log nδ −1 ). Hence the
total runtime of Frank-Wolfe is O(ε−1 n2 δ −3 α−1 T 2 log nδ −1 ).
As specified in the main text, relating an approximate solution of (15) to a solution of (14) is
nontrivial. Assume ρ∗ has distinct elements separated by η, and chose ε to be less than η 2 αδ/8. If
ρ is ε-suboptimal, then by αδ-strong convexity we must have kρ − ρ∗ k2 < η/2, and therefore kρ −
ρ∗ k∞ < η/2. Since the smallest consecutive gap between elements of ρ∗ is η, this implies that ρ and
ρ∗ have the same ordering, and therefore admit the same solution x after thresholding. Accounting
for this choice in ε, we have an exact solution to (14) in total runtime of O(η −2 n2 δ −4 α−2 T 2 log nδ −1 ).
24
D
Expectation and Variance of the Influence Function
We wish to study the influence I(X, y), its expectation and its variance as a function of y. By
definition, the influence function is given by
X
Y
y(s)
1 −
Xst .
(41)
I(X, y) =
t∈T
(s,t)∈E
Before we prove the stated results, we will simplify the functions involved.
Maximizing I(X, y) is equivalent to minimizing the function
X Y
y(s)
Xst
(42)
t∈T (s,t)∈E
and vice-versa. The particular properties we are interested in, namely convexity and submodularity,
are preserved under sums. Moreover, expectation is linear and variances add, so for our purposes
we can focus on only one term of the above sum. After reindexing in terms of i = 1, . . . , n instead
of (s, t) ∈ E, we are left studying functions of the form
f (y) =
n
Y
Xiyi .
(43)
i=1
If f (y) is always convex (or supermodular), then I(X, y) is always concave (submodular) in y, and
similarly for their expectations and variances.
Expectation
By independence,
E[f (y)] =
n
Y
E[Xiyi ].
(44)
i=1
Suppose that each Xi ∼ Beta(αi , βi ), so that
Γ(αi + βi )Γ(αi + yi )
Γ(αi + βi + yi )
Γ(αi + βi )Γ(αi + yi )Γ(βi )
=
Γ(αi + βi + yi )Γ(βi )
B(αi + yi , βi )
=
.
B(αi , βi )
E[Xiyi ] =
Then,
E[f (y)] =
n
Y
B(αi + yi , βi )
i=1
B(αi , βi )
∝
n
Y
B(αi + yi , βi ),
(45)
(46)
(47)
(48)
i=1
where by ∝ we mean that the product of the denominators is a positive constant, dependent on
the problem data but independent of y.
25
Variance
The variance of f (y) can be written as
"
Var
n
Y
#
Xiyi
=E
" n
Y
i=1
#
Xi2yi
−E
i=1
=
=
#2
Xiyi
(49)
i=1
n
n
h
i Y
Y
2
E Xi2yi −
E [Xiyi ]
i=1
(50)
i=1
n
Y
B(αi + 2yi , βi )
i=1
D.1
" n
Y
B(αi , βi )
−
n
Y
B(αi + yi , βi ) 2
i=1
B(αi , βi )
.
(51)
Gradient of Expected Influence
Recall the identity
∂
B(a, b) = B(a, b)(ψ(a) − ψ(a + b)),
(52)
∂a
where ψ is the digamma function. We can then compute each component of the gradient of E[f (y)]:
n
Y
Y
∂
∂
1
·
B(αj + yj , βj ) ·
(E[f (y)]) =
(B(αi + yi , βi ))
∂yi
B(αi , βi )
∂yi
=
i=1
n
Y
i=1
(53)
j6=i
Y
1
·
B(αj + yj , βj ) · B(αi + yi , βi ) · (ψ(αi + yi ) − ψ(αi + yi + βi ))
B(αi , βi )
j6=i
(54)
= E[f (y)] · (ψ(αi + yi ) − ψ(αi + yi + βi )).
D.2
(55)
Counterexample for Fact 2.1
p
We give a specific choice of parameters n, αi , βi and yi for which the resulting function Var(f (y))
is non-convex, non-concave, non-submodular and non-supermodular for various points y ∈ Rn+ . For
the p
case T = 1, the function 1−f (y) is a valid influence function, so we have a valid counterexample
for Var(I(X, y)).
Consider the case n = 2, with α1 = α2 = 1 and β1 = β2 = 1. This corresponds to the Budget
Allocation problem where we have two sources each with an edge to one customer, and we have
only our prior (i.e. no data)
p on either of the edge probabilities. Using equation (51), we can directly
compute the Hessian of Var(f (y)) at any point y, e.g. using Mathematica.
In particular, for
p
y1 = y2 = 1, the Hessian has a positive and a negative eigenvalue, so Var(f (y)) is neither
p convex
nor concave at this point. Also for y1 = y2 = 1, the off-diagonal element is negative, so Var(f (y))
is not supermodular over all of R2+ . However, for y1 = y2 = 3, the off-diagonal element is positive,
so our function is also not submodular.
26
| 2cs.AI
|
arXiv:1512.07067v1 [cs.PL] 22 Dec 2015
Transforming Javascript Event-Loop Into a Pipeline
Etienne Brodu, Stéphane Frénot
Frédéric Oblé
{etienne.brodu, stephane.frenot}@insa-lyon.fr
[email protected]
Univ Lyon, INSA Lyon, Inria, CITI, F-69621 Villeurbanne,
France
Worldline, Bât. Le Mirage, 53 avenue Paul Krüger
CS 60195, 69624 Villeurbanne Cedex
ABSTRACT
1. INTRODUCTION
The development of a real-time web application often starts
with a feature-driven approach allowing to quickly react to
users feedbacks. However, this approach poorly scales in
performance. Yet, the user-base can increase by an order
of magnitude in a matter of hours. This first approach is
unable to deal with the highest connections spikes. It leads
the development team to shift to a scalable approach often linked to new development paradigm such as dataflow
programming. This shift of technology is disruptive and
continuity-threatening. To avoid it, we propose to abstract
the feature-driven development into a more scalable highlevel language. Indeed, reasoning on this high-level language
allows to dynamically cope with user-base size evolutions.
We propose a compilation approach that transforms a
Javascript, single-threaded real-time web application into a
network of small independent parts communicating by message streams. We named these parts fluxions, by contraction
between a flow1 and a function. The independence of these
parts allows their execution to be parallel, and to organize
an application on several processors to cope with its load,
in a similar way network routers do with IP traffic. We test
this approach by applying the compiler to a real web application. We transform this application to parallelize the
execution of an independent part and present the result.
“Release early, release often”, “Fail fast”. The growth of
a real-time web service is partially due to Internet’s capacity to allow very quick releases of a minimal viable product
(MVP). It is crucial for the prosperity of such project to
quickly validate that it meets the needs of its users. Indeed,
misidentifying the market needs is the first reason for startup
failure2 . Hence the development team quickly concretizes an
MVP using a feature-driven approach and iterates on it.
The service needs to be scalable to be able to respond to
the growth of its user-base. However, feature-driven development best practices are hardly compatible with the required parallelism. The features are organized in modules
which disturb the organization of a parallel execution [6, 12,
16]. Eventually the growth requires to discard the initial approach to adopt a more efficient processing model. Many of
the most efficient models decompose applications into execution units [10, 20, 7]. However, these tools are in disruption
from the initial approach. This shift causes the development
team to spend development resources in background to start
over the initial code base, without adding visible value for
the users. It is a risk for the evolution of the project. Running out of cash and missing the right competences are the
second and third reasons for startup failures2 .
The risk described above comes from a disruption between
the two levels of application expression, the feature level and
the execution level. To avoid this risk and allow a continuous
development process, we propose a tool to automatically
map one level onto the other, and make the transition.
We focus on web applications driven by users requests
and developed in Javascript using the Node.js 3 execution
environment. Javascript is widely adopted45 to develop web
applications, and its event-loop model is very similar to a
pipeline architecture. So we propose a compiler to transform
an application into a pipeline of parallel stages communicating by message streams. We named these stages fluxions, by
contraction between a flux and a function.
We present a proof of concept for this compilation approach. Section 2 describes the execution environment targeted by this compiler. Then, section 3 presents the compiler, and section 4 its evaluation. Section 5 compare our
work with related works. And finally, we conclude this paper.
Categories and Subject Descriptors
Software and its engineering [Software notations and tools]:
Compilers—Runtime environments
Keywords
Flow programming, Web, Javascript
1
flux in french
2
https://www.cbinsights.com/blog/startup-failure-post-mortem/
https://nodejs.org/
4
http://githut.info/
5
http://stackoverflow.com/tags
3
ACM ISBN 978-1-4503-2138-9.
DOI: 10.1145/1235
2.
2.2 Example
FLUXIONAL EXECUTION MODEL
This section presents an execution model to provide scalability to web applications with a granularity of parallelism
at the function level. Functions are encapsulated in autonomous execution containers with their state, so as to be
mobile and parallel, similarly to the actors model. The communications are similar to the dataflow programming model,
which allows to reason on the throughput of these streams,
and to react to load increases [4].
The fluxional execution model executes programs written
in our high-level fluxionnal language, whose grammar is presented in figure 1. An application hprogrami is partitioned
into parts encapsulated in autonomous execution containers
named fluxions hflxi. The following paragraphs present the
fluxions and the messaging system to carry the communications between fluxions, and then an example application
using this execution model.
2.1 Fluxions and Messaging System
3
var app = require ( ’ express ’) () ,
fs = require ( ’ fs ’) ,
count = 0;
4
5
6
7
8
9
10
app . get ( ’/ ’ , function handler ( req , res ) {
fs . readFile ( __filename , function reply ( err , data ) {
count += 1;
res . send ( err || template ( count , data ) ) ;
}) ;
}) ;
11
12
app . listen (8080) ;
Listing 1: Example web application
The fluxional execution model is illustrated with an example application presented in listing 1. This application
reads a file, and sends it back along with a request counter.
The handler function, line 5 to 10, receives the input stream
of requests. The count variable at line 3 counts the requests,
and needs to be saved between two messages receptions. The
template function formats the output stream to be sent back
to the client. The app.get and res.send functions, lines 5
and 8, interface the application with the clients. Between
these two interface functions is a chain of three functions to
process the client requests : app.get →
→ handler → reply.
This chain of functions is transformed into a pipeline, expressed in the high-level fluxionnal language in listing 2. The
transformation process between the source and the fluxional
code is explained in section 3.
grp_res
reply
count
template
main
handler
2 start
messaging
system
4 exec
5 post
3 dequeue 6 enqueue
serialization
A fluxion hflxi is named by a unique identifier hidi to receive messages, and might be part of one or more groups
indicated by tags htagsi. A fluxion is composed of a processing function hfni, and a local memory called a context
hctxi.
At a message reception, the fluxion modifies its context,
and sends messages to downstream fluxions on its output
streams hstreamsi. The context stores the state on which a
fluxion relies between two message receptions. The messaging system queues the output messages for the event loop to
process them later by calling the downstream fluxions.
In addition to message passing, the execution model allows fluxions to communicate by sharing state between their
contexts. The fluxions that need this synchronization are
grouped with the same tag, and loose their independence.
There are two types of streams, start and post, which correspond to the nature of the rupture point producing the
stream. A variable created within a chain of post streams
requires more synchronization than a variable created upstream a start stream. The two types and implications of
rupture points are further detailed in section 3. Start rupture points are indicated with a double arrow (→
→ or >>) and
post rupture points with a simple arrow (→ or ->).
1
2
messaging
system
4
5
3
6
1 register
hprogrami
hflxi
htagsi
hstreamsi
hstreami
hdesti
hctxi
hmsgi
hlisti
htypei
hidi
hfni
|=
|=
|=
|=
|=
|=
|=
|=
|=
|=
|=
|=
hflxi | hflxi eol hprogrami
flx hidi htagsi hctxi eol hstreamsi eol hfni
& hlisti | empty string
null | hstreami | hstreami eol hstreamsi
htypei hdesti [hmsgi]
hlisti
{hlisti}
[hlisti]
hidi | hidi , hlisti
>> | ->
Identifier
Source language with hstreami placeholders
Figure 1: Syntax of a high-level language to represent a program in the fluxionnal form
program
Figure 2: The fluxionnal execution model in details
The execution is illustrated in figure 2. The dashed arrows
between fluxions represent the message streams as seen in
the fluxionnal application. The plain arrows represent the
operations of the messaging system during the execution.
These steps are indicated by numeroted circles. The program registers its fluxions in the messageing system, 1 . The
fluxion reply has a context containing the variable count and
template. When the application receives a request, the first
fluxion in the stream, main, queues a start message containing the request, 2 . This first message is to be received
by the next fluxion handler, 3 , and triggers its execution,
4 . The fluxion handler sends back a message, 5 , to be
enqueued, 6 . The system loops through steps 3 through
6 until the queue is empty. This cycle starts again for each
new incoming request causing another start message.
1
2
3
4
5
flx main & grp_res
>> handler [ res ]
var app = require ( ’ express ’) () ,
fs = require ( ’ fs ’) ,
count = 0;
source
esprima
Analyzer
pipeline
representation
escope
scopes
representation
Pipeliner
AST
flx source
6
app . get ( ’/ ’ , >> handler ) ;
app . listen (8080) ;
7
8
Figure 3: Compilation chain
9
10
11
12
13
14
flx handler
-> reply [ res ]
function handler ( req , res ) {
fs . readFile ( __filename , -> reply );
}
The chain of compilation is described in figure 3. The
compiler extracts an AST from the source with esprima.
From this AST, the Analyzer step identifies the limits of
the different application parts and how they relate to form
a pipeline. This first step outputs a pipeline representation
of the application. Section 3.1 explains this first compilation
step. In the pipeline representation, the stages are not yet
independent and encapsulated into fluxions. From the AST,
escope produces a representation of the memory scopes. The
Pipeliner step analyzes the pipeline representation and the
scopes representation to distribute the shared memory into
independent groups of fluxions. Section 3.2 explains this
second compilation step.
15
16
17
18
19
20
21
flx reply & grp_res { count , template }
-> null
function reply ( error , data ) {
count += 1;
res . send ( err || template ( count , data ) ) ;
}
Listing 2: Example application expressed in the
high-level fluxional language
The chain of functions from listing 1 is expressed in the
fluxional language in listing 2. The fluxion handler doesn’t
have any dependencies, so it can be executed in a parallel
event-loop. The fluxions main and reply belong to the group
grp_res, indicating their dependency over the variable res.
The group name is arbitrarily chosen by the compiler. All
the fluxions inside a group are executed sequentially on the
same event-loop, to protect against concurrent accesses.
The variable res is created and consumed within a chain
of post stream. Therefore, it is exclusive to one request and
cannot be propagated to another request. It doesn’t prevent
the whole group from being replicated. However, the fluxion
reply depends on the variable count created upstream the
start stream, which prevents this replication. If it did not
rely on this state, the group grp_res would be stateless, and
could be replicated to cope with the incoming traffic.
This execution model allows to parallelize the execution
of an application as a pipeline, as with the fluxion handler.
And some parts are replicated, as could be the group grp_res.
This parallelization improves the scalability of the application. Indeed, as a fluxion contains its state and expresses
its dependencies, it can be migrated. It allows to adapt the
number of fluxions per core to adjust the resource usage in
function of the desired throughput.
Our goal, as described in the introduction, is not to propose a new high-level language but to automate the architectural shift. We present the compiler to automate this
architectural shift in the next section.
3.
FLUXIONNAL COMPILER
The source languages we focus on should offer higher-order
functions and be implemented as an event-loop with a global
memory. Javascript is such a language and is often implemented on top of an event-loop, like in Node.js. We developed a compiler that transforms a Node.js application into a
fluxional application compliant with the execution model described in section 2. Our compiler uses the estools 6 suite to
parse, manipulate and generate source code from Abstract
Syntax Tree (AST). And it is tailored for – but not limited
to – web applications using Express 7 , the most used Node.js
web framework.
6
7
https://github.com/estools
http://expressjs.com/
3.1 Analyzer step
The limit between two application parts is defined by a
rupture point. The analyzer identifies these rupture points,
and outputs a representation of the application in a pipeline
form. Application parts are the stages, and rupture points
are the message streams of this pipeline.
3.1.1 Rupture points
A rupture point is a call of a loosely coupled function. It
is an asynchronous call without subsequent synchronization
with the caller. In Node.js, I/O operations are asynchronous
functions and indicate rupture points between two application parts. Figure 4 shows a code example of a rupture point
with the illustration of the execution of the two application
parts isolated into fluxions. The two application parts are
the caller of the asynchronous function call on one hand, and
the callback provided to the asynchronous function call on
the other hand.
upstream fluxion
Main thread
downstream fluxion
I/O thread
Main thread
arguments evaluation
asyncCall execution
Following statements 1
asynchronous operation
result
callback execution 2
1
2
asyncCall ( arguments , function callback ( result ) { 2
// Following statements 1
}) ;
Figure 4: Rupture point interface
A callback is a function passed as a parameter to a function call. It is invoked by the callee to continue the execution
with data not available in the caller context. There are three
kinds of callbacks, but only two are asynchronous: listeners
and continuations. The two corresponding types of rupture
points are start and post.
Start rupture points (listeners) are on the border between the application and the outside, continuously receiving incoming user requests. An example of a start rupture
Stream.
point is in listing 1, between the call to app.get(), and its
listener handler. These rupture points indicate the input of
a data stream in the program, and the beginning of a chain
of fluxions to process this stream.
Post rupture points (continuations) represent a continuity in the execution flow after an asynchronous operation
yielding a unique result, such as reading a file, or a database.
An example of a post rupture points is in listing 1, between
the call to fs.readFile(), and its continuation reply.
If a modified variable is read by downstream application
parts, then the pipeliner makes the upstream fluxion add
this variable to the message stream to be sent to the downstream fluxions. It is impossible to send variables to upstream fluxions, without causing inconsistencies. If the fluxion retro propagates the variable for an upstream fluxion to
read, the upstream fluxion might use the old version while
the new version is on its way.
In figure 5, the variable b is set in the function onReq,
and read in the function add. The pipeliner step makes the
fluxion onReq send the updated variable b, in addition to the
variable v, in the message sent to the fluxion add.
Exceptionally, if a variable is defined inside a post chain,
like b, then this variable can be streamed inside this post
chain without restriction on the order of modification and
read. Indeed, the execution of the upstream fluxion for the
current post chain is assured to end before the execution of
the downstream fluxion. Therefore, no reading of the variable by the upstream fluxion happens after the modification
by the downstream fluxion.
3.1.2 Detection
The compiler uses a list of common asynchronous callees,
like the express and file system methods. This list can be
augmented to match asynchronous callees individually for
any application. To identify the callee, the analyzer walks
the AST to find a call expression matching this list.
After the identification of the callee, the callback needs to
be identified as well, to be encapsulated in the downstream
fluxion. For each asynchronous call detected, the compiler
tests if one of the arguments is of type function. Some
callback functions are declared in situ, and are trivially detected. For variable identifiers, and other expressions, the
analyzer tries to detect their type. The analyzer walks back
the AST to track their assignations and modifications, so as
to determine their last value.
Share.
If a variable is needed for modification by several application parts, or is read by an upstream application part, then
it needs to be synchronized between the fluxions. To respect
the semantics of the source application, we cannot tolerate
inconsistencies. Therefore, the pipeliner groups all the fluxions sharing this variable with the same tag. And it adds
this variable to the contexts of each fluxions.
In figure 5, the variable c is set in the function end, and
read in the function add. As the fluxion add is upstream
of end, the pipeliner step groups the fluxion add and end
with the tag grp_c to allow the two fluxions to share this
variable.
3.2 Pipeliner step
A rupture point eventually breaks the chain of scopes between the upstream and downstream fluxion. The closure
in the downstream fluxion cannot access the scope in the
upstream fluxion as expected. The pipeliner step replaces
the need for this closure, allowing application parts to rely
only on independent memory stores and message passing. It
determines the distribution using the scope representation,
which represents the variables’ dependencies between application parts. Depending on this representation, the compiler
can replace the broken closures in three different ways. We
present these three alternatives in figure 5.
4. REAL TEST CASE
This section presents a test of the compiler on a real application, gifsockets-server8 . This test proves the possibility for
an application to be compiled into a network of independent
parts. It shows the current limitations of this isolation and
the modifications needed on the application to circumvent
them. This section then presents future works.
flx main
var a = 0;
var c = 0;
var a = 0;
var c = 0;
get(>> onReq);
req
get(function onReq(req) {
var b = req.count;
read(function add(v) {
a += b + c + v;
update(a, function end(updt) {
c = updt;
});
});
});
flx onReq
var b = req.count;
read(-> add);
flx add
a += b + c + v;
update(a, -> end);
flx end
c = updt;
b
1
v,b
grp_c
a,c
updt
2
3
4
5
6
7
c
8
9
10
Figure 5: Variable management from Javascript to
the high-level fluxionnal language
11
12
13
14
15
Scope.
If a variable is modified inside only one application part
in the current post chain, then the pipeliner adds it to the
context of its fluxion.
In figure 5, the variable a is updated in the function add.
The pipeliner step stores this variable in the context of the
fluxion add.
var express = require ( ’ express ’) ,
app = express () ,
routes = require ( ’ gifsockets - middleware ’) ,
getRawBody = require ( ’raw - body ’);
16
function bodyParser ( limit ) {
return function saveBody ( req , res , next ) {
getRawBody ( req , {
expected : req . headers [ ’ content - length ’] ,
limit : limit
} , function ( err , buffer ) {
req . body = buffer ;
next () ;
}) ;
};
}
17
18
19
app . post ( ’/ image / text ’ , bodyParser (1 * 1024 * 1024) ,
routes . writeTextTo Im ag es ) ;
app . listen (8000) ;
Listing 3: Simplified version of gifsockets-server
8
https://github.com/twolfson/gifsockets-server
4.2 Isolation
This application, simplified in listing 3, is a real-time chat
using gif-based communication channels. It was selected in
a previous work [5] from the npm registry because it depends
on express, it is tested, working, and simple enough to illustrate this evaluation. The server transforms the received
text into a gif frame, and pushes it back to a never-ending
gif to be displayed on the client.
On line 18, the application registers two functions to process the requests received on the url /image/text. The closure saveBody, line 7, returned by bodyParser, line 6, and the
method routes.writeTextToImages from the external module gifsockets-middleware, line 3. The closure saveBody
calls the asynchronous function getRawBody to get the request body. Its callback handles the errors, and calls next
to continue processing the request with the next function,
routes.writeTextToImages.
In listing 4, the fluxion anonymous_1000 modifies the object req, line 23, to store the text of the received request,
and it calls next to continue the execution, line 24. These
operations produce side-effects that should propagate in the
whole application, but the isolation prevents this propagation. Isolating the fluxion anonymous_1000 produces runtime
exceptions. We detail in the next paragraph, how we handle
this situation to allow the application to be parallelized.
4.2.1 Variable req
The variable req is read in fluxion main, lines 10 and 11.
Then its property body is associated to buffer in fluxion
anonymous_1000, line 23. The compiler is unable to identify
further usages of this variable. However, the side effect resulting from this association impacts a variable in the scope
of the next callback, routes.writeTextToImages. We modified the application to explicitly propagate this side-effect
to the next callback through the function next. We explain
further modification of this function in the next paragraph.
4.1 Compilation
We compile this application with the compiler detailed in
section 3. Listing 4 presents the compilation result. The
function call app.post, line 18, is a rupture point. However,
its callbacks, bodyParser and routes.writeTextToImages are
evaluated as functions only at runtime. For this reason, the
compiler ignores this rupture point, to avoid interfering with
the evaluation.
1
2
3
4
5
6
flx main & express { req }
>> anonymous_100 0 [ req , next ]
var express = require ( ’ express ’) ,
app = express () ,
routes = require ( ’ gifsockets - middleware ’) ,
getRawBody = require ( ’raw - body ’) ;
7
8
9
10
11
12
13
14
15
function bodyParser ( limit ) {
return function saveBody ( req , res , next ) {
getRawBody ( req , {
expected : req . headers [ ’ content - length ’] ,
limit : limit
} , >> anonymous_1000 ) ;
};
}
16
17
18
app . post ( ’/ image / text ’ , bodyParser (1 * 1024 * 1024) ,
routes . writeTextTo Im ag es ) ;
app . listen (8000) ;
19
20
21
22
23
24
25
flx anonymous_1000
-> null
function ( err , buffer ) {
req . body = buffer ;
next () ;
}
Listing 4: Compilation result of gifsockets-server
The compiler detects a rupture point : the function getRawBody and its anonymous callback, line 11. It encapsulates this callback in a fluxion named anonymous_1000. The
callback is replaced with a stream placeholder to send the
message stream to this downstream fluxion. The variables
req and next are appended to this message stream, to propagate their value from the main fluxion to the anonymous_1000
fluxion.
When anonymous_1000 is not isolated from the main fluxion, as if they belong to the same group, the compilation
result works as expected. The variables used in the fluxion, req and next, are still shared between the two fluxions.
Our goal is to isolate the two fluxions, to be able to safely
parallelize their executions.
4.2.2 Closure next
The function next is a closure provided by the express
Router to continue the execution with the next function to
handle the client request. Because it indirectly relies on the
variable req, it is impossible to isolate its execution with the
anonymous_1000 fluxion. Instead, we modify express, so as
to be compatible with the fluxionnal execution model. We
explain the modifications below.
1
2
3
4
5
6
flx anonymous_1000
-> e x p r e s s _ d i s p a tc he r
function ( err , buffer ) {
req . body = buffer ;
next_placehol de r ( req , -> e x p r e s s _ d i s p a t ch er) ;
}
7
8
9
10
11
12
13
flx e x p r e s s _ d i s p at c he r & express { req }
-> null
function ( modified_req ) {
merge ( req , modified_req );
next () ;
}
Listing 5: Simplified modification on the compiled
result
In listing 3, the function next is a continuation allowing
the anonymous callback, line 11, to call the next function to
handle the request. To isolate the anonymous callback into
anonymous_1000, next is replaced by a rupture point. This
replacement is illustrated in listing 5. The express Router
registers a fluxion named express_dispatcher, line 8, to continue the execution after the fluxion anonymous_1000. This
fluxion is in the same group express as the main fluxion,
hence it has access to the original variable req, and to the
original function next. The call to the original next function is replaced by a placeholder to push the stream to the
fluxion express_dispatcher, line 5. The fluxion express_dispatcher receives the stream from the upstream fluxion
anonymous_1000, merges back the modification in the variable req to propagate the side effects, and finally calls the
original function next to continue the execution, line 12.
After the modifications detailed above, the server works
as expected. The isolated fluxion correctly receives, and returns its serialized messages. The client successfully receives
a gif frame containing the text.
4.3 Future works
We intend to implement the compilation process presented
into the runtime. A just-in-time compiler would allow to
identify callbacks dynamically evaluated, and to analyze the
memory to identify side-effects propagations instead of relying only on the source code. Moreover, this memory analysis
would allow the closure serialization required to compile application using higher-order functions.
5.
CONCLUSION
In this paper, we presented our work on a high-level language allowing to represent a web application as a network
of independent parts communicating by message streams.
We presented a compiler to transform a Node.js web application into this high-level representation. To identify two
independent parts, the compiler spots rupture points in the
application, possibly leading to memory isolation and thus,
parallelism. We presented an example of a compiled application to show the limits of this approach. The parallelism
of this approach allows code-mobility which may lead to a
better scalability. We believe it can enable the scalability
required by highly concurrent web applications without discarding the familiar, feature-driven programming models.
9
[1]
[2]
[3]
[4]
RELATED WORKS
Splitting a task into independent parts goes back to the
Actor’s model, functional programming [12] and the following works on DataFlow leading to Flow-based Programming (FBP) and Functional Reactive programming (FRP)
[8]. Both FBP and FRP, recently got some attention in the
Javascript community with NoFlo 9 , Bacon.js 10 and react 11 .
The execution model we presented in section 2, is inspired
by works on scalability for very large systems, like the Staged
Event-Driven Architecture (SEDA) by Matt Welsh [20] and
by the MapReduce architecture [7]. It also drew its inspiration from more recent work following SEDA like Spark [21],
MillWheel [1], Naiad [15] and Storm [19]. The first part of
our work stands upon these thorough studies. However, we
believe that it is difficult for most developers to distribute
the state of an application. This belief motivated us to propose a compiler from an imperative programming model to
these more scalable, distributed execution engines.
The transformation of an imperative programming model
to be executed onto a parallel execution engine was recently
addressed by Fernandez et. al. [9]. However, as in similar
works [17], it requires annotations from developers, therefore
partially conserves the disruption with the feature-based development. Our approach discards the need for annotations,
thus targets a broader range of developers than only ones
experienced with parallel development.
A great body of work focuses on parallelizing sequential
programs [3, 13, 14, 18]. Because of the synchronous execution of a sequential program, the speedup of parallelization
is inherently limited [2, 11]. On the other hand, our approach
is based on an asynchronous programming model. Hence the
attainable speedup is not limited by the main synchronous
thread of execution.
6.
References
http://noflojs.org/
10
https://baconjs.github.io/
11
https://facebook.github.io/react/
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
T Akidau and A Balikov. “MillWheel: Fault-Tolerant
Stream Processing at Internet Scale”. In: VLDB (2013).
G Amdahl. “Validity of the Single Processor Approach
to Achieving Large Scale Computing Capabilities”. In:
(1967).
U Banerjee. Loop parallelization. 2013.
T Bartenstein and Y Liu. “Rate Types for Stream Programs”. In: OOPSLA (2014).
E Brodu, S Frénot, and F Oblé. “Toward automatic
update from callbacks to Promises”. In: AWeS (2015).
A Clements, M Kaashoek, N Zeldovich, R Morris, and
E Kohler. “The scalable commutativity rule”. In: SOSP
(2013).
J Dean and S Ghemawat. “MapReduce: Simplified Data
Processing on Large Clusters”. In: OSDI (2004).
C Elliott and P Hudak. “Functional reactive animation”. In: SIGPLAN (1997).
R Fernandez, M Migliavacca, E Kalyvianaki, and P
Pietzuch. “Making state explicit for imperative big data
processing”. In: USENIX ATC (2014).
A Fox, S Gribble, Y Chawathe, E Brewer, and P Gauthier. “Cluster-based scalable network services”. In: SOSP
(1997).
N Gunther. “A general theory of computational scalability based on rational functions”. In: ArXiv e-prints
(2008).
J Hughes. “Why functional programming matters”. In:
The computer journal (1989).
F Li, A Pop, and A Cohen. “Automatic Extraction of
Coarse-Grained Data-Flow Threads from Imperative
Programs”. In: IEEE Micro (2012).
N Matsakis. “Parallel Closures A new twist on an old
idea”. In: HotPar (2012).
F McSherry, R Isaacs, M Isard, and D Murray. “Composable Incremental and Iterative Data-Parallel Computation with Naiad”. In: Microsoft Research (2012).
D Parnas. “On the criteria to be used in decomposing systems into modules”. In: Communications of the
ACM (1972).
R Power and J Li. “Piccolo: Building Fast, Distributed
Programs with Partitioned Tables.” In: OSDI (2010).
C Radoi, S Fink, R Rabbah, and M Sridharan. “Translating imperative code to MapReduce”. In: OOPSLA
(2014).
A Toshniwal, J Donham, N Bhagat, S Mittal, D Ryaboy,
S Taneja, A Shukla, K Ramasamy, J Patel, S Kulkarni, J Jackson, K Gade, and M Fu. “Storm@ twitter”.
In: SIGMOD (2014).
M Welsh, S Gribble, E Brewer, and D Culler. “A design
framework for highly concurrent systems”. In: University of California, Berkeley (2000).
M Zaharia, T Das, H Li, S Shenker, and I Stoica. “Discretized streams: an efficient and fault-tolerant model
for stream processing on large clusters”. In: Proceedings of the 4th USENIX conference on Hot Topics in
Cloud Ccomputing (2012).
| 6cs.PL
|
The ParallelEye Dataset: Constructing Large-Scale
Artificial Scenes for Traffic Vision Research
arXiv:1712.08394v1 [cs.CV] 22 Dec 2017
Xuan Li, Kunfeng Wang, Member, IEEE, Yonglin Tian, Lan Yan, and Fei-Yue Wang, Fellow, IEEE
Abstract—Video image datasets are playing an essential role
in design and evaluation of traffic vision algorithms. Nevertheless, a longstanding inconvenience concerning image datasets is
that manually collecting and annotating large-scale diversified
datasets from real scenes is time-consuming and prone to error.
For that virtual datasets have begun to function as a proxy
of real datasets. In this paper, we propose to construct largescale artificial scenes for traffic vision research and generate a
new virtual dataset called “ParallelEye”. First of all, the street
map data is used to build 3D scene model of Zhongguancun
Area, Beijing. Then, the computer graphics, virtual reality, and
rule modeling technologies are utilized to synthesize large-scale,
realistic virtual urban traffic scenes, in which the fidelity and
geography match the real world well. Furthermore, the Unity3D
platform is used to render the artificial scenes and generate accurate ground-truth labels, e.g., semantic/instance segmentation,
object bounding box, object tracking, optical flow, and depth. The
environmental conditions in artificial scenes can be controlled
completely. As a result, we present a viable implementation
pipeline for constructing large-scale artificial scenes for traffic
vision research. The experimental results demonstrate that this
pipeline is able to generate photorealistic virtual datasets with
low modeling time and high accuracy labeling.
I. I NTRODUCTION
The publicly available video image datasets have received
much attention in recent years, due to its indispensability in
design and evaluation of computer vision algorithms [1]. In
general, a computer vision algorithm needs a large amount
of labeled images for training and evaluation. The datasets
can be divided into two types: unlabeled datasets used for
unsupervised learning and labeled datasets used for supervised
This work was partly supported by National Natural Science Foundation
of China under Grant 61533019, Grant 71232006, and Grant 91520301.
Xuan Li is with the School of Automation, Beijing Institute of Technology,
Beijing 100081, China, and also with The State Key Laboratory for Management and Control of Complex Systems, Institute of Automation, Chinese
Academy of Sciences, Beijing 100190, China (e-mail: [email protected]).
Kunfeng Wang (Corresponding author) is with The State Key Laboratory
for Management and Control of Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China, and also with
Qingdao Academy of Intelligent Industries, Qingdao 266000, China (e-mail:
[email protected]).
Yonglin Tian is with the Department of Automation, University of Science
and Technology of China, Hefei 230027, China, and also with The State Key
Laboratory for Management and Control of Complex Systems, Institute of
Automation, Chinese Academy of Sciences, Beijing 100190, China.
Lan Yan is with The State Key Laboratory for Management and Control
of Complex Systems, Institute of Automation, Chinese Academy of Sciences,
Beijing 100190, China.
Fei-Yue Wang is with The State Key Laboratory for Management and Control of Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China, and also with the Research Center for Computational Experiments and Parallel Systems Technology, National University of
Defense Technology, Changsha 410073, China (e-mail: [email protected]).
learning. However, manually annotating the images is timeconsuming and labor-intensive, and participants often lack
professional knowledge, making some annotation tasks difficult to execute. Experts are always sparse and should be
properly identified. As we known, the human annotators are
subjective, and their annotations should be re-examined if two
or more annotators have disagreements about the label of one
entity. By contrast, the computer is objective in processing
data and particularly good at batch processing, so why not let
the computer annotate the images automatically?
At present, most publicly available datasets are obtained
from real scenes. As the computer vision field enters the
big data era, researchers begin to look for better ways to
annotate large-scale datasets [2]. At the same time, the development of virtual datasets has a long history, starting at
least from Bainbridge’s work [3]. Bainbridge used Second Life
and World of Warcraft as two distinct examples of virtual
worlds to predict the scientific research potential of virtual
worlds, and introduced the virtual worlds into a lot of research
fields that scientists are now exploring, including sociology,
computer science, and anthropology. In fact, synthetic data
has been used for decades to benchmark the performance of
computer vision algorithms. The use of synthetic data has
been particularly significant in object detection [4], [5] and
optical flow estimation [6]-[8], but most virtual data are not
photorealistic or akin to the real-world data, and lack sufficient
diversity [9]. The fidelity of some virtual data is close to the
real-world [10]. However, the synthesized virtual worlds are
seldom equivalent to the real world in geographic position,
and seldom annotate the virtual images automatically. Richter
et al. [11] used a commercial game engine to extract virtual
images, with no access to the source code or the content. The
SYNTHIA dataset [12] provided a realistic virtual city as well
as synthetic images with automatically generated pixel-level
annotations, but in that dataset there lacks other annotation
information such as object bounding box and object tracking.
Gaidon et al. [13] proposed a virtual dataset called “Virtual
KITTI” as a proxy for tracking algorithm evaluation. While
this dataset was cloned from “KITTI”, it cannot extend easily
to arbitrary traffic networks. Due to the above limitations, new
virtual datasets that match the real world and provide detailed
ground truth annotations are still desirable.
Manually annotating pixel-level semantics for images is
time-consuming and not accurate enough. For example, annotating high-quality semantics with 10-20 categories in one
image usually takes 30-60 minutes [14]. This is known as
the “curse of dataset annotation” [15]. The more detailed the
Sky
Vegetation
Building
Fence
Sidewalk
Road
Trafficsign
Trafficlight
Pole
Tree
Cyclist
Pedestrian
Chair
Billboard
Fig. 1. Examples of our generated ParallelEye dataset. From left to right: a general view of the constructed artificial scenes, its semantic labels, a sample
frame with tracking bounding boxes generated automatically, and its semantic labels. Best viewed with zooming.
Perception and
understanding
Perception and
understanding
Artificial scenes
Computing and
observation
Experiment and
evaluation
Observation
and evaluation
Data-driven
In order to solve these problems, this paper proposes
a pipeline for constructing artificial scenes and generating
virtual images. First of all, we use map data to build the
3D scene model of Zhongguancun Area, Beijing. Then, we
use the computer graphics, virtual reality, and rule modeling
technologies to create a realistic, large-scale virtual urban
traffic scene, in which the fidelity and geographic information
can match the real world well. Furthermore, we use the
Unity3D development platform for rendering the scene and automatically annotating the ground truth labels including pixellevel semantic/instance segmentation, object bounding box,
object tracking, optical flow, and depth. The environmental
conditions in artificial scenes can be controlled completely. In
consequence, we generate a new virtual image dataset, called
“ParallelEye” (see Fig. 1). We will build a website and make
this dataset publicly available before the publication of this
paper. The experimental results demonstrate that our proposed
implementation pipeline is able to generate photorealistic
Real scene
Model-driven
semantics, the more labor-intensive the annotation process. As
a result, many datasets do not provide semantic segmentation
annotations. For example, ImageNet [16], [17] has 14 million
images, in which more than one million images have definite
class and the images are annotated with object bounding box
for object recognition. However, ImageNet does not have
semantic segmentation annotations. Some datasets provide
only limited semantic segmentation annotations. For example,
NYU-Depth V2 [18] has 1449 densely labelled images, KITTI
[1] has 547 images, CamVid [19], [20] has 600 images,
Urban LabelMe [21] has 942 images, and Microsoft COCO
[22] has three hundred thousand images. These datasets play
an important role in the study of semantic segmentation.
However, these datasets cannot be used directly in intelligent
transportation, especially in automobile navigation, because
the number of labeled images is insufficient and the segmented
semantics have different categories. Currently, computer vision algorithms that exploit context for pattern recognition
would benefit from datasets with many annotated categories
embedded in images from complex scenes. Such datasets
should contain a wide variety of environmental conditions with
annotated object instances co-occurring in the same scenes.
However, the real scenes are unrepeatable and the captured
images are expensive to annotate, making it difficult to obtain
large-scale, diversified datasets with precise annotations.
Learning and
training
Fig. 2. Basic framework and architecture for parallel vision [23].
virtual images with low modeling time and high fidelity.
The rest of this paper is organized as follows. Section
II introduces the significance of parallel vision and virtual
dataset. Section III presents our approach to constructing
artificial scenes and generating virtual images with groundtruth labels. Section IV reports the experimental results and
analyzes the performance. Finally, the concluding remarks are
made in section V.
II. PARALLEL V ISION AND V IRTUAL DATASET
Parallel vision [23]-[25] is an extension of the ACP (Artificial systems, Computational experiments, and Parallel execution) theory [26]-[30] into the computer vision field. For parallel vision, photo-realistic artificial scenes are used to model
and represent complex real scenes, computational experiments
are utilized to learn and evaluate a variety of vision models,
and parallel execution is conducted to online optimize the
vision system and realize perception and understanding of
complex scenes. The basic framework and architecture for
parallel vision [23] is shown in Fig. 2. Based on the parallel
vision theory, this paper constructs a large-scale virtual urban
network and synthesizes a large number of realistic images.
The first stage of parallel vision is to construct photorealistic
artificial scenes by simulating a variety of environmental
conditions occurring in real scenes, and accordingly to synthesize large-scale diversified datasets with precise annotations
generated automatically. Generally speaking, the construction
of artificial scenes can be regarded as “video game design”,
i.e., using the computer animation-like techniques to model
the artificial scenes. The main technologies used in this
stage include computer graphics, virtual reality, and microsimulation. Computer graphics and computer vision, on the
whole, can be thought of as a pair of forward and inverse
problems. The goal of computer graphics is to synthesize
image measurements given the description of world parameters
according to physics-based image formation principles (forward inference), while the focus of computer vision is to map
the pixel measurements to 3D scene parameters and semantics
(inverse inference). Apparently their goals are opposite, but
can converge to a common point: parallel vision.
From the parallel vision perspective, we design the ParallelEye dataset. ParallelEye is synthesized by referring to
the urban network of Zhongguancun Area, Beijing. Using
OpenStreetMap (OSM), an urban network with length 3km
and width 2km is extracted. Artificial scenes are constructed
on this urban network. Unity3D is used to control the environmental conditions in the scene. There are 15 object classes in
ParallelEye, reflecting the common elements of traffic scenes,
including sky, buildings, cars, roads, sidewalks, vegetation,
fence, traffic signs, traffic lights, lamp poles, billboards, trees,
cyclists, pedestrians, and chairs. These object classes can be
automatically annotated to generate pixel-level semantics. For
traffic vision research, we pay special attention to instance
segmentation, with each object of interest segmented automatically. In addition, ParallelEye provides accurate ground truth
for object detection and tracking, depth, and optical flow.
OpenStreetMap
CityEngine
Unity3D
、
、
、
、、
、
III. A PPROACH
Our pipeline for generating the ParallelEye dataset is shown
in Fig. 3. Firstly, the OSM data released by OpenStreetMap
is used to achieve the correspondence in geographic location
between the virtual and real world. Secondly, CityEngine
is used to write CGA (Computer Generated Architecture)
rules and design a realistic artificial scene, including roads,
buildings, cars, trees, sidewalks, etc. Thirdly, the artificial
scene is imported into Unity3D and gets rendered by using
the script and the shader. In the dataset, accurate ground truth
annotations are generated automatically, and environmental
conditions can be controlled completely and flexibly.
A. Correspondence of Artificial and Real Scenes
In order to increase the fidelity, we choose to import
geographic data from OpenStreetMap. Although Google Maps
occupies an important position in geographic information, it
is not an open-source software. By contrast, OpenStreetMap
is an open-source, online map editing program with the goal
of creating a world where content is freely accessible to
everyone. In OpenStreetMap, the ways denote a directional
node sequence. Each node of the network can connect 2-2000
paths, and then arrive at another node. The road information
includes direction, lane number, lane width, street name, and
speed limit. Each path can form three combinations: nonclosed paths, closed paths, and regions. The non-closed paths
correspond to the roads, rivers, and railways in the real world.
The closed paths correspond to subway, bus routes, residential
roads, and so on. The regions correspond to buildings, parks,
Fig. 3. Pipeline for generating the ParallelEye dataset with OpenStreetMap,
CityEngine, and Unity3D.
lakes, and so on. Based on the properties of OSM data, it is
easy to relate the real world to the geographic information of
the artificial scene. Fig. 4 shows the real Automation Building
of CASIA (Institute of Automation, Chinese Academy of
Sciences) and its virtual proxy generated by CGA rules. They
are similar in appearance.
B. Generation of Ground-Truth Annotations
As stated above, ground-truth annotations are essential
for vision algorithm design and evaluation. Traditionally, the
images were annotated by hand. The manual annotation is
time-consuming and prone to error. Taking semantic/instance
segmentation as an example, it usually takes 30-60 minutes
to annotate an image with 10-20 object categories. Besides,
manual annotation is more or less subjective, so that different
annotators can make different semantic labels for the same image, especially near the object boundaries. Instead of manual
annotation, this paper uses Unity3D to automatically generate
accurate ground-truth labels. Fig. 5 shows some examples of
ground-truth annotations, including depth, optical flow, object
tracking, object detection, instance segmentation, and semantic
segmentation.
Generating ground truth with Unity3D is accurate and
efficient. Semantic segmentation ground truth can be directly
Fig. 4. The real Automation Building of CASIA (top) and its virtual proxy
(bottom).
generated by using unlit shaders on the materials of the
objects, with each category outputting a unique color. Instance
segmentation ground truth is generated using the same method,
but assigns a unique color tag to each object of interest.
The modified shaders output a color which is not affected
by the lighting and shading conditions. Depth ground truth is
generated using built-in depth buffer information to get depth
data for screen coordinates. The depth ranges from 0 to 1
with a nonlinear distribution, with 1 representing “infinitely
distant”. Optical flow ground truth is generated by calculating
the instantaneous velocity of moving objects on the imaging
plane and using the pixel changes in the image sequence to find
the correspondence between the previous frame and the current
frame. Given a pixel point (x, y) in the image, at any time the
brightness of that point is E(x + 4x, y + 4y, t + 4t). Let
∂y
(u, v) = ( ∂x
∂t , ∂t ) represent the instantaneous velocity of the
point in the horizontal and vertical directions, the brightness
change occurs when the point moves. We use the Taylor
formula to represent the pixel brightness:
E(x + 4x, y + 4y, t + 4t)
∂E
∂E
∂E
= E(x, y, t) +
4x +
4y +
4t + ε.
∂x
∂y
∂t
(1)
For any 4t → 0, let ω = (u, v) , the optical flow constraint
equation is given by
−
∂E
∂E ∂x ∂E ∂y
=
+
= ∇E · ω,
∂t
∂x ∂t
∂y ∂t
(2)
where ω is the optical flow of E(x, y, t).
We generate multi-object tracking ground truth based on
four rules: 1) when the object appears within the field of
Fig. 5. Examples of ground-truth annotations generated automatically by
Unity3D. Top: depth (left) and optical flow (right). Middle: object tracking
(left) and object detection (right). Bottom: pixel-level instance segmentation
(left) and semantic segmentation (right). Best viewed with zooming.
view of the camera, the three-dimensional bounding box of
the object is converted to a two-dimensional bounding box;
2) when the object appears or disappears from the image
boundary, we perform special handling for the bounding box;
3) we do not draw bounding boxes for objects that have less
than 15 pixels in width or less than 10 pixels in height; 4)
when occlusion occurs and the occlusion rate is higher than
a threshold, we do not draw bounding boxes for the occluded
object.
C. Diversity of Artificial Scenes
In order to increase the diversity and fidelity of artificial
scenes, we control the parameters in the script, the material,
and the simulated environmental conditions. Specifically, the
controllable parameters include: 1) number, type, trajectory,
speed, and direction of the vehicles; 2) position and configuration of the camera; 3) weather (sunny, cloudy, rainy, foggy,
etc) and illumination (daytime, dawn, dusk, etc).
Traditionally, video image datasets are collected by capturing in the real world or retrieving from the Internet. It
is impossible to control the environmental conditions and
repeat the scene layout under different environments, and thus
difficult to isolate the effects of environmental conditions on
the performance of computer vision algorithms. By contrast,
it is easy to control the environmental conditions in artificial
scenes. In this work, we are able to flexibly control the
camera’s location, height, and orientation to capture different
Fig. 7. Continuous images captured by an onboard camera: a sample image
(left), another image annotated with object bounding boxes (middle), and the
third image annotated with tracking bounding boxes of different colors (right).
Best viewed with zooming.
Fig. 6. Illustration of the diversity of artificial scenes. Top: Virtual images
with illumination at 6:00 am (left) and 12:00 pm (right) in a sunny day.
Bottom: Virtual images with weather of fog (left) and rain (right).
contents of the artificial scene. We are also able to dynamically change the illumination (from sunrise to sunset) and
weather conditions (sunny, cloudy, and foggy). Although we
can change the environmental conditions in artificial scene, the
ground-truth annotations are always easy to generate, no matter how adverse the illumination and weather conditions are
and how blurred the image details are. This makes it possible
to quantitatively analyze the impacts of each environmental
condition on algorithm performance, usually called “ceteris
paribus analysis”. Fig. 6 illustrates the diversity of artificial
scenes in terms of illumination and weather conditions.
IV. E XPERIMENTS
Based on the proposed approach, we construct the artificial
scene and configure virtual cameras to capture images from
the scene. The virtual cameras can be moving or stationary.
For automobile applications, the virtual cameras are installed
on moving vehicles. For visual surveillance applications, the
virtual cameras are fixed on the roadside or at intersections.
The experiments are conducted to verify that the artificial
scenes are repeatable and that the camera’s position, height,
and orientation can be configured flexibly.
A. Onboard Camera
In this experiment, an onboard camera is configured at
a height of 2 meters, mimicking the camera installed on
the vehicle roof. There are totally 67 vehicles on the road,
including 52 vehicles parking on the roadside (3 buses, 4
trucks, and 45 cars ) and another 15 vehicles in motion. We
turn the camera orientation from left to right and get five
orientations (i.e., -30, -15, 0, 15, and 30 degrees with respect
to the lane direction). The distance between two cameras
on adjacent lanes is 5 meters. These configurations lead to
substantial changes in object appearance. Fig. 7 shows three
continuous images captured by the onboard camera.
Fig. 8. Continuous images captured by a surveillance camera: images
annotated with object bounding boxes (top row), original images (middle
row), and images annotated with tracking bounding boxes of different colors
(bottom row). Best viewed with zooming.
B. Surveillance Camera
In this experiment, a surveillance camera is installed at an
intersection. We rotate the camera and control the rotation
speed at 10 degrees per second, and the rotation range is 180
degrees. We also change the camera height, with the lifting
speed of 0.1 meters per second and the lifting range of 2-5
meters. Such settings can fully simulate the role of surveillance
cameras. Based on this experiment, the artificial scene provides
virtual video images for intersection monitoring. Fig. 8 shows
images captured by the surveillance camera.
In order to increase diversity of virtual images and record
the ground truth, we adopt the same operations for both
the onboard camera and the surveillance camera. To record
the ground truth, we use a green bounding box to record
the detection ground truth for each object. We also assign
a bound box of unique color to record the tracking ground
truth for each object instance. To increase diversity, we dynamically change the illumination (daytime, dawn, and dusk)
and weather (sunny, cloudy, rainy, and foggy) conditions in
the artificial scenes. These subtle changes simulate different
environmental conditions in the virtual world, and would
otherwise need the expensive process of re-acquiring and relabeling images of the real world. The advantage of this setting
is that it can increase diversity of the ParallelEye dataset. In
the experiments, with image resolution of 500*375 pixels for
ParallelEye, the pipeline for artificial scene construction and
ground truth generation runs at 8-12 fps (frames per second)
on a workstation computer. We have collected a total of 31,000
image frames, each of which has been annotated with accurate
ground truth. We will build a website and make the dataset
publicly available before the publication of this paper.
V. C ONCLUDING R EMARKS
In this paper, we propose a new virtual image dataset called
“ParallelEye”. For that we present a dataset generation pipeline
that uses street map, computer graphics, virtual reality, and
rule modeling technologies to construct a realistic, large-scale
virtual urban traffic scene. The artificial scene matches the real
world well in terms of fidelity and geographic information. In
the artificial scene, we flexibly configure the camera (including
its position, height, and orientation) and the environmental
conditions, to collect diversified images. Each image has
been annotated automatically with ground truth including
semantic/instance segmentation, object bounding box, object
tracking, optical flow, and depth.
In the future, we will improve the diversity of ParallelEye
by introducing moving pedestrians and cyclists, which are
harder to animate. We will increase the scale of ParallelEye.
In addition, we will combine ParallelEye and the existing real
datasets (e.g., PASCAL VOC, MS COCO, and KITTI) to learn
and evaluate traffic vision models, in order to improve the
accuracy and robustness of traffic vision models when applied
to complex traffic scenes.
R EFERENCES
[1] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets Robotics:
The KITTI dataset,” International Journal of Robotics Research, vol. 32,
no. 11, pp. 1231-1237, 2013.
[2] A. Handa, T. Whelan, J. McDonald, and A. J. Davison, “A benchmark for
RGB-D visual odometry, 3D reconstruction and SLAM,” in Proceedings
of the International Conference on Robotics and Automation, IEEE, pp.
1524-1531, 2014.
[3] W. S. Bainbridge, “The scientific research potential of virtual worlds,”
Science, vol. 317, no. 5837, pp. 472-476, 2007.
[4] J. Marin, D. Vazquez, D. Geronimo, and A. M. Lopez, “Learning
appearance in virtual scenarios for pedestrian detection,” in Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition,
IEEE, pp. 137-144, 2010.
[5] J. Papon and M. Schoeler, “Semantic pose using deep networks trained
on synthetic RGB-D,” in Proceedings of the IEEE Conference on International Conference on Computer Vision, IEEE, pp. 774-782, 2015.
[6] J. L. Barron, D. J. Fleet, and S. S. Beauchemin, “Performance of optical
flow techniques,” International Journal of Computer Vision, vol. 12, no.
1, pp. 43-77, 1994.
[7] B. McCane, K. Novins, D. Crannitch, and B. Galvin, “On benchmarking
optical flow,” Computer Vision and Image Understanding, vol. 84, no. 1,
pp. 126-143, 2001.
[8] S. Baker, D. Scharstein, J. P. Lewis, S. Roth, M. J. Black, and R. Szeliski,
“A database and evaluation methodology for optical flow,” International
Journal of Computer Vision, vol. 92, no. 1, pp. 1-31, 2011.
[9] G. Ros, S. Ramos, M. Granados, A. Bakhtiary, D. Vazquez, and A. M.
Lopez, “Vision-based offline-online perception paradigm for autonomous
driving,” in Proceedings of the IEEE Conference on Applications of
Computer Vision, IEEE, pp. 231-238, 2015.
[10] H. Prendinger, K. Gajananan, A. Bayoumy Zaki, A. Fares, R. Molenaar,
D. Urbano, H. Van Lint, and W. Gomaa, “Tokyo Virtual Living Lab:
Designing smart cities based on the 3d internet,” in Proceedings of
the IEEE Conference on Applications of Computer Vision, Springer
International Publishing,vol. 17, no. 6, pp. 30-38, 2013.
[11] S. R. Richter, V. Vineet, S. Roth, and V. Koltun, “Playing for data:
Ground truth from computer games,” in Proceedings of the European
Conference on Computer Vision, IEEE Internet Computing, pp. 102C118,
2016.
[12] G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and A. M. Lopez, “The
SYNTHIA dataset: A large collection of synthetic images for semantic
segmentation of urban scenes,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, IEEE, pp. 3234-3243, 2016.
[13] A. Gaidon, Q. Wang, Y. Cabon, and E. Vig, “Virtual worlds as proxy for
multi-object tracking analysis,” in Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, IEEE, pp. 4340-4349, 2016.
[14] A. Kundu, Y. Li, F. Dellaert, F. Li, and J. M. Rehg, “Joint semantic segmentation and 3D reconstruction from monocular video,” in Proceedings
of the European Conference on Computer Vision, pp. 703-718, 2014.
[15] J. Xie, M. Kiefel, M.-T. Sun, and A. Geiger, “Semantic instance
annotation of street scenes by 3D to 2D label transfer,” 2016 IEEE
Conference on Computer Vision and Pattern Recognition, pp. 3688-3697,
2016.
[16] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and F. F.
Li, “Large-scale video classification with convolutional neural networks,”
in Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pp. 1725-1732, 2014.
[17] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z.
Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and F.-F.
Li, “ImageNet Large Scale Visual Recognition Challenge,” International
Journal of Computer Vision, vol. 115, no. 3, pp. 211-252, 2015.
[18] N. Silberman, D. Hoiem, P. Kohli, and R. Fergus, “Indoor segmentation
and support inference from RGBD images,” in Proceedings of the
Conference European Conference on Computer Vision, pp. 746-760, 2012.
[19] G. J. Brostow, J. Fauqueur, and R. Cipolla, “Semantic object classes
in video: A high-definition ground truth database,” Pattern Recognition
Letters, vol. 30, no. 2, pp. 88-97, 2009.
[20] G. J. Browstow, J. Shotton, J. Fauqueur, and R. Cipolla, “Segmentation
and recognition using structure from motion point clouds,” in Proceedings
of the European Conference on Computer Vision, Heidelberg: Springer
Berlin Heidelberg, vol. 5302, pp. 44-57, 2008.
[21] B. C. Russell, A. Torralba, K. P. Murphy, and W. T. Freeman, “LabelMe:
A database and web-based tool for image annotation,” International
Journal of Computer Vision, vol. 77, no. 1, pp. 157-173, 2008.
[22] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan,
P. Dollar, and C. L. Zitnick, “Microsoft COCO: Common objects in
context,” in Proceedings of the European Conference on Computer Vision,
pp. 740-755, 2014.
[23] K. Wang, C. Gou, and F.-Y. Wang, “Parallel vision: An ACP-based
approach to intelligent vision computing,” Acta Automatica Sinica, vol.
42, no. 10, pp. 1490-1500, 2016.
[24] K. Wang, C. Gou, N. Zheng, J. M. Rehg, and F.-Y. Wang, “Parallel
vision for perception and understanding of complex scenes: Methods,
framework, and perspectives,” Artificial Intelligence Review, vol. 48, no.
3, pp. 298-328, 2017.
[25] K. Wang, Y. Lu, Y. Wang, Z. Xiong, and F.-Y. Wang, “Parallel imaging:
A new theoretical framework for image generation,” Pattern Recognition
and Artificial Intelligence, vol. 30, no. 7, pp. 577-587, 2017.
[26] F.-Y. Wang, “Parallel control and management for intelligent transportation systems: Concepts, architectures, and applications,” IEEE Transactions on Intelligent Transportation Systems, vol. 11, no. 3, pp. 630-638,
2010.
[27] F.-Y. Wang, “Parallel control: A method for data-driven and computational control,” Acta Automatica Sinica, vol. 39, no. 4, pp. 293-302,
2014.
[28] F.-Y. Wang, X. Wang, L. Li, and L. Li, “Steps toward parallel intelligence,” IEEE/CAA Journal of Automatica Sinica, vol. 3, no. 4, pp.345348, 2016.
[29] L. Li, Y. Lin, D. Cao, N. Zheng, and F.-Y. Wang, “Parallel learning —
A new framework for machine learning,” Acta Automatica Sinica, vol.
43, no. 1, pp. 1-8, 2017.
[30] X. Liu, X. Wang, W. Zhang, J. Wang, and F.-Y. Wang, “Parallel data:
From big data to data intelligence,” Pattern Recognition and Artificial
Intelligence, vol. 30, no. 8, pp. 673-682, 2017.
| 1cs.CV
|
1
Simulation of Quantum Circuits
via Stabilizer Frames
Héctor J. Garcı́a
Igor L. Markov
University of Michigan, EECS, Ann Arbor, MI 48109-2121
{hjgarcia, imarkov}@eecs.umich.edu
arXiv:1712.03554v1 [cs.DS] 10 Dec 2017
F
Abstract—Generic quantum-circuit simulation appears intractable for
conventional computers and may be unnecessary because useful quantum circuits exhibit significant structure that can be exploited during
simulation. For example, Gottesman and Knill identified an important
subclass, called stabilizer circuits, which can be simulated efficiently
using group-theory techniques and insights from quantum physics.
Realistic circuits enriched with quantum error-correcting codes and
fault-tolerant procedures are dominated by stabilizer subcircuits and
contain a relatively small number of non-Clifford components. Therefore,
we develop new data structures and algorithms that facilitate parallel
simulation of such circuits. Stabilizer frames offer more compact storage
than previous approaches but require more sophisticated bookkeeping.
Our implementation, called Quipu, simulates certain quantum arithmetic circuits (e.g., reversible ripple-carry adders) in polynomial time
and space for equal superpositions of n-qubits. On such instances,
known linear-algebraic simulation techniques, such as the (state-ofthe-art) BDD-based simulator QuIDDPro, take exponential time. We
simulate quantum Fourier transform and quantum fault-tolerant circuits
using Quipu, and the results demonstrate that our stabilizer-based technique empirically outperforms QuIDDPro in all cases. While previous
high-performance, structure-aware simulations of quantum circuits were
difficult to parallelize, we demonstrate that Quipu can be parallelized
with a nontrivial computational speedup.
1
I NTRODUCTION
Quantum information processing manipulates quantum
states rather than conventional 0-1 bits. It has been
demonstrated with a variety of physical technologies
(NMR, ion traps, Josephson junctions in superconductors, optics) and used in recently developed commercial
products. Examples of such products include MagiQ’s
quantum key distribution system and ID-Quantique’s
quantum random number generator. Shor’s factoring
algorithm [24] and Grover’s search algorithm [13] apply the principles of quantum information to carry out
computation asymptotically more efficiently than conventional computers. These developments fueled research
efforts to design, build and program scalable quantum
computers. Due to the high volatility of quantum information, quantum error-correcting codes (QECC) and
effective fault-tolerant (FT) architectures are necessary
to build reliable quantum computers. For instance, the
work in [22] describes practical FT architectures for
quantum computers, and [15] explores architectures for
constructing reliable communication channels via distribution of high-fidelity EPR pairs in a large quantum
computer. Most quantum algorithms are described in
terms of quantum circuits and, just like conventional digital circuits, require functional simulation to determine
the best FT design choices given limited resources. In
particular, high-performance simulation is a key component
in quantum design flows [25] that facilitates analysis
of trade-offs between performance and accuracy. Simulating quantum circuits on a conventional computer is
a difficult problem. The matrices representing quantum
gates, and the vectors that model quantum states grow
exponentially with an increase in the number of qubits
– the quantum analogue of the classical bit. Several
software packages have been developed for quantumcircuit simulation including Oemer’s Quantum Computation Language (QCL) [21] and Viamontes’ Quantum
Information Decision Diagrams (QuIDD) implemented
in the QuIDDPro package [26]. While QCL simulates
circuits directly using state vectors, QuIDDPro uses a
variant of binary decision diagrams to store state vectors
more compactly in some cases. Since the state-vector representation requires excessive computational resources
in general, simulation-based reliability studies (e.g. faultinjection analysis) of quantum FT architectures using
general-purpose simulators has been limited to small
quantum circuits [5]. Therefore, designing fast simulation
techniques that target quantum FT circuits facilitates more
robust reliability analysis of larger quantum circuits.
Stabilizer circuits and states. Gottesman [12] and Knill
identified an important subclass of quantum circuits,
called stabilizer circuits, which can be simulated efficiently on classical computers. Stabilizer circuits are exclusively composed of Clifford gates – Hadamard, Phase
and controlled-NOT gates (Figure 1) followed by onequbit measurements in the computational basis. Such cir-
1
H= √
2
1
1
1
−1
1
P =
0
0
i
1
0
CN OT =
0
0
0
1
0
0
0
0
0
1
0
0
1
0
Fig. 1. Clifford gates: Hadamard (H), Phase (P) and
controlled-NOT (CNOT).
2
cuits are applied to a computational basis state (usually
|00...0i) and produce output states known as stabilizer
states. Because of their extensive applications in QECC
and FT architectures, stabilizer circuits have been studied
heavily [1], [12]. Stabilizer circuits can be simulated in
polynomial-time by keeping track of the Pauli operators
that stabilize1 the quantum state. Such stabilizer operators are maintained during simulation and uniquely
represent stabilizer states up to an unobservable global
phase.2 Thus, this technique offers an exponential improvement over the computational resources needed to simulate stabilizer circuits using vector-based representations.
Aaronson and Gottesman [1] proposed an improved
technique that uses a bit-vector representation to simulate stabilizer circuits. Aaronson implemented this simulation approach in his CHP software package. Compared
to other vector-based simulators (QuIDDPro, QCL) the
technique in [1] does not maintain the global phase of a
state and simulates each Clifford gate in Θ(n) time using
Θ(n2 ) space. The overall runtime of CHP is dominated
by measurements, which require O(n2 ) time to simulate.
for input states consisting of equal superpositions of
computational-basis states. On such instances, wellknown generic simulation techniques take exponential
time. We simulate various quantum Fourier transform
and quantum fault-tolerant circuits, and the results
demonstrate that our data structure leads to ordersof-magnitude improvement in runtime and memory as
compared to state-of-the-art simulators.
In the remaining part of this document, we assume at
least a superficial familiarity with quantum computing.
Section 2 describes key concepts related to quantumcircuit simulation and the stabilizer formalism. In Section 3, we introduce stabilizer frames and describe
relevant algorithms. Section 4 describes in detail our
simulation flow implemented in Quipu, and Section 5
discusses a parallel implementation of our technique.
In Section 6, we describe empirical validation of Quipu
and in single- and multi-threaded variants, as well as
comparisons with state-of-the-art simulators. Section 7
closes with concluding remarks.
Stabilizer-based simulation of generic circuits. We propose a generalization of the stabilizer formalism that
admits simulation of non-Clifford gates such as Toffoli3
gates. This line of research was first outlined in [1],
where the authors describe a stabilizer-based representation that stores an arbitrary quantum state as a sum
of density-matrix4 terms. In contrast, we store arbitrary
states as superpositions5 of pure stabilizer states. Such superpositions are stored more compactly than the approach
from [1], although we do not handle mixed stabilizer
states. The key obstacle to the more efficient pure-state
approach has been the need to maintain the global phase
of each stabilizer state in a superposition, where such
phases become relative. We develop a new algorithm to
overcome this obstacle. We store stabilizer-state superpositions compactly using our proposed stabilizer frame
data structure. To speed up relevant algorithms, we
store generator sets for each stabilizer frame in row-echelon
form to avoid expensive Gaussian elimination during
simulation. The main advantages of using stabilizer-state
superpositions to simulate quantum circuits are:
(1) Stabilizer subcircuits are simulated with high efficiency.
(2) Superpositions can be restructured and compressed on the
fly during simulation to reduce resource requirements.
(3) Operations performed on such superpositions can be
computed in parallel and lend themselves to distributed
or asynchronous processing.
Our stabilizer-based technique simulates certain quantum arithmetic circuits in polynomial time and space
2
1. An operator U is said to stabilize a state iff U |ψi = |ψi.
2. According to quantum physics, the global phase exp(iθ) of a
quantum state is unobservable and does not need to be simulated.
3. The Toffoli gate is a 3-bit gate that maps (a, b, c) to (a, b, c ⊕ (ab)).
4. Density matrices are self-adjoint positive-semidefinite matrices of
trace 1.0, that describe the statistical state of a quantum system [19].
5. A superposition is a norm-1 linear combination of terms.
BACKGROUND
AND
P REVIOUS W ORK
Quantum information processes, including quantum algorithms, are often modeled using quantum circuits and,
just like conventional digital circuits, are represented
by diagrams [19], [26]. Quantum circuits are sequences
of gate operations that act on some register of qubits –
the basic unit of information in a quantum system. The
quantum state |ψi of a single qubit is described by a twodimensional complex-valued vector. In contrast to classical bits, qubits can be in a superposition of the 0 and 1
states. Formally, |ψi = α0 |0i + α1 |1i, where |0i = (1, 0)>
and |1i = (0, 1)> are the two-dimensional computational
basis states and αi are probability amplitudes that satisfy
|α0 |2 +|α1 |2 = 1. An n-qubit register is the tensor product
of n single qubits and thus is modeled
by a complex
P2n −1
vector |ψ n i = |ψ1 i ⊗ · · · ⊗ |ψn i =
i=0 αi |bi i, where
each bi is a binary string representing the value
i of each
P2n −1
basis state. Furthermore, |ψ n i satisfies i=0 |αi |2 = 1.
Each gate operation or quantum gate is a unitary matrix
that operates on a small subset of the qubits in a register.
For example, the quantum analogue of a NOT gate is the
operator X = ( 01 10 ),
X⊗I
α0 |00i + α1 |10i 7−−−→ α0 |10i + α1 |00i
Similarly, the two-qubit CNOT operator flips the second
qubit (target) iff the first qubit (control) is set to 1, e.g.,
CN OT
α0 |00i + α1 |10i 7−−−−→ α0 |00i + α1 |11i
Another operator of particular importance is the
Hadamard (H), which is frequently used to put a qubit
in a superposition of computational-basis states, e.g.,
I⊗H
α0 |00i + α1 |10i 7−−−→
α0 (|00i + |01i) + α1 (|10i + |11i)
√
2
3
Note that the H gate generates unbiased superpositions
in the sense that the squares of the absolute value of the
amplitudes are equal.
The dynamics involved in observing a quantum state
are described by non-unitary measurement operators [19,
Section 2.2.3]. There are different types of quantum measurements, but the type most pertinent to our discussion
comprises projective measurements in the computational basis, i.e., measurements with respect to the |0i or |1i basis
states. The corresponding measurement operators are
P0 = ( 10 00 ) and P1 = ( 00 01 ), respectively. The probability
p(x) of obtaining outcome x ∈ {0, 1} on the j th qubit
of state |ψi is given by the inner product hψ| Pxj |ψi,
where hψ| is the conjugate transpose of |ψi. For example,
the probability of obtaining |1i upon measuring |ψi =
α0 |0i + α1 |1i is
and amplitudes. The symmetries are operators for which
these states are 1-eigenvectors. Algebraically, symmetries
form group structures, which can be specified compactly
by group generators [14].
p(1) = (α0∗ , α1∗ )P1 (α0 , α1 )> = (0, α1∗ )(α0 , α1 )> = |α1 |2
Observe that I stabilizes all states and −I does not stabi√
lize any state. Thus, the entangled state (|00i + |11i)/ 2
is stabilized by the Pauli operators X ⊗X, −Y ⊗Y , Z ⊗Z
and I ⊗I. As shown in Table 1, it turns out that the Pauli
matrices along with I and the multiplicative factors ±1,
±i, form a closed group under matrix multiplication [19].
Formally, the Pauli group Gn on n qubits consists of the nfold tensor product of Pauli matrices, P = ik P1 ⊗ · · · ⊗ Pn
such that Pj ∈ {I, X, Y, Z} and k ∈ {0, 1, 2, 3}. For
brevity, the tensor-product symbol is often omitted so
that P is denoted by a string of I, X, Y and Z characters or Pauli literals, and a separate integer value k
for the phase ik . This string-integer pair representation
allows us to compute the product of Pauli operators
without explicitly computing the tensor products,6 e.g.,
(−IIXI)(iIY II) = −iIY XI. Since | Gn |= 4n+1 , Gn
can have at most log2 | Gn |= log2 4n+1 = 2(n + 1)
irredundant generators [19]. The key idea behind the
stabilizer formalism is to represent an n-qubit quantum
state |ψi by its stabilizer group S(|ψi) – the subgroup of
Gn that stabilizes |ψi.
Theorem 1: For an n-qubit pure state |ψi and k ≤ n,
S(|ψi) ∼
= Zk2 . If k = n, |ψi is specified uniquely by S(|ψi)
and is called a stabilizer state.
Proof: (i) To prove that S(|ψi) is commutative, let
P, Q ∈ S(|ψi) such that P Q |ψi = |ψi. If P and Q
anticommute, -QP |ψi = -Q(P |ψi) = -Q |ψi = -|ψi =
6 |ψi.
Thus, P and Q cannot both be elements of S(|ψi).
(ii) To prove that every element of S(|ψi) is of degree
2, let P ∈ S(|ψi) such that P |ψi = |ψi. Observe that
Cofactors of quantum states. The output states obtained
after performing computational-basis measurements are
called cofactors, and are states of the form |0i |ψ0 i and
|1i |ψ1 i. These states are orthogonal to each other and
add up to the original state. The norms of cofactors
and the original state are subject to the Pythagorean
theorem. We denote the |0i- and |1i-cofactor by ψ c=0
and ψ c=1 , respectively, where c is the index of the
measured qubit. One can also consider iterated cofactors,
such as double cofactors ψ qr=00 , ψ qr=01 , ψ qr=10 and
ψ qr=11 . Cofactoring with respect to all qubits produces
amplitudes of individual basis vectors. Readers familiar
with cofactors of Boolean functions can use intuition
from logic optimization and Boolean function theory.
2.1
Quantum circuits and simulation
To simulate a quantum circuit C, we first initialize the
quantum system to some desired state |ψi (usually a
basis state). |ψi can be represented using a fixed-size
data structure (e.g., an array of 2n complex numbers)
or a variable-size data structure (e.g., algebraic decision
diagram). We then track the evolution of |ψi via its
internal representation as the gates in C are applied one
at a time, eventually producing the output state C |ψi [1],
[19], [26]. Most quantum-circuit simulators [8], [20], [21],
[26] support some form of the linear-algebraic operations
described earlier. The drawback of such simulators is
that their runtime and memory requirements grows exponentially in the number of qubits. This holds true not
only in the worst case but also in practical applications
involving quantum arithmetic and quantum FT circuits.
Gottesman developed a simulation method involving
the Heisenberg model [12] often used by physicists to
describe atomic-scale phenomena. In this model, one keeps
track of the symmetries of an object rather than represent
the object explicitly. In the context of quantum-circuit
simulation, this model represents quantum states by
their symmetries, rather than complex-valued vectors
2.2
The stabilizer formalism
A unitary operator U stabilizes a state |ψi iff |ψi is
a 1–eigenvector of U , i.e., U |ψi = |ψi. We are interested in operators U derived
from the Pauli matrices:
1 0
X = ( 01 10 ) , Y = 0i −i
0 −1 , and the identity
0 ,Z =
1
0
I = ( 0 1 ). The one-qubit states stabilized by the Pauli
matrices are:
√
√
X : (|0i + |1i)/ √2 −X : (|0i − |1i)/ √2
Y : (|0i + i |1i)/ 2 −Y : (|0i − i |1i)/ 2
Z :
|0i
−Z :
|1i
6. This holds true due to the identity: (A⊗B)(C ⊗D) = (AC ⊗BD).
TABLE 1
Multiplication table for Pauli matrices. Shaded
cells indicate anticommuting products.
I
X
Y
Z
I
I
X
Y
Z
X
X
I
−iZ
iY
Y
Y
iZ
I
−iX
Z
Z
−iY
iX
I
4
P 2 = il I for l ∈ {0, 1, 2, 3}. Since P 2 |ψi = P (P |ψi) =
P |ψi = |ψi, we obtain il = 1 and P 2 = I.
(iii) From group theory, a finite Abelian group with
identity element e such that a2 = e for every element
a in the group must be ∼
= Zk2 .
(iv) We now prove that k ≤ n. First note that each independent generator P ∈ S(|ψi) imposes the linear constraint P |ψi = |ψi on the 2n -dimensional vector space.
The subspace of vectors that satisfy such a constraint
has dimension 2n−1 , or half the space. Let gen(|ψi) be
the set of generators for S(|ψi). We add independent
generators to gen(|ψi) one by one and impose their
linear constraints, to limit |ψi to the shared 1-eigenvector.
Thus the size of gen(|ψi) is at most n. In the case
|gen(|ψi)| = n, the n independent generators reduce the
subspace of possible states to dimension one. Thus, |ψi
is uniquely specified.
The proof of Theorem 1 shows that S(|ψi) is specified
by only log2 2n = n irredundant stabilizer generators.
Therefore, an arbitrary n-qubit stabilizer state can be represented by a stabilizer matrix M whose rows represent
a set of generators Q1 , . . . , Qn for S(|ψi). (Hence we use
the terms generator set and stabilizer matrix interchangeably.) Since each Qi is a string of n Pauli literals, the
size of the matrix is n × n. The fact that Qi ∈ S(|ψi)
implies that the leading phase of Qi can only be ±1
and not ±i.7 Therefore, we store the phases of each Qi
separately using a binary vector of size n.
The storage cost for M is Θ(n2 ), which is an exponential
improvement over the O(2n ) cost often encountered in
vector-based representations.
√
Example 1. The state |ψi = (|00i + |11i)/ 2 is uniquely
XX
specified by any of the following matrices: M1 = +
+ [ ZZ ],
+ XX
− YY
M2 = − [ Y Y ], M3 = + [ ZZ ]. One obtains M2 from M1
by left-multiplying the second row by the first. Similarly,
M3 is obtained from M1 or M2 via row multiplication.
Observe that multiplying any row by itself yields II,
which stabilizes |ψi. However, II cannot be used as a
generator because it is redundant and carries no information about the structure of |ψi. This holds true in general
for M of any size.
Theorem 1 suggests that Pauli literals can be represented using two bits, e.g., 00 = I, 01 = Z, 10 = X and
11 = Y . Therefore, a stabilizer matrix can be encoded
using an n × 2n binary matrix or tableau. This approach
induces a linear map Z2n
2 7→ Gn because vector addition
in Z22 is equivalent to multiplication of Pauli operators
up to a global phase. The tableau implementation of the
stabilizer formalism is covered in [1], [19].
Observation 1. Consider a stabilizer state |ψi represented by a set of generators of its stabilizer group
S(|ψi). Recall from the proof of Theorem 1 that, since
S(|ψi) ∼
= Zn2 , each generator imposes a linear constraint
on |ψi. Therefore, the set of generators can be viewed as
7. Suppose the phase of Qi is ±i, then Q2i =-I ∈ S(|ψi) which is
not possible since -I does not stabilize any state.
(a)
(b)
Fig. 2. (a) Stabilizer-matrix structure for basis states.
(b) Row-echelon form for stabilizer matrices. The X-block
contains a minimal set of generators with X/Y literals.
Generators with Z/I literals only appear in the Z-block.
a system of linear equations whose solution yields the 2k
(for some k between 0 and n) non-zero computationalbasis amplitudes that make up |ψi. Thus, one needs
to perform Gaussian elimination to obtain such basis
amplitudes from a generator set.
Canonical stabilizer matrices. Observe from Example 1
that, although stabilizer states are uniquely determined
by their stabilizer group, the set of generators may be
selected in different ways. Any stabilizer matrix can be
rearranged by applying sequences of elementary row
operations in order to obtain the row-reduced echelon
form structure depicted in Figure 2b. This particular
stabilizer-matrix structure defines a canonical representation for stabilizer states [10], [12]. The elementary row
operations that can be performed on a stabilizer matrix
are transposition, which swaps two rows of the matrix,
and multiplication, which left-multiplies one row with
another. Such operations do not modify the stabilizer
state and resemble the steps performed during a Gaussian elimination8 procedure. Several row-echelon (standard) forms for stabilizer generators along with relevant
algorithms to obtain them have been introduced in the
literature [3], [12], [19].
Stabilizer-circuit simulation. The computational-basis
states are stabilizer states that can be represented by
the stabilizer-matrix structure depicted in Figure 2a. In
this matrix form, the ± sign of each row along with
its corresponding Zj -literal designates whether the state
of the j th qubit is |0i (+) or |1i (−). Suppose we
want to simulate circuit C. Stabilizer-based simulation
first initializes M to specify some basis state. Then, to
simulate the action of each gate U ∈ C, we conjugate
each row Qi of M by U .9 We require that the conjugation
U Qi U † maps to another string of Pauli literals so that the
resulting matrix M0 is well-formed. It turns out that the
H, P and CNOT gates have such mappings, i.e., these
gates conjugate the Pauli group onto itself [12], [19].
Table 2 lists mappings for the H, P and CNOT gates.
Example 2. Suppose we simulate a CNOT gate on |ψi =
8. Since Gaussian elimination essentially inverts the n × 2n matrix, this
could be sped up to O(n2.376 ) time by using fast matrix inversion algorithms.
However, O(n3 )- time Gaussian elimination seems more practical.
9. Since Qi |ψi = |ψi, the resulting state U |ψi is stabilized by
U Qi U † because (U Qi U † )U |ψi = U Qi |ψi = U |ψi.
5
√
(|00i+|11i)/ 2. Using the stabilizer representation, Mψ =
+XX CN OT
0
7−−−−→ M0ψ = +XI
+IZ . The rows of Mψ stabilize
+ZZ
√
CN OT
|ψi 7−−−−→ (|00i + |10i)/ 2 as required.
Since H, P and CNOT gates are directly simulated via
the stabilizer formalism, these gates are also known as
stabilizer gates and any circuit composed exclusively of
such gates is called a unitary stabilizer circuit. Table 2
shows that at most two columns of M are updated when
a Clifford (stabilizer) gate is simulated. Therefore, such
gates are simulated in Θ(n) time. Furthermore, for any
pair of Pauli operators P and Q, P QP † = (−1)c Q, where
c = 0 if P and Q commute, and c = 1 otherwise. Thus,
Pauli gates can also be simulated in linear time as they
only permute the phase vector of the stabilizer matrix.
Theorem 2: An n-qubit stabilizer state |ψi can be ob⊗n
tained by applying a stabilizer circuit to the |0i
computational-basis state.
Proof: The work in [1] represents the generators
using a tableau, and then shows how to construct a
canonical stabilizer circuit C from the tableau. We refer
the reader to [1, Theorem 8] for details of the proof. Algorithms for obtaining more compact canonical circuits
are discussed in [11].
Corollary 3: An n-qubit stabilizer state |ψi can
be transformed by Clifford gates into the |00 . . . 0i
computational-basis state.
Proof: Since every stabilizer state can be produced by
⊗n
applying some unitary stabilizer circuit C to the |0i
state, it suffices to reverse C to perform the inverse
transformation. To reverse a stabilizer circuit, reverse the
order of gates and replace every P gate with P P P .
The stabilizer formalism also admits measurements
in the computational basis [12]. Conveniently, the formalism avoids the direct computation of measurement
operators and inner products (Section 2). However, the
updates to M for such gates are not as efficient as for
Clifford gates. Note that any qubit j in a stabilizer state is
either in a |0i (|1i) state or in an unbiased superposition
of both. The former case is called a deterministic outcome
and the latter a random outcome. We can tell these cases
apart in Θ(n) time by searching for X or Y literals in
the j th column of M. If such literals are found, the qubit
must be in a superposition and the outcome is random
with equal probability (p(0) = p(1) = .5); otherwise the
outcome is deterministic (p(0) = 1 or p(1) = 1).
TABLE 2
Conjugation of Pauli-group elements by Clifford
gates [19]. For the CN OT case, subscript 1
indicates the control and 2 the target.
G ATE
H
P
I NPUT
X
Y
Z
X
Y
Z
O UTPUT
Z
-Y
X
Y
-X
Z
G ATE
CN OT
I NPUT
I1 X2
X1 I2
I1 Y2
Y1 I2
I1 Z2
Z1 I2
O UTPUT
I1 X2
X1 X2
Z1 Y2
Y1 X2
Z1 Z2
Z1 I2
Case 1 – randomized outcomes: one flips an unbiased
coin to decide the outcome x ∈ {0, 1} and then updates
M to make it consistent with the outcome. Let Rj be a
row in M with an X/Y literal in its j th position, and
let Zj be the Pauli operator with a Z literal in its j th
position and I everywhere else. The phase of Zj is set
to +1 if x = 0 and −1 if x = 1. Observe that Rj and Zj
anticommute. If any other rows in M anticommute with
Zj , multiply them by Rj to make them commute with
Zj . Then, replace Rj with Zj . Since this process requires
up to n row multiplications, the overall runtime is O(n2 ).
Case 2 – deterministic outcomes: no updates to M are
necessary but we need to figure out whether the qubit is
in the |0i or |1i state, i.e., whether the qubit is stabilized
by Z or -Z. One approach is to perform Gaussian elimination (GE) to put M in row-echelon form. This removes
redundant literals from M and makes it possible to
identify the row containing a Z in its j th position and I’s
everywhere else. The ± phase of such a row decides the
outcome of the measurement. Since this is a GE-based
approach, it takes O(n3 ) time in practice.
The work in [1] improved the runtime of deterministic
measurements by doubling the size of M to include
n destabilizer generators in addition to the n stabilizer
generators. Such destabilizer generators help identify
which specific row multiplications to compute in order
to decide the measurement outcome. This approach
avoids GE and thus deterministic measurements are
computed in O(n2 ) time. In Section 3, we describe a
different approach that computes such measurements in
linear time without extra storage but with an increase in
runtime when simulating Clifford gates.
In quantum mechanics, the states eiθ |ψi and |ψi are
considered phase-equivalent because eiθ does not affect the statistics of measurement. Since the stabilizer
formalism simulates Clifford gates via their action-byconjugation, such global phases are not maintained.
Example 3. Suppose we have state |1i, which is stabilized by -Z. Conjugating the stabilizer by the Phase
gate yields P (-Z)P † =-Z. However, in the state-vector
P
representation, |1i 7−
→ i |1i. Thus the global phase i is
not maintained by the stabilizer.
Since global phases are unobservable, they do not need
to be maintained when simulating a single stabilizer
state. However, in Section 4, we show that such phases
must be maintained when dealing with stabilizer-state
superpositions, where global phases become relative.
3
S TABILIZER F RAMES : DATA S TRUCTURE
AND A LGORITHMS
The Clifford gates by themselves do not form a universal
set for quantum computation [1], [19]. However, the
Hadamard and Toffoli (T OF ) gates do [2]. To simulate
T OF and other non-Clifford gates, we extend the formalism to include the representation of arbitrary quantum
6
states as superpositions of stabilizer states.
Example 4. Recall from Section 2.2 that computationalbasis states are stabilizer states. Thus, any one-qubit state
|ψi = α0 |0i + α1 |1i is a superposition of the stabilizer
states |0i and |1i. In general, any state decomposition in
a computational basis is a stabilizer superposition.
Suppose |ψi in Example 4 is an unbiased state such
that α0 = ik α1 where k = {0, 1, 2, 3}. Then |ψi can
be represented using a single stabilizer state instead of
two (up to a global phase). The key idea behind our
technique is to identify and compress large unbiased
superpositions on the fly during simulation to reduce
resource requirements. To this end, we leverage the following observation and derive a compact data structure
for representing stabilizer-state superpositions.
Observation 2. Given an n-qubit stabilizer state |ψi,
there exists an orthonormal basis including |ψi and
consisting entirely of stabilizer states. One such basis is
obtained directly from the stabilizer of |ψi by changing
the signs of an arbitrary, non-empty subset of generators of S(|ψi), i.e., by modifying the phase vector of
the stabilizer matrix for |ψi.10 Thus, one can produce
2n −1 additional orthogonal stabilizer states. Such states,
together with |ψi, form an orthonormal basis. This basis
is unambiguously specified by a single stabilizer state,
and any one basis state specifies the same basis.
Definition 1. An n-qubit stabilizer frame F is a set of
k ≤ 2n stabilizer states {|ψj i}kj=1 that forms an orthogonal subspace basis in the Hilbert space. We represent F
by a pair consisting of (i) a stabilizer matrix M and (ii)
a set of distinct phase vectors {σj }kj=1 , where σj ∈ {±1}n .
We use Mσj to denote the ordered assignment of the
elements in σj as the (±1)-phases of the rows in M.
Therefore, state |ψj i is represented by Mσj . The size of
the frame, which we denote by |F|, is equal to k.
Each phase vector σj can be viewed as a binary (0-1)
encoding of the integer index that denotes the respective basis vector. Thus, when dealing with 64 qubits
or less, a phase vector can be compactly represented
by a 64-bit integer (modern CPUs also support 128-bit
integers). To represent an arbitrary state |Ψi using F, one
additionally maintains a vector of complex amplitudes
a = (a1 , . . . , ak ), which corresponds to the decomposition of |Ψi in the basis {|ψj i}kj=1 defined by F, i.e.,
Pk
Pk
2
|Ψi =
j=1 |aj | = 1. Observe that
j=1 aj |ψj i and
each aj forms a pair with phase vector σj in F since
|ψj i ≡ Mσj . Any stabilizer state can be viewed as a
one-element frame.
We now describe several frame operations that are useful for manipulating stabilizer-state superpositions.
ROTATE(F, U ). Consider the stabilizer basis {|ψj i}kj=1
defined by frame F. A stabilizer or Pauli gate U acting
on F maps such a basis to {U |ψj i = eiθj |ϕj i}kj=1 , where
eiθj is the global phase of stabilizer state |ϕj i. Since
we obtain a new stabilizer basis that spans the same
subspace, this operation effectively rotates the stabilizer
frame. Computationally, we perform a frame rotation as
follows. First, update the stabilizer matrix associated
with F as per Section 2.2. Then, iterate over the phase
vectors in F and update each one accordingly (Table 2).
Let a = (a1 , . . . , ak ) ∈ Ck be the decomposition of |Ψi
onto F. Frame rotation simulates the action of Clifford
gate U on |Ψi since,
U |Ψi =
k
X
j=1
aj U |ψj i =
k
X
aj eiθj |ϕj i
(1)
j=1
Observe that the global phase eiθj of each |ϕj i becomes
relative with respect to U |Ψi. Therefore, our approach
requires that we compute such phases explicitly in order
to maintain a consistent representation.
Global phases of states in F. Recall from Theorem 2 that
any stabilizer state |ψi = C |0 . . . 0i for some stabilizer
circuit C. To compute the global phase of |ψi, one keeps
track of the global factors generated when each gate in
C is simulated in sequence. In the context of frames, we
maintain the global phase of each state in F using the
amplitude vector a. Let M be the matrix associated with
F and let σj be the (±)-phase vector in F that forms a
pair with aj ∈ a. When simulating Clifford gate U , each
aj is updated as follows:
1) Set the leading phases of the rows in M to σj .
2) Obtain a basis state |bi from M and store its non-zero
amplitude β. If U is the Hadamard gate, it may be
necessary to sample a sum of two non-zero basis
amplitudes (one real, one imaginary).
3) Consider the matrix representation of U and the vector
representation of β |bi, and compute U (β |bi) = β 0 |b0 i via
matrix-vector multiplication.
4) Obtain |b0 i from U MU † and store its amplitude γ 6= 0.
5) Compute the global factor generated as aj = (aj · β 0 )/γ.
By Observation 1, M needs to be in row-echelon
form (Figure 2b) in order to sample the computational-
Example 5. Let |Ψi = a1 (|00i + |01i) + a2 (|10i + |11i).
Then |Ψi can be represented by the stabilizer frame F
depicted in Figure 3.
10. Let S(|ψi) and S(|ϕi) be the stabilizer groups for |ψi and |ϕi,
respectively. If there exist P ∈ S(|ψi) and Q ∈ S(|ϕi) such that P =
-Q, |ψi and |ϕi are orthogonal since |ψi is a 1-eigenvector of P and
|ϕi is a (−1)-eigenvector of P .
Fig. 3.
A two-qubit stabilizer state |Ψi whose frame
representation uses two phase vectors.
7
basis amplitudes |bi and |b0 i. Thus, simulating gates with
global-phase maintenance would take O(n3 |F|) time for
n-qubit stabilizer frames. To improve this, we introduce
a simulation invariant.
Invariant 1: The stabilizer matrix M associated with F
remains in row-echelon form during simulation.
Since Clifford gates affect at most two columns of
M, Invariant 1 can be repaired with O(n) row multiplications. Since each row multiplication takes Θ(n),
the runtime required to update M during global-phase
maintenance simulation is O(n2 ). Therefore, for an nqubit stabilizer frame, the overall runtime for simulating
a single Clifford gate is O(n2 + n|F|) since one can
memoize the updates to M required to compute each aj .
Another advantage of maintaining this invariant is that
the outcome of deterministic measurements (Section 2.2)
can be decided in time linear in n since it eliminates the
need to perform Gaussian elimination.
4
S IMULATING Q UANTUM C IRCUITS
S TABILIZER F RAMES
WITH
Let F be the stabilizer frame used to represent the nqubit state |Ψi. Following our discussion in Section 3,
any stabilizer or Pauli gate can be simulated directly via
frame rotation.
Suppose we want to simulate the action of T OFc1 c2 t ,
where c1 and c2 are the control qubits, and t is the
target. First, we decompose |Ψi into all four of its double
cofactors (Section 2) over the control qubits to obtain the
following equal superposition of orthogonal states:
Ψc1 c2 =00 + Ψc1 c2 =01 + Ψc1 c2 =10 + Ψc1 c2 =11
2
Here, we assume the most general case where all the
c1 c2 cofactors are non-empty. The number of states in the
superposition obtained could also be two (one control
qubit has an empty cofactor) or one (both control qubits
have empty cofactors). After cofactoring, we compute
the action of the Toffoli as,
|Ψi =
COFACTOR(F, c). This operation facilitates measurement
of stabilizer-state superpositions and simulation of nonClifford gates using frames. (Recall from Section 2 that
post-measurement states are also called cofactors.) Here,
c ∈ {1, 2, . . . , n} is the cofactor index. Let {|ψj i}kj=1 be the
stabilizer basis defined by F. Frame cofactoring maps
such a basis to { ψjc=0 , ψjc=1 }kj=1 . Therefore, after a
frame is cofactored, its size either remains the same
(qubit c was in a deterministic state and thus one of its
cofactors is empty) or doubles (qubit c was in a superposition state and thus both cofactors are non-empty).
We now describe the steps required to cofactor F.
where Xt is the Pauli gate (NOT) acting on target t. We
simulate Equation 2 with the following frame operations.
(An example of the process is depicted in Figure 4.)
1) Check M associated with F to determine whether
qubit c is a random (deterministic) state. (Section 2.2)
2a) Suppose qubit c is in a deterministic state. Since M is
maintained in row-echelon form (Invariant 1) no frame
updates are necessary.
2b) In the randomized-outcome case, apply the
measurement algorithm from Section 2.2 to M while
forcing the outcome to x ∈ {0, 1}. This is done
twice – once for each |xi-cofactor, and the row operations
performed on M are memoized each time.
3) Let σj be the (±)-phase vector in F that forms a pair
with aj . Iterate over each hσj , aj i pair and update its
elements according to the memoized operations.
4) For each hσj , aj i, insert a new phase vector-amplitude
pair corresponding to the cofactor state added to the
stabilizer basis.
Controlled-phase gates R(α)ct can also be simulated
using stabilizer frames. This gate applies a phase-shift
factor of eiα if both the control qubit c and target qubit
t are set. Thus, we compute the action of R(α)ct as,
Similar to frame rotation, the runtime of cofactoring is
linear in the number of phase vectors and quadratic in
the number of qubits. However, after this operation, the
number of phase vectors (states) in F will have grown by
a (worst case) factor of two. Furthermore, any state |Ψi
represented by F is invariant under frame cofactoring.
Example 6. Figure 4 shows how |Ψi = |000i + |010i +
|100i + |110i is cofactored with respect to its first two
qubits. A total of four cofactor states are obtained.
T OFc1 c2 t |Ψi = ( Ψc1 c2 =00 + Ψc1 c2 =01
+ Ψc1 c2 =10 + Xt Ψc1 c2 =11 )/2
(2)
1) COFACTOR(F, c1 ).
2) COFACTOR(F, c2 ).
3) Let Zj be the Pauli operator with a Z literal in its j th
position and I everywhere else. Due to Steps 1 and 2, the
matrix M associated with F must have two rows of the
form Zc1 and Zc2 . Let u and v be the indices of such
rows, respectively. For each phase vector σj∈{1,...,|F |} ,
if the u and v elements of σj are both -1 (i.e., if the phase
vector corresponds to the Ψc1 c2 =11 cofactor), flip the
value of element t in σj (apply Xt to this cofactor).
R(α)ct |Ψi = ( Ψct=00 + Ψct=01
+ Ψct=10 + eiα Ψct=11 )/2
(3)
Fig. 4. Simulation of T OFc1 c2 t |Ψi using a stabilizer-state
superposition (Equation 2). Here, c1 = 1, c1 = 2 and t = 3.
Amplitudes are omitted for clarity and the ± phases are
prepended to matrix rows. The X gate is applied to the
third qubit of the Ψc1 c2 =11 cofactor.
8
Equation 3 can be simulated via frame-based simulation
using a similar approach as discussed for T OF gates.
Let (a1 , . . . , a|F | ) be the decomposition of |Ψi onto F.
First, cofactor F over the c and t qubits. Then, for any
phase vector σj∈{1,...,|F |} that corresponds to the Ψct=11
cofactor, set aj = aj eiα . Observe that, in contrast to T OF
gates, controlled-R(α) gates produce biased superpositions. The Hadamard and controlled-R(α) gates are used
to implement the quantum Fourier transform circuit,
which plays a key role in Shor’s factoring algorithm.
Measuring F. Since the states in F are orthogonal, the
outcome probability when measuring F is calculated as
the sum of the normalized outcome probabilities of each
state. The normalization is with respect to the amplitudes
stored in a, and thus the overall measurement outcome
may have a non-uniform distribution. Formally, let |Ψi =
P
i ai |ψi i be the superposition of states represented by
F, the probability of observing outcome x ∈ {0, 1} upon
measuring qubit m is,
p(x)Ψ =
k
X
i=1
|ai |2 hψi | Pxm |ψi i =
k
X
|ai |2 p(x)ψi
i=1
Pxm
where
denotes the measurement operator in the
computational basis x as discussed in Section 2. The
outcome probability for each stabilizer state p(x)ψi is
computed as outlined in Section 2.2. Once we compute
p(x)Ψ , we flip a (biased) coin to decide the outcome
and cofactor the frame such that only the states that are
consistent with the measurement remain in the frame.
Prior work on simulation of non-Clifford gates using
the stabilizer formalism can be found in [1] where the
authors represent a quantum state as a sum of O(42dk )
density-matrix terms while simulating k non-Clifford
operations acting on d distinct qubits. In contrast, the
number of states in our technique is O(2k ) although
we do not handle density matrices and perform more
sophisticated bookkeeping.
Technology-dependent gate decompositions. Stabilizerframe simulation can be specialized to quantum technologies that are represented by libraries of primitive
gates (and, optionally, macro gates) with restrictions
on qubit interactions as well as gate parallelism and
scheduling. The work in [17] describes several primitive gate libraries for different quantum technologies
including quantum dot, ion trap and superconducting
systems. Such libraries can be incorporated into framebased simulation by decomposing the primitive gates
into linear combinations of Pauli or Clifford operators,
subject to qubit-interaction
constraints. For example, the
0
T = 10 eiπ/4
gate and its inverse, which are primitive
gates in most quantum machine descriptions,√ can be
simulated as Tt |Ψi = ( Ψt=0 + e±iπ/4 Ψt=1 )/ 2.
Multiframe simulation. Although a single frame is sufficient to represent a stabilizer-state superposition |Ψi,
one can sometimes tame the exponential growth of states
in |Ψi by constructing a multiframe representation. Such
a representation cuts down the total number of states
required to represent |Ψi, thus improving the scalability
of our technique. Our experiments in Section 6 show
that, when simulating ripple-carry adders, the number of
states in |Ψi grows linearly when multiframes are used
but exponentially when a single frame is used. To this
end, we introduce an additional frame operation.
COALESCE(F). One derives a multiframe representation
directly from a single frame F by examining the set
of phase vectors and identifying candidate pairs that
can be coalesced into a single phase vector associated
with a different stabilizer matrix. Since we maintain the
stabilizer matrix M of a frame in row-echelon form
(Invariant 1), examining the phases corresponding to
Zk -rows (Z-literal in k th column and I’s in all other
columns) allows us to identify the columns in M that
need to be modified in order to coalesce candidate pairs.
More generally, suppose hσr , σj i are a pair of phase
vectors from the same n-qubit frame. Then hσr , σj i is
considered a candidate iff it has the following properties:
(i) σr and σj are equal up to m ≤ n entries corresponding
to Zk -rows (where k is the qubit the row stabilizes), and
(ii) ar = id aj for some d ∈ {0, 1, 2, 3} (where ar and
aj are the frame amplitudes paired with σr and σj ). Let
e = {e1 , . . . , em } be the indices of a set of differing phasevector elements, and let v = {v1 , . . . , vm } be the qubits
stabilized by the Zk -rows identified by e. The steps in
our coalescing procedure are:
1) Sort phase vectors such that candidate pairs with
differing elements e are next to each other.
2) Coalesce candidates into a new set of phase vectors σ 0 .
3) Create a new frame F 0 consisting of σ 0 and matrix
CMC † , where C=CNOTv1 ,v2 CNOTv1 ,v3 · · · CNOTv1 ,vm Pdv1 Hv1 .
4) Repeat Steps 2–3 until no candidate pairs remain.
During simulation, we execute the coalescing operation
after the set of phase vectors in a frame is expanded
via cofactoring. Therefore, the choice of e (and thus
v) is driven by the Zk -rows produced after a frame is
cofactored. (Recall that cofactoring modifies M such that
each cofactored qubit k is stabilized by a ±Zk operator.)
The output of our coalescing operation is a list of n-qubit
frames F = {F10 , F20 , . . . , Fs0 } (i.e., a multiframe) that
together represent the same superposition as the original
input frame F. The size of the multiframe generated is
half the number of phase vectors in the input frame. The
runtime of this procedure is dominated by Step 1. Each
phase-vector comparison takes Θ(n) time. Therefore, the
runtime of Step 1 and our overall coalescing procedure
is O(nk log k) for a single frame with k phase vectors.
Example 7. Suppose we coalesce the frame F depicted
in Figure 5. Candidate pairs are hσ1 , σ2 i and hσ3 , σ4 i,
with e = {2} and e = {2, 3}, respectively. To obtain F1 ,
9
Fig. 5. Example of how a multiframe representation is
derived from a single-frame representation.
conjugate the second column of M by an H gate (Step 3),
which will coalesce hσ1 , σ2 i into a single phase vector
σ1 . Similarly, to obtain F2 , conjugate the second column
by H, then conjugate the second and third columns by
CNOT, which will coalesce hσ3 , σ4 i. Observe that no P
gates are applied since d = 0 for all pairs in a.
Candidate pairs can be identified even in the absence
of Zk -rows in an n-qubit M. By Corollary 3, one can
always find a stabilizer circuit C that maps M to the
matrix structure depicted in Figure 2a, whose rows are
all of Zk form. Several O(n3 )-time algorithms exist for
obtaining C [1], [3], [11]. We leverage such algorithms to
extend our coalescing operation as follows:
1)
2)
3)
4)
Find C that maps M to computational-basis form.
ROTATE(F, C).
{F10 , F20 , . . . , Fs0 } ← COALESCE(F).
ROTATE(Fi0 , C † ) for i ∈ {1, . . . , s}.
To simulate stabilizer, Toffoli and controlled-R(α)
gates using multiframe F, we apply single-frame operations to each frame in the list independently. For Toffoli
and controlled-R(α) gates, additional steps are required:
1) Apply the coalescing procedure to each frame and
insert the new “coalesced” frames in the list.
2) Merge frames with equivalent stabilizer matrices.
3) Repeat Steps 1–2 until no new frames are generated.
Orthogonality of multiframes. We introduce the following invariant to facilitate simulation of quantum
measurements on multiframes.
Invariant 2: The stabilizer frames that represent a superposition of stabilizer states remain mutually orthogonal during simulation, i.e., every pair of (basis) vectors
from any two frames are orthogonal.
Given multiframe F = {F1 , . . . , Fk }, one needs to consider two separate tasks in order to maintain Invariant 2.
The first task is to verify the pairwise orthogonality of
the states in F. The orthogonality of two n-qubit stabilizer states can be checked using the inner-product algorithm describe in [11], which takes O(n3 ) time. To improve this, we derive a heuristic based on Observation 2,
which takes advantage of similarities across the (canonical) matrices in F to avoid expensive inner-product computations in many cases. We note that, when simulating
quantum circuits that exhibit significant structure, F
contains similar stabilizer matrices with equivalent rows
(Pauli operators). Let M = {M1 , . . . , Mk } be the set of
n-qubit stabilizer matrices in F. Our heuristic keeps track
of a set of Pauli operators P = {P1 , P2 , . . . , Pk≤n }, that
form an intersection across the matrices in M.
Example 8. Consider the multiframe from Figure 5. The
intersection P consists of ZII (first row of both M).
By Observation 2, if two phase vectors (states) have
different entries corresponding to the Pauli operators in
P, then the states are orthogonal and no inner-product
computation is required. For certain practical instances,
including the benchmarks described in Section 6, we
obtain a non-empty P and our heuristic proves effective.
When P is empty or the phase-vector pair is equivalent,
we use the algorithm from [11] to verify orthogonality.
Therefore, in the worst case, checking pairwise orthogonality of the states in F takes O(n3 k 2 ) time for a
multiframe that represents a k-state superposition.
The second task to consider when maintaining Invariant 2 is the orthogonalization of the states in F =
{F1 , . . . , Fk } when our check fails. To accomplish this,
we iteratively apply the COFACTOR operation to each
frame in F in order to decompose F into a single frame.
At each iteration, we select a pivot qubit p based on
the composition of Pauli literals in the corresponding
column. We apply the COFACTOR(F, p) operation only if
there exists a pair of matrices in F that contain a different
set of Pauli literals in the pivot column.
1)
2)
3)
4)
Find pivot qubit p.
COFACTOR(Fi , p) for i ∈ {1, . . . , k}.
Merge frames with equivalent stabilizer matrices.
Repeat Steps 1–3 until a single frame remains.
Observe that the order of pivot selection does not impact the performance of orthogonalization. Each iteration
of the algorithm can potentially double the number of
states in the superposition. Since the algorithm terminates when a single frame remains, the resulting states
Fig. 6. Overall simulation flow for Quipu.
10
are represent by distinct phase vectors and are therefore
pairwise orthogonal. In the worst case, all n qubits are
selected as pivots and the resulting frame constitutes a
computational-basis decomposition of size 2n .
The overall simulation flow of our frame-based techniques is shown in Figure 6 and implemented in our
software package Quipu.
Example 9. Figure 7 depicts the main steps of the Quipu
simulation flow for a small non-Clifford circuit.
5
PARALLEL S IMULATION
Unlike other techniques based on compact representations of quantum states (e.g., using BDD data structures [26]), most frame-based operations are inherently
parallel and lend themselves to multi-threaded implementation. The only step in Figure 6 that presents a
bottleneck for a parallel implementation is the orthogonalization procedure, which requires communication
across frames. All other processes at both the single- and
multi-frame levels can be executed on large subsets of
phase vectors independently.
We implemented a multithreaded version of Quipu
using the C++11 thread support library. Each worker
thread is launched via the std::async() function.
Figure 8 shows our wrapper function for executing
calls to std::async(). The async_launch() function takes as input: (i) a frame operation (Func f),
(ii) a range of phase-vector elements defined by Iter
begin and Iter end, and (iii) any additional parameters (Params... p) required for the frame operation. Furthermore, the function returns a vector of
std::future – the C++11 mechanism for accessing the
result of an asynchronous operation scheduled by the
C++ runtime support system. The workload (number
of phase vectors) of each thread is distributed evenly
across the number of cores in the system (MTHREAD).
The results from each thread are joined only when
orthogonalization procedures are performed since they
require communication between multiple threads. This
is accomplished by calling the std::future::get()
function on each future. All Clifford gates and measurements are simulated in parallel.
6
E MPIRICAL VALIDATION
We tested a single-threaded and multi-threaded versions
of Quipu on a conventional Linux server using several
benchmark sets consisting of stabilizer circuits, quantum
ripple-carry adders, quantum Fourier transform circuits
and quantum fault-tolerant (FT) circuits.
We used a straightforward implementation of the
state-vector model using an array of complex amplitudes
to perform functional verification of: (i) all benchmarks
with < 30 qubits and (ii) Quipu output for such
benchmarks. We simulated such circuits and checked
for equivalence among the resultant states and operators [26]. In the case of stabilizer circuits, we used the
equivalence-checking method described in [1], [11].
Stabilizer circuits. We compared the runtime performance of single-threaded Quipu against that of CHP
using a benchmark set similar to the one used in [1].
We generated random stabilizer circuits on n qubits, for
n ∈ {100, 200, . . . , 1500}. The use of randomly generated
benchmarks is justified for our experiments because
(i) our algorithms are not explicitly sensitive to circuit
topology and (ii) random stabilizer circuits have been
considered representative [16]. For each n, we generated
the circuits as follows: fix a parameter β > 0; then
choose βdn log2 ne random unitary gates (CNOT, P or
H) each with probability 1/3. Then measure each qubit
a ∈ {0, . . . , n − 1} in sequence. We measured the number
of seconds needed to simulate the entire circuit. The
entire procedure was repeated for β ranging from 0.6 to
1.2 in increments of 0.1. Figure 9 shows the average time
needed by Quipu and CHP to simulate this benchmark
set. The purpose of this comparison is to evaluate the
overhead of supporting generic circuit simulation in
Quipu. Since CHP is specialized to stabilizer circuits, we
do not expect Quipu to be faster. When β = 0.6, the
simulation time appears to grow roughly linearly in n for
both simulators. However, when the number of unitary
gates is doubled (β = 1.2), the runtime of both simulators
grows roughly quadratically. Therefore, the performance
of both simulators depends strongly on the circuit being
simulated. Although Quipu is 5× slower than CHP, we
note that Quipu maintains global phases whereas CHP
does not. Nonetheless, Figure 9 shows that Quipu is
asymptotically as fast as CHP when simulating stabilizer
circuits that contain a linear number of measurements.
The multithreaded speedup in Quipu for non-Clifford
circuits is not readily available for stabilizer circuits.
Ripple-carry adders. Our second benchmark set consists
of n-bit ripple-carry (Cuccaro) adder [7] circuits, which
often appear as components in many arithmetic circuits
[18]. The Cuccaro circuit for n = 3 is shown in Figure
10. Such circuits act on two n-qubit input registers, one
ancilla qubit and one carry qubit for a total of 2(n + 1)
qubits. We applied H gates to all 2n input qubits in
order to simulate addition on a superposition of 22n
computational-basis states. Figure 11 shows the average
runtime needed to simulate this benchmark set using
Quipu. For comparison, we ran the same benchmarks
on an optimized version of QuIDDPro, called QPLite11 ,
specific to circuit simulation [26]. When n < 15, QPLite
is faster than Quipu because the QuIDD representing
the state vector remains compact during simulation.
However, for n > 15, the compactness of the QuIDD
is considerably reduced, and the majority of QPLite’s
11. QPLite is up to 4× faster since it removes overhead related to
QuIDDPro’s interpreted front-end for quantum programming.
11
Fig. 7. Example simulation flow for a small non-Clifford circuit (top left) using Quipu. The multiframes obtained are
pairwise orthogonal and thus no orthogonalization is required.
runtime is spent in non-local pointer-chasing and memory (de)allocation. Thus, QPLite fails to scale on such
benchmarks and one observes an exponential increase
in runtime. Memory usage for both Quipu and QPLite
was nearly unchanged for these benchmarks. Quipu
consumed 4.7MB on average while QPLite consumed
almost twice as much (8.5MB).
We ran the same benchmarks using single and multi-
template<class Func, class Iter, class... Params>
auto async_launch( Func f, const Iter begin, const Iter end,
Params... p )
-> vector< decltype( async(f, begin, end, p...) ) >
{
vector< decltype( async(f, begin, end, p...) ) > futures;
int size = distance(begin, end);
int n = size/MTHREAD;
futures.reserve(MTHREAD);
for(int i = 0; i < MTHREAD; i++)
{
Iter first = begin + i*n;
Iter last = (i < MTHREAD - 1) ? begin + (i+1)*n : end;
futures.push_back( async( f, first, last, p...) );
}
return futures;
}
H
•
•
|a0 i
|0i
|b1 i
H
•
•
|a1 i
H
|b2 i
H
|a2 i
|zi
H
|b0 i
•
•
H
• •
|s0 i
•
|a0 i
|0i
|s1 i
• •
|a1 i
• •
•
• •
•
•
|s2 i
• •
•
•
|a2 i
|z ⊕ s3 i
Fig. 10. Ripple-carry (Cuccaro) adder for 3-bit numbers
a = a0 a1 a2 and b = b0 b1 b2 [7, Figure 6]. The third qubit
from the top is an ancilla and the z qubit is the carry. The
b-register is overwritten with the result s0 s1 s2 .
frames. The number of states in the superposition grows
exponentially in n for a single frame, but linearly in n
when multiple frames are allowed. This is because T OF
gates produce large equal superpositions that are effectively compressed by our coalescing technique. Since
our frame-based algorithms require poly(k) time for
k states in a superposition, Quipu simulates Cuccaro
Fig. 8. Our C++11 template function for parallel execution
of the frame operations described in Section 4.
15
10
100
β = .6
β = .7
β = .8
β = .9
β= 1.0
β= 1.1
β = 1.2
CHP
60
40
5
0
200
80
β = .6
β = .7
β = .8
β = .9
β= 1.0
β= 1.1
β = 1.2
Quipu
20
400
600
800
0
1000 1200 1400 1600 200
Runtime (secs)
Avg. Runtime (secs)
80
20
QPLite
70
60
5
Quipu
50
zoomed-in
40
30
0
20
10
15
10
400
600
800 1000 1200 1400 1600
Number of qubits
Fig. 9. Average time needed by Quipu and CHP to
simulate an n-qubit stabilizer circuit with βn log n gates
and n measurements. Quipu is asymptotically as fast as
CHP but is not limited to stabilizer circuits.
Quadratic fit
0
5
10
15
20
25
n-bit Cuccaro adder (2n + 2 qubits)
Fig. 11. Average runtime and memory needed by Quipu
and QuIDDPro to simulate n-bit Cuccaro adders on an
equal superposition of all computational basis states.
Quantum Fourier transform (QFT) circuits. Our third
benchmark set consists of circuits that implement the nqubit QFT, which computes the discrete Fourier transform of the amplitudes in the input quantum state. Let
|x1 x2 . . . xn i, xi P
∈ {0, 1} be a computational-basis state
m
and x1,2,...,m = k=1 xk 2−k . The action of the QFT on
this input state can be expressed as:
400
Runtime (secs)
circuits in polynomial time and space for input states
consisting of large superpositions of basis states. On such
instances, known linear-algebraic simulation techniques
(e.g., QuIDDPro) take exponential time while Quipu’s
runtime grows quadratically (best quadratic fit f (x) =
0.5248x2 − 15.815x + 123.86 with R2 = .9986).
The work in [18] describes additional quantum arithmetic circuits that are based on Cuccaro adders (e.g.,
subtractors, conditional adders, comparators). We used
Quipu to simulate such circuits and observed similar
runtime performance as that shown in Figure 11.
Peak memory (MB)
12
350
300
QPLite
Quipu
250
200
150
QuipuMT
100
50
0
10
12
14
16
18
20
22
24
700
600
QPLite
Quipu
500
400
300
200
QuipuMT
100
0
10
12
14
16
18
20
22
24
n-qubit QFT circuit
Fig. 13. Average runtime and memory needed by Quipu
(single and multi-threaded) and QuIDDPro to simulate nqubit QFT circuits, which contain n(n + 1)/2 gates. We
used the |11 . . . 1i input state for all benchmarks.
plex amplitudes. Such non-compact data structures can
be streamlined to simulate most quantum gates (e. g.,
Hadamard, controlled-R(α)) with limited runtime overhead, but scale to only around 30 qubits due to poor
1
2iπ·xn−1,n
2iπ·xn
memory
scaling. Our results showed that Quipu was
|1i ⊗
|1i ⊗ |0i + e
|0i + e
|x1 . . . xn i = √
2n
approximately
3× slower than an array-based imple
2iπ·x1,2,...,n
mentation
when
simulating QFT instances. However,
· · · ⊗ |0i + e
|1i
(4)
such implementations cannot take advantage of circuit
The QFT is used in many quantum algorithms, notably structure and, unlike Quipu and QPLite, do not scale to
Shor’s factoring and discrete logarithm algorithms. Such instances of stabilizer and arithmetic circuits with > 30
circuits are composed of a network of Hadamard and qubits (Figures 9 and 11).
controlled-R(α) gates, where α = π/2k and k is the
distance over which the gate acts. The three-qubit QFT Fault-tolerant (FT) circuits. Our next benchmark set
circuit is shown in Figure 12. In general, the first qubit consists of circuits that, in addition to preparing encoded
requires one Hadamard gate, the next qubit requires a quantum states, implement procedures for performing
Hadamard and a controlled-R(α) gate, and each follow- FT quantum operations [9], [19], [23]. FT operations limit
ing qubit requires an additional controlled-R(α) gate. the propagation errors from one qubit in a QECC-register
Summing up the number of gates gives O(n2 ) for an (the block of qubits that encodes a logical qubit) to
n-qubit QFT circuit. Figure 13 shows average runtime another qubit in the same register, and a single faulty
and memory usage for both Quipu and QPLite on QFT gate damages at most one qubit in each register. One
instances for n = {10, 12, . . . , 20}. Quipu runs approx- constructs FT stabilizer circuits by executing each Clifimately 10× faster than QPLite on average and con- ford gate transversally12 across QECC-registers as shown
sumes about 96% less memory. For these benchmarks, in Figure 14. Non-Clifford gates need to be implemented
we observed that the number of states in our multiframe using a FT architecture that often requires ancilla qubits,
data structure was 2n−1 . This is because controlled-R(α) measurements and correction procedures conditioned on
gates produce biased superpositions (Section 4) that measurement outcomes. Figure 15 shows a circuit that
cannot be effectively compressed using our coalescing implements a FT-Toffoli operation [23]. Each line repreprocedure. Therefore, as Figure 13 shows, the runtime sents a 5-qubit register based on the DiVincenzo/Shor13
and memory requirements of both Quipu and QPLite code, and each gate is applied transversally. The state
√
grow exponentially in n for QFT instances. However, |cati = ( 0⊗5 + 1⊗5 )/ 2 is obtained using a stabilizer
Quipu scales to 24-qubit instances while QPLite scales subcircuit (not shown). The arrows point to the set of
to only 18 qubits. The multi-threaded version of Quipu gates that is applied if the measurement outcome is
exhibited roughly a 2× speedup and used a comparable 1; no action is taken otherwise. Controlled-Z gates are
amount of memory on a four-core Xeon server.
implemented as Hj CN OTi,j Hj with control i and target
We compared Quipu to a straightforward implemen- j. Z gates are implemented as P 2 .
tation of the state-vector model using an array of comWe implemented FT benchmarks for the half-adder
and full-adder circuits (Figure 16) as well as for computing f (x) = bx mod 15. Each circuit from Figure 17
•
•
|x2 i
|y0 i
H
implements f (x) with a particular coprime base value
•
|x1 i
|x0 i
H
R(π/2)
R(π/2)
H
R(π/4)
Fig. 12. The three-qubit QFT circuit.
|y1 i
|y2 i
12. In a transversal operation, the ith qubit in each QECC-register
interacts only with the ith qubit of other QECC-registers [12], [19], [23].
13. The DiVincenzo/Shor 5-qubit code functions successfully in the
presence of both bit-flip and phase-flip errors even if they occur during
correction procedures [9].
13
TABLE 3
Average time and memory needed by Quipu and QPLite to simulate several quantum FT circuits. The second
column shows the QECC used to encode k logical qubits into n physical qubits. Benchmarks with (∗ ) use the 5-qubit
DiVincenzo/Shor code [9] instead of the 3-qubit bit-flip code. We used the |00 . . . 0i input state for all benchmarks.
The top numbers from each row correspond to direct simulation of Toffoli gates (Section 4) and the bottom numbers
correspond to simulation via the decomposition from Figure 18. Shaded rows are Clifford circuits for mod-exp.
FAULT- TOLERANT
CIRCUIT
NUM .
QUBITS
QECC
[n, k]
toffoli∗
[15, 3]
45
halfadd∗
[15, 3]
45
fulladd∗
[20, 4]
80
2x mod15
[18, 6]
81
4x mod15∗
[30, 6]
30
7x mod15
[18, 6]
81
8x mod15
[18, 6]
81
11x mod15∗
[30, 6]
30
[18, 6]
81
[30, 6]
30
13x mod15
14x mod15∗
|xi
|yi
•
H
|xi
|yi
P
(a) Logical operation
CLIFF
H
H
P
P
NUM . GATES
& MEAS .
NON - CLIFF .
155
15
305
90
160
15
310
90
320
30
620
180
396
36
756
216
30
0
402
36
762
216
399
36
759
216
25
0
399
36
759
216
40
0
•
•
H
•
P
(b) Transversal execution
Fig. 14. Transversal implementation of a stabilizer circuit
acting on three-qubit QECC registers.
|0i
|0i
|0i
|cati
|cati
|cati
|xi
|yi
|zi
H
r
r
r
H
r
r
r
H
r
r
e ee
e e
Z
e
r
e e
H
r
r
e e
H
H
r
r
r
e
|xi
|yi
e |z ⊕ xyi
e r
Z e
6 6 6 Meas.
r
RUNTIME ( SECS )
QPLite
43.68
72.45
43.80
75.05
84.96
1173.48
4.81hrs
> 24hrs
0.01
11.25hrs
> 24hrs
11.37hrs
> 24hrs
0.02
11.28hrs
> 24hrs
0.02
Quipu
0.20
0.83
0.20
0.84
0.88
1.61
1.48
6.08
< 0.01
1.52
4.98
1.52
6.08
< 0.01
1.56
4.64
< 0.01
MEMORY
QPLite
98.45
137.02
94.82
137.03
91.86
139.44
11.85
118.17
6.14
12.41
134.05
12.48
135.18
6.14
11.85
135.23
6.14
(MB)
Quipu
12.76
12.78
12.76
12.78
12.94
13.52
12.96
14.23
12.01
13.29
14.77
13.29
14.77
12.01
12.25
14.62
12.01
MAX SIZE (Ψ)
SINGLE F
MULTI
2816
8192
2816
8192
2816
16384
22528
222
1
22528
222
22528
222
1
22528
222
1
F
32
32
32
32
32
32
64
64
1
64
64
64
64
1
64
64
1
more robust 5-qubit code in our larger benchmarks. Our
results in Table 3 show that Quipu is typically faster than
QPLite by several orders of magnitude and consumes
8× less memory for the toffoli, half-adder and full-adder
benchmarks. For FT benchmarks that consist of stabilizer circuits (shaded rows), the QuIDD representation
remains compact and utilizes half as much memory as
our frame representation.
Table 3 also shows that our coalescing technique is
very effective as the maximum size of the stabilizer-state
superposition is much smaller when multiple frames
are used. Since the total number of states observed is
relatively small, the multithreaded version of Quipu
exhibited similar runtime and memory requirements for
these benchmarks.
Meas.
r
Meas.
e
Meas.
e
Meas.
r H
Meas.
Fig. 15. Fault-tolerant implementation of a Toffoli gate.
b as a (2, 4) look-up table (LUT).14 The Toffoli gates
in all our FT benchmarks are implemented using the
architecture from Figure 15. Since FT-Toffoli operations
require 6 ancilla registers, a circuit that implements t
FT-Toffolis using a k-qubit QECC, requires 6tk ancilla
qubits. Therefore, to compare with QPLite, we used
the 3-qubit bit-flip code [19, Ch. 10] instead of the
14. A (k, m)-LUT takes k read-only input bits and m > log2 k
ancilla bits. For each 2k input combination, an LUT produces a
pre-determined m-bit value, e.g., a (2, 4)-LUT is defined by values
(1, 2, 4, 8) or (1, 4, 1, 4).
Technology-dependent circuits. Section 4 outlines how
Quipu supports primitive gate libraries, especially for
quantum-optical systems where Clifford gates are considered primitive [17]. Therefore, to simulate FT circuits
for photonic systems, it suffices to decompose T OF F
gates into sequences of Hadamard, CNOT and T gates
as shown in Figure 18. Table 3 reports simulations of
our FT benchmarks using such decompositions. Since
the total number of gates is larger, Quipu is roughly
|xi
|yi
|0i
•
•
•
|xi
|sumi
|ai
|bi
|zi
|carryi
|0i
(a) Half adder
•
•
•
•
•
•
|xi
|yi
|sumi
|carryi
(b) Full adder
Fig. 16. Adder circuits implemented in our benchmarks.
14
|x0 i
H
|x1 i
H
•
•
H
• •
H
•
•
H
H
•
•
H
•
•
H
• •
•
|x0 i
|x1 i
|0i
|y0 i
|0i
|y1 i
|0i
|y2 i
|y3 i
|0i
b=2
b=4
b=7
b=8
Fig. 17. Mod-exp with M = 15 implemented as (2, 4)LUTs [18] for several coprime base values. Negative
controls are shown with hollow circles.
4× slower as compared to direct simulation of T OF F
gates. QPLite takes > 24 hours to simulate several
of these benchmarks while Quipu takes only several
seconds since the majority of the gates introduced by
the decomposition from Figure 18 are Clifford gates.
7
new Clifford miters – linear combinations of Clifford operators that represent a specific quantum circuit. Clifford
miters can speed up formal verification by exploiting
similarities in circuits and the fast equivalence-checking
algorithms from [1], [11].
Section 4 described how Quipu incorporates quantum
machine descriptions in the form of primitive gate libraries to simulate technology-dependent circuits. This
method can benefit from new decompositions for library
gates into linear combinations of Pauli or Clifford operators. Such decompositions can be obtained on the fly
when simulating original gates one-by-one in sequence.
They can also be precomputed and used for a compiled
version of the original circuit, where scheduling can be
optimized for parallelism and architecture constraints.
Acknowledgements. This work was sponsored in part
by the Air Force Research Laboratory under agreement
FA8750-11-2-0043.
C ONCLUSIONS
In this work, we developed new techniques for
quantum-circuit simulation based on superpositions of
stabilizer states, avoiding shortcomings in prior work [1].
To represent such superpositions compactly, we designed a new data structure called a stabilizer frame. We
implemented stabilizer frames and relevant algorithms
in our software package Quipu. Current simulators
based on the stabilizer formalism, such as CHP, are limited to simulation of stabilizer circuits. Our results show
that Quipu performs asymptotically as fast as CHP on
stabilizer circuits with a linear number of measurement
gates, but simulates certain quantum arithmetic circuits
in polynomial time and space for input states consisting of equal superpositions of computational-basis
states. In contrast, QuIDDPro takes exponential time on
such instances. We simulated quantum Fourier transform and quantum fault-tolerant circuits with Quipu,
and the results demonstrate that our stabilizer-based
technique leads to orders-of-magnitude improvement in
runtime and memory as compared to QuIDDPro. While
our technique uses more sophisticated mathematics and
quantum-state modeling, it is significantly easier to implement and optimize. In particular, our multithreaded
implementation of Quipu exhibited a 2× speed up on a
four-core server.
Future Directions. The work in [27] describes an
equivalence-checking method for quantum circuits
based on the notion of a reversible miter – a counterpart of
miter circuits used in equivalence-checking of digital circuits. An attractive direction for future work is deriving
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
•
•
•
•
•
≡
H
•
T†
T
•
T†
T†
T
•
T†
T
P
H
Fig. 18. The Toffoli gate and its decomposition into onequbit and CNOT gates [19, Figure 4.9].
[20]
[21]
[22]
[23]
S. Aaronson, D. Gottesman, “Improved Simulation of Stabilizer
Circuits,” Phys. Rev. A, vol. 70, no. 052328 (2004).
D. Aharonov. “A Simple Proof that Toffoli and Hadamard are
Quantum Universal,” arXiv:0301040 (2003).
K. M. R. Audenaert, M. B. Plenio, “Entanglement on Mixed
Stabilizer States: Normal Forms and Reduction Procedures,” New
J. Phys., vol. 7, no. 170 (2005).
A. Barenco et al., “Approximate Quantum Fourier Transform and
Decoherence,” Phys. Rev. A, vol. 54, pp. 139–146 (1996).
O. Boncalo et al., “Using Simulated Fault Injection for Fault
Tolerance Assessment of Quantum Circuits,” Proc. Sim. Symp.,
pp. 213–220 (2007).
D. Coppersmith, “An Approximate Fourier Transform Useful in
Quantum Factoring,” arXiv:0201067 (2002).
S. A. Cuccaro et al., “A New Quantum Ripple-carry Addition
Circuit,” arXiv:0410184v1 (2004).
K. De Raedt et al., “Massively Parallel Quantum Computer Simulator”, Comp. Phys. Comm., vol. 176, no. 2 (2007).
D. P. DiVincenzo, P. W. Shor, “Fault-Tolerant Error Correction with
Efficient Quantum Codes”, Phys. Rev. Lett., vol. 77, no. 3260 (1996).
I. Djordjevic, Quantum Information Processing and Quantum Error
Correction: an Engineering Approach, Academic press (2012).
H. J. Garcia, I. L. Markov, A. W. Cross, “On the Geometry of
Stabilizer States,” Quant. Info. and Comp., vol. 14, no. 7–8 (2014).
D. Gottesman, “The Heisenberg Representation of Quantum
Computers,” arXiv:9807006v1 (1998).
L. Grover, “A Fast Quantum Mechanical Algorithm for Database
Search,” Symp. on Theory of Comp., pp. 212–219 (1996).
T. W. Hungerford, Algrebra, Springer (1974).
N. Isailovic et al., “Interconnection Networks for Scalable Quantum Computers,” Inter. Symp. on Comp. Arch., pp. 366–377 (2006).
E. Knill et al., “Randomized Benchmarking of Quantum Gates,”
Phys. Rev. A., vol. 77, no. 1 (2007).
C. Lin, A. Chakrabarti, N.K. Jha, “Optimized Quantum Gate
Library for Various Physical Machine Descriptions,” IEEE Trans.
on VLSI Sys., vol. 21, no. 11 (2013).
I. L. Markov, M. Saeedi, “Constant-optimized Quantum circuits
for Modular Multiplication and Exponentiation,” Quant. Info. and
Comp., vol. 12, no. 5 (2012).
M. A. Nielsen, I. L. Chuang, Quantum Computation and Quantum
Information, Cambridge University Press (2000).
K. M. Obenland, A. M. Despain, “A Parallel Quantum Computer
Simulator, ” arXiv:9804039 (1998).
B. Oemer, http://tph.tuwien.ac.at/∼oemer/qcl.html (2003).
M. Oskin, F. T. Chong, I L. Chuang, “A Practical Architecture
for Reliable Quantum Computers,” IEEE Computer, vol. 35, no. 1
(2002).
J. Preskill, “Fault Tolerant Quantum Computation,” Introduction
to Quantum Computation, World Scientific (1998). arXiv:9712048.
15
[24] P. Shor, “Polynomial-time Algorithms for Prime Factorization and
Discrete Logarithms on a Quantum Computer,” SIAM J. Comput,
vol. 26, no. 5 (1997).
[25] K. M. Svore, A. V. Aho, A. W. Cross, I. L. Chuang, I. L. Markov, “A
Layered Software Architecture for Quantum Computing Design
Tools,” IEEE Computer, vol. 39, no. 1 (2006).
[26] G. F. Viamontes, I. L. Markov, J. P. Hayes, Quantum Circuit
Simulation, Springer (2009).
[27] S. Yamashita, I. L. Markov, “Fast Equivalence-checking for Quantum Circuits” Quant. Info. and Comp., vol. 9, no. 9–10 (2010).
| 8cs.DS
|
Global algorithms for maximal eigenpair
Mu-Fa Chen
arXiv:1706.07584v1 [math.PR] 23 Jun 2017
Assisted by Yue-Shuang Li
(Beijing Normal University)
April 29, 2017
Abstract
This paper is a continuation of [5] where an efficient algorithm for
computing the maximal eigenpair was introduced first for tridiagonal
matrices and then extended to the irreducible matrices with nonnegative
off-diagonal elements. This paper introduces two global algorithms for
computing the maximal eigenpair in a rather general setup, including
even a class of real (with some negative off-diagonal elements) or complex
matrices.
2000 Mathematics Subject Classification: 60J60, 34L15
Key words and phrases. Maximal eigenpair, shifted inverse iteration, global algorithm.
1
Introduction
To compute the maximal eigenpair of the tridiagonal matrices with positive
sub-diagonal elements, an efficient algorithm was introduced [5; §3]. In the
tridiagonal case, the construction of the initials for the algorithm is explicit.
In some sense, the results are more or less complete (a modified algorithm,
Algorithm 17, is included in §4.4). Next, the algorithm was extended to the
general case in [5; §4] which is still efficient for tridiagonally dominant matrices.
Note that the initial v0 constructed in [5; §4.2] may not be efficient enough,
since the shape of the maximal eigenvector can be rather arbitrary, could be
quite far away from v0 constructed in [5; §4.2]. Thus, we are worrying about
the efficiency of the extended algorithm and moreover a global algorithm is
still missed in our general setup. This is the aim of this paper. In §3, a
part of the off-diagonal elements of the matrices are allowed to be negative.
We can even handle with some complex matrices. Let us concentrate on the
nonnegative matrices from now on, unless otherwise is stated.
By a shift if necessary, unless otherwise stated, we assume that the given
matrix A “ paij : 0 ď i, j ď N q is irreducible and nonnegative: aij ě 0. We
now state our algorithms. To guarantee the convergence of the iterations in
the paper, we assume that the matrix is irreducible having positive trace, or
equivalently,
An ą 0
for each n ě some n0 .
(1)
2
Mu-Fa Chen
We mention that in the present nonnegative case, the condition having positive trace is not serious, otherwise, simply adopt a shift as mentioned at the
beginning of [5].
Inřwhat follows, we omit, without mention time by time, the trivial case
that j aij ” constant m ą 0. Since then the maximal eigenpair of A becomes
pm, 1q, where 1 is the constant function having components 1 everywhere.
Recall that the choice of the initials is quite essential for the Rayleigh
Quotient Iteration (RQI), a special shifted inverse iteration. In general, it
seems no hope at the moment to have such explicit analytic formulas as used in
[5; §3]. Instead, as suggested in many textbooks, one may use other approach
to obtain in a numerical way the required initials, say use the power iteration
for instance. The last approach is safe, but rather slow as shown at the
beginning of [5]. This leads us to come back to the shifted inverse iterations
which is a fast cubic algorithm. The ratio of the numbers of iterations for
these two algorithms can be thousands. Now, a critical point is to avoid
the dangerous pitfalls, i.e., the region p0, ρpAqq, where ρpAq is the maximal
eigenvalue of A. The answer is given in part (1) of the next two algorithms.
At the moment, we are interesting in the generality and safety, do not take
care much about the convergence speed, perhaps, maybe some price we have
to pay here. We will see soon what happen in the next section.
Algorithm 1 (Specific Rayleigh quotient iteration)
Let A “ paij q be given.
(1) Define column vectors
wp0q “ p1, 1, . . . , 1q˚ ,
and set
v p0q “ wp0q
L?
N ` 1,
`
˘
z p0q “ max Awp0q i .
0ďiďN
(2) For given v :“ v pn´1q and z :“ z pn´1q , let w :“ wpnq solve the equation
As in step p1q, define v pnq
xpnq “ min
0ďjďN
pAwpnq qj
pnq
wj
,
pzI ´ Aqw “ v.
L?
“ w w˚ w . Next, define
y pnq “ max
0ďjďN
pAwpnq qj
pnq
wj
,
(2)
˚
z pnq “ v pnq Av pnq .
(3) If at some n ě 1, y pnq ´ xpnq ă 10´6 (or |z pnq ´ z pn`1q`| ă 10´6˘)(say!),
then stop the computation. At the same time, regard z pnq , v pnq as an
approximation of the maximal eigenpair.
The algorithm was presented in [5; §4.1: Choice I]. The simplest choice v0 is
reasonable in the sense that it enables us to cover the general case. We did not
pay enough attention on this algorithm since it looks less efficient. However,
as some examples will be illustrated below, this algorithm is actually rather
powerful. It is the place to state the main new algorithm of the paper.
3
Maximal Eigenpair
Algorithm 2 (Shifted inverse iteration) Everything is the same as in Algorithm 1, except y pnq and z pnq defined in parts p2q and p3q there are exchanged.
Moreover, the resulting z pnq (resp. xpnq ) is decreasing (resp. increasing) in n.
Let us repeat the sequences z pnq , y pnq and xpnq defined in Algorithm 2:
xpnq “ min
0ďjďN
pAwpnq qj
pnq
wj
,
˚
y pnq “ v pnq Av pnq ,
z pnq “ max
0ďjďN
pAwpnq qj
pnq
.
wj
It is obvious that
xpnq ď y pnq ď z pnq .
In general, Algorithm 1 is a little effective than Algorithm 2, saving one iteration for instance, but in Algorithm 2, each iteration is safe, never failed into
the pitfall. This is based on the following dual variational formula.
Proposition 3 [11; Theorem (8)] For a nonnegative irreducible matrix A, the
Collatz–Wielandt formula holds:
sup min
xą0 iPE
pAxqi
pAxqi
“ ρpAq “ inf max
.
xą0
iPE
xi
xi
Actually, suppose that we have wpn´1q ą 0 in Algorithm 2. Then by Proposition 3 and step (2) of Algorithm 2, we have z pn´1q ą ρpAq and then the
solution wpnq to the equation (2) should be positive: wpnq ą 0. Otherwise,
if z pn´1q ă ρpAq, then the solution wpnq is negative. This is the main reason
why we choose such a z pn´1q for each n ě 1 in Algorithm 2 and in the case of
n “ 0 in Algorithm 1 as our shift, avoiding the change of signs. Note that in
Algorithm 1 we adopt y pnq at each step n ě 1, hence the solution wpnq changes
its sign often. This seems dangerous because y pnq is located in the dangerous
region, but up to now, we have not meet trouble. Therefore, it is still regarded
as one of our two main algorithms. Certainly, if necessary, you can replace z pkq
pkq
defined in Algorithm 1 by z pkq “ max0ďjďN pAwpkq qj {wj for some k “ 1, 2, 3
or so. Or, simply use Algorithm 2 instead.
A careful comparison of Algorithm 1 and the powerful one introduced in
[5; §3] is delayed to the Appendix.
An easier way to see the efficiency of Algorithms 1 and 2 is comparing them
with the one given in [5; §4.2]. Suppose that we have used three iterations in
computing a model using the method introduced in [5; §4.2], this means on
the one hand we have solved the linear equations in three times. On the other
hand, we have solved three more times in advance to figure out the initials v p0q
and z p0q in terms of the triple pψ, h, µq. Altogether, we have solved six linear
equations. Or in other words, we have used 6 iterations in the computation for
the specific model. Thus, Algorithms 1 and 2 should be regarded as efficient
one if no more than 6 iterations are used in the computation for the same
model. As we will see soon, we are actually in such a successful situation.
4
Mu-Fa Chen
To conclude this section, we rewrite Algorithms 1 and 2 to a class of matrices with nonnegative off-diagonal elements and negative diagonal elements:
Q “ pqij q:
N
ÿ
qij ď 0, 0 ď i ď N.
qij ě 0, i ‰ j;
j“0
In this case, we are studying the maximal eigenpair of Q, or alternatively, the
minimal eigenpair of ´Q. To which,
the next two algorithms are devoted.
ř
Again, the trivial case that N
q
j“0 ij equals a constant is ignored throughout
the paper.
Algorithm 4 (Specific Rayleigh quotient iteration)
Let Q “ pqij q be given.
(1) Define column vectors
wp0q “ p1, 1, . . . , 1q˚ ,
v p0q “ wp0q
L?
N ` 1,
and set z p0q “ 0.
(2) For given v :“ v pn´1q and z :“ z pn´1q , let w :“ wpnq solve the equation
As in step p1q, define v pnq
xpnq“ min
0ďjďN
pp´Qqwpnq qj
pnq
wj
p´Q ´ zIqw “ v.
L?
“ w w˚ w . Next, define
, y pnq“ max
0ďjďN
pp´Qqwpnq qj
pnq
wj
(3)
˚
, z pnq“ v pnq p´Qqv pnq .
(3) If at some n ě 1, y pnq ´ xpnq ă 10´6 (or |z pnq ´ z pn`1q`| ă 10´6˘)(say!),
then stop the computation. At the same time, regard z pnq , v pnq as an
approximation of the minimal eigenpair.
Algorithm 5 (Shifted inverse iteration) Everything is the same as in Algorithm 4, except xpnq and z pnq defined in parts p2q and p3q there are exchanged.
Moreover, the resulting z pnq (resp. xpnq ) is increasing (resp. decreasing) in n.
Algorithms 4 and 5 are based on [5; Corollary 12], a corollary of Proposition
3.
2
Examples
To illustrate the power of the algorithms introduced in the last section, we
examine some typical examples in this section.
To go to practical computation for concrete models, our readers are urged
to prepare enough patience, one may have a large number of iterations since
the initials given in part (1) are quite rough.
The efficient application of Algorithm 1 was illustrated by [5; Examples
13–16]. To have a concrete comparison of the present algorithms with the one
introduced in [5; §4.2], let us consider a simple example.
5
Maximal Eigenpair
Example 6 [5; Example 21] Let
¨
˛
´3 2
0
1
0
˚ 4 ´7 3
‹
0
0
˚
‹
˚
‹.
Q“˚ 0
5 ´5
0
0
‹
˝ 10 0
‚
0 ´16
6
0
0
0
11 ´11 ´ b4
Corresponding to different b4 , the minimal eigenvalue λ0 of ´Q and its approximation are as follows. Here and in what follows, we stop at z p2q once the
outputs z pkq “ z p2q for every k ě 2.
Table 1. The outputs by Algorithm 1
b4
0.01
1
100
104
z p1q
0.000278773
0.0251531
0.191729
0.201695
z p2q
0.000278686
0.0245175
0.182822
0.195019
z p3q “ λmin p´Qq
0.182819
0.195015
Table 2. The outputs by Algorithm 2
b4
0.01
1
100
104
p1q
z
0.000278637
0.0241546
0.168776
0.179525
z p2q
0.000278686 “ λmin p´Qq
0.0245175
0.18275
0.194932
z p3q “ λmin p´Qq
0.182819
0.195015
While the outputs by the algorithm given in [5; §4.2] are the following.
Table 3. The outputs by the algorithm given in [5]
b4
0.01
1
100
z p1q
0.000278573
0.0236258
0.200058
z p2q
0.000278686
0.0245174
0.182609
z p3q “ λmin p´Qq
0.0245175
0.182819
These tables show that the three algorithms are more or less at the same
level of effectiveness. However, the first two are actually more economic since
the last one requires an extra work computing the initial v0 .
Comparing [5; Example 15] with the corrected version of [5; Example 20]
and its improvements given in [5; Tables 11, 12] (see the author’s homepage),
we see that the extended algorithm introduced in [5; §4.2] can be less efficient
than Algorithm 1, it has some limitation for general non-symmetrizable (nonsymmetric) matrices. We call a matrix A “ paji q is symmetrizable, if there
exists a positive measure pµi q such that
µi aij “ µj aji ,
i ‰ j.
6
Mu-Fa Chen
A simple necessary condition for the symmetrizability is
aij ą 0 ðñ aji ą 0,
i ‰ j.
Refer to [3; Chapter 7] and references within for the solution to the symmetrizability problem.
Let us start at a class of non-symmetrizable matrices which are taken from
the so-called single birth Q-matrix (cf. [3] and references within). Define
¨
˛
´1
1
0
0 ¨¨¨¨¨¨
0
0
˚ a1
‹
´a1 ´2
2
0 ¨¨¨¨¨¨
0
0
˚
‹
˚ a2
‹
0
´a2 ´3 3 ¨ ¨ ¨ ¨ ¨ ¨
0
0
˚
‹
Q“˚ .
‹ . (4)
.
.
.
..
..
.. ¨ ¨ ¨ ¨ ¨ ¨
˚ ..
‹
N ´2
0
˚
‹
˝aN ´1
0
0
0
¨ ¨ ¨ ´aN ´1 ´N `1
N ´1 ‚
aN
0
0
0 ¨¨¨¨¨¨
0
´aN ´N
For this matrix, we have computed several cases:
ak “ 1{pk ` 1q, ak ” 1, ak “ k, ak “ k2 .
Among them, the first one is hardest and is hence presented below.
Example 7 Let Q be defined by (4). For different N , the outputs of Algorithm 5 (equivalently, Algorithm 2) are as follows.
N
8
16
32
50
100
500
1000
5000
104
Table 4.
z p1q
0.276727
0.222132
0.187826
0.171657
0.152106
0.121403
0.111879
0.0947429
0.0888963
The outputs for different N by Algorithm 5
z p2q
z p3q
z p4q
z p5q
z p6q
0.427307 0.451902 0.452339
0.367827 0.399959 0.400910
0.329646 0.370364 0.372308 0.372311
0.311197 0.357814 0.360776 0.360784
0.287996 0.343847 0.349166 0.349197
0.247450 0.321751 0.336811 0.337186
0.233257 0.313274 0.334155 0.335009 0.335010
0.205212 0.293025 0.328961 0.332609 0.332635
0.194859 0.284064 0.326285 0.332113 0.332188
The last line shows that when N “ 104 , λmin p´Qq « 0.332188. If we use
the shifted matrix A “ Q ` mI, then ρpAq « 9999.67. From which, we get
λmin p´Qq « 104 ` 10´4 ´ 9999.67. Clearly, the second approach has a less
precise output. That is the main difference between Algorithms 1, 2 and 4, 5,
even though they are equivalent analytically.
It should be meaningful to have a comparison of the present results with
those produced by [5; §4.2]. The outputs listed in the table below come from
the algorithm without using δ1 defined in that section. For the outputs using
δ1 , one more iteration is needed for those N from 16 to 100 listed in the table.
7
Maximal Eigenpair
Table 5. The outputs for different
N
z p1q
8
0.450694
16
0.399520
32
0.371433
64
0.355722
100 0.349501
500 0.340666
1000 0.340871
5000 0.347505
104 0.352643
N by the
z p2q
0.452338
0.400910
0.372311
0.355940
0.349197
0.337185
0.335003
0.332536
0.331975
algorithm given in [5; §4.2]
z p3q
0.452339
0.337186
0.335010
0.332635
0.332188
Clearly, the general algorithm introduced in [5; §4.2] is efficient for this nonsymmetrizable model. We have seen that the present algorithms require more
iterations than the earlier one, this is reasonable since the computations of the
initials are excluded from the last table. Actually, the computations of the
last table cost double time than the previous one.
The next example is motivated from the classical branching process. Denote by ppk : k ě 0q a given probability measure with p1 “ 0. Let
¨
´1 p2 p3 p4 ¨ ¨ ¨ ¨ ¨ ¨
˚ 2p0 ´2 2p2 2p3 ¨ ¨ ¨ ¨ ¨ ¨
˚
˚ 0 3p0 ´3 3p2
¨¨¨
˚
.
.
.
.
..
Q“˚
..
..
..
˚ ..
.
˚
˚ ..
..
..
.
.
..
..
˝ .
.
.
0
0
0
0 ¨¨¨¨¨¨
˛
ř
p
k
kěN
ř
2 řkěN ´1 pk ‹
‹
3 kěN ´2 pk ‹
‹
‹,
..
‹
.
‹
‹
´pN ´ 1q pN ´ 1qp2 ‚
N p0
´N p0
pN ´1
2pN ´2
3pN ´3
..
.
In the original model, the state 0 is an absorbing one. Here we regard it as a
killing boundary. Hence it is ruled out from our
ř state space. Thus, the matrix
is defined on E :“ t1, 2, . . . , N u. Set M1 “ kPE kpk . When N “ 8, in the
subcritical case that M1 ă 1, with a little modification at 0, it is known that
the process generated by Q is ergodic, and is indeed exponentially ergodic
(cf. [8; Theorem 1.4 (iii)]). Hence the exponential convergence rate should be
positive. Otherwise, the process is not ergodic and so the convergence rate
should be zero.
From now on, fix
p0 “ α{2, p1 “ 0, p2 “ p2 ´ αq{22 , . . . pn “ p2 ´ αq{2n , ¨ ¨ ¨ ,
α P p0, 2q.
Then M1 “ 3p2 ´ αq{2 and hence we are in the subcritical case iff α P p4{3, 2q.
Example 8 Set α “ 1. Then the outputs of the approximation for the minimal eigenvalue of ´Q by Algorithm 2 (or 5) are as follows.
8
Mu-Fa Chen
Table 6. The outputs in the supercritical case
N
8
16
z p1q
0.0311491
0.00256281
z p2q
0.0346044
0.00260088
z p3q
0.0346310
When N ě 50, z p1q ă 10´6 . Hence, z pnq decays quite quick to zero when
N Ñ 8 (for n ě 2). This is reasonable since we are now away from the
subcritical region.
Example 9 Set α “ 7{4. We are now in the subcritical case and so the
maximal eigenvalue should be positive. We want to know how fast the local
maximal eigenvalue becomes stable (i.e., close enough to the converge rate at
N “ 8). Again, we adopt Algorithm 2 (or 5). Up to N “ 104 , the steps of the
iterations we need are no more than 6. To fasten the convergence, we adopt
a convex combination, as we did several times in [5]. Replace the original
z p0q “ max0ďjďN pAwp0q qj by
z p0q “ ξ min pAwp0q qj ` p1 ´ ξqpv p0q q˚ Av p0q .
0ďjďN
In view of the practice on N “ 8, we make the choice that ξ “ 0.69. Then the
outputs of the approximation of the minimal eigenvalue of ´Q for different N
are as follows.
Table 7. The outputs in the subcritical case
N
8
16
50
100
500
1000
5000
104
z p1q
0.637800
0.621430
0.609976
0.606948
0.604409
0.604082
0.603817
0.603784
z p2q
0.638153
0.625490
0.624052
0.623377
0.622116
0.621688
0.620838
0.620511
z p3q
z p4q
0.625539
0.624997
0.624991
0.624962
0.624944
0.62489
0.624861
0.625000
0.625000
0.625000
0.625000
0.625000
0.625000
From the above table, we see that for N varies from 8 to 104 , in each case, we
need at most 4 iterations only. The computation in each case costs no more
than one minute. Besides, starting from N “ 50, the final outputs are all the
same: 0.625, which then can be regarded as a very good approximation of the
maximal eigenvalue at infinity N “ 8.
Hopefully, we have already shown the power of our algorithms.
9
Maximal Eigenpair
3
A class of real or complex matrices
This section is out of the scope of [5] which depends heavily on probabilistic
idea. Thanks are given to the extended Perron–Frobenius theory ([10]–[12])
which makes this section possible.
First, we consider the real case. The special case that all off-diagonal
elements of A are negative has been treated above, using ´Q instead of A
here. Thus, we are now mainly interested in the case that a part of the offdiagonal elements are negative. Again, we are concentrated in the study of
the maximal eigenpair.
Proposition 10
Let A be a real matrix satisfying (1). Then Algorithms 1
and 2 are available.
Proof. By [10; Theorem 2.2], condition (1) implies that the matrix A possesses
the strong Perron–Frobenius property. Hence it has the maximal eigenvalue
ρpAq which is simple, positive and corresponds to a positive eigenvector. Besides, by [10; Theorem 2.6], the Collatz–Wielandt formula given in Proposition
3 holds. These facts are enough to use Algorithms 1 and 2. l
The next simple observation is helpful.
Lemma 11 Condition (1) holds iff
Ak ą 0
for k “ n0 , n0 ` 1, . . . , 2n0 ´ 1.
Proof. Given n ě n0 , write
n “ rn0 ` s
for some integer r ě 1 and s “ 0, 1, . . . , n0 ´ 1. If r “ 1, then the conclusion
holds by assumption. Otherwise, let r ě 2. Then express
n “ pr ´ 1qn0 ` pn0 ` sq.
It follows that
as required.
l
`
˘r´1 n `s
An “ An0
A 0 ą0
We now illustrate our algorithms by a simple example.
Example 12 [11; Example (7)] Let
¨
˛
´1 8 ´1
A “ ˝ 8 8 8 ‚.
´1 8 8
10
Mu-Fa Chen
Then
¨
˛
66 48 57
A2 “ ˝48 192 120‚ ą 0,
57 120 129
¨
˛
261 1368 774
A3 “ ˝1368 2880 2448‚ ą 0.
774 2448 1935
By Lemma 11, condition (1) holds with n0 “ 2. The eigenvalues of A are as
follows.
17.5124, ´7.4675, 4.95513.
The corresponding maximal eigenvector is
p0.486078, 1.24981, 1q
which is positive.
Here are the outputs of our algorithms. Both algorithms are started at
z p0q “ 24.
Table 8. The outputs for a matrix with more negative elements
n
1
2
3
z pnq : Algorithm 1
17.3772
17.5124
z pnq : Algorithm 2
18.5316
17.5416
17.5124
Next, we turn to study the complex case. Instead of (1), we assume that
` ˘
Re An ą 0
for n ě some n0 .
(5)
Certainly, as usual RepAq means the real part of a complex matrix A. This
condition is based on [12; Theorems 2.3 and 2.2], from which we know that A
has the maximal, simple, positive eigenvalue. Then we have a weak extension
of the Collatz–Wielandt formula as follows.
Proposition 13 [12; Theorems 2.3 and 2.4] Let Ak ‰ 0 for each k ě 1 and
RepAn q ě 0 for every large enough n. Then we have for each x ą 0
min
0ďjďN
pRepAqxqj
pRepAqxqj
ď ρpAq ď max
.
0ďjďN
xj
xj
Since for the complex conjugate x̄˚ of x, the quantity x̄˚ Ax may still be
complex, in view of this, Proposition 13 and the positivity of ρpAq by (5),
it seems not reasonable to use x̄˚ Ax{px̄˚ xq as a shift. In this sense, we do
not have a modified version of Algorithm 1. Fortunately, Algorithm 2 is still
meaningful.
Algorithm 14 (Shifted inverse iteration)
Assume (5).
11
Maximal Eigenpair
(1) Define column vectors
wp0q “ p1, 1, . . . , 1q˚ ,
and set
v p0q “ wp0q
L?
N ` 1,
`
˘
z p0q “ max RepAqwp0q i .
0ďiďN
(2) For given v :“ v pn´1q and z :“ z pn´1q , let w :“ wpnq solve the equation
pzI ´ Aqw “ v.
As in step p1q, define v pnq “ w
z pnq “ max
0ďjďN
(6)
L?
w
s˚ w . Next, define
pRepAqRepwpnq qqj
,
Repwpnq qj
y pnq “ pv̄ pnq q˚ Av pnq .
(3) If at some n ě 1, |y pn`1q ´`y pnq | ă 10˘´6 (say!), then stop the computation.
At the same time, regard y pnq , v pnq as an approximation of the maximal
eigenpair.
(
(
Note that in Algorithm 14, the sequence y pnq ně0 , but not z pnq ně0 ,
converges to ρpAq. To illustrate the use of the algorithm, we consider the
following example.
Example 15 [12; Example 2.1] Let
¨
0.75 ´ 1.125 i
A “ ˝ ´0.5 ´ i
2.75 ´ 0.125 i
0.5882 ´ 0.1471 i
2.1765 ` 0.7059 i
0.5882 ´ 0.1471 i
˛
1.0735 ` 1.4191 i
2.1471 ´ 0.4118 i ‚,
´0.9265 ` 0.4191 i
where the coefficients are all accurate, to four decimal digits. Then A has
eigenvalues
3, ´2 ´ i, 1 ` i
with maximal eigenvector
p0.408237, 0.816507, 0.408237q.
The outputs of Algorithm 14 are as follows.
Table 9. The outputs for a complex matrix
y p1q
3.03949 ´ 0.0451599 i
y p2q
3.00471 ´ 0.0015769 i
y p3q
3
12
4
4.1
Mu-Fa Chen
Appendix
Proof of the last assertion in Algorithm 2
Proposition 16 The sequence
z pnq “ max
0ďjďN
ˆ
pAwpnq qj
pnq
wj
resp. xpnq “ min
0ďjďN
pAwpnq qj
pnq
wj
˙
defined in Algorithm 2 is decreasing (resp. increasing ) in n.
Proof. Let w ą 0 and define
ρ̄ “ max
0ďjďN
pAwqj
.
wj
Then pAwqj ď ρ̄wj for every j. That is,
pAz wqj ď ρ̄z wj @j,
Az :“ A{z, ρ̄z “ ρ̄{z.
Since Az ě 0, it follows that
A
8
ÿ
n“0
˙
ˆ
8
8
8
ÿ
ÿ
ÿ
Anz w.
ρ̄Anz w “ ρ̄
Anz w ď ρ̄w `
Anz w ď A w ` ρ̄z
n“1
n“0
n“0
This means that
ApI ´ Az q´1 w ď ρ̄pI ´ Az q´1 w
since z ą ρpAq by assumption and then ρpAz q ă 1. Hence
max
0ďjďN
pAppI ´ Az q´1 vqqj
ď ρ̄,
ppI ´ Az q´1 vqj
?
v :“ w{ w˚ w.
Regarding w “ wpn´1q and v “ v pn´1q , this gives us
z pnq “ max
0ďjďN
pAwpnq qj
pnq
wj
ď ρ̄ “ max
0ďjďN
pAwpn´1q qj
pn´1q
wj
“ z pn´1q .
Here we have assumed that z pn´1q ą ρpAq, otherwise, the computation should
be finished at the step n ´ 1. We have thus proved the assertion on z pnq .
Dually, we have the assertion on xpnq . l
13
Maximal Eigenpair
4.2
Proof of the last assertion in Algorithm 5
.
Recall the sequence tz pnq u used in Algorithm 2 is given in Proposition 16.
Denote by tz̃ pnq u. Then, by the
ř relation of Q and A used in Algorithm 5:
A “ Q ` mI, where m “ maxi j aij . Hence
z p0q I ´ A “ ´Q ´ pm ´ z p0q qI.
This means not only z̃ p0q “ 0, but also
`
˘´1
`
˘´1 p0q
wp1q “ z p0q I ´ A v p0q “ ´ Q ´ z̃ p0q I
v “: w
rp1q ,
where w̃p1q is obtained by the first iteration of Algorithm 5. Furthermore,
similar to the proof of [5; Corollary 12], we have
z̃ p1q “ min
i
p´Qw
rp1q qi
p1q
w
ri
“ m ´ max
i
pAwp1q qi
Recursively, we obtain the required assertion.
4.3
p1q
wi
“ m ´ z p1q .
l
Comparison of Algorithms 1 and 4 with the one given in
[5; §3]
.
Since Algorithms 1 and 4 are equivalent, we need only to compare Algorithm 4 with the one given in [5; §3]. The main difference is their initial
pv p0q , z p0q q. Clearly, the initial v p0q used in [5; §3] is finer than the one used in
Algorithm 4. Hence, we need only to compare their z p0q .
Next, let v :“ v p0q be the initial vector used in [5; §3]. Denote by w be the
solution of the ordinary inverse iteration (that is the first step of Algorithm 4
or equivalently, Algorithm 1):
´Qw “ v.
Then
vj
p´Qwqj
“
“ IIj pvq´1 .
(7)
wj
pp´Qq´1 vqj
Here in the last equality of (7), we have used the first formula in the proof of
[5; Proposition 23]. Hence
inf
j
p´Qwqj
“ inf IIj pvq´1 .
j
wj
(8)
The right-hand side of (8) is just δ1´1 used in [5; §3] as its initial z p0q . The lefthand side of (8) should be positive, due to the inverse iteration algorithm, it is
certainly bigger than 0 used as the initial z p0q in Algorithm 4. In conclusion,
both initials used in [5; §3] are better than those used in Algorithm 4. This
completes the comparison of Algorithm 4 and the one given in [5; §3].
Naturally, this comparison leads to the next remark.
14
4.4
Mu-Fa Chen
Modification of the algorithm defined in [5; §3]
Step 1. By a shift if necessary, we may assume that we are given a matrix Q
having the form
¨
´pb0 ` c0 q
b0
0
˚
a
´pa
`
b
`
c
q
b
1
1
1
1
1
˚
˚
0
a2
´pa2 ` b2 ` c2 q
Q“˚
˚
..
..
..
˝
.
.
.
0
0
0
0
0
b2
..
.
¨¨¨
¨¨¨
¨¨¨
..
.
aN
´paN ` cN q
˛
‹
‹
‹
‹,
‹
‚
where ai ą 0, bi ą 0, ci ě 0 but ci ı 0. Note that the maximal eigenvalue of
Q is shifted from the original one but the corresponding eigenvector remain
the same.
Step 2. Following [5; §3], assume for a moment that some of ci pi “ 0, 1, . . . , N ´
1q is positive. Then, define
an ` cn
an
c0
, rn “ 1 `
´
,
1 ď n ă N,
b0
bn
bn rn´1
n´1
ź
h0 “ 1, hn “ hn´1 rn´1 “
rk ,
1 ď n ď N,
r0 “ 1 `
k“0
and additionally,
hN `1 “ cN hN ` aN phN ´ hN ´1 q.
We remark that in the special case that
c0 “ ¨ ¨ ¨ “ cN ´1 “ 0,
by induction, it is easy to check that
r0 “ ¨ ¨ ¨ “ rN ´1 “ 1
and hence
h0 “ ¨ ¨ ¨ “ hN “ 1.
Furthermore, hN `1 “ cN . Thus, in this special case, we simply ignore the
sequence thk u but replace cN by bN . Note that here we use all of the three
sequence pak q, pbk q and pck q given in Q but no extra thing. The role of the
sequence thk u is reducing the former case to the last special one and keep the
r
same spectrum, in terms of the H-transform Q:
r “ Diagphi q´1 Q Diagphi q.
Q
(9)
` ` ˘
˘
r “ ρpQq, Diagphi q´1 g .
The maximal eigenpair pρpQq, gq is transformed to ρ Q
15
Maximal Eigenpair
Step 3. In view of Step 2 above, it suffices to consider the following matrix
¨
˛
´b0
b0
0
0
¨¨¨
˚ a1 ´pa1 ` b1 q
‹
b1
0
¨¨¨
˚
‹
˚ 0
‹
a
´pa
`
b
q
b
¨
¨
¨
2
2
2
2
Q“˚
(10)
‹,
˚ ..
‹
..
..
..
..
˝ .
‚
.
.
.
.
0
0
0
aN
´paN ` bN q
where ai , bi ą 0. This step is changed from the original, where everything we
are working here is transfer into the original matrix Q rather than the simpler
one here. It seems a direct treatment of the present matrix Q is slightly
simpler.
Define the sequence pµi q as usual:
µ0 “ 1, µn “ µn´1
b0 b1 ¨ ¨ ¨ bn´1
bn´1
“
,
an
a1 a2 ¨ ¨ ¨ an
1 ď n ď N.
Next, define
ϕn “
and
δ1 “ max
0ďnďN
„
?
N
ÿ
1
,
µ b
k“n k k
ϕn
n
ÿ
µk
?
k“0
0 ď n ď N.
1
ϕk ` ?
ϕn
ÿ
(11)
3{2
µj ϕj
n`1ďjďN
.
(12)
Having these preparations at hand, we can now start our iterations.
Step 4. As in [5; §3], choose
wp0q “
?
v p0q “ wp0q {}wp0q }µ,2 ,
ϕ,
z p0q “ δ1´1 ,
(13)
where } ¨ }µ,2 denotes the L2 pµq-norm. Note that here in the non-symmetric
case, the use of the measure pµi q cannot be ignored since in this case, we are
based on, δk for instance, the L2 pµq setup.
Step 5. For given v “ v pn´1q and z “ z pn´1q , let w “ wpnq solve the linear
equation
p´Q ´ zIqw “ v
(14)
and then define v pnq “ w{}w}µ,2 . An explicit solution of this w is now available,
refer to [6; Algorithm 3].
Step 6. At the kth (k ě 1) iteration, in addition to the one pv pkq , ´Qv pkq qµ,2
used in [5; §3], one may also adopt z pkq “ δk´1 :
δk “ max
0ďiďN
1
pkq
vi
„
ϕi
i
ÿ
j“0
pkq
µ j vj
`
ÿ
i`1ďjďN
pkq
µj ϕj vj
.
(15)
16
Mu-Fa Chen
This is the main new point in the modified algorithm. Since [4; Theorems
2.4 (3), 3.2 (1) and (3.6)], we have
δk´1 ď λmin p´Qq ď pv pkq , ´Qv pkq qµ,2 for each k and n.
By [5; Proposition 23] and [4; Theorem 3.2 (1)], we have known that the
sequence tδk´1 u, deduced in the theorem just cited using the approximating
eigenvectors obtained by the ordinary inverse iteration (without shift), is increasing to λmin p´Qq. It should be clear that the present sequence tδk´1 u produced by the advanced shifted inverse iteration should converge to λmin p´Qq
more faster. Thus the new z pkq pk ě 1q not only avoids the dangerous region but may also accelerate the convergence of the algorithm. Certainly, the
computation of δk needs more work than the one of pv pkq , ´Qv pkq qµ,2 .
The use of the quantity (15) is motivated from the remark above on “Comparison of Algorithms 1 and 4 with the one given in [5; §3]”. The formula (15)
is a corollary of [4; Theorem 2.4 (3)] which depends on the form (10) of Q. For
general Q such the one in Step 1, we do not have an analog of [4; Theorem
2.4 (3)], and so (15) is not applicable in such a general situation.
Step 7. To go back to the original matrix A, denote its maximal eigenpair by
pρpAq, gq. Recall that the matrix
řQ at the beginning is obtained from A by a
shift: Q “ A ´ mI, m :“ maxi j aij . Let pz, vq be the output from the last
iteration in Step 6. Then we have
ρpAq « m ´ z,
g « Diagphi qv.
(16)
We now summery the above discussions as a modified algorithm.
Algorithm 17 For tridiagonal matrix, the Step 1–Step 7 above consist a modified algorithm of the one introduced in [5; §3].
We are now ready to study a randomly chosen example, introduced to
the author by Tao Tang, to justify the power of our algorithms and also to
compare their efficiency.
Example 18 Let
¨
˛
2.334 0.9962
0
0
0
0
˚0.5142 2.6725 0.1111
0
0
0 ‹
˚
‹
˚ 0
0.2115 2.263 0.1405
0
0 ‹
˚
‹.
A“˚
0
0.8442 2.8457 0.7595
0 ‹
˚ 0
‹
˝ 0
0
0
0.2347 2.2257 0.0781‚
0
0
0
0
0.9837 2.1582
Then the eigenvalues of A are
3.26753, 3.16247, 2.40182, 2.12632, 1.80416, 1.73679.
The outputs of our algorithms are given in the table below.
17
Maximal Eigenpair
Table 10. Comparison of four algorithms
Algorithm
Algorithm 1
Algorithm 2
Algorithm 17a
Algorithm 17b
z p1q
3.30193
3.64033
3.2618
3.27947
z p2q
3.26737
3.32623
3.26752
3.2685
z p3q
3.26754
3.26937
3.26753
3.26754
z p4q
3.26753
3.26756
z p5q
3.26753
3.26753
where the algorithms in the last two lines mean that
Algorithm 17a: take z pkq “ pv pkq , ´Qv pkq qµ,2 for each k ě 1.
Algorithm 17b: take z pkq “ δk´1 defined by (15) for each k ě 1.
Proof. To apply Algorithm 17, take m “ 4.4494. Then Q “ A ´ mI:
¨
˛
´2.1154 0.9962
0
0
0
0
˚ 0.5142 ´1.7769 0.1111
‹
0
0
0
˚
‹
˚
‹
0
0.2115
´2.1864
0.1405
0
0
‹.
Q“˚
˚
‹
0
0
0.8442 ´1.6037 0.7595
0
˚
‹
˝
0
0
0
0.2347 ´2.2237 0.0781 ‚
0
0
0
0
0.9837 ´2.2912
We have h “ p2.12347, 29.3339, 453.284,
of Q becomes
¨
´2.1154
2.1154
0
˚0.242151 ´1.7769
1.53475
˚
˚
0
0.0153104 ´2.1864
r
Q“˚
˚
0
0
0.0546316
˚
˝
0
0
0
0
0
0
924.514, 24961q. The H-transform
˛
0
0
0
‹
0
0
0
‹
‹
2.17109
0
0
‹.
‹
´1.6037
1.54907
0
‹
0.115072 ´2.2237 2.10863 ‚
0
0.0364346 ´2.2912
r and
Then we are ready to use Algorithm 17 for the maximal eigenpair of Q
finally return to the one for A by (16). l
To explain the word “modified” in detail, we transfer Algorithm 17 to the
one presented in [5; §3]. To do so, we keep the notation Q, µ, ϕ, δ1 and so
on used in [5; §3], but add superscript r to those notation used in Steps 3,
4 above. Let µ̃ “ h2 µ (i.e., µ̃i “ h2i µi ). Then, as mentioned in [5; §5], the
2
2
˜
`mapping f Ñ f :“ ˘f {h gives us not only an isometry from L2 pµq to L pµ̃q
r on
i.e., }f }µ,2 “ }f˜}µ̃,2 , and then also an isospectrum of Q on L pµq and Q
2
L pµ̃q:
r f˜qµ̃ ,
pf, Qf qµ “ pf˜, Q
}f }µ,2 “ 1.
Now, from L2 pµ̃q to L2 pµq, we have
ϕ̃n “
N
ÿ
1
1
Ñ
“ ϕn ,
h h µ b
k“n k k`1 k k
k“n µ̃k b̃k
N
ÿ
0 ď n ď N.
18
Mu-Fa Chen
Here the transform µ̃k b̃k Ñ hk hk`1 µk bk for each k ď N ´ 1 is regular, except
˘´1
`
, where b̃N is actually the element c̃N which
the last term in the sum µ̃N b̃N
r and hN `1 and bN are specified in [5;
is obtained from the transform Q Ñ Q,
§3] to make the unified expression in the second sum. We mention here that
hN `1 is the original paper [5] should be replaced by
hN `1 “ cN hN ` aN phN ´ hN ´1 q
since the sequence pci q used in [5] and [7] have different sign. Next,
„
n
ÿ
a ÿ
a
1
3{2
µ̃j ϕ̃j
ϕ̃n
µ̃k ϕ̃k ` ?
δ̃1 “ max
0ďnďN
ϕ̃n n`1ďjďN
k“0
„
n
ÿ
? ÿ
1
2?
2 3{2
Ñ δ1 “ max
ϕn
µk hk ϕk ` ?
µj hj ϕj .
0ďnďN
ϕn n`1ďjďN
k“0
At the same time,
`
˘
r ´ z̃I w̃ “ ṽ
´Q
`
˘
ðñ ´ Diagphq´1 Q Diagphq ´ z̃I w̃ “ ṽ
ðñ p´Q ´ z̃IqDiagphqw̃ “ Diagphqṽ
ðñ p´Q ´ zIqw “ v.
Here in the last line, z̃ is replaced by z, this is due to the isospectrum: an
r is also the one of ´Q. The fact that
lower bound of the spectrum of ´Q
Diagphqw̃ “ w comes from the definition of our mapping f Ñ f˜. Finally, since
the isometry, we have }w}µ,2 “ }w̃}µ̃,2 . We have thus deduced the algorithm
presented in [5; §3] from the modified one.
4.5
Modification of the algorithm introduced in [5; §4.2]
In parallel to §4.4, we may introduce a modification of the algorithm presented
in [5; §4.2]. The main idea is: once we obtain the function h, it can be ignored
r defined in (9) instead of the original
since we can use the general transform Q
Q to continue the procedure of the algorithm constructed in [5; §4.2]. Since
this modification is only a mimic of the one for tridiagonal matrix (§4.4),
something may be lost. For instance, the sequence tδk´1 u formally defined by
(15) may no longer be the lower bound of λmin p´Qq, one has to take care in
practice.
To conclude this paper, we remark some possible extension of the algorithms given here to a more general setup. For a larger class of Markov
generators, the algorithms are meaningful. Actually, the Perron–Frobenius
property as well as the the Collatz–Wielandt formula have been generalized
by a number of authors. In particular, the part of the Collatz–Wielandt formula used in Algorithm 5 as z pnq was extended by [9; ψ2 pV q in the Theorem].
Maximal Eigenpair
19
See also [13; (1.1) i) and §2] and more recently, [1; Theorem 2.1]. Note the
difference: we are working on λmin p´Lq here rather than λmax pLq in the cited
papers.
In the nonlinear case, the shifted inverse iteration (Algorithms 2 or 5)
is more essential, actually Algorithm 1 may no longer be applicable since
equation (2) often has no real solution. This point is illustrated in [6] where
the shift is based on a generalization of (15). In view of [2; Theorem 2.3 and
Corollary 2.5], it seems that Algorithm 2 and its variations could be applied
to a more general setup.
Acknowledgments
The author thanks Ms Yue-Shuang Li for her assistance in
computing the large matrices using MatLab, and also pointed out the error on hN `1
mentioned in §4.4. The author also acknowledges Mr Xu Zhu for constructing Example 18 which leads us to find out the error just mentioned. Research supported in
part by National Natural Science Foundation of China (No. 11626245), the project
from the Ministry of Education in China, and the Project Funded by the Priority
Academic Program Development of Jiangsu Higher Education Institutions.
References
[1] Arapostathis, A., Borkar, V.S. and Kumar, K.S. (2016). Risk–sensitive control
and an abstract Collatz–Wielandt formula. J. Theor. Probab. 29(4), 1458–1484.
[2] Chang, K.C. (2014). Nonlinear extensions of the Perron–Frobenius theorem and
the Krein–Rutman theorem. J. Fixed Point Theory Appl. 15, 433–457.
[3] Chen, M.F. (2004). From Markov Chains to Non-Equilibrium Particle Systems.
World Scientific, Singapore, 2nd Ed. (1st Ed., 1992).
[4] Chen, M.F. (2010). Speed of stability for birth–death processes. Front. Math.
China 5(3), 379–515.
[5] Chen, M.F. (2016). Efficient initials for computing the maximal eigenpair. Front.
Math. China 11(6): 1379–1418. See also volume 4 in the middle of the author’s
homepage:
http://math0.bnu.edu.cn/˜chenmf
[6]
[7]
[8]
[9]
[10]
A package based on the paper is available on CRAN now. One may check it
through the link:
https://cran.r-project.org/web/packages/EfficientMaxEigenpair/index.html
Chen, M.F. (2017) Efficient algorithm for principal eigenpair of discrete pLaplacian. Preprint.
Chen, M.F. and Zhang, X. (2014) Isospectral operators. Commu Math Stat 2,
17–32.
Chen, R.R. (1997). An Extended Class of Time-Continuous Branching Processes.
J. Appl. Probab. 34(1), 14-23
Donsker, W.D. and Varadhan, S.R.S. (1975). On a variational formula for the
principal eigenvalue for operators with maximum principle. Proc. Natl. Acad.
Sci. 72(3), 780–783.
Noutsos, D. (2006). On Perron-Frobenius property of matrices having some negative entries. Linear Algebra Appl. 412, 132–153.
20
Mu-Fa Chen
[11] Noutsos, D. (2008). Perron Frobenius theory and some extensions.
http://www.pdfdrive.net/perron-frobenius-theory-and-some-extensionse10082439.html
[12] Noutsos, D. and Varga, R.S. (2012). On the Perron–Frobenius theory for complex
matrices. Linear Algebra and its Applications 437, 1071–1088.
[13] Sheu, S.J. (1984). Stochastic control and principal eigenvalue. Stochastics 11(3–
4), 191–211.
Mu-Fa Chen
School of Mathematical Sciences, Beijing Normal University, Laboratory of Mathematics and Complex Systems (Beijing Normal University), Ministry of Education,
Beijing 100875, The People’s Republic of China.
E-mail: [email protected]
Home page: http://math0.bnu.edu.cn/˜chenmf/main eng.htm
| 10math.ST
|
arXiv:1712.00519v4 [math.PR] 4 Jan 2018
An Elementary Analysis of the Probability
That a Binomial Random Variable
Exceeds Its Expectation
Benjamin Doerr
Laboratoire d’Informatique (LIX)
École Polytechnique
Palaiseau
France
January 8, 2018
Abstract
We give an elementary proof of the fact that a binomial random variable
X with parameters n and 0.29/n ≤ p < 1 with probability at least 1/4
strictly exceeds its expectation. We also show that for 1/n ≤ p < 1 − 1/n,
X exceeds its expectation by more than one with probability at least 0.0370.
Both probabilities approach 1/2 when np and n(1 − p) tend to infinity.
1
Introduction
Let X be a random variable following
a binomial distribution with parameters n
and p, that is, we have Pr[X = i] = ni pi (1 − p)n−i for all i ∈ [0..n]. Then, apart
from maybe extreme cases, it seems very natural that with reasonable probability
X is at least its expectation E[X] or even exceeds it. Surprisingly, and despite
the fact that such statements are very important in the machine learning literature, only very recently rigorous proofs of such statements appeared. We refer to
Greenberg and Mohri [GM14] for a detailed discussion on the previous lack of such
results.
Prior to the work of Greenberg and Mohri, apart from general bounds like
those in Slud [Slu77], apparently only a result of Rigollet and Tong [RT11] was
known. This result is stated as Pr[X ≥ E[X]] ≥ min{p, 41 } for all p ≤ 21 in the
1
paper (Lemma 6.4), but the proof shows the stronger statement
Pr[X ≥ E[X]] ≥
1
4
p
if p ∈ [ n1 , 12 ]
if p < n1 .
(1)
The main work in the proof is showing another interesting result, namely that for
all k ∈ [2.. n2 ] one has
) ≥ k].
Pr[Bin(n, nk ) ≥ k + 1] ≥ Pr[Bin(n, k−1
n
(2)
The proof of this result uses a connection between binomial distributions and order
statistics of uniform distributions (to be found in Section 7.2 of the second volume
of Feller [Fel71]) and then proceeds by showing the inequality
k
Z
0
k−1
n
k−1
t
(1 − t)
n−k
dt ≤ (n − k)
Z
k
n
0
tk (1 − t)n−k−1 dt.
It is not clear how to extend (1) to p > 21 . Note that neither (2) nor this
equation with the inequality reversed are true for all k ∈ [ n2 ..n − 1]. Hence the
following relatively recent result of Greenberg and Mohri appears to be the first
one treating the problem in full generality.
Lemma 1 (Greenberg and Mohri [GM14]). Let n ∈ N and
X ∼ Bin(n, p). Then
1
n
< p ≤ 1. Let
Pr[X ≥ E[X]] > 14 .
This result has found applications not only in machine learning, but also in
randomized algorithms, see, e.g., [KKK16, BCN+ 17, MM17]. While the result is
very simple, the proof is not and uses the Camp-Paulson normal approximation
to the binomial cumulative distribution function.
Via a different, again non-elementary proof technique, using among others the
hazard rate order and the likelihood ratio order of integer-valued distributions, the
following result was shown by Pelekis and Ramon [PR16].
Lemma 2 (Pelekis and Ramon [PR16]). Let n ∈ N and
X ∼ Bin(n, p). Then
q
1
n
≤ p ≤ 1 − n1 . Let
np(1 − p)
1
Pr[X ≥ E[X]] ≥ √ q
.
2 2 np(1 − p) + 1 + 1
2
1
1
0.5
0.5
0
0
0.2
0.4
0.6
0.8
1
0
0
0.2
p
0.4
0.6
0.8
1
p
Figure 1: Comparison of the previous bounds of [GM14] (Lemma 1, sparsely dotted
lines) and [PR16] (Lemma 2, dashed lines), of our bound (equation (3), solid lines),
and of the true value for Pr[Bin(n, p) > np] (dotted lines) for n = 10 (left) and
n = 100 (right). To increase the readability, for n = 100 the true value is only
depicted at the local minima {0, n1 , n2 , ..., 1}.
Lemma 2 improves the bound of Lemma 1 when np(1 − p) > 8, which in
particular requires n > 32 and E[X] = np > 8. It however never gives a bound
better than 2√1 2 ≈ 0.3536.
In this work, we show that also truly elementary arguments give interesting
results for this problem. We prove in Lemma 8 that for n1 ≤ p < 1 and k := ⌊np⌋,
we have
1
Pr[X > E[X]] > −
2
s
n
.
2πk(n − k)
(3)
This bound is not perfectly comparable to the previous, but Figure 1 indicates
that it is often superior. It has the particular advantage that it tends to 12 when
np and n(1 − p) tend to infinity. Our bound does not immediately imply the
1
bound of Greenberg and Mohri [GM14], however elementary analyses of a few
4
“small cases” suffice to obtain in Theorem 10 that Pr[X > E[X]] ≥ 14 for all
0.2877 n1 ≈ ln( 43 ) n1 ≤ p < 1. The strict version Pr[X > E[X]] > 41 of the claim is
also valid except when n = 2 and p = 12 .
We also show that for n1 ≤ p < 1 − n1 , X exceeds its expectation by more than
one with probability at least 0.0370, again with better bounds available when np
and n(1 − p) are larger, see Theorem 12. Such a statement was recently needed in
the analysis of an evolutionary algorithm (in the proof of Lemma 3 of the extended
version of [DGWY17]).
3
2
Preliminaries
All notation we shall use is standard and should need not much additional explanation. We denote by N := {1, 2, . . . } the positive integers. For intervals of
integers, we write [a..b] := {x ∈ Z | a ≤ x ≤ b}. We use the standard definition
00 := 1 (and not 00 = 0).
It is well-known that (1 − 1r )r is monotonically increasing and that (1 − 1r )r−1 is
monotonically decreasing in r (and that both converge to 1e ). We need two slightly
stronger statements in this work (Lemma 3 (a) and (c)).
Lemma 3. (a) For all α ≥ 0, the expression (1 − x1 )x−0.5+α is increasing for
x ≥ 1.
(b) For all α ≥ 0, the expression (1 + x1 )x+0.5+α is decreasing for x > 0.
1
(c) The expression (1 − x1 )x + (1 − x1 )x−1 = 2(1 − 2x
)(1 − x1 )x−1 is decreasing for
x ≥ 1.
Proof. To prove the first part, it suffices to show that f (x) := (1 − x1 )x−0.5 is
increasing for x ≥ 1. It is obvious that 0 = f (1) < f (x) for all x > 1, so we can
concentrate on the case x > 1. We show that ln(f (x)) is increasing for x > 1.
Using the series expansion of the natural logarithm, we compute ln(f (x)) = (x −
P∞
P∞ 1
P∞
1
1
1 P∞
1
1
i−1 1
)
ln(1−
)
=
(x−
)
)
=
)+
=
−1−
(−
(−
i
i
i
i=1
i=0
i=1
i=1 2i(i+1) xi ,
2
x
2
ix
(i+1)x
2ix
which is a sum of constant and increasing functions.
The second claim follows from noting that the reciprocal of our expression,
1 (x+1)−0.5+α
1
)
, is increasing for x ≥ 0 by the first part.
= (1 − x+1
(1+ x1 )x+0.5+α
The third claim follows along similar arguments as the first one. Since f (x) =
1
2(1− 2x
)(1− x1 )x−1 is continuous in x = 1, it suffices to show the claim for x > 1. For
1
x > 1, we regard ln( 12 f (x)) and compute ln( 12 f (x)) = ln(1− 2x
)+(x−1) ln(1− x1 ) =
P∞
1
−1 + i=1 ( i+1
− 21i ) 1i x1i , which is a sum of constant and decreasing functions.
From the following version of Stirling’s formula we obtain an estimate for binomial coefficients and from that an estimate for Pr[Bin(n, nk ) = k].
Theorem 4 (Robbins [Rob55]). For all n ∈ N,
√
n! = 2πn( ne )n Rn ,
1
1
where 1 < exp( 12n+1
) < Rn < exp( 12n
) < 1.08690405.
Corollary 5. For all n ∈ N and k ∈ [1..n − 1],
!
1
n
=√
k
2π
s
n
k(n − k)
4
n
k
!k
n
n−k
!n−k
Rnk ,
f (x) = (1 − x1 )x−1
f (x) = 12 (1 − x1 )x−1 + 12 (1 − x1 )x
f (x) = (1 − x1 )x−0.5
f (x) = (1 − x1 )x
f (x)
0.8
0.6
0.4
0.2
2
4
6
x
8
10
Figure 2: Plots related to Lemma 3.
1
1
where 0.88102729... = exp(− 16 + 25
) ≤ exp(− 12k
−
1
1
1
exp(− 12k+1 − 12(n−k)+1 + 12n ) < 1.
1
12(n−k)
Lemma 6. Let n ∈ N and k ∈ [1..n − 1]. Then Pr[X = k] <
Proof. With Corollary 5 we estimate
Pr[X = k] ≤ Pr[Y = k]
!k
!
k
k
n
1−
=
n
n
k
s
1
n
.
<√
2π k(n − k)
+
1
)
12n+1
√1
2π
q
< Rnk <
n
.
k(n−k)
!n−k
A simple coupling argument establishes the natural fact that a binomial distribution with smaller p-value is dominated (in a strict sense) by one with larger
p-value. Very similar results were used and proven also in [GM14] (Lemma 1)
and [PR16] (Lemma 2.4), however, with more complicated and less intuitive proofs
(via differentiating the function p 7→ Pr[Bin(n, p) ≥ np] in [GM14] and via arguing
that Bin(n, p) is smaller than Bin(n, q) in the likelihood order in [PR16]).
Lemma 7. Let n ∈ N and 0 ≤ p < q ≤ 1. Let X ∼ Bin(n, p) and Y ∼ Bin(n, q).
Then for all k ∈ [0..n], Pr[X ≥ k] < Pr[Y ≥ k].
Proof. Let R1 , . . . , Rn be independent random variables uniformly distributed in
[0, 1). Let
X̃ := |{i ∈ [1..n] | Ri < p}|,
5
Ỹ := |{i ∈ [1..n] | Ri < q}|.
By construction, we have X̃ ≤ Ỹ and thus Pr[X̃ ≥ k] ≤ Pr[Ỹ ≥ k]. Since the
event “X̃ = k − 1 and Ỹ = k” appears with positive probability, we even have
Pr[X̃ ≥ k] < Pr[Ỹ ≥ k]. Also by construction, X̃ ∼ X and Ỹ ∼ Y , so the previous
statement is also valid for X and Y .
3
Proofs of Our Results
We are now ready to prove our results.
Lemma 8. Let n ∈ N and
1
n
≤ p < 1. Let X ∼ Bin(n, p). Let k = ⌊np⌋. Then
1
Pr[X > E[X]] > −
2
s
n
=: g(n, k).
2πk(n − k)
(4)
Proof. We compute
Pr[X > E[X]] = Pr[X ≥ k + 1]
≥ Pr[Bin(n, nk ) ≥ k + 1]
= Pr[Bin(n, nk ) ≥ k] − Pr[Bin(n, nk ) = k]
s
1
n
.
> −
2
2πk(n − k)
Here the first inequality stems from the natural stochastic domination relation
between binomial distributions with different success probabilities (Lemma 7).
Last estimate uses (i) the well-known fact that a binomial distribution with integral
expectation has this expectation as (unique) median [Neu66] and (ii) the estimate
from Lemma 6.
Note that when n is fixed, then g(n, k) is minimal for k = 1 and k = n − 1.
Also, g(n, 1) = g(n, n − 1) is monotonically increasing in n. Hence for all n ≥ 3,
1
≤ p < 1, and k := ⌊np⌋, we have Pr[X > E[X]] > g(n, k) ≥ g(n, 1) ≥ g(3, 1) >
n
0.0113. For n = 2 and consequently p ≥ 21 , we compute Pr[X > E[X]] ≥ Pr[X =
2] = p2 ≥ 41 . Hence Lemma 8 immediately gives a constant lower bound for the
probability to exceed the expectation. Since we expect that the precise constant of
1
shown below is not important in several applications, e.g., in the runtime analysis
4
of algorithms, we formulate this elementary result explicitly. We add the trivial
observation that this result, by possibly lowering the constant, can be extended to
smaller values of p as long as they are at least nε for some ε > 0.
6
Corollary 9. Let ε > 0. Then for all n ∈ N,
have Pr[X > E[X]] > min{e−ε , 0.0113}.
ε
n
≤ p < 1, and X ∼ Bin(n, p), we
Proof. For p ≥ n1 , the claim follows from the above discussion. For 0 < p < n1 , we
compute Pr[X > E[X]] = 1 − Pr[X = 0] = 1 − (1 − p)n > 1 − exp(−pn), using the
well-known estimate 1 + x < ex valid for all x 6= 0.
We now show how to improve the lower bound to 14 . Since g(n, k) approaches 12
when k and n − k tend to infinity, it is clear that we only have to deal with “small
cases”. More specifically, since g(20, 3) = g(20, 17) ≥ 0.2501, we need to regard for
arbitrary n the cases that p < n3 and p ≥ 1 − n3 . In addition, we need to consider
the finite number of cases where n ≤ 19. This will give the following result.
Theorem 10. Let n ∈ N and αn ≤ p < 1, where α := ln( 34 ) < 0.2877. Let
X ∼ Bin(n, p). Then Pr[X > E[X]] ≥ 14 .
Proof. For p ∈ [ αn , n1 ) with α := ln( 34 ) < 0.2877, we compute Pr[X > E[X]] =
Pr[X ≥ 1] = 1 − (1 − p)n > 1 − exp(−pn) ≥ 1 − exp(−α) ≥ 41 by choice of α.
We assume from now on that p ≥ n1 . Let Y ∼ Bin(n, nk ) with k = ⌊np⌋. By
Lemma 7,
Pr[X > E[X]] = Pr[X ≥ k + 1] ≥ Pr[Y ≥ k + 1].
Hence it suffices to show
Pr[X ≥ k + 1] ≥
1
4
with X ∼ Bin(n, nk )
(5)
for all n ∈ N and k ∈ [1..n − 1]. To this aim, let us assume that X ∼ Bin(n, nk ) in
the remainder of this proof. We start by treating the “small” cases k ∈ {1, 2, n −
3, n − 2, n − 1}.
For k = 1 and n ≥ 3, we compute Pr[X ≥ 2] = 1 − Pr[X = 0] − Pr[X = 1] =
7
1 − (1 − n1 )n − (1 − n1 )n−1 ≥ 27
> 0.2592, where the first inequality stems from the
fact that n 7→ (1 − n1 )n + (1 − n1 )n−1 is decreasing (Lemma 3) and n ≥ 3.
For k = 2 and n ≥ 5, in a similar fashion we compute Pr[X ≥ 3] = 1−(1− n2 )n −
4
4
2(1− n2 )n−1 −2(1− n1 )(1− n2 )n−2 = 1−(1− n2 )n−1 (5+ (n−2)n
) > 1−e−2 (5+ (n−2)n
)≥
4
1 n/2−1/2 2
1−e−2 (5+ 15
) > 0.2872 by noting that (1− n2 )n−1 = ((1− n/2
)
) is increasing
−2
(Lemma 3) and tending to e .
For k = n−1 and n ≥ 2, we estimate Pr[X ≥ n] = (1− n1 )n ≥ 41 using Lemma 3
and n ≥ 2.
For k = n − 2 and n ≥ 4, we estimate Pr[X ≥ n − 1] = 2(1 − n2 )n−1 + (1 − n2 )n ≥
2( 12 )3 + ( 12 )4 = 0.3125 using Lemma 3 and n ≥ 4.
For k = n − 3 and n ≥ 6, we estimate Pr[X ≥ n − 2] = 29 (1 − n1 )(1 − n3 )n−2 +
3(1 − n3 )n−1 + (1 − n3 )n > 92 (1 − n3 )n−1 + 3(1 − n3 )n−1 + (1 − n3 )n ≥ 0.25, using
Lemma 3 and n ≥ 6.
7
With this case distinction, we have proven (5) for all n ∈ N and k ∈ {1, 2, n −
3, n − 2, n − 1} ∩ [1..n − 1]. Since g(20, 3) = g(20, 17) > 0.25, the concavity of
k 7→ g(n, k) gives g(20, k) > 0.25 for all k ∈ [3..17]. Since for all n ∈ N and
k ∈ [1..n − 1] we have g(n + 1, k) > g(n, k) and g(n + 1, k + 1) > g(n, k), we have
g(n, k) > 0.25 for all n ≥ 20 and k ∈ [3..n − 3], which proves (5) for all n ≥ 20.
Hence it remains to show (5) for all n ≤ 19 and k ∈ [3..n − 4]. In principle,
these 91 cases can easily be checked in an automated fashion. If we prefer a
human-readable proof, we can argue as follows.
For the case k = 3 and n ≥ 7, we compute Pr[X ≥ 4] ≥ 1 − (1 − n3 )n −
3(1 − n3 )n−1 − 29 (1 − n1 )(1 − n3 )n−2 − 29 (1 − n1 )(1 − n2 )(1 − n3 )n−3 . By Lemma 3,
(1 − n3 )n and 3(1 − n3 )n−1 are increasing in n and tend to e−3 . We shall argue
that 92 (1 − n1 )(1 − n3 )n−2 + 29 (1 − n1 )(1 − n2 )(1 − n3 )n−3 is decreasing, hence it is at
most 124416
for n ≥ 7. Hence Pr[X ≥ 4] ≥ 1 − 4e−3 − 124416
≥ 0.2720. To see
235298
235298
9
1
3 n−2
9
1
2
3 n−3
that 2 (1 − n )(1 − n )
+ 2 (1 − n )(1 − n )(1 − n )
is decreasing, we rewrite this
5
term as f (n) = 9(1 − n1 )(1 − 2n
)(1 − n3 )n−3 . In a similar fashion as in the proof
3i+1 −( 5 )i (i+1)−(i+1)
2
n−i , which is a
of Lemma 3, we see that ln( 19 f (n)) = −3 + ∞
i=1
i(i+1)
sum of constant and decreasing function.
3
For the case k = n − 4 and n ≥ 8, we compute Pr[X ≥ n − 3] = 46 (1 − n1 )(1 −
3
2
)(1 − n4 )n−3 + 8(1 − n1 )(1 − n4 )n−2 + 4(1 − n4 )n−1 + (1 − n4 )4 > 46 (1 − n1 )(1 − n4 )n−2 +
n
8(1 − n1 )(1 − n4 )n−2 + 4(1 − n4 )n−1 + (1 − n4 )4 , which is increasing in n by Lemma 3.
Using n ≥ 8, we conclude Pr[X ≥ n − 3] ≥ 0.2903.
We now note that g(12, 4) = g(12, 8) > 0.25 and that g(11, 5) = g(11, 6) >
0.25. With the same monotonicity arguments as above, this solves all cases (n, k)
which can be written as (n, k) = (n0 + i + j, k0 + j) with i, j ∈ N ∪ {0} and
(n0 , k0 ) ∈ {(12, 4), (11, 5), (11, 6), (12, 8)}. This leaves to check only the cases
(n, k) ∈ {(9, 4), (10, 4), (11, 4), (10, 5)}. These are best computed by hand, e.g.,
Pr[Bin(9, 49 ) ≥ 5] ≥ Pr[Bin(9, 49 ) = 5] + Pr[Bin(9, 49 ) = 6] ≥ 0.2081 + 0.1110 =
0.3191.1
P
Since [GM14] showed a probability of strictly more than 14 (albeit for the weaker
event X ≥ E[X]), let us remark that our proof above also shows Pr[X > E[X]] > 41
for all n ∈ N and ln( 43 )/n ≤ p < 1 except (n, p) = (2, 12 ), where indeed Pr[X >
E[X]] = 41 . The proof above does not exclude Pr[Bin(n, nk ) = k + 1] = 41 only
for the case that n ≥ 2 and k = n − 1. The success probability shown above is
(strictly) increasing in n, so the one remaining case is (n, k) = (2, 1), where indeed
we do not have Pr[X > E[X]] > 14 . By the strict domination result of Lemma 7,
for p ∈
/ { n1 , n2 , . . . , n−1
} we always have Pr[X > E[X]] > 41 .
n
More precisely, for (n, k) = (9, 4), (10, 4), (11, 4), (10, 5) we have Pr[Bin(n, nk ) ≥ k + 1] ≈
0.3655, 0.3668, 0.3678, 0.3769, respectively, where all values have been rounded down to the nearest multiple of 0.0001.
1
8
0.3
0.1
0.2
5 · 10−2
0
0.1
0
0.2
0.4
0.6
0.8
0
1
0.2
0.4
p
0.6
0.8
p
Figure 3: Comparison of the lower bounds for Pr[Bin(n, p) > E[X] + 1] stemming
from Theorem 11 [Pel16] (dotted lines) and our Theorem 12 (a) (solid lines), (b)
(dashed lines), and (c) (sparsely dotted lines) for n = 20 (left) and n = 100 (right).
Exceeding the Expectation by More Than One
We end this section with a short proof of the fact that binomial random variables exceed their expectation also by more than one with constant probability
(obviously only when p < 1 − n1 ). For this problem, not much previous work exists. Pelekis [Pel16] shows the following estimate for exceeding the expectation by
general amounts.
⌋, we
Theorem 11. Let n ∈ N, 0 < p < 1, and np < k ≤ n − 1. With ℓ := ⌊ k−np
1−p
have
n
p2ℓ+2 ℓ+1
.
Pr[X ≥ k] ≥
k
2
ℓ+1
For the case of exceeding the expectation by more than one, we show the
following results. As Figure 3 indicates, Pelekis’ and our bounds do not compare
easily, but ours give the more uniform results.
Theorem 12. Let n ∈ N≥3 and
Then
1
n
≤ p < 1 − n1 . Let X ∼ Bin(n, p). Let k = ⌊np⌋.
(a) Pr[X > E[X] + 1] ≥ 0.5 − 2
q
n
;
2πk(n−k)
(b) Pr[X > E[X] + 1] ≥ 0.25 −
q
n
2π(k+1)(n−k−1)
(c) Pr[X > E[X] + 1] ≥ 0.0370.
9
1−
1
k+1
k+1
1+
1
n−k−1
n−k−1
;
Proof. Let Y ∼ Bin(n, nk ). By Lemma 7, we have
Pr[X > E[X] + 1] = Pr[X ≥ k + 2] ≥ Pr[Y ≥ k + 2],
so we can assume in the following that p = nk with k ∈ [1..n − 2] and show our
claims for Pr[X ≥ k + 2].
Similar to the proof of Lemma 8, we compute
Pr[X ≥ k + 2] = Pr[X ≥ k] − Pr[X = k] − Pr[X = k + 1]
≥ Pr[X ≥ k] − 2 Pr[X = k]
s
1
n
> −2
,
2
2πk(n − k)
where the estimate Pr[X = k] ≥ Pr[X = k + 1] used in the penultimate inequality
either follows from a simple computation or from the well-known fact that the
mode of Bin(n, nk ) is k.
By using Theorem 10 and Corollary 5, we compute
Pr[X ≥ k + 2] = Pr[X > E[X]] − Pr[X = E[X] + 1]
1
≥ −
4
s
n
1
1−
2π(k + 1)(n − k − 1)
k+1
k+1
1
1+
n−k−1
:= h(n, k).
n−k−1
To prove the last claim, recall that we have to show
Pr[Bin(n, nk ) ≥ k + 2] ≥ 0.0370
(6)
for all n ≥ 3 and k ∈ [1..n − 2]. To this aim, we first exploit in two different ways
the lower bound h(n, k).
By rewriting
s
n
1
1+
n−k−1
n−k−1
n−k−1
=
s
n
1
1+
n−k
n−k−1
n−k−1+0.5
as product of two terms which are both decreasing in n for n > k + 1, see
Lemma 3 (b), we observe that h(n, k) is increasing for k fixed and n ≥ k + 2
growing. Since h(n, k) > 0.0370 for (n, k) ∈ {(6, 1), (9, 2), (9, 3), (10, 4), (10, 5)},2
we have shown (6) for all
(n, k) ∈ {(6 + i, 1), (9 + i, 2), (9 + i, 3), (10 + i, 4), (10 + i, 5) | i ≥ 0}.
2
More precisely, we have h(6, 1) ≈ 0.0391, h(9, 2) ≈ 0.0392, h(9, 3) ≈ 0.0392, h(10, 4) ≈
0.0442, and h(10, 5) ≈ 0.0394, where the value given is the true value rounded down to the
nearest multiple of 0.0001.
10
We now argue
q that h(n, n − 5) is increasing in n for n ≥ 10. We rewrite the
n
1 n−4
4 1/2
1
1 (n−4)−1
(1 − n−4
)
= (1 + n−4
) (1 − n−4
)(1 − n−4
)
. The
relevant part n−4
last factor is known to be decreasing for n > 4. For the first two factors, we
4 1/2
1
2
7
4
compute ((1 + n−4
) (1 − n−4
))2 = 1 + n−4
− (n−4)
2 + (n−4)3 , which is decreasing
from n − 4 ≥ 6 on. Since h(10, 5) > 0.0370, we have h(n, k) > 0.0370 for all
(n, k) ∈ {(10 + i, 5 + i) | i ≥ 0}, and via the first monotonicity statement for all
(n, k) ∈ {(10 + i + j, 5 + i) | i, j ≥ 0}.
For the remaining cases, we estimate directly the probability q(n, k) :=
Pr[Bin(n, nk ) ≥ k + 2] in (6). For k ∈ {n − 2, n − 3, n − 4}, we simply compute
q(n, n − 2) = (1 − n2 )n ≥ 0.0370,
q(n, n − 3) = (1 − n3 )n + 3(1 − n3 )n−1 ≥ 0.0507,
q(n, n − 4) = (1 − n4 )n + 4(1 − n4 )n−1 + 8(1 − n1 )(1 − n4 )n−2 ≥ 0.0579,
where the last estimates stem from noting that all the terms involved are increasing
in n by Lemma 3 (a) and evaluating the expressions for n = 3, n = 4, and n = 5,
respectively. This shows (6) for all
(n, k) ∈ {(3 + i, 1 + i), (4 + i, 1 + i), (5 + i, 1 + i) | i ≥ 0}.
For the last four cases (n, k) ∈ {(7, 2), (8, 2), (8, 3), (9, 4)}, equation (6) is easily
checked by hand: We have q(7, 2) ≈ 0.1082, q(8, 2) ≈ 0.1138, q(8, 3) ≈ 0.1374, and
q(9, 4) ≈ 0.1573.
We note that, as in Corollary 9, smaller values of p can be admitted at the
price of a smaller probability for the event “X > E[X] + 1”. For p = αn < n1 , we
have Pr[X > E[X]+ 1] = 1 −(1 −p)n −np(1 −p)n−1 = 1 −(1 − αn )n −α(1 − αn )n−1 ≥
1 − exp(−α) − α exp(−α n−1
).
n
Acknowledgment
The author would like to thank Philippe Rigollet (Massachusetts Institute of Technology (MIT)) and Carsten Witt (Danish Technical University (DTU)) for useful
discussions and pointers to the literature.
References
[BCN+ 17]
Luca Becchetti, Andrea E. F. Clementi, Emanuele Natale, Francesco
Pasquale, Riccardo Silvestri, and Luca Trevisan. Simple dynamics for
plurality consensus. Distributed Computing, 30:293–306, 2017.
11
[DGWY17] Benjamin Doerr, Christian Gießen, Carsten Witt, and Jing
Yang. The (1+λ) evolutionary algorithm with self-adjusting mutation rate.
In Genetic and Evolutionary Computation Conference, GECCO 2017. ACM, 2017.
Full version available at
http://arxiv.org/abs/1704.02191.
[Fel71]
William Feller. An Introduction to Probability Theory and Its Applications, volume II. John Wiley & Sons Inc., New York, second edition,
1971.
[GM14]
Spencer Greenberg and Mehryar Mohri. Tight lower bound on the
probability of a binomial exceeding its expectation. Statistics and
Probability Letters, 86:91–98, 2014.
[KKK16]
Matti Karppa, Petteri Kaski, and Jukka Kohonen. A faster subquadratic algorithm for finding outlier correlations. In Proceedings
of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA 2016, pages 1288–1305. SIAM, 2016.
[MM17]
Michael Mitzenmacher and Tom Morgan. Reconciling graphs and sets
of sets. CoRR, abs/1707.05867, 2017.
[Neu66]
P. Neumann. Über den Median der Binomial- and Poissonverteilung.
Wissenschaftliche Zeitschrift der Technischen Universität Dresden,
19:29–33, 1966.
[Pel16]
Christos Pelekis. A lower bound on binomial tails: an approach via
tail conditional expectations. ArXiv e-prints, 1609.06651, 2016.
[PR16]
Christos Pelekis and Jan Ramon. A lower bound on the probability
that a binomial random variable is exceeding its mean. Statistics and
Probability Letters, 119:305–309, 2016.
[Rob55]
Herbert Robbins. A remark on Stirling’s formula. American Mathematical Monthly, 62:26–29, 1955.
[RT11]
Philippe Rigollet and Xin Tong. Neyman-Pearson classification, convexity and stochastic constraints. Journal of Machine Learning Research, 12:2831–2855, 2011.
[Slu77]
Eric V. Slud. Distribution inequalities for the binomial law. Annals
of Probability, 5:404–412, 1977.
12
| 9cs.NE
|
ZEILBERGER’S KOH THEOREM AND THE STRICT
UNIMODALITY OF q-BINOMIAL COEFFICIENTS
arXiv:1311.4480v2 [math.CO] 1 Apr 2014
FABRIZIO ZANELLO
Abstract. A recent nice result due to I. Pak and G. Panova is the strict unimodality
of the q-binomial coefficients a+b
b q (see [2] and also [3] for a slightly revised version of
their theorem). Since their proof used representation theory and Kronecker coefficients, the
authors also asked for an argument that would employ Zeilberger’s KOH theorem. In this
note, we give such a proof. Then, as a further application of our method, we also provide a
short proof of their conjecture that the difference between consecutive coefficients of a+b
b q
can get arbitrarily large, when we assume that b is fixed and a is large enough.
A sequence c1 , c2 , . . . , ct is unimodal if it does not increase strictly after a strict decrease.
It is symmetric if ci = ct−i for all i. The unimodality of the q-binomial coefficient
(1 − q)(1 − q 2 ) · · · (1 − q a+b )
a+b
=
,
(1 − q)(1 − q 2 ) · · · (1 − q a ) · (1 − q)(1 − q 2 ) · · · (1 − q b )
b
q
which is easily proven to be a symmetric polynomial in q, is a classical and highly nontrivial
result in combinatorics. It was first shown in 1878 by J.J. Sylvester, and has since received
a number of other interesting proofs (see e.g. [4, 5, 7]). In particular, a celebrated paper
of K. O’Hara [1] provided a combinatorial proof for the unimodality of a+b
. O’Hara’s
b q
argument was subsequently expressed in algebraic terms by D. Zeilberger [8] by means of
the beautiful KOH identity. This identity decomposes a+b
into a finite sum of polynomials
b q
with nonnegative integer coefficients, which are all unimodal and symmetric about ab/2.
More precisely, fix integers a ≥ b ≥ 2. For any given partition λ = (λ1 , λ2 , . . . ) of b, set
P
Yi = ij=1 λj for all i ≥ 1, and Y0 = 0. Then the KOH theorem can be stated as follows:
P
Lemma 1 (KOH). a+b
= λ⊢b Fλ (q), where
b q
Y j(a + 2) − Yj−1 − Yj+1
P
2 i≥1 (λ2i )
.
Fλ (q) = q
λj − λj+1
q
j≥1
A recent nice result shown by I. Pak and G. Panova is a characterization of the strict
unimodality of q-binomial coefficients; i.e., they determined when a+b
strictly increases
b q
2010 Mathematics Subject Classification. Primary: 05A15; Secondary: 05A17.
Key words and phrases. q-binomial coefficient; Gaussian polynomial; unimodality.
1
2
FABRIZIO ZANELLO
from degree 1 to degree ⌊ab/2⌋ (see [2], and also [3] for a slightly revised version of the
theorem). Since their argument employed the algebraic machinery of Kronecker coefficients,
the authors asked whether a proof could also be given that uses Zeilberger’s KOH identity.
We do this in the present note. Then, as a further pithy application of this method, using
the KOH theorem we also give a very short proof of a conjecture stated in the same papers,
on the unbounded growth of the difference between consecutive coefficients of a+b
.
b q
The next lemma is a trivial and probably well-known fact of which we omit the proof.
Lemma 2. Let c and d be positive integers such that the q-binomial coefficient
c+d
d q
is
strictly unimodal. Then, for any positive integer t ≤ cd such that t 6= cd − 2, the product
t+1
c+d
is strictly unimodal (in all nonnegative degrees).
d q 1 q
Theorem 3 ([2, 3]). The q-binomial coefficient
a+b
b q
a = b = 2 or b ≥ 5, with the exception of
is strictly unimodal if and only if
(a, b) = (6, 5), (10, 5), (14, 5), (6, 6), (7, 6), (9, 6), (11, 6), (13, 6), (10, 7).
Proof. We can assume that b ≥ 5, otherwise, as it is also noted in [2, 3], the result is easy to
show. By Lemma 1, since all terms in the KOH decomposition of a+b
are unimodal and
b q
symmetric with respect to ab/2, in order to show that a+b
is strictly unimodal, it clearly
b q
suffices to determine, for each positive degree up to ab/2, some suitable KOH term that is
strictly increasing in that degree. We begin by showing that, for any a ≥ b ≥ 2, a+b
b q
strictly increases up to degree ab/2 − a for b even, and up to degree ab/2 − a/2 for b odd.
Let b = 2m be even. Then the KOH term contributed by the partition λ = (λ1 =
2, . . . , λm−1 = 2, λm = 1, λm+1 = 1) of b is given by:
Fλ (q) = q
2(m−1)
(m + 1)(a + 2) − (2m − 1) − 2m
(m − 1)(a + 2) − 2(m − 2) − (2m − 1)
1
1
q
q
=q
b−2
ab/2 + a − b + 3
ab/2 − a − b + 3
.
1
1
q
q
Notice that the product of the last two q-binomial coefficients is strictly increasing (by
1) from degree 0 to degree ab/2 − a − b + 2. Also, a+b
is clearly strictly increasing
b q
from degree 1 to degree b − 2, since so is the usual partition function p(n) (see e.g. [6],
Chapter 1). From this, we easily have that a+b
strictly increases from degree 1 to degree
b q
(ab/2 − a − b + 2) + (b − 2) = ab/2 − a.
The proof for b = 2m + 1 odd, giving us that
a+b
b q
is strictly increasing up to degree
ab/2 − a/2, is similar (using λ = (λ1 = 2, . . . , λm = 2, λm+1 = 1)) and thus will be omitted.
ZEILBERGER’S KOH THEOREM AND THE STRICT UNIMODALITY OF q-BINOMIAL COEFFICIENTS3
Now, in order to show that for the desired values of a and b,
a+b
b q
strictly increases in
each of the remaining degrees up to ab/2, we consider three cases depending on the residue
of b modulo 3. We start with b ≡ 0 modulo 3, and assume that b ≥ 15. The KOH term
corresponding to the partition λ = (b/3, b/3, b/3) of b is given by:
Fλ (q) = q
6(b/3
2 )
3(a + 2) − 2b/3 − b
b(b−3)/3 (3a − 2b + 6) + b/3
.
=q
b/3
b/3
q
q
Notice that b(b−3)/3 < ab/2−a, and 3a−2b+6 ≥ 15. Thus, it easily follows by induction
a+b
implies
that
of
, as desired.
that, for b ≥ 15, the strict unimodality of (3a−2b+6)+b/3
b q
b/3
q
Let now b ≡ 1 modulo 3, and assume b ≥ 19. By considering the partition λ = ((b −
1)/3, (b − 1)/3, (b − 1)/3, 1) of b, we get:
Fλ (q) = q
(b−1)(b−4)/3
4a − 2b + 9
3a − 2b + 8 + (b − 4)/3
.
1
(b − 4)/3
q
q
It is easy to check that, under the current assumptions on a and b, we have (b−1)(b−4)/3 <
ab/2 − a and (3a − 2b + 8)(b − 4)/3 ≥ (4a − 2b + 8) + 3. In particular, we are under the
hypotheses of Lemma 2. Thus, since 3a−2b+ 8 ≥ 15, the strict unimodality of a+b
follows
b q
3a−2b+8+(b−4)/3
, for all b ≥ 19, as we wanted to show.
by induction from that of
(b−4)/3
q
The treatment of the case b ≡ 2 modulo 3, b ≥ 20, is analogous so we will omit the details.
We only remark here that one considers the partition λ = ((b − 2)/3, (b − 2)/3, (b − 2)/3, 1, 1)
of b, whose contribution to the KOH expansion of a+b
is:
b q
Fλ (q) = q
(b−2)(b−5)/3
5a − 2b + 11
3a − 2b + 10 + (b − 5)/3
.
1
(b − 5)/3
q
q
The strict unimodality of a+b
, for b ≥ 20, then follows in a similar fashion from that of
b q
3a−2b+10+(b−5)/3
, by employing Lemma 2 and induction.
(b−5)/3
q
Therefore, it remains to show the theorem for 5 ≤ b ≤ 17 (b 6= 15). We will assume for
simplicity that a ≥ 2b + 13, the result being easy to verify directly for the remaining values
of a. The KOH term contributed by the partition (b) of b in the expansion of a+b
is:
b q
F(b) (q) = q
2(2b)
a+2−b
b(b−1) (a − 2b + 2) + b
.
=q
b
b
q
q
Clearly, since a ≥ 2b + 13, we have b(b − 1) < ab/2 − a and a − 2b + 2 ≥ 15. Thus, by
induction, the strict unimodality of (a−2b+2)+b
implies that of a+b
, as desired.
b
b q
q
In Remark 3.6 of [2, 3], the authors also conjectured that, roughly speaking, the difference
between consecutive coefficients of a q-binomial coefficient is eventually larger than any
fixed integer. As a further nice, and very brief, application of our method, we answer this
4
FABRIZIO ZANELLO
conjecture in the positive using the KOH identity. (Just notice that unlike in the original
formulation of the conjecture, our proof will assume that b is fixed and only a is large enough.)
Proposition 4. Fix any integer d ≥ 1. Then there exist integers a0 , b and L such that, if
Pab
a+b
i
=
i=0 ci q , then ci − ci−1 ≥ d, for all indices L ≤ i ≤ ab/2 and for all a ≥ a0 .
b q
[k]
[k]
[k]
Proof. Consider the partition λ[k] = (λ1 = b − k, λ2 = 1, . . . , λk+1 = 1) of b, where k ≥ 1.
It is easy to see that its contribution to the KOH identity for a+b
is given by:
b q
(k + 1)(a + 2) − 2b + 1
(b−k)(b−k−1) a − 2b + 2k + 2 + (b − k − 1)
.
Fλ[k] (q) = q
1
b−k−1
q
q
Set for instance b = 2d + 4 and a0 = (d + 2)(d + 3) + 6, where we can assume d ≥ 2. A
standard computation gives that, for any a ≥ a0 and k ≤ b/2 − 2 = d, we are under the
hypotheses of Lemma 2. Hence, by Theorem 3, each polynomial Fλ[k] (q) is strictly unimodal
from degree (b − k)(b − k − 1) on, and the theorem now immediately follows by choosing
P
L = (b − 1)(b − 2) + 1 = 4d2 + 10d + 7 and considering the coefficients of dk=1 Fλ[k] (q).
1. Acknowledgements
The idea of this paper originated during a visit to UCLA in October 2013, for whose
invitation we warmly thank Igor Pak. We wish to thank the referee for a careful reading of
our manuscript and for helpful suggestions that improved the presentation, and Igor Pak,
Greta Panova, and Richard Stanley for several helpful discussions. We are also very grateful
to Doron Zeilberger for comments (and for calling this a “proof from the book”). This work
was done while the author was partially supported by a Simons Foundation grant (#274577).
References
[1] K. O’Hara: Unimodality of Gaussian coefficients: a constructive proof, J. Combin. Theory Ser. A 53
(1990), no. 1, 29–52.
[2] I. Pak and G. Panova: Strict unimodality of q-binomial coefficients, C. R. Math. Acad. Sci. Paris 351
(2013), no. 11-12, 415–418.
[3] I. Pak and G. Panova: Strict unimodality of q-binomial coefficients (new version), preprint. Available
on the arXiv.
[4] R. Proctor: Solution of two difficult combinatorial problems using linear algebra, Amer. Math. Monthly
89 (1982), no. 10, 721–734.
[5] R. Stanley: Weyl groups, the hard Lefschetz theorem, and the Sperner property, SIAM J. Algebraic
Discrete Methods 1 (1980), no. 2, 168–184.
[6] R. Stanley: “Enumerative Combinatorics”, Vol. I, Second Ed., Cambridge University Press, Cambridge,
U.K. (2012).
[7] J.J. Sylvester: Proof of the hitherto undemonstrated fundamental theorem of invariants, Collect. Math.
papers, Vol. 3, Chelsea, New York (1973), 117–126.
ZEILBERGER’S KOH THEOREM AND THE STRICT UNIMODALITY OF q-BINOMIAL COEFFICIENTS5
[8] D. Zeilberger: Kathy O’Hara’s constructive proof of the unimodality of the Gaussian polynomials, Amer.
Math. Monthly 96 (1989), no. 7, 590–602.
Department of Mathematics, MIT, Cambridge, MA 02139-4307
matical Sciences, Michigan Tech, Houghton, MI 49931-1295
E-mail address: [email protected], [email protected]
and
Department of Mathe-
| 0math.AC
|
Improved Training of Generative Adversarial Networks
Using Representative Features
Duhyeon Bang 1 Hyunjung Shim 1
arXiv:1801.09195v1 [cs.CV] 28 Jan 2018
Abstract
Despite of the success of Generative Adversarial Networks (GANs) for image generation tasks,
the trade-off between image diversity and visual
quality are an well-known issue. Conventional
techniques achieve either visual quality or image
diversity; the improvement in one side is often the
result of sacrificing the degradation in the other
side. In this paper, we aim to achieve both simultaneously by improving the stability of training
GANs. A key idea of the proposed approach is to
implicitly regularizing the discriminator using a
representative feature. For that, this representative
feature is extracted from the data distribution, and
then transferred to the discriminator for enforcing slow updates of the gradient. Consequently,
the entire training process is stabilized because
the learning curve of discriminator varies slowly.
Based on extensive evaluation, we demonstrate
that our approach improves the visual quality and
diversity of state-of-the art GANs.
1. Introduction
Generative models aim to solve the problem of density estimation by learning the model distribution Pmodel , which
approximates the true but unknown data distribution of Pdata
using a set of training examples drawn from Pdata (Goodfellow, 2016). Generative adversarial networks (GANs)
(Goodfellow et al., 2014) are a family of generative models
capable of implicitly estimating a data distribution without
an analytic formula or variational bound of Pmodel . Existing
GANs have been mainly used for image generation tasks,
with which they have showed impressive results and produced sharp and realistic images of natural scenes. Because
Preliminary work. Under review by the International Conference on Machine Learning (ICML). Do not distribute. 1 School
of Integrated Technology, Yonsei University, South Korea. Correspondence to: Hyunjung Shim <[email protected]>.
Proceedings of the 35th International Conference on Machine
Learning, Stockholm, Sweden, PMLR 80, 2018. Copyright 2018
by the author(s).
of the flexible nature of the model definition and high quality results, GANs have been applied to many real-world
applications, including super-resolution, colorization, face
generation, image completion, etc. (Bao et al., 2017; Ledig
et al., 2016; Yeh et al., 2017; Cao et al., 2017)
Training a GAN involves two separate networks with competitive goals: a discriminator, D, to distinguish between
the real and fake data and a generator, G, to create data to be
as real as possible to fool the discriminator. Consequently,
the generator implicitly models Pmodel which approximates
Pdata . In (Goodfellow et al., 2014), this problem is formulated by the following minimax game as
min max E log(D(x)) + E log(1 − D (G (z)) .
G
D
x∼Pdata
z∼Pz
When the generator produces perfect samples (meaning
that Pmodel is identical to Pdata ), the discriminator cannot
distinguish between the real and fake data. Subsequently,
this game ends because it reaches a Nash equilibrium.
Although GANs have been successful in the field of image generation, the instability of training processes such as
extreme sensitivity of a network structure and parameter
tuning are well-known disadvantages. Training instability
has yielded two major problems, namely gradient vanishing
and mode collapse. As discussed in (Arjovsky & Bottou,
2017), gradient vanishing becomes a serious problem when
any subset of Pdata and Pmodel are disjointed such that the
discriminator separates real and fake data perfectly; i.e. the
generator no longer improves the data because the discriminator has reached its optimum. This yields poor results
because training stops even though Pmodel has not learned
Pdata properly. Mode collapse is another common problem
of GANs; the generator repeatedly produces the same or
similar output because Pmodel only encapsulates the major
or single modes of Pdata in order to easily fool the discriminator.
Unfortunately, the trade-off between image quality and
mode collapse are theoretically and empirically investigated
in previous work (Berthelot et al., 2017; Fedus et al., 2017).
Existing studies often achieved either visual quality or image diversity, but not both. Their results can be interpreted
as the visual quality can be achieved by minimizing the
reverse KL divergence while the image diversity is strongly
Submission and Formatting Instructions for ICML 2018
correlated with minimizing the forward KL divergence. To
break the trade-off, recent techniques (Kodali et al., 2017;
Gulrajani et al., 2017; Fedus et al., 2017) introduce the gradient penalty for regularizing the divergence (or distance)
for training GANs. The gradient penalty smooths out the
learning curve, thus improving the stability of training. As a
result of stabilized training, the gradient penalty is effective
to improve both visual quality and image diversity, and this
idea is evaluated in various architectures of GANs.
The goal of the proposed approach is similar to the gradient
penalty in that we also aim to stabilize the training in order
to break the trade-off between the visual quality and image
diversity. To this end, we propose an unsupervised approach
to implicitly regularizing the discriminator using representative features. Unlike the gradient penalty term, our approach
does not modify the objective function of GANs (i.e., use
the same divergence or loss definition as a baseline GAN).
Instead, we introduce the representative features from a
pre-trained autoencoder (AE) and transfer them to a discriminator for training GANs. Representative features are a good
representation for reconstructing the overall data distribution Pdata , but not as distinguishable as the discriminative
features (i.e., features learnt for classification). Thereby,
the discriminator is implicitly interrupted by representative
features for the discrimination while encouraged to account
for the entire data distribution.
It is worthwhile noting that this pre-trained AE learns from
samples of Pdata and is then fixed. By isolating the representative feature extraction from training GANs, it is possible
to guarantee that the embedding space of pre-trained AE
and corresponding features exhibit the representative power.
Because the representative features are derived from the
pre-trained network, they are much informative at the early
stage of training discriminator; accelerating the early stage
training of GANs. However, from the second half of training, the representative features no longer distinguish real
and fake images. At this stage, the overall loss of our model
consists of the loss of representative features and that of
discriminative features. Because the loss of representative
feature is nearly constant at the second half of training, the
overall loss of discriminator varies slowly over iterations.
As a result, we stabilize GAN training, thereby it is effective to improve both the visual quality and image diversity
of generated samples. We name this new architecture as a
Representative Feature based on a Generative Adversarial
Network (RFGAN).
The major contributions of this paper are as follows:
1. We employ additional representative features extracted
from a pre-trained AE for implicitly constraining the
discriminator’s update. In this way, the learning curve
of GANs varies slowly, thus the GAN training is stabilized. As a result, we simultaneously achieve the visual
quality and image diversity of GANs in an unsupervised manner.
2. Our framework of combining the pre-trained AE is
easily extendable to various GANs using different divergences or structures. Also, the proposed model is
robust against the parameter selections; all results in
this paper use the same hyper-parameters suggested by
a baseline GAN.
3. We conduct extensive experimental evaluations to show
the effectiveness of RFGAN; our approach improves
existing GANs including GANs with the gradient
penalty.
In Section 2, we review recent studies and analyze the relationships of RFGAN with them. Next, the architecture and
distinctive characteristics of RFGAN are demonstrated in
Section 3. In Section 4, we summarize the results of extensive experiments using both simulated and real data. Based
on both quantitative and qualitative evaluations, we show
that the RFGAN improved image quality and also achieved
diversity in data generation.
2. Related Work
A variety of techniques have been proposed in the past for
improving the stability of GAN training which mostly aim
to resolve gradient vanishing and mode collapse. For that,
existing studies can be categorized into two groups.
A. GAN training by modifying the network design
To avoid gradient vanishing, (Goodfellow et al., 2014) proposed modifying a minimax game-based GAN formulation into a non-saturating game. Consequently, the objective function of the generator was changed from J(G) =
Ez∼Pz log(1 − D(G(z))) to J(G) = − 12 Ez∼Pz log(D(G(z))).
This simple modification effectively resolved the problem of
gradient vanishing, and several studies (Arjovsky & Bottou,
2017; Fedus et al., 2017) have confirmed the same conclusion based on theoretical and empirical analysis. (Radford
et al., 2015) first introduced GAN with a stable deep convolutional architecture (DCGAN), and their visual quality was
quantitatively superior to a variant of GANs proposed later,
according to (Lucic et al., 2017).
Unfortunately, mode collapse was a major weakness of
DCGAN. To tackle this problem, unrolled GAN was proposed by (Metz et al., 2016) to adjust the gradient update
of the generator by introducing a surrogate objective function, which simulated the response of discriminator upon
the changes of the generator. As a result, unrolled GAN
successfully solved model collapse. InfoGAN (Chen et al.,
2016) achieved an unsupervised disentangled representation by minimizing the mutual information of auxiliary loss
(i.e., matching the semantic information) and adversarial
loss. Additionally, (Salimans et al., 2016) proposed various ways to stabilize the training process of GAN using
Submission and Formatting Instructions for ICML 2018
semi-supervised learning and smoothing labeling.
B. Effects of various divergences
In (Nowozin et al., 2016), the authors showed that the
Jensen-Shannon divergence used in the original GAN formulation (Goodfellow et al., 2014) can be extended to different
divergences, including f-divergence (f-GAN). (Arjovsky &
Bottou, 2017) and (Arjovsky et al., 2017) analyzed a source
of instability when training GAN and theoretically showed
that the Kullback–Leibler (KL) divergence was one of the
causes. Later, they proposed the Wasserstein distance to
measure the similarity between Pmodel and Pdata in order to
overcome the instability when training GAN. In order to
implement the Wasserstein distance to produce Wasserstein
GAN, (Arjovsky et al., 2017) introduced the k-Lipschitz constraint into the discriminator, which is enforced by weight
clipping. However, (Gulrajani et al., 2017) pointed out that
weight clipping often fails to capture the higher moments
of Pdata , and the authors suggested a gradient penalty for
better modeling of Pdata . (Kodali et al., 2017) showed that
the discriminator became closer to the convex set by using
the gradient penalty as a regularization term, and this is effective in improving the stability of training GAN. For Least
Squares GAN (LSGAN), (Mao et al., 2017) replaced the
Jenson-Shannon divergence defined by the sigmoid crossentropy loss term to the least squares loss term, and showed
that GAN with least squares loss is essentially interpreted
as the minimization of the Pearson χ̃2 divergence.
To summarize, most existing approaches for GAN have
suggested finding stable architecture or additional layers
to stabilize the discriminator’s updates, to change the divergence, or to add a regularization term to stabilize the
discriminator. The proposed algorithm, RFGAN, can be
classified into the first category, which modifies the architecture of GAN. RFGAN is distinguishable from other GANs
in that features from the encoder layers of the pre-trained
AE are transferred while training the discriminator. Several
existing approaches also use AE or encoder architectures,
and they are explained as follows.
ALI(Dumoulin et al., 2016), BiGAN(Donahue et al., 2016)
and MDGAN(Che et al., 2016) suggested that the samples
from Pdata should be mapped to the latent space of the generator using the encoder structure. This is to enforce the
latent space of the generator to learn the entire distribution
of Pdata , thus solving mode collapse. Although these are
similar approaches to ours in that the encoder layers are
employed to develop GAN, our RFGAN uses an AE for
a completely different purpose by an alternative method
to extract representative features, and furthermore, those
features are to stabilize the discriminator. EBGAN(Zhao
et al., 2016) and BEGAN(Berthelot et al., 2017) suggested
an energy-based function to develop the discriminator, by
which their networks showed stable convergence and were
less sensitive to parameter selection. In both studies, they
borrowed the AE architecture for defining the energy-based
function, which served the discriminator. Our model employs representative features from the encoder layers and
still includes the conventional discriminator architecture
to maintain the discriminative power of the discriminator.
(Warde-Farley & Bengio, 2016) extracted the features from
the discriminator layers, applied a denoising AE, and used
its output to regularize the adversarial loss. Based on this
denoising feature matching, they improved the quality of
image generation and employed the denoising AE to ensure
the robust features of the discriminator. The proposed model
is different in that the AE is used as the feature extractor.
Additionally, unlike all of the existing approaches that have
trained the AE or encoder layers as part of the GAN architecture, the proposed algorithm separately trains the AE to
learn Pdata in an unsupervised manner. In this way, we disconnect the feedback from Pmodel when training the AE, and
so focus on learning the feature space for representing Pdata .
Moreover, we do not utilize the decoder, thus preventing
problems such as image blur.
3. The RFGAN model
To resolve the instability of training GAN, we extract the
representative features from an pre-trained AE and transfer
them to the discriminator; this implicitly enforces that the
discriminator is slowly updated once the generator produces
beyond the certain level of visual quality.
The aim of an AE is to learn a reduced representation of
the given data because it is formulated by reconstructing the
input data after passing through the network. Consequently,
feature spaces learnt by the AE are powerful representations
for reconstructing the Pdata distribution.
By focusing on the functionality of the AE as a feature
extractor, several studies have utilized it for classification
tasks through fine-tuning (Zhou et al., 2012; Chen et al.,
2015). On the contrary, the argument by (Alain & Bengio,
2014; Wei et al., 2015) noted that good representation for
reconstruction does not guarantee good classification. This
is a reasonable statement because the features of the reconstruction model and those of the discriminative model are
derived from different objectives, thus should be applied to
the appropriate tasks for optimal performance.
When training a GAN, the discriminator operates as a binary
classifier (Radford et al., 2015), so the features extracted
from the discriminator specialize in distinguishing whether
the input is real or fake. This means that the discriminator’s
features are totally different properties from the AE’s features. By focusing on these different properties, we denote
the AE’s features and those of the discriminator as representative and discriminative features, respectively. While
Submission and Formatting Instructions for ICML 2018
Figure 1. Graphical Model of RFGAN. Xreal and X f ake are input
and generated images; E, G, and D are encoder, generator, and
discriminator networks; Z is the latent vector; Y is binary output
representing the real or synthesized image; h1 and h2 are representative and discriminative features; and w1 , w2 , wD , and wG
are network parameters. The blue solid lines and red dash lines
represent forward/backward propagation.
the original formulation of GAN evaluates the quality of
data generation purely based on discriminative features, we
propose leveraging both representative and discriminative
features to implicitly regularize the discriminator in order
to stabilize GAN training.
h1 and h2 represent the representative and discriminative
feature vectors, respectively. These two feature vectors are
concatenated and transformed to a single sigmoid output
Y through a fully connected layer. The output is evaluated
with the ground truth label based on sigmoid cross entropy,
and then the gradient of the loss function is delivered to the
discriminator via backpropagation to update the parameters.
Note that this feedback is not propagated to the encoder
because its parameters are already trained and then fixed.
The procedure for gradient updates is defined as follows:
D(x) = − log Y for x ∼ Pdata , Y = σ(h1 w1 + h2 w2 ),
∇wi =
∂D(x)
1
= − · Y (1 − Y) · hi = (Y − 1) hi ,
∂wi
Y
∇wD =
∂D(x)
= (Y − 1) · w2 · u (wD ) .
∂wD
The GAN objective function represented by parameters
J (θG , θD ) = E log D (x; θD ) + E log (1 − D (G (z; θG ) ; θD )) ,
x∼Pdata
In this section, we describe the proposed model, RFGAN,
and the effect of the modified architecture for training the
discriminator. Furthermore, we investigate how this effect
could overcome mode collapse and improve visual quality.
which is updated by
θt+1
D
←
θtD
− ηD
3.1. The architecture of RFGAN
The main contribution of our model is in adopting representative features from pre-trained AE to develop GAN,
thus our model is built based on various GAN architectures.
Hence, RFGAN refers to a set of GANs using the representative features. For simplicity, we use DCGAN (Radford et al.,
2015) employing non-saturated loss as a baseline GAN, and
apply representative features to discriminator to construct
DCGAN-RF. In Section 4, we use various other GANs as
baselines to develop RFGAN variants. Throughout this
paper, we use exactly the same hyper-parameters, metrics,
and settings as suggested for a baseline GAN, to show that
our model is indeed insensitive to parameter selection. Instead, we supply representative features extracted from the
encoder layer (part of pre-trained AE) to the discriminator.
Note that the AE is pre-trained in an unsupervised manner using samples from Pdata and isolated from the GAN
training.
More specifically, we construct the AE in a way that its
encoder and decoder share the same architecture as the discriminator and generator, respectively. Next, we concatenate
two feature vectors, one extracted from the last convolution
layer of the encoder and the other from the discriminator.
Given the concatenated feature vector, final weights are
trained for deciding between real or fake input. This process
is demonstrated by the simplified graphical model shown
in Figure 1, in which the input data passes through two
networks, encoder E and discriminator D. In this figure,
z∼Pz
θt+1
← θtG − ηG
G
d f θG , θtD
dθt
D
d f θG , θtD
dθtG
= θtD − ηD (∇wD + ∇wi ) ,
,
where σ and u are a sigmoid and step function, respectively.
Since the encoder is pre-determined, we only consider the
update of the discriminator. After this, we derive the gradient toward the discriminator by calculating the partial
derivative of loss term with respect to wD , which indicates
the network parameters as shown in Figure 1. From this formulation, we observe that ∇wD depends on h1 . In this way,
the representative features affect the discriminator update.
This procedure is derived by the case where x is real. In the
case of a fake sample, the same conclusion is deduced except
that D(x) is now changed to −log(1 − σ(h1 w1 + h2 w2 )) .
Based on the aforementioned formulation, the generator
is trained by considering both the representative and discriminative features because it should fool the discriminator
by maximizing −log D(G(z)) . Again, it is notable that our
representative features keep their own properties such as a
global representation for reconstructing the data distribution
by fixing the encoder parameters.
3.2. Mode collapse
Based on probabilistic interpretation, the decoder of the
AE estimates the parameters of a distribution P(x|En(x))
that may generate x with high probability formulated by
cross-entropy loss (Vincent et al., 2010). We could deem
Submission and Formatting Instructions for ICML 2018
that the AE follows the forward KL divergence between
Pdata and Pmodel (i.e. KL(Pdata ||Pmodel )). Since the model
approximated by the forward KL divergence is evaluated
using every true data sample (i.e. any x for Pdata (x) > 0), it
tends to average out all modes of Pdata (Goodfellow, 2016).
Hence, we can expect that the representative features extracted from the AE are similar in that they are effective at
representing entire modes of Pdata .
On the contrary, the aim of DCGAN with a non-saturated
loss (the base architecture of our model) is to optimize
the reverse KL divergence between Pdata and Pmodel : more
specifically, KL(Pmodel ||Pdata ) − 2JSD (Arjovsky & Bottou,
2017), where JSD is the notion of the Jensen–Shannon
divergence. Because the model based on the reverse KL
objective is examined for every fake sample (i.e. any x
where Pmodel (x) > 0), it has no penalty for covering the
entire true data distribution; hence, it is likely to focus on
the single or partial modes of the true data distribution. This
phenomenon is a well-known mode collapse problem.
Our RFGAN performs the optimization of the reverse KL divergence because our framework is built upon non-saturated
GAN. At the same time, we introduce representative features from the AE, which encourages the model to cover
the entire modes of Pdata , like optimizing the forward KL
divergence. This leads to suppressing the tendency toward
falling mode collapse.
3.3. Improving visual quality
While the representative features are useful to advance the
discriminator in the early stage, they become less informative when approaching the second half of training. This
behavior is expected because the AE shows limited performance during the discrimination task . Note that the AE is
built upon minimizing the reconstruction error (e.g. L2 or
L1 loss); this model cannot learn multiple different, correct
answers, which causes the model to choose the averaged (or
median) output (Goodfellow, 2016). This property is still
useful to distinguish between poor fake and real input. However, when the generator starts producing good fake input,
the representative features from the AE are less discriminative, thus rather interfere with decisions by the discriminator
in the later stages of training. Figure 2 shows the output
when several real and fake examples passed through the
pre-trained AE. At the beginning of training, it is easy to distinguish whether the input is real or fake. Yet, after several
iterations, we observed that both look quite similar. These
experimental results demonstrate the discriminative power
of the AE for different levels of fake examples.
Consequently, it is difficult to improve the visual quality of
data generation beyond a certain level using representative
features alone. This is why the proposed model employs
both representative and discriminative features to train the
Figure 2. Reconstruction comparison after iteration with the pretrained AE. The first row shows generated images. These along
with real images are passed through the pre-trained AE, as shown
in the second and third rows, respectively.
discriminator. Although the representative features interfere
with discrimination between the real and fake input after the
training goes on, the discriminator using the proposed model
still has discriminative features, which allows the training
to continue. As a result, the generator consistently receives
sufficient feedback (i.e. the gradient from the discriminator
is increased.) from the discriminator to learn Pdata . Furthermore, because these two features are two opposing forces
and disagree with each other, in the end the discriminator
slowly approaches saturation. By slowing down the growth
of the discriminator, we observe that our model generates
high quality data, thereby improving the existing GAN.
4. Experimental results
For both the quantitative and qualitative evaluations, we conducted experiments using simulated and three real datasets;
CelebA (Liu et al., 2015), LSUN-bedroom (Yu et al., 2015),
and CIFAR-10 (Krizhevsky & Hinton, 2009) by normalizing between -1 and 1. During the experiment, denoising
AE (Vincent et al., 2008) is used to improve the robustness
of the feature extraction and slight quality improvement is
achieved compared to the conventional AE.
4.1. Handling mode collapse
To evaluate how well the proposed model could achieve the
diversity of data generation (i.e. solving mode collapse), we
trained our network with a simple 2D mixture of 8 Gaussians
suggested by (Metz et al., 2016). The mean of each Gaussian
formed a ring and each Gaussian distribution had a standard
deviation of 0.1. In Figure 3, we compare our RFGAN
with GAN and unrolled GAN. From this experiment, we
Submission and Formatting Instructions for ICML 2018
Figure 3. Mode collapse testing by learning a mixture of eight
Gaussian spreads in a circle.
confirmed the observation from (Metz et al., 2016) in that
GAN easily suffered from mode collapse while unrolled
GAN is effective in solving this problem.
We observed an interesting behavior in the proposed model.
Those of existing studies (Arjovsky et al., 2017) and (Donahue et al., 2016) solved the mode collapse similarly to
unrolled GAN; they covered the entire region of the distribution and then gradually localized the modes. However,
RFGAN learned each mode first and then escaped from
mode collapse when representative and features are balanced. This phenomenon is related to the characteristics
of the proposed model in that the formulation of RFGAN
minimizes the reverse KL divergence but, at the same time,
is influenced by the representative features derived from
the forward KL divergence. When the representative features no long distinguished between real and fake input,
we can interpret this as the generator at this stage had the
representation power of the representative features. In other
words, the generator learned the entire mode as much as
the representative features, and then escaped mode collapse.
This is why RFGAN first responded similarly to GAN but
gradually produced the entire mode.
4.2. Quantitative evaluation
Since RFGAN is built upon the baseline architecture with
their suggested hyper-parameters, the input dimensionality
is set at (64, 64, 3), which is acceptable for handling the
CelebA and LSUN datasets. Yet, for CIFAR-10 dataset,
we modified the network dimensions to fit the input into
(32, 32, 3) for a fair and coherent comparison with previous studies. In addition, we drew 500 k images randomly
from the LSUN-bedroom dataset for efficient training and
comparison purposes.
We used two different metrics to measure 1) the visual
quality and 2) the diversity of data generation. An inception
score (Salimans et al., 2016) is introduced to measure the
visual quality of the GANs using CIFAR-10 datasets with a
larger score representative of higher quality. To evaluate the
Inception score
DCGAN DCGAN-RF LSGAN LSGAN-RF
6.5050
6.6349
5.9843
6.2791
Figure 4. CIFAR10 Inception score DCGAN and LSGAN
with/without the representative feature.
diversity of the GANs, the MS-SSIM (Odena et al., 2016) is
popularly used, with which the smaller the MS-SSIM value,
the better the performance in producing diverse samples.
The inception score correlates well with quality evaluation
of human annotators (Salimans et al., 2016), and so is widely
used for assessing the visual quality of the generated samples from GANs. Analogous to (Salimans et al., 2016),
we computed the inception score for 50 k samples generated from GANs. To compare with existing GANs, we
just used the DCGAN-based architecture. Yet, in order to
show that the proposed algorithm is extendable to different GAN architectures, we applied our framework to other
state-of-the-art GANs as well. For that, we modified their
discriminators by adding representative features and trained
with the same hyper-parameters initially proposed in their
respective studies. From among the most recent works, we
chose LSGAN (Mao et al., 2017), DRAGAN (Kodali et al.,
2017), and WGAN-GP (Gulrajani et al., 2017) for comparison purposes. In the case of WGAN-GP, the generator is
updated once after the discriminator is updated five times.
Following the reference code1 , other networks are trained
by updating the generator twice and the discriminator once.
This scheme made sure that discriminator loss did not vanish
(i.e. the loss did not become zero), which generally yields
better performance.
Figure 4 shows plots of the inception scores as a function
of epoch. We compared DCGAN, DCGAN-RF, LSGAN,
and LSGAN-RF, where any extension with our representative feature is named with the postfix of RF. From this
comparison, we observed that DCGAN-RF and LSGANRF outperformed the DCGAN and LSGAN, respectively,
in terms of inception scores. Moreover, we consistently
1
https://github.com/carpedm20/DCGAN-tensorflow
Submission and Formatting Instructions for ICML 2018
Table 1. MS-SSIM values. Real dataset scores 0.372669. Note
that MS-SSIM value ranged between 0.0 and 1.0 and scores low
when samples generated from GANs with the higher diversity.
O RIGINAL
WITH RF
DCGAN LSGAN DRAGAN WGAN-GP
0.4432
0.3907
0.3869
0.3813
0.4038
0.3770
0.3683
0.3773
DRAGAN and DRAGAN-RF, we observed that the score
gap between them increased, which is expected because the
performance of DRAGAN approached that of DCGAN as
the effect of the gradient penalty significantly decreased.
However, since WGAN-GP replaces the weight clipping
in WGAN with a gradient penalty, it does not satisfy the
k-Lipschitz constraint with a low gradient penalty, which
led to the degraded performance of WGAN-GP. Thus, it
is difficult to confirm the tendency of our representative
feature upon various coefficients for WGAN-GP.
GP DRAGAN DRAGAN-RF WGAN-GP WGAN-GP-RF
Inception 0.1
6.3191
6.5314
5.7951
5.9141
Score
10
6.4783
6.4905
6.2680
6.2699
Figure 5. CIFAR10 Inception score DRAGAN (top) and WGANGP (bottom) with/without the representative feature with two gradient penalty coefficients: 0.1 and 10.
found that the inception score of DCGAN-RF and LSGANRF grew faster than DCGAN and LSGAN, respectively,
which justified that the training efficiency is improved by
applying our representative feature; the proposed algorithm
approached faster than the baseline GAN to achieve the
same visual quality.
The other baseline GANs are DRAGAN and WGAN-GP,
which have recently been proposed using the gradient
penalty as a regularization term for training the discriminator. We also extended them using our representative feature and compared our extensions with the original baseline
GANs (Figure 5 shows the results). Our modification still
improved the inception scores although the improvement
is not as significant as with DCGAN-RF and LSGAN-RF
if the coefficient of gradient penalty is 10. Interestingly,
we found that this coefficient played an important role in
increasing the inception score: the greater the coefficient,
the stronger the gradient penalty. When the gradient penalty
term became stronger, training the discriminator is disturbed
because the gradient update is directly penalized. When
setting the coefficient of gradient penalty to 0.1 for both
By comparing the inception scores, we observed that
DCGAN-RF produced the best score among the others,
including existing GANs with or without our representative
feature. This is analogous to the observations by (Lucic
et al., 2017) in that DCGAN is the most effective model for
high quality image generation. On average, our achievement
over the baseline GAN is approximately 0.128, which is a
considerable gain because this value is either greater than
or similar to the differences between the different GANs.
Moreover, the improvement is noticeable between LSGAN
and LSGAN-RF.
MS-SSIM computes the similarity between image pairs randomly drawn from generated images. (Odena et al., 2016)
first introduced it as a suitable measure for diversity of image
generation. As discussed in (Fedus et al., 2017), it is important to note that MS-SSIM is meaningless if the dataset
is already highly diverse. Therefore, we used the CelebA
dataset only to compare MS-SSIM values since CIFAR10 is composed of different classes, thus already includes
highly diverse samples, and LSUN-bedroom also exhibits
various views and structures, so has a diverse data distribution. For the comparison of MS-SSIM, we again chose four
existing GANs as baseline algorithms (DCGAN, LSGAN,
DRAGAN, and WGAN-GP), and compared them with their
RFGAN variants (DCGAN-RF, LSGAN-RF, DRAGANRF, and WGAN-GP-RF). RFGAN considerably improved
the diversity compared with the baseline GANs, and this
tendency is consistent over all cases, even in the presence
of the gradient penalty term; their MS-SSIM scores are
summarized in Table 1.
From Table 1, the scores from LSGAN-RF, WGAN-GPRF, and DRAGAN-RF are close to the score of the real
dataset diversity, meaning that the generator is able to produce diverse samples reasonably well. Among all of them,
DRAGAN-RF attained the best MS-SSIM performance by
Submission and Formatting Instructions for ICML 2018
generating the most diverse samples. A comparison of DCGAN and DCGAN-RF demonstrated the most notable improvement with the representative feature because DCGAN
most frequently suffers from mode collapse, and so based
on this experimental study, we can confirm that RFGAN is
effective in solving this problem.
4.3. Qualitative evaluation
We compared the generated images from DCGAN and RFGAN from the same training iteration and found that our
RFGAN clearly produced better results, as shown in Figure 6. Our achievement can be interpreted as speeding up
the training process because the visual quality of RFGAN
is similar to the results from the later iterations of DCGAN.
This result is consistent with the quantitative evaluations in
Figure 4.
Because we reused the training data for extracting the representative features, it is possible to question if the gain
came from overfitting of the training data. To justify that
our achievement is not the result of data overfitting, we
generated samples by walking in latent space, which is
demonstrated in Figure 7. According to (Radford et al.,
2015), (Bengio et al., 2013), and (Dinh et al., 2016), the
interpolated images between two images in latent space do
not have meaningful connectivity (i.e. there is a lack of
smooth transitions). From Figure 7, we can confirm that
RFGAN learned the meaningful landscape in latent space
because of it producing natural interpolations of various
examples. Consequently, we conclude that RFGAN did not
overfit the training data.
Figure 6. Stepwise visual quality comparison between generated
images using DCGAN and RFGAN trained with CelebA(Left) and
LSUN(right).
5. Conclusions
In this study, we develop an improved technique for stabilizing GAN training for breaking the trade-off between the
visual quality and image diversity. While existing GANs explicitly add regularization terms (e.g., the gradient penalty)
for improving the training stability, our approach implicitly
hinder the fast growth of the discriminator updates, thus
achieving stable training. To this end, we employ the representative features from representative features using an
AE pre-trained with the real data. Because representative
features slow down the updates of discriminator, they are
effective to stabilize GAN training. As a result, we successfully improve the visual quality of generated sample and
solve mode collapse. Moreover, we demonstrated that RFGAN is easily extendable to various different architectures
and is robust to parameter selection.
From extensive experimental studies, we are able to justify
that representative features are useful for training the discriminator of GANs. Although our representative features
are trained from the AE, we believe that there could be
Figure 7. Latent space interpolations from LSUN and CelebA
dataset. Left and right-most columns show randomly generated
samples by RFGAN, and the intermediate columns are the results
of linear interpolation in the latent space between them.
other types of representative features or training schemes
for further improving GAN performance. We hope that our
work can serve as a basis for further work by employing
various features or prior information to better design the
discriminator.
Submission and Formatting Instructions for ICML 2018
References
Alain, Guillaume and Bengio, Yoshua. What regularized
auto-encoders learn from the data-generating distribution.
The Journal of Machine Learning Research, 15(1):3563–
3593, 2014.
Arjovsky, Martin and Bottou, Léon. Towards principled
methods for training generative adversarial networks.
arXiv preprint arXiv:1701.04862, 2017.
Arjovsky, Martin, Chintala, Soumith, and Bottou, Léon.
Wasserstein gan. arXiv preprint arXiv:1701.07875, 2017.
Bao, Jianmin, Chen, Dong, Wen, Fang, Li, Houqiang,
and Hua, Gang. Cvae-gan: Fine-grained image generation through asymmetric training. arXiv preprint
arXiv:1703.10155, 2017.
Bengio, Yoshua, Mesnil, Grégoire, Dauphin, Yann, and
Rifai, Salah. Better mixing via deep representations.
In Proceedings of the 30th International Conference on
Machine Learning (ICML-13), pp. 552–560, 2013.
Dumoulin, Vincent, Belghazi, Ishmael, Poole, Ben, Lamb,
Alex, Arjovsky, Martin, Mastropietro, Olivier, and
Courville, Aaron. Adversarially learned inference. arXiv
preprint arXiv:1606.00704, 2016.
Fedus, William, Rosca, Mihaela, Lakshminarayanan, Balaji, Dai, Andrew M, Mohamed, Shakir, and Goodfellow, Ian. Many paths to equilibrium: Gans do not need
to decrease adivergence at every step. arXiv preprint
arXiv:1710.08446, 2017.
Goodfellow, Ian. Nips 2016 tutorial: Generative adversarial
networks. arXiv preprint arXiv:1701.00160, 2016.
Goodfellow, Ian, Pouget-Abadie, Jean, Mirza, Mehdi, Xu,
Bing, Warde-Farley, David, Ozair, Sherjil, Courville,
Aaron, and Bengio, Yoshua. Generative adversarial nets.
In Ghahramani, Z., Welling, M., Cortes, C., Lawrence,
N. D., and Weinberger, K. Q. (eds.), Advances in Neural Information Processing Systems 27, pp. 2672–2680.
Curran Associates, Inc., 2014.
Berthelot, David, Schumm, Tom, and Metz, Luke. Began:
Boundary equilibrium generative adversarial networks.
arXiv preprint arXiv:1703.10717, 2017.
Gulrajani, Ishaan, Ahmed, Faruk, Arjovsky, Martin, Dumoulin, Vincent, and Courville, Aaron. Improved training
of wasserstein gans. arXiv preprint arXiv:1704.00028,
2017.
Cao, Yun, Zhou, Zhiming, Zhang, Weinan, and Yu, Yong.
Unsupervised diverse colorization via generative adversarial networks. arXiv preprint arXiv:1702.06674, 2017.
Kodali, Naveen, Abernethy, Jacob, Hays, James, and Kira,
Zsolt. On convergence and stability of gans. arXiv
preprint arXiv:1705.07215, 2017.
Che, Tong, Li, Yanran, Jacob, Athul Paul, Bengio, Yoshua,
and Li, Wenjie. Mode regularized generative adversarial
networks. arXiv preprint arXiv:1612.02136, 2016.
Krizhevsky, Alex and Hinton, Geoffrey. Learning multiple
layers of features from tiny images. 2009.
Chen, Lin, Rottensteiner, Franz, and Heipke, Christian.
Feature descriptor by convolution and pooling autoencoders. The International Archives of Photogrammetry,
Remote Sensing and Spatial Information Sciences, 40(3):
31, 2015.
Chen, Xi, Chen, Xi, Duan, Yan, Houthooft, Rein, Schulman,
John, Sutskever, Ilya, and Abbeel, Pieter. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. In Lee, D. D., Sugiyama,
M., Luxburg, U. V., Guyon, I., and Garnett, R. (eds.),
Advances in Neural Information Processing Systems 29,
pp. 2172–2180. Curran Associates, Inc., 2016.
Dinh, Laurent, Sohl-Dickstein, Jascha, and Bengio, Samy.
Density estimation using real nvp. arXiv preprint
arXiv:1605.08803, 2016.
Donahue, Jeff, Krähenbühl, Philipp, and Darrell, Trevor. Adversarial feature learning. CoRR, abs/1605.09782, 2016.
URL http://arxiv.org/abs/1605.09782.
Ledig, Christian, Theis, Lucas, Huszár, Ferenc, Caballero,
Jose, Cunningham, Andrew, Acosta, Alejandro, Aitken,
Andrew, Tejani, Alykhan, Totz, Johannes, Wang, Zehan, et al. Photo-realistic single image super-resolution
using a generative adversarial network. arXiv preprint
arXiv:1609.04802, 2016.
Liu, Ziwei, Luo, Ping, Wang, Xiaogang, and Tang, Xiaoou.
Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer
Vision, pp. 3730–3738, 2015.
Lucic, Mario, Kurach, Karol, Michalski, Marcin, Gelly,
Sylvain, and Bousquet, Olivier. Are gans created equal?
a large-scale study. arXiv preprint arXiv:1711.10337,
2017.
Mao, Xudong, Li, Qing, Xie, Haoran, Lau, Raymond YK,
Wang, Zhen, and Smolley, Stephen Paul. Least squares
generative adversarial networks. In 2017 IEEE International Conference on Computer Vision (ICCV), pp. 2813–
2821. IEEE, 2017.
Submission and Formatting Instructions for ICML 2018
Metz, Luke, Poole, Ben, Pfau, David, and Sohl-Dickstein,
Jascha. Unrolled generative adversarial networks. arXiv
preprint arXiv:1611.02163, 2016.
Nowozin, Sebastian, Cseke, Botond, and Tomioka, Ryota.
f-gan: Training generative neural samplers using variational divergence minimization. In Lee, D. D., Sugiyama,
M., Luxburg, U. V., Guyon, I., and Garnett, R. (eds.),
Advances in Neural Information Processing Systems, pp.
271–279. Curran Associates, Inc., 2016.
Odena, Augustus, Olah, Christopher, and Shlens, Jonathon.
Conditional image synthesis with auxiliary classifier gans.
arXiv preprint arXiv:1610.09585, 2016.
Radford, Alec, Metz, Luke, and Chintala, Soumith. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint
arXiv:1511.06434, 2015.
Salimans, Tim, Goodfellow, Ian, Zaremba, Wojciech, Cheung, Vicki, Radford, Alec, Chen, Xi, and Chen, Xi.
Improved techniques for training gans. In Lee, D. D.,
Sugiyama, M., Luxburg, U. V., Guyon, I., and Garnett, R.
(eds.), Advances in Neural Information Processing Systems 29, pp. 2234–2242. Curran Associates, Inc., 2016.
Vincent, Pascal, Larochelle, Hugo, Bengio, Yoshua, and
Manzagol, Pierre-Antoine. Extracting and composing
robust features with denoising autoencoders. In Proceedings of the 25th International Conference on Machine Learning, ICML ’08, pp. 1096–1103, New York,
NY, USA, 2008. ACM. ISBN 978-1-60558-205-4. doi:
10.1145/1390156.1390294.
Vincent, Pascal, Larochelle, Hugo, Lajoie, Isabelle, Bengio,
Yoshua, and Manzagol, Pierre-Antoine. Stacked denoising autoencoders: Learning useful representations in a
deep network with a local denoising criterion. Journal of
Machine Learning Research, 11(Dec):3371–3408, 2010.
Warde-Farley, David and Bengio, Yoshua. Improving generative adversarial networks with denoising feature matching. 2016.
Wei, Hao, Seuret, Mathias, Chen, Kai, Fischer, Andreas,
Liwicki, Marcus, and Ingold, Rolf. Selecting autoencoder
features for layout analysis of historical documents. In
Proceedings of the 3rd International Workshop on Historical Document Imaging and Processing, pp. 55–62, New
York, NY, USA, 2015. ACM.
Yeh, Raymond A, Chen, Chen, Lim, Teck Yian, Schwing,
Alexander G, Hasegawa-Johnson, Mark, and Do, Minh N.
Semantic image inpainting with deep generative models.
In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pp. 5485–5493, 2017.
Yu, Fisher, Seff, Ari, Zhang, Yinda, Song, Shuran,
Funkhouser, Thomas, and Xiao, Jianxiong. Lsun: Construction of a large-scale image dataset using deep
learning with humans in the loop. arXiv preprint
arXiv:1506.03365, 2015.
Zhao, Junbo, Mathieu, Michael, and LeCun, Yann. Energybased generative adversarial network. arXiv preprint
arXiv:1609.03126, 2016.
Zhou, Guanyu, Sohn, Kihyuk, and Lee, Honglak. Online
incremental feature learning with denoising autoencoders.
In Lawrence, Neil D. and Girolami, Mark (eds.), Artificial
Intelligence and Statistics, pp. 1453–1461, La Palma,
Canary Islands, 2012. PMLR.
| 1cs.CV
|
1
On Error Detection in Asymmetric Channels
arXiv:1706.04540v2 [cs.IT] 9 Sep 2017
Mladen Kovačević, Member, IEEE
Abstract—We study the error detection problem in q -ary
asymmetric channels wherein every input symbol xi is mapped
to an output symbol yi satisfying yi ≥ xi . A general setting
is assumed where the noise vectors are (potentially) restricted
in: 1) the amplitude, yi − xi ≤ a
a, 2) the P
Hamming weight,
P
n
n
1
≤
h
h,
and
3)
the
total
weight,
{y
=
6
x
}
i=1
i
i
i=1 (yi − xi ) ≤ tt.
Optimal codes detecting these types of errors are described for
certain sets of parameters a, h, tt, both in the standard and in the
mod qq) version of the problem. It is also demonstrated
cyclic (mod
that these codes are optimal in the large alphabet limit for every
a, h, t and every block-length n .
(a) (2, 1, ·)-asymmetric errors.
(b) (·, ·, 2)-asymmetric errors.
(c) (2, ·, 3)-asymmetric errors.
(d) (2, ·, ·)-asymmetric errors.
Index Terms—Asymmetric channel, limited-magnitude error,
flash memory, noisy typewriter, optimal code, lattice packing.
I. I NTRODUCTION
P
HYSICAL properties of certain optical and flash memory
devices that are used for information storage and transmission are such that the received symbol (voltage, number of
photons, etc.) can never be larger than the corresponding transmitted symbol.1 For this reason, the resulting communication
models are usually referred to as asymmetric channels. In the
present paper we study the problem of error detection in such
channels. In particular, we analyze the effect of the amplitude,
the Hamming weight, and the total weight of noise vectors on
the size of optimal error-detecting codes. This approach unifies
and generalizes several known error models from the literature.
Our main results are proofs of optimality of a family of errordetecting codes for some classes of channel parameters, and
a proof that the same family is optimal in the limit of large
alphabets for all channel parameters and all block-lengths.
The paper is structured as follows. In Section II we give
a description of the type of asymmetric channels we have in
mind and the problem that will be analyzed in the sequel, as
well as a brief overview of the relevant literature. Sections
III, IV, V contain our main results concerning optimal errordetecting codes for asymmetric channels with infinite, finite,
and cyclic alphabets, respectively. A brief conclusion and some
pointers for further work are stated in Section VI.
II. M ODEL DESCRIPTION
AND PROBLEM FORMULATION
Let A denote the channel alphabet, which we shall take to be
either Z—the set of all integers—or a subset of Z of the form
{0, 1, . . . , q−1}. For any input vector x = (x1 , . . . , xn ) ∈ An ,
the channel outputs a vector y = (y1 , . . . , yn ) ∈ An satisfying
Date: September 9, 2017.
This work was supported by the Singapore Ministry of Education (MoE)
Tier 2 grant “Network Communication with Synchronization Errors: Fundamental Limits and Codes” (Grant number R-263-000-B61-112).
The author is with the Department of Electrical & Computer Engineering,
National University of Singapore (email: [email protected]).
1 For convenience, we assume that the received symbol cannot be smaller
than the transmitted one, which is of course equivalent.
Fig. 1: The region Output(0)—the set of channel outputs
that can be produced by the input vector 0 (depicted as black
dot)—for various channel parameters and block-length n = 2.
the following conditions:
1) 0P≤ yi − xi ≤ a,
n
2) Pi=1 1{yi 6=xi } ≤ h,
n
3)
i=1 (yi − xi ) ≤ t.
Hence, we consider asymmetric channels with additional constraints imposed on: 1) the amplitude of the noise at each
coordinate (so-called limited-magnitude errors), 2) the number
of symbols hit by noise, i.e., the Hamming weight of the error
vector y − x, and 3) the total weight of the error vector. The
error vectors satisfying the above constraints will be referred
to as (a, h, t)-asymmetric errors. Note that the situations where
only some of the above three constraints are imposed on
the noise are special cases of our setting. Namely, by taking
a = t the constraint 1) is effectively excluded (i.e., becomes
redundant), as it is when a = q − 1 in the finite alphabet case.
Similarly, one can exclude constraint 2) by setting h = n,
and constraint 3) by setting t = ah. We shall refer to the
(q − 1, n, t)-asymmetric errors (when only the constraint 3)
is in effect) as the (·, ·, t)-asymmetric errors, and similarly for
the other cases. Hence, the ‘·’ indicates that the corresponding
constraint is either redundant, or is not being considered at all.
Convention: To avoid discussing trivial cases, as well as
to simplify the exposition, we shall assume hereafter that
q, n, a, h, t are positive integers satisfying a ≤ q − 1 (in the
finite alphabet case), a ≤ t ≤ ah, and h ≤ n.
N
For x ∈ An , denote by Output(x) the set of all channel
outputs y ∈ An that can be produced by the input x and
the noise vectors satisfying the constraints 1)–3) (see Figure
1). The dependence of Output(x) on the parameters a, h, t is
suppressed for notational simplicity; this should not cause any
confusion.
2
A code C ⊆ An is said to detect certain error patterns if
no codeword x can produce another codeword y 6= x at the
channel output, when any of those error patterns occur. In
symbols, if x ∈ C and y ∈ Output(x), y 6= x, then y ∈
/ C.
This ensures that the received vector is either the transmitted
codeword, or not a codeword at all, meaning that the receiver
can unambiguously determine whether an error has happened
during transmission. Our object of study in this paper are codes
detecting (a, h, t)-asymmetric errors;2 they will be referred to
as (a, h, t)-asymmetric-error-detecting codes, or (a, h, t)-AED
codes for short.
Related work
The work most closely related to ours, in which optimal
(·, ·, t)-AED codes over q-ary alphabets were determined, is
[3]. In fact, we show that the same family of codes that
was studied there remains optimal in the case of (a, h, t)asymmetric errors in some instances. We also extend and
generalize the results of [3] to infinite alphabets.
The mentioned construction from [3] was also used in [1]
for the (a, ·, ·) case, and a very similar construction (also for
the (a, ·, ·) case, but in a quite different setting) appears in
[13]. Systematic (a, ·, ·)-AED codes (as well as (a, h, ·)-AED
codes) were studied in [9].
Some other works on the error detection problem in asymmetric channels should also be mentioned, e.g., [2], [4], [5],
[6], [7], [14]. All of these works are focused on special cases of
the model introduced above—(·, ·, t), (a, h, ·), binary alphabet,
etc. In [16], a generalization of the (·, ·, t) error model is
studied where both positive and negative errors are allowed.
Finally, for a study of the error correction problem in
asymmetric channels we refer the reader to [10]; see also [8],
[11], [12], [15], [17] for some of the more recent works.
III. A SYMMETRIC CHANNELS WITH INFINITE ALPHABET
We first consider the case when the channel alphabet is the
set of all integers Z. Though this is clearly not practically
motivated, there are several reasons why these results are
relevant for the study of communication over asymmetric
channels. First, this will provide geometric intuition about the
problem and allow us to “visualize” AED codes. Second, the
codes described below will be shown optimal for all channel
parameters in the infinite alphabet case, suggesting that the
corresponding codes in the finite alphabet case are nearly
optimal, at least in some asymptotic regimes. In fact, as we
already mentioned, they will be proven optimal in some special
instances of the finite alphabet case as well. Third, infinite
alphabet can be seen as a limiting case and an approximation
of a finite alphabet; this is relevant in situations where the
alphabet size q is large compared to the “error radius” t.
Before stating the results we need a few definitions. To
quantify what it means for a code in Zn to be optimal, we
define the density of C ⊆ Zn as follows:
|C ∩ {−k, . . . , k}n |
.
(1)
µ(C) := lim
k→∞
(2k + 1)n
2 Codes
detecting asymmetric errors can also detect so-called unidirectional
errors—errors which are either positive or negative within a codeword, but
the sign is not known in advance; see, e.g., [5].
This parameter represents the infinite-space analog of the
cardinality of codes in finite spaces. In case the limit in (1)
does not exist, one can naturally define the upper (µ(C)) and
the lower (µ(C)) density by replacing lim with lim sup and
lim inf, respectively. We say that C is an optimal (a, h, t)-AED
code in Zn if no (a, h, t)-AED code in Zn has upper density
larger than µ(C). A code C ⊆ Zn is said to be linear if it is
a sublattice of Zn , i.e., if (C, +) is a subgroup of (Zn , +).
For a linear code we have µ(C) = |Zn1/C| , where Zn /C is the
quotient group of the lattice C.
For S ⊂ Zn , we say that (S, C) is a packing in Zn if the
translates x + S and y + S are disjoint for any two distinct
codewords x, y ∈ C (here x + S = {x + s : s ∈ S}). If (S, C)
is a packing and 0 ∈ S, then each of these translates contains
1
exactly one codeword, and so we must have µ(C) ≤ |S|
.
The following claim gives an upper bound on the density
of (a, h, t)-AED codes in Zn .
Theorem 1: Let C be an (a, h, t)-AED code in Zn . Then
1
µ(C) ≤ t+1
.
Proof: Let ei be the unit vector having a 1 at the i’th
coordinate and 0’s elsewhere. Observe the following vectors:
0,
e1 ,
2 · e1 ,
a · e1 + e2 ,
..
.
...,
...,
a · e1 ,
a · e1 + a · e2 ,
(2)
a · e1 + a · e2 + · · · + a · ek−1 + αk · ek
where the list extends until one of the constraints on the noise
is violated.
Ph In other words, the last vector on the list is of the
form i=1 αi · ei , where, for some k ∈ {1, . . . , h}, αi = a
Pk
for all i < k, αi = 0 for all i > k, and
i=1 αi = t.
Denote the set of all vectors on the resulting list by S, and
note that |S| = t + 1. An important observation about this set
is that, for any two vectors f , g ∈ S, where f precedes g on
the list (2),Pthe vector g − f satisfies all the noise
Phconstraints:
h
g−f =
i=1 γi ≤ t.
i=1 γi · ei with 0 ≤ γi ≤ a and
This means that, for any two such vectors f , g we must have
g − f ∈ Output(0).
Now, let C ⊆ Zn be an (a, h, t)–AED code, |C| ≥ 2. We claim
that (S, C) is a packing in Zn . Suppose that this is not the case,
i.e., that x + f = y + g for two distinct codewords x, y ∈ C
and two distinct vectors f , g ∈ S. Without loss of generality,
we can assume that f precedes g on the list (2). We then have
g−f ∈ Output(0) and so x = y+g−f ∈ Output(y), which
means that the code C is not (a, h, t)-AED, a contradiction.
Therefore, any (a, h, t)-AED code defines a packing of the set
1
1
= t+1
.
S in Zn , and so its density cannot exceed |S|
We next give an explicit construction of linear codes
achieving the upper bound just derived. A particular such code
is depicted in Figure 2.
Theorem 2: The code
)
(
n
X
n
xi ≡ 0 mod (t + 1)
(3)
C(n; t) := x ∈ Z :
i=1
is an optimal (a, h, t)-AED code in Zn .
3
construction. In other words, we consider codes in [q]n of the
form (z + C(n; t)) ∩ [q]n , for an arbitrary vector z ∈ Zn . The
resulting family of codes can be written as
)
(
n
X
(j)
n
xi ≡ j mod (t + 1) , (5)
Cq (n; t) := x ∈ [q] :
i=1
(j)
Fig. 2: The code C(2; 2) ⊆ Z2 and an illustration of its (2, 1, ·)AED property. Codewords are depicted as black dots.
Proof: Due to our assumptions about the noise, each
of the allowed error vectors from P
Output(0) can change
n
the sum of the codeword symbols,
i=1 xi , by at most t.
Therefore, no codeword of C(n; t), other than the one that
was transmitted, can be produced at the output of the channel,
proving that this code is indeed (a, h, t)-AED. To demonstrate
its optimality, observe that the quotient group of the lattice
C(n; t) is Zn /C(n; t) ∼
= Zt+1 . The density of C(n; t) is
1
1
therefore µ (C(n; t)) = |Zt+1
| = t+1 , which is by Theorem
1 the largest possible value.
also be written in the form C(n; t) =
The code C(n; t) can
ξ · G(n; t) : ξ ∈ Zn , where G(n; t) is an n × n generator
matrix:
t+ 1 0 0 ··· 0
−1 1 0 · · · 0
(4)
G(n; t) = −1 0 1 · · · 0 .
..
..
.. .. . .
.
.
.
. .
−1
0
0 ···
1
Remark 3: Note that the codes C(n; t)—optimal (a, h, t)AED codes in Zn —do not depend on the parameters a, h,
a somewhat counter-intuitive fact. Hence, detecting (·, ·, t)asymmetric errors incurs no loss in code efficiency compared
to the case of detecting more restrictive (a, h, t)-asymmetric
errors. The corresponding statement for error-correcting codes
is in general false. Note that the density of C(n; t) is independent of the block-length n as well.3
N
IV. A SYMMETRIC
CHANNELS WITH FINITE ALPHABET
The main idea in constructing codes over finite alphabets
is simple: take an (a, h, t)-AED code in Zn and restrict it to
the hypercube {0, 1, . . . , q − 1}n . This will clearly yield an
(a, h, t)-AED code with alphabet [q] := {0, 1, . . . , q − 1}. We
say that C is an optimal (a, h, t)-AED code in [q]n if it has
the largest cardinality among all (a, h, t)-AED codes in [q]n .
Since the code C(n; t) from (3) is optimal in Zn , it is
natural to take it, or any of its translations, as the basis for
3 In the finite alphabet case, it is known that the number of redundant
symbols of an optimal (·, ·, t)-AED code depends only on t and not on the
block-length n, see for example [3], [4], [5].
where j ∈ {0, 1, . . . , t}. The cardinality of Cq (n; t) is maximized (with respect to j) for j = j ∗ := ⌊ (q−1)n
⌋ mod (t + 1),
2
see [3].
(j ∗ )
The codes Cq (n; t) are known [3] to be optimal (·, ·, t)AED codes for every q, n, t. We prove below their optimality
in some other cases as well. The proof method from [3],
however, does not seem to be applicable to these cases due to
a different shape of the regions Output(x).
Theorem 4: Let C be a (·, h, t)-AED code in [q]n . Then
q
|C| ≤ q n−1 ⌈ t+1
⌉.
Proof: Partition the space [q]n into q n−1 “lines”, each
containing q points whose coordinates 2, . . . , n are fixed and
the first coordinate varies through [q]. If x = (x1 , x2 , . . . , xn )
is a codeword of a (·, h, t)-AED code, then (x1 +k, x2 , . . . , xn )
cannot be a codeword for any 1 ≤ k ≤ t, because
(x1 + k, x2 , . . . , xn ) ∈ Output(x). This implies that each of
q
⌉ codewords, and
the mentioned lines contains at most ⌈ t+1
q
n−1
therefore |C| ≤ q
⌈ t+1 ⌉.
Theorem 5: Suppose that t + 1 divides q. Then the codes
(j)
Cq (n; t) are optimal (·, h, t)-AED codes in [q]n . Their cardi(j)
qn
nality is Cq (n; t) = t+1
, ∀j ∈ {0, 1, . . . , t}.
Proof: For every choice of the values x1 , . . . , xn−1 ∈ [q]
q
there are t+1
Pnpossible values of xn ∈ [q] satisfying the
congruence
i=1 xi ≡ j mod (t + 1), for any fixed j.
(j)
q
for all j ∈ {0, 1, . . . , t},
Therefore, |Cq (n; t)| = q n−1 · t+1
which is by Theorem 4 the largest possible value.
The following statement refers to codes detecting a single
(h = 1) asymmetric error, with no bound on the amplitude
of the error, other than the implicit one a ≤ q − 1. It
is a special case of Theorem 5 but we state it separately
nonetheless because the corresponding problem for error
correction has been studied in some detail in the literature
[11], [12], [18].
(j)
Corollary 6: The codes Cq (n; q − 1) are optimal (·, 1, ·)(j)
AED codes in [q]n . Their cardinality is Cq (n; q−1) = q n−1 ,
∀j ∈ {0, 1, . . . , q − 1}.
Proof: Take t = q − 1 in Theorem 5 and notice that
(·, 1, q − 1)-asymmetric errors are in fact (·, 1, ·)-asymmetric
errors.
(j)
It should be noted that the codes Cq (n; t) are not optimal
for general (a, h, t)-asymmetric errors—counterexamples can
be constructed for small values of these parameters (see [1]
for a counterexample for the (a, ·, ·) case).
V. A SYMMETRIC CHANNELS WITH CYCLIC ALPHABET
In this section we discuss briefly the cyclic version of the
asymmetric channel (as in, e.g., [11], [17]). Our motivating
example is the so-called noisy typewriter channel wherein each
transmitted symbol xi is received as either xi , or xi +1 mod q.
Let +q denote addition modulo q. The cyclic asymmetric
channel we have in mind is defined as follows: for any input
4
vector x ∈ [q]n the channel outputs y = x +q f ∈ [q]n ,
where f ∈ [q]n is an arbitrary noise vector satisfying the
constraints 1)–3) described in Section II. Hence, the model
is the same as before, the only difference being that the sum
of the input vector and the noise vector is now taken mod q;
in other words, we now allow the errors to “wrap around”. To
distinguish between cyclic and non-cyclic cases, we shall refer
to the errors just described as (a, h, t)◦ -asymmetric errors, and
similarly for the corresponding codes.
The code space in this setting can be represented as the torus
Znq in which there are no “boundary effects” that are present
in the non-cyclic case. This enables one to derive a simple
upper bound on the cardinality of optimal codes by using a
method identical to the one used for the infinite alphabet case.
Theorem
7: Let C be an (a, h, t)◦ -AED code in [q]n . Then
qn
.
|C| ≤ t+1
Proof: Analogous to the proof of Theorem 1.
We next identify a class of parameters for which the above
bound is tight. (As we shall point out in Section VI, it cannot
be tight in general.)
Theorem 8: Suppose that t + 1 divides q. Then the codes
(j)
Cq (n; t) are optimal (a, h, t)◦ -AED codes in [q]n .
(j)
Proof: We have shown in Theorem 5 that Cq (n; t) =
n
q
t+1 when t+1 divides q, which is by Theorem 7 the maximum
possible cardinality of an (a, h, t)◦ -AED code in [q]n . It is left
(j)
to prove that Cq (n; t) are indeed (a, h, t)◦ -AED. We prove
this fact below for j = 0; the statement for an arbitrary j is
an easy consequence.
The key observation is that, when t + 1
(0)
divides q, x ∈ Zn : ∃x′ ∈ Cq (n; t) s.t. x ≡ x′ mod q =
C(n; t), which follows from the definition of the codes C(n; t)
(0)
and Cq (n; t). In other words, the code C(n; t) is a periodic
(0)
extension to Zn of the code Cq (n; t). With this interpretation
(0)
in mind it is easy to see that the statement that Cq (n; t) is
◦
(a, h, t) -AED is equivalent to the statement that C(n; t) is
(a, h, t)-AED, which we already know is true.
(j)
Corollary 9: The codes Cq (n; q − 1) are optimal (·, 1, ·)◦ n
AED codes in [q] .
VI. C ONCLUDING
REMARKS AND FURTHER WORK
The problem we have addressed in this letter is that of
finding optimal error-detecting codes for asymmetric channels
with various constraints on the noise. The solution has been
obtained in the infinite alphabet case for all parameters, but for
finite alphabets the general question is still open. The task of
settling it for every q, n, a, h, t may turn out to be too difficult
and it is instructive to focus on asymptotic optimality instead.
For example, observe the regime where q, a, h, t are fixed
(j)
and n → ∞. Note that Cq (n; t) : j ∈ {0, 1, . . . , t}
n
is a partition of [q] , meaning that the codes in this set
are mutually disjoint and their union is all of [q]n . It is
not difficult to argue that the members of this partition are
of “approximately the same cardinality” for large n, and
(j)
qn
for any j ∈ {0, 1, . . . , t}
consequently Cq (n; t) ∼ t+1
(here an ∼ bn is a shorthand for limn→∞ abnn = 1). Whether
this family of codes is asymptotically optimal is an interesting
question which we summarize below. In light of Theorem 2
one may conjecture that the answer is positive.
Problem: Fix q, a, h, t, and let Dq (n; a, h, t) denote the
size of an optimaln (a, h, t)-AED code in [q]n . Is it true that
q
as n → ∞?
N
Dq (n; a, h, t) ∼ t+1
As for the cyclic case, nwe note that the cardinality of optimal
q
in general. For example, a code is
codes cannot scale as t+1
◦
(·, h, ·) -AED if and only if its minimum Hamming distance is
> h, and it is known that such codes cannot have size ∼ cq n
for h ≥ 2 (this follows from the sphere packing bound in the
q-ary Hamming space).
ACKNOWLEDGMENT
The author would like to thank Vincent Y. F. Tan for reading
a preliminary version of this work and for several helpful
discussions on the subject matter.
R EFERENCES
[1] R. Ahlswede, H. Aydinian, L. H. Khachatrian, and L. M. G. M.
Tolhuizen, “On q-Ary Codes Correcting All Unidirectional Errors of
a Limited Magnitude,” in Proc. Ninth Int’l Workshop Algebr. Combin.
Coding Theory, pp. 20–26, Jun. 2004.
[2] S. Al-Bassam and B. Bose, “Asymmetric/Unidirectional Error Correcting and Detecting Codes,” IEEE Trans. Comput., vol. 43, no. 5, pp.
590–597, May 1994.
[3] J. M. Borden, “Optimal Asymmetric Error Detecting Codes,” Inform.
and Control, vol. 53, no. 1-2, pp. 66–73, Apr.-May 1982.
[4] B. Bose, S. Elmougy, and L. G. Tallini, “Systematic t-Unidirectional
Error-Detecting Codes over Zm ,” IEEE Trans. Comput., vol. 56, no. 7,
pp. 876–880, Jul. 2007.
[5] B. Bose and D. J. Lin, “Systematic Unidirectional Error-Detecting
Codes,” IEEE Trans. Comput., vol. 34, no. 11, pp. 1026–1032, Nov.
1985.
[6] B. Bose and D. K. Pradhan, “Optimal Unidirectional Error Detecting/Correcting Codes,” IEEE Trans. Comput., vol. 31, no. 6, pp. 564–
568, Jun. 1982.
[7] B. Bose and T. R. N. Rao, “Theory of Unidirectional Error Correcting/Detecting Codes,” IEEE Trans. Comput., vol. 31, no. 6, pp. 521–530,
Jun. 1982.
[8] Y. Cassuto, M. Schwartz, V. Bohossian, and J. Bruck, “Codes for
Asymmetric Limited-Magnitude Errors with Application to Multilevel
Flash Memories,” IEEE Trans. Inf. Theory, vol. 56, no. 4, pp. 1582–
1595, Apr. 2010.
[9] N. Elarief, B. Bose, and S. Elmougy, “Limited Magnitude Error Detecting Codes over Zq ,” IEEE Trans. Comput., vol. 62, no. 5, pp. 984–989,
May 2013.
[10] T. Kløve, “Error Correcting Codes for the Asymmetric Channel,” Technical Report, Dept. of Informatics, University of Bergen, 1981. (Updated
in 1995.)
[11] T. Kløve, B. Bose, and N. Elarief, “Systematic, Single Limited Magnitude Error Correcting Codes for Flash Memories,” IEEE Trans. Inf.
Theory, vol. 57, no. 7, pp. 4477–4487, Jul. 2011.
[12] T. Kløve, J. Luo, I. Naydenova, and S. Yari, “Some Codes Correcting
Asymmetric Errors of Limited Magnitude,” IEEE Trans. Inf. Theory,
vol. 57, no. 11, pp. 7459–7472, Nov. 2011.
[13] M. Kovačević, M. Stojaković, and V. Y. F. Tan, “Zero-Error Capacity
of P -ary Shift Channels and FIFO Queues,” arXiv:1605.02441 [cs.IT].
[14] I. Naydenova and T. Kløve, “Generalized Bose-Lin Codes, a Class of
Codes Detecting Asymmetric Errors,” IEEE Trans. Inf. Theory, vol. 53,
no. 3, pp. 1188–1193, Mar. 2007.
[15] M. Schwartz, “Quasi-Cross Lattice Tilings with Applications to Flash
Memory,” IEEE Trans. Inf. Theory, vol. 58, no. 4, pp. 2397–2405, Apr.
2012.
[16] L. G. Tallini and B. Bose, “On L1 -Distance Error Control Codes,” in
Proc. 2011 IEEE Int. Symp. Inf. Theory (ISIT), pp. 1061–1065, SaintPetersburg, Russia, Aug. 2011.
[17] S. Yari, T. Kløve, and B. Bose, “Some Codes Correcting Unbalanced
Errors of Limited Magnitude for Flash Memories,” IEEE Trans. Inf.
Theory, vol. 59, no. 11, pp. 7278–7287, Nov. 2013.
[18] T. Zhang and G. Ge, “New Results on Codes Correcting Single Error
of Limited Magnitude for Flash Memory,” IEEE Trans. Inf. Theory, vol.
62, no. 8, pp. 4494–4500, Aug. 2016.
| 7cs.IT
|
Learning to Represent Haptic Feedback for Partially-Observable Tasks
arXiv:1705.06243v1 [cs.RO] 17 May 2017
Jaeyong Sung1,2 , J. Kenneth Salisbury1 and Ashutosh Saxena3
Abstract— The sense of touch, being the earliest sensory
system to develop in a human body [1], plays a critical part of
our daily interaction with the environment. In order to successfully complete a task, many manipulation interactions require
incorporating haptic feedback. However, manually designing a
feedback mechanism can be extremely challenging. In this work,
we consider manipulation tasks that need to incorporate tactile
sensor feedback in order to modify a provided nominal plan. To
incorporate partial observation, we present a new framework
that models the task as a partially observable Markov decision
process (POMDP) and learns an appropriate representation of
haptic feedback which can serve as the state for a POMDP
model. The model, that is parametrized by deep recurrent
neural networks, utilizes variational Bayes methods to optimize
the approximate posterior. Finally, we build on deep Q-learning
to be able to select the optimal action in each state without
access to a simulator. We test our model on a PR2 robot for
multiple tasks of turning a knob until it clicks.
Learned Haptic Representation
Learned POMDP
Haptic Sensor
Q-Learning
Modify
Path
Unknown
Internal Mechanism
I. I NTRODUCTION
Many tasks in human environments that we do without
much effort require more than just visual observation. Very
often they require incorporating the sense of touch to complete the task. For example, consider the task of turning a
knob that needs to be rotated until it clicks, like the one
in Figure 1. The robot could observe the consequence of
its action if any visible changes occur, but such clicks can
often only be directly observed through the fingers. Many
of the objects that surround us are explicitly designed with
feedback — one of the key interaction design principles —
otherwise “one is always wondering whether anything has
happened” [2].
Recently, there has been a lot of progress in making robots
understand and act based on images [3], [4], [5] and pointclouds [6]. A robot can definitely gain a lot of information
from visual sensors, including a nominal trajectory plan for
a task [6]. However, when the robot is manipulating a small
object or once the robot starts interacting with small parts
of appliances, self-occlusion by its own arms and its endeffectors limits the use of the visual information.
However, building an algorithm that can examine haptic
properties and incorporate such information to influence a
motion is very challenging for multiple reasons. First, haptic
feedback is a dynamic response that is dependent on the
action the robot has taken on the object as well as internal
states and properties of the object. Second, every haptic
sensor produces a vastly different raw sensor signal.
Moreover, compared to the rich information that can be
extracted about a current state of the task from few images
1 Department of Computer Science, Stanford University. 2 Department
of Computer Science, Cornell University. 3 Brain Of Things, Inc.
{jysung,jks,asaxena}@cs.stanford.edu
Fig. 1: Haptic feedback from a tactile sensor being used
to modify a nominal plan of manipulation. Our framework learns an appropriate representation (embedding space)
which in turn is used to learn to find optimal control.
(e.g. position and velocity information of an end-effector and
an object [5], [3]), a short window of haptic sensor signal
is merely a partial consequence of the interaction and of
the changes in an unobservable internal mechanism. It also
suffers from perceptual aliasing — i.e. many segments of a
haptic signal at different points of interaction can produce
a very similar signal. These challenges make it difficult to
design an algorithm that can incorporate information from
haptic modalities (in our case, tactile sensors).
In this work, we introduce a framework that can learn to
represent haptic feedback for tasks requiring incorporation of
a haptic signal. Since a haptic signal only provides a partial
observation, we model the task using a partially observable
Markov decision process (POMDP). However, since we do
not know of definition of states for a POMDP, we first
learn an appropriate representation from a haptic signal to
be used as continuous states for a POMDP. To overcome
the intractability in computing the posterior, we employ a
variational Bayesian method, with a deep recurrent neural
network, that maximizes lower bound of likelihood of the
training data.
Using a learned representation of the interaction with
feedback, we build on deep Q-learning [5] to identify an
appropriate phase of the action from a provided nominal
plan. Unlike most other applications of successful reinforce-
ment learning [5], [7], the biggest challenge is a lack of a
robotics simulation software that can generate realistic haptic
signals for a robot to safely simulate and explore various
combinations of states with different actions.
To validate our approach, we collect a large number
of sequences of haptic feedback along with their executed
motion for the task of ‘turning a knob until it clicks’ on
objects of various shapes. We empirically show on a PR2
robot that we can modify a nominal plan and successfully
accomplish the task using the learned models, incorporating
tactile sensor feedback on the fingertips of the robot. In
summary, the key contributions of this work are:
• an algorithm which learns task relevant representation
of haptic feedback
• a framework for modifying a nominal manipulation plan
for interactions that involves haptic feedback
• an algorithm for learning optimal actions with limited
data without simulator
II. R ELATED W ORK
Haptics. Haptic sensors mounted on robots enable many
different interesting applications. Using force and tactile
input, a food item can be classified with characteristics which
map to appropriate class of motions [8]. Haptic adjectives
such as ‘sticky’ and ‘bumpy’ can be learned with biomimetic
tactile sensors [9]. Whole-arm tactile sensing allows fast
reaching in dense clutter. We focus on tasks with a nominal
plan (e.g. [6]) but requires incorporating haptic (tactile)
sensors to modify execution length of each phase of actions.
For closed-loop control of robot, there is a long history of
using different feedback mechanisms to correct the behavior
[10]. One of the common approaches that involves contact
relies on stiffness control, which uses the pose of an endeffector as the error to adjust applied force [11], [12]. The
robot can even self-tune its parameters for its controllers
[13]. A robot also uses the error in predicted pose for force
trajectories [14] and use vision for visual servoing [15].
Haptic sensors have also been used to provide feedback. A
human operator with a haptic interface device can teleoperate
a robot remotely [16]. Features extracted from tactile sensors
can serve as feedback to planners to slide and roll objects
[17]. [18] uses tactile sensor to detect success and failure of
manipulation task to improve its policy.
Partial Observability. A POMDP is a framework for a
robot to plan its actions under uncertainty given that the
states are often only obtained through noisy sensors [19].
The framework has been successfully used for many tasks
including navigation and grasping [20], [21]. Using wrist
force/torque sensors, hierarchical POMDPs help a robot
localize certain points on a table [22]. While for some
problems [20], states can be defined as continuous robot
configuration space, it is unclear what the ideal state space
representation is for many complex manipulation tasks.
When the knowledge about the environment or states is
not sufficient, [23] use a fully connected DBN for learning
factored representation online, while [24] employ a two step
method of first learning optimal decoder then learning to
encode. While many of these work have access to a good
environment model, or is able to simulate environment where
it can learn online, we cannot explore or simulate to learn
online. Also, the reward function is not available. For training
purposes, we perform privileged learning [25] by providing
an expert reward label only during the training phase.
Representation Learning. Deep learning has recently
vastly improved the performance of many related fields such
as compute vision (e.g. [26]) and speech recognition (e.g.
[27]). In robotics, it has helped robots to better classify
haptic adjectives by combining images with haptic signals
[28], predict traversability from long-range vision [29], and
classify terrains based on acoustics [30].
For controlling robots online, a deep auto-encoder can
learn lower-dimensional embedding from images and modelpredictive-control (MPC) is used for optimal control [31].
DeepMPC [14] predicts its future end-effector position with
a recurrent network and computes an appropriate amount of
force. Convolutional neural network can be trained to directly
map images to motor torques [3], [32]. As mentioned earlier,
we only take input of haptic signals, which suffers from
perceptual aliasing, and contains a lot less information in
a single timestep compared to RGB images.
Recently developed variational Bayesian approach [33],
[34], combined with a neural network, introduces a recognition model to approximate intractable true posterior. Embedto-Control [4] learns embedding from images and transition
between latent states representing unknown dynamical system. Deep Kalman Filter [35] learns very similar temporal
model based on Kalman Filter but is used for counterfactual
inference on electronic health records.
Reinforcement learning (RL), also combined with a neural
network, has recently learned to play computer games by
looking at pixels [5], [36]. Applying standard RL to a
robotic manipulation task, however, is challenging due to
lack of suitable state space representation [32]. Also, most
RL techniques rely on trial and error [37] with the ability
to try different actions from different states and observe
reward and state transition. However, for many of the robotic
manipulation tasks that involve physical contact with the
environment, it is too risky to let an algorithm try different
actions, and reward is not trivial without instrumentation of
the environment for many tasks. In this work, the robot learns
to represent haptic feedback and find optimal control from
limited amount of haptic sequences despite lack of good
robotic simulator for haptic signal.
III. O UR A PPROACH
Our goal is to build a framework that allows robots to
represent and reason about haptic signals generated by its
interaction with an environment.
Imagine you were asked to turn off the hot plate in
Figure 1 by rotating the knob until it clicks. In order to
do so, you would start by rotating the knob clockwise or
counterclockwise until it clicks. If it doesn’t click and if
you feel the wall, you would start to rotate it in the opposite
direction. And, in order to confirm that you have successfully
𝒒𝝓,𝟏
𝒒𝝓,𝟐
𝒒𝝓,𝟑
𝒒𝝓,𝒕
not observed
𝝁
observed
𝜮
𝑸(𝒔𝒕 , 𝒂𝒕 )
⋯
observed only during training
𝒂𝟐
𝒔𝟏
𝒐𝟏
𝒂𝑻
⋯
𝒔𝟐
𝒓𝟏
𝒐𝟐
𝒓𝟐
𝒐𝟏
𝒐𝟐
𝒐𝟑
⋯
𝒔𝑻
𝒐𝑻
𝒐𝒕
𝒓𝑻
(a) Graphical Model Rep. of POMDP Model
𝒔𝒕−𝟏
𝒂𝒕
(b) Tran. Network
𝒂𝟏
𝒂𝟐
𝒂𝟑
𝒔𝒕
𝒂𝒕
(c) Deep Recurrent Recognition Network
𝒂𝒕
(d) Deep Q Network
Fig. 2: Framework Overview. We model the task that requires incorporation of tactile feedback in a partially observable
MDP (a) which its transition and emission functions are parametrized by neural networks (b). To find an appropriate
representation of states for the POMDP, we approximate the posterior with a Deep Recurrent Recognition Network (c),
consisting of two LSTM (square blocks) recurrent networks. Deep Q-Network (d), consisting of two fully connected layers,
utilizes a learned representation from (c) and a learned transition model from (a) to train Deep Q-Network (d).
completed the task or hit the wall, you would use your sense
of touch on your finger to feel a click. There could also be
a sound of a click as well as other observable consequences,
but you would not feel very confident about the click in the
absence of haptic feedback.
However, such haptic signal itself does not contain sufficient information for a robot to directly act on. It is unclear
what is the best representation for a state of the task, whether
it should only be dependent on states of internal mechanisms
of the object (which are unknown) or it should incorporate
information about the interaction as well. The haptic signal
is merely a noisy partial observation of latent states of the
environment, influenced by many factors such as a type of
interaction that is involved and a type of grasp by the robot.
To learn an appropriate representation of the state, we first
define our manipulation task as a POMDP model. However,
posterior inference on such latent state from haptic feedback
is intractable. In order to approximate the posterior, we
employ variational Bayes methods to jointly learn model
parameters for both a POMDP and an approximate posterior model, each parametrized by a deep recurrent neural
network.
Another big challenge is the limited opportunity to explore
with different policies to fine-tune the model, unlike many
other applications that employs POMDP or reinforcement
learning. Real physical interactions involving contact are too
risky for both the robot and the environment without lots
of extra safety measures. Another common solution is to
explore in a simulated environment; however, none of the
available robot simulators, as far as we are aware, are capable
of generating realistic feedback for objects of our interest.
Instead, we learn offline from previous experiments by utilizing a learned haptic representation along with its transition
model to explore offline and learn Q-function.
A. Problem Formulation
Given a sequence of haptic signals (~o = o1 , ..., ot ) up
to current time frame t along with a sequence of actions
taken (~a = a1 , ..., at ), our goal is to output a sequence of
stirrer
speaker
fan
Fig. 3: Samples of haptic signals from three different objects
with a PR2 fingertip tactile sensor. Notice a large variation
in feedback produced by what humans identify as a ‘click’.
appropriate state representations (~s = s1 , ..., st ) such that
we can take an optimal next action at+1 inferred from the
current state st .
B. Generative Model
We formulate the task that requires haptic feedback as
a POMDP model, defined as (S, A, T, R, O). S represents
a set of states, A represents a set of actions, T represents
a state transition function, R represents a reward function,
and O represents an observation probability function. Fig. 2a
represents a graphical model representation of a POMDP
model and all notations are summarized in Table I.
Among the required definitions of a POMDP model, most
importantly, state S and its representation are unknown.
Thus, all functions T, R, O that rely on states S are also
not available.
We assume that all transition and emission probabilities
are distributed as Gaussian distributions; however, they can
take any appropriate distribution for the application. Mean
and variance of each distribution are defined as a function
with input as parent nodes in the graphical model (Fig. 2a):
s1 ∼ N (0, I)
st ∼ N (fsµ (st−1 , at ), fsΣ (st−1 , at )2 I)
ot ∼ N (foµ (st ), foΣ (st )2 I)
rt ∼ N (frµ (st ), frΣ (st )2 I)
We parametrize each of these functions as a neural network.
Fig. 2b shows a two layer network for parametrization of
the transition function, and emission networks take a similar
structure. The parameters of these networks form the parameters of the generative model θ = {sµ , sΣ , oµ , oΣ , rµ , rΣ }.
TABLE I: Summary of Notations.
Notations
S
O
T
A
R
pθ
θ
qφ
φ
Q(s, a)
γ
Descriptions
continuous state space (a learned representation)
observation probability (S → O) of haptic signal
conditional probability between states (S × A → S)
a set of possible actions to be taken at each time step
a reward function (S → R)
a generative model for O and R
parameters of generative model
an approximate posterior distribution
(a recognition network for representing haptic signal)
parameters of recognition network (recurrent neural network)
an approximate action-value function (S × A → R)
a discount factor
C. Deep Recurrent Recognition Network
Due to non-linearity of multi-layer neural network, computing the posterior distribution p(~s|~o, ~r, ~a) becomes intractable [35]. The variational Bayes method [33], [34]
allows us to approximate the real posterior distribution with
a recognition network (encoder) qφ (~s|~o, ~r, ~a).
Although it is possible to build a recognition network
qφ (~s|~o, ~r, ~a) that takes the reward ~r as a part of the input,
such recognition network would not be useful during a test
time when the reward ~r is not available. Since a reward is not
readily available for many of the interaction tasks, we assume
that the sequence of rewards ~r is available only during a
training phase given by a expert. Thus, we build an encoder
qφ (~s|~o, ~a) without a reward vector while our goal will be to
reconstruct a reward ~r as well (Sec. III-D).
Among many forms and structures qφ could take,
through validation with our dataset, we chose to define
qφ,t (st |o1 , ..., ot , a1 , ..., ot ) as a deep recurrent network with
two long short-term memory (LSTM) layers as shown in
Fig. 2c.
D. Maximizing Variational Lower-bound
To jointly learn parameters for the generative θ and
the recognition network φ, our objective is to maximize
likelihood of the data:
maxθ log pθ (~o, ~r|~a)
Using a variational method, a lower bound on conditional
log-likelihood is defined as:
log pθ (~o, ~r|~a) = DKL (qφ (~s|~o, ~r, ~a)||pθ (~s|~o)) + L(θ, φ)
Algorithm 1 Deep Q-Learning in Learned Latent State Space
Dgt = {}
. “ground-truth” transitions by qφ
for all timestep t of (~o, ~a) in training data (i) do
st , st+1 ← qφ,µ + qφ,Σ where ∼ p()
(i)
(i)
(i)
(i)
Dgt ← Dgt ∪ hst , at+1 , rt+1 , st+1 i
end for
loop
Dexplore = {}
. explore with learned transition
(i)
for all st
in training data that succeeded do
rand(a ∈ A)
with prob.
at+1 =
(i)
argmax
Q(s
,
a)
otherwise
a∈A
t
(i)
(i)
r
if
a
==
a
t+1
t
t+1
rt+1 =
−1 otherwise
(i)
st+1 ← T (st , at )
(i)
Dexplore ← Dexplore ∪ hst , at+1 , rt+1 , st+1 i
end for
D ← Dgt ∪ Dexplore
. update deep Q-network
for all minibatch from D do
yt ← rt + γmaxa0 Q(st+1 , a0 )
Take gradient with loss [yt − Q(st , at+1 )]2
end for
end loop
We jointly back-propagate on neural networks for both sets
of encoder φ and decoder θ parameters with mini-batches to
maximize the lower bound using AdaDelta [38].
E. Optimal Control in Learned Latent State Space
After learning a generative model for the POMDP and a
recognition network using a variational Bayes method, we
need an algorithm for making an optimal decision in learned
representation of haptic feedback and action. We employ
a reinforcement learning method, Q-Learning, which learns
to approximate an optimal action-value function [37]. The
algorithm computes a score for each state action pair:
Q:S×A→R
The Q function is approximated by a two layer neural
network as shown in Fig. 2d.
In a standard reinforcement learning setting, in each
state st , an agent learns by exploring the selected action
argmaxa∈A Q(st , a) with a current Q function. However,
doing so requires an ability to actually take or simulate the
chosen action from st and observe rt+1 and st+1 . However,
≥ L(θ, φ)
there does not exist a good robotics simulation software that
Thus, to maximize maxθ log pθ (~o, ~r|~a) , the lower bound
can simulate complex interactions between a robot and an
L(θ, φ) can instead be maximized.
object and generate different haptic signals. Thus, we cannot
L(θ, φ) = −DKL qφ (~s|~o, ~r, ~a)||pθ (~s|~a)
freely explore any states.
+ Eqφ (~s|~o,~r,~a) log pθ (~o, ~r|~s, ~a)
(1)
Instead, we first take all state transitions and rewards
(i)
(i) (i)
(i)
Using a reparameterization trick [33] twice, we arrive at hst , at+1 , rt+1 , st+1 i from the i-th training data sequence
following lower bound (refer to Appendix for full derivation): and store in Dgt . Both st(i) and s(i)
t+1 are computed by the
recognition network qφ with a reparameterization technique
L(θ, φ) ≈ −DKL qφ (s1 |~o, ~r, ~a)||p(s1 )
(similar to Sec. III-D).
T
L
1 XX
(l)
At each iteration, we first have an exploration stage. For
−
DKL qφ (st |st−1 , ~o, ~r, ~a)||p(st |st−1 , ut−1 )
(i)
L t=2
explorations,
we start from states st of training sequences
l=1
L
that resulted in successful completion of the task and choose
1 X
+
log pθ (~o|~s(l) ) + log pθ (~r|~s(l) )
an action at+1 with -greedy. With the learned transition
L
l=1
function T (Sec. III-B), the selected action at+1 is executed
(i)
where ~s(l) = gφ ((l) , ~o, ~r, ~a) and (l) ∼ p()
(2) from st . However, since we are using a learned transition
Learning System
Nominal
Trajectory
Unsupervised
Pre-training
Dataset
Tactile-based
Grasping
Task
Planner
JTCartesian
Controller
Noise
Suppressor
Deep
Q-Network
Generative
Model
ethernet
Normalizer
Joints
Speaker
Fan
Deep
Q-Learning
Variational
Bayes
Recognition
Network
Stirrer/Hot Plate
recurrent
state
Tactile
Sensors
Model on GPU
Robot
Fig. 4: System Details of our system for learning and robotic
experiments.
function, any deviation from the distribution of training data
could result in unexpected state, unlike explorations in a real
or a simulated environment.
Thus, if the optimal action at+1 using a current Q-function
(i)
deviates from the ground-truth action at+1 , the action is
penalized with a negative reward to prevent deviations into
unexplored states. If the optimal action is same as the
ground-truth, the same reward as the original is given. For
such cases, the only difference from the ground-truth would
be in st+1 , which is inferred by the learned transition
function. All exploration steps are recorded in Dexplore .
After the exploration step in each iteration, we take
minibatches from D = Dgt ∪ Dexplore and backpropagate
on the deep Q-network with the loss function:
[rt + γmaxa0 Q(st+1 , a0 )] − Q(st , at )]2
The algorithm is summarized in Algorithm 1.
IV. S YSTEM D ETAILS
Robotic Platform. All experiments were performed on
a PR2 robot, a mobile robot with two 7 degree-of-freedom
arms. Each two-fingered end-effector has an array of tactile
sensors located at its tips. We used a Jacobian-transpose
based JTCartesian controller [39] for controlling its arm
during experiments.
For stable grasping, we take advantage of the tactile
sensors to grasp an object. The gripper is slowly closed until
certain thresholds are reached on both sides of the sensors,
allowing the robot to easily adapt to objects of different sizes
and shapes. To avoid saturating the tactile sensors, the robot
does not grasp the object with maximal force.
Tactile Sensor. Each side of the fingertip of a PR2 robot
is equipped with RoboTouch tactile sensor, an array of 22
tactile sensors covered by protective silicone rubber cover.
The sensors are designed to detect range of 0 – 30 psi
(0 – 205 kPa) with sensitivity of 0.1 psi (0.7 kPa) at the
rate of 35 Hz.
We observed that each of the 44 sensors has a significant variation and noise in raw sensor readings with drifts
over time. To handle such noise, values are first offset
by starting values when interaction between an object and
Fig. 5: A set of objects used for experiment. All three objects
have different surface area and shape, which results in vastly
different types of ‘clicks’ when observed via a tactile sensor.
the robot started (i.e. when a grasp occurred). Given the
relative signals, we find a normalization value for each of
44 sensors such that none of the values goes above 0.05
when stationary and all data is clipped to the range of −1
and 1. Normalization takes place by recording few seconds
of sensor readings after grasping.
Learning Systems. For fast computation and executions,
we offload all of our models onto a remote workstation with
a GPU connected over a direct ethernet connection. Our
models run on a graphics card using Theano [40], and our
high level task planner sends a new goal location at the rate
of 20 Hz. The overall system detail is shown in Figure 4.
V. E XPERIMENTS & R ESULTS
In order to validate our approach, we perform a series of
experiments on our dataset and on a PR2 robot.
A. Dataset
In order to test our algorithm that learns to represent haptic
feedback, we collected a dataset of three different objects —
a stirrer, a speaker, and a desk fan (Fig. 5) — each of which
have a knob with a detent structure (an example CAD model
shown in Fig. 1). Although these objects internally have
some type of a detent structure that produce a feedback that
humans would identify as a ‘click’, each ‘click’ from each
object is very distinguishable. As shown in Fig. 3, different
shapes of objects and the flat surface of the two fingers result
in vastly differently tactile sensor readings.
In our model, for the haptic signals ~o, we use a vector of
normalized 44 tactile sensor array as described in Sec. IV.
The reward ~r is given as one of three classes at each time
step, representing a positive, a negative and a neutral reward.
For every object, action ~a is an array of binary variables, each
representing a phase in its nominal plan.
In more detail, the stirrer (hot plate) has a knob with
a diameter of 22.7mm with a depth of 18.7mm, which
produces a haptic feedback that lasts about 30◦ rotations
when it is turned on or off. Our robot starts from both the
left (off state) and the right side (on state) of the knob.
The speaker has a tiny cylindrical knob that decreases
in its diameter from 13.1mm to 9.1mm with height of
12.8mm and requires 30◦ degree rotation. However, since
PR2 fingertips are parallel plates and measure 23mm with
silicon covers, grasping a 9.1mm knob results in drastically
different sensor readings at every execution of the task. The
desk fan has a rectangular knob with a width of 25.1mm
and a large surface area. It has a two-step detent control
TABLE II: Result of haptic signal prediction and robotic experiment. The prediction experiment reports the average
L2-norm from the haptic signal (44 signals in newtons) and the robotic experiment reports the success rate. It shows the
results of more than 200 robotic experiments.
Chance
Non-recurrent Recognition Network [4]
Recurrent-network as Representation [14]
Our Model without Exploration
Our Model
0.05secs
6.68 (±0.18)
1.39 (±2.51)
0.33 (±0.01)
0.72 (±0.08)
Haptics Prediction
0.25secs
0.50secs
6.68 (±0.17)
6.69 (±0.18)
5.03e5 (±5.27e7)
3.23e7 (±1.07e10)
1.01 (±0.09)
1.76 (±0.03)
0.79 (±0.09)
0.78 (±0.10)
with a click that lasts 45◦ degree rotation and has a narrow
stoppable window of about ±20◦ degrees.
The stirrer and the speaker can both be rotated clockwise
and counterclockwise and have a wall at both ends. The desk
fan has three stoppable points (near 0◦ , 45◦ , and 90◦ ) to
adjust fan speed and can get stuck in-between if a rotation
is not enough or exceeds a stopping point.
Each object is provided with a nominal plan with multiple
phases, each defined as a sequence of smoothly interpolated
waypoints consisting of end-effector position and orientation
along with gripper actions of grasping similar to [6]. For
each of the objects, we collected at least 25 successes and
25 failures. The success cases only includes rotations that
resulted in successful transition of states of objects (e.g.
from off to on state). The failures include slips, excessive
rotations beyond acceptable range, rotation even after hitting
a wall, and near breaking of the knob. There also exists less
dramatic failures such as insufficient rotations. Especially for
the desk fan, if a rotation results in two clicks beyond the first
stopping point, it is considered a failure. Each data sequence
consists of a sequence of trajectory (phases) as well as tactile
sensor signal after an execution of each waypoint.
To label the reward for each sequence, an external camera with a microphone was placed nearby the object. By
reviewing the audio and visually inspecting haptic signal
afterwards, an expert labeled the timeframe that each sequence succeeded or failed. These extra recordings were only
used for labeling the rewards, and such input is not made
available to the robot during our experiments. For sequences
that turned the knob past the successful stage but did not
stop the rotation, only negative rewards were given.
Among multiple phases of a nominal plan, which includes
pre-grasping and post-interaction trajectories, we focus on
three phases (before-rotation/rotation/stopped). These phases
occur after grasping and success is determined by ability to
correctly rotate and detect when to shift to the final phase.
B. Baselines
We compare our model against several baseline methods
on our dataset and in our robotic experiment. Since most
of the related works are applied to problems in different
domains, we take key ideas (or key structural differences)
from relevant works and fit them to our problem.
1) Chance: It follows a nominal plan and makes a transition
between phases by randomly selecting the amount of degree
to rotate a knob without incorporating haptic feedback.
2) Non-recurrent Recognition Network: Similar to [4], we
take non-recurrent deep neural network of only observations
Robotic Experiment
Stirrer
Speaker
Desk Fan
31.6%
38.1%
28.5%
52.9%
57.9%
62.5%
63.2%
68.4%
70.0%
35.0%
33.3%
52.6%
80.0%
73.3%
86.7%
without actions. However, it has access to a short history in a
sliding window of haptic signal at every frame. For control,
we apply the same Q-learning method as our full model.
3) Recurrent Network as Representation: Similar to [14], we
directly train a recurrent network to predict future haptic
signals. At each time step t, a LSTM network takes concatenated observation ot and previous action at as input,
and the output of the LSTM is concatenated with at+1 to
predict ot+1 . However, while [14] relies on hand-coded MPC
cost function to choose an action, we apply same Q-learning
that was applied to our full model. For haptic prediction
experiment, transitions happen by taking the output of the
next time step as input to the next observation.
4) Our Model without Exploration: During the final deep QLearning (Sec. III-E) stage, it skips the exploration step that
uses a learned transition model and only uses sequences of
representation from the recognition network.
C. Results and Discussion
To evaluate all models, we perform two types of experiments — haptic signal prediction and robotic experiment.
Haptic Signal Prediction. We first compare our model
against baselines on a task of predicting future haptic signal. For all sequences that either eventually succeeded or
failed, we take every timestep t, and predict timestep t + 1
(0.05secs), t + 5 (0.25secs) and t + 10 (0.5secs). The
prediction is made by encoding (recognition network) a
sequence up to time t and then transiting encoded states with
a learned transition model to the future frames of interest.
We take the L2-norm of the prediction of 44 sensor values
(which are in newtons) and take the average of that result.
The result is shown in the middle column of Table II.
Robotic Experiment. On a PR2 robot, we test the task of
turning a knob until it clicks on three different objects: stirrer,
speaker, and desk fan (Fig. 5). The right hand side of Table II
shows the result of over 200 executions. Each algorithm was
tested on each object at least 15 times.
Can it predict future haptic signals? When it predicts randomly (chance), regardless of the timestep, it has an average
of 6.7. When the primary goal is to be able to perform
the next haptic signal prediction, for one step prediction,
recurrent-network as representation baseline performs best
of 0.330 among all models, while ours performed 0.718. On
the other hand, our model does not diverge and performs
consistently well. After 0.5secs, when other models started
to diverge to an error of 1.757 or much larger, our model
still had prediction error of 0.782.
TABLE III: Time difference between the time the robot
stopped and the time the expert indicated it ‘clicked’.
Stirrer
0.180 secs (±0.616)
Speaker
0.539 secs (±1.473)
Desk Fan
−0.405 secs (±0.343)
What does learned representation represent? We visualize
our learned embedding space of haptic feedback using tSNE [41] in Fig. 6. Initially, both successful (blue paths) and
unsuccessful (red paths) all starts from similar states but they
quickly diverge into different clusters of paths much before
they eventually arrive at states that were given positive or
negative rewards shown as blue and red dots.
Does good representation lead to successful execution?
Our model allows robot to successfully execute on the three
objects 80.0%, 73.3%, and 86.7% respectively, performing
the highest compared to any other models. The next best
model which uses recurrent network as representation performed at 63.2%, 68.4%, and 70.0%. However, note that this
baseline still take advantage of our Q-learning method. Our
model that did not take advantage of simulated exploration
performed much poorly (35.0%, 33.3%, and 52.6%), showing that good representation combined with our Q-learning
method leads to successful execution of the tasks.
Is recurrent network necessary for haptic signals? Nonrecurrent recognition network quickly diverged to extremely
large number of 3.2e7 even though it successfully predicted
1.389 for a single step prediction. Note that it takes windowed haptic sequence of last 5 frames as input. Unlike images, short window of data does not hold enough information
about haptic sequence which lasts much longer timeframe.
For robotic experiment, non-recurrent network performed
52.9%, 57.9%, and 62.5% even with our Q learning method.
How accurately does it perform the task? When our full
model was being tested on three objects, we also had one of
the author observe (visually and audibly) very closely and
press a button as soon as a click occurs. On successful execution of the task, we measure the time difference between the
time the robot stops turning and the time the expert presses
the key, and the results are shown in Table III.
The positive number represents that the model was delayed
than the expert and the negative number represents that the
model transitioned earlier. Our model only differed from
human with an average of 0.37 seconds. All executions of
tasks were performed at same translational and rotational
velocity as the data collection process.
Note that just like a robot has a reaction time to act on
perceived feedback, an expert has a reaction time to press the
key. However, since the robot was relying on haptic feedback
while the observer was using every possible human senses
available including observation of the consequences without
touch, some differences are expected. We also noticed that
fan especially had a delay in visible consequences compared
to the haptic feedback because robot was rotating these knobs
slower than normal humans would turn in daily life; thus, the
robot was able to react 0.4 seconds faster.
Video of robotic experiments are available at this website:
http://jaeyongsung.com/haptic_feedback/
VI. C ONCLUSION
In this work, we present a novel framework for learning
to represent haptic feedback of an object that requires sense
of touch. We model such tasks as partially observable model
with its generative model parametrized by neural networks.
To overcome intractability of computing posterior, variational
Bayes method allows us to approximate posterior with a
deep recurrent recognition network consisting of two LSTM
layers. Using a learned representation of haptic feedback,
we also introduce a Q-learning method that is able to learn
optimal control without access to simulator in learned latent
state space utilizing only prior experiences and learned
generative model for transition. We evaluate our model on a
task of rotating a knob until it clicks against several baseline.
With more than 200 robotic experiments on the PR2 robot,
we show that our model is able to successfully manipulate
knobs that click while predicting future haptic signals.
APPENDIX
A. Lowerbound Derivation
To continue our derivation of the lower bound from
Sec. III-D. The
second term of equation 1:
Eqφ (~s|~o,~r,~a) log pθ (~o, ~r|~s, ~a)
= Eqφ (~s|~o,~r,~a) log pθ (~o|~s) + log pθ (~r|~s)
≈
L
1 X
log pθ (~o|~s(l) ) + log pθ (~r|~s(l) )
L
l=1
L T
1 XX
(l)
(l)
log pθ (ot |st ) + log pθ (rt |st )
=
L
t=1
l=1
where ~s(l) = qφ,µ + (l) qφ,Σ and (l) ∼ p()
Reparametrization trick ([33], [34]) at last step samples from
the inferred distribution by a recognition network qφ .
And, for the first term from
equation 1:
DKL qφ (~s|~o, ~r, ~a)||pθ (~s|~a)
Z
Z
qφ (~s|~o, ~r, ~a)
=
···
qφ (~s|~o, ~r, ~a) log
pθ (~s|~a)
s1
sT
= DKL qφ (s1 |~o, ~r, ~a)||p(s1 )
+
T
X
Est−1 ∼qφ (st−1 |~o,~r,~a)
t=2
DKL qφ (st |st−1 , ~o, ~r, ~a)||p(st |st−1 , at−1 )
using reparameterazation trick again,
= DKL qφ (s1 |~o, ~r, ~a)||p(s1 )
+
L
T
X
1 X
(l)
DKL qφ (st |st−1 , ~o, ~r, ~a)||p(st |st−1 , at−1 )
L
t=2
l=1
(l)
where st−1 = qφ,t−1,µ + (l) qφ,t−1,Σ and (l) ∼ p()
Combining these two terms, we arrive at equation 2.
We do not explain each step of the derivation at length
since similar ideas behind the derivation can be found at
[35] although exact definition and formulation are different.
Acknowledgment. We thank Ian Lenz for useful discussions.
This work was supported by Microsoft Faculty Fellowship
and NSF Career Award to Saxena.
Not Successful
Not Successful
Successful
Successful
Fig. 6: Projection of learned representation of haptic feedback using t-SNE [41] for ‘stirrer’ and ‘fan’. Each dot represents
an inferred state at each time frame, and blue and red dots represents positive and negative reward at those time frame. Here
we show some of successful (blue) and unsuccessful (red) sequences. For both objects, notice both classes initially starts
from similar state and then diverges, forming clusters. Several successful and unsuccessful haptic signals are shown as well.
R EFERENCES
[1] A. Montagu, “Touching: The human significance of the skin.” 1971.
[2] D. A. Norman, The design of everyday things, 1988.
[3] S. Levine, C. Finn, T. Darrell, and P. Abbeel, “End-to-end training of
deep visuomotor policies,” arXiv preprint arXiv:1504.00702, 2015.
[4] M. Watter, J. Springenberg, J. Boedecker, and M. Riedmiller, “Embed
to control: A locally linear latent dynamics model for control from
raw images,” in NIPS, 2015.
[5] V. Mnih, et al., “Playing atari with deep reinforcement learning,” arXiv
preprint arXiv:1312.5602, 2013.
[6] J. Sung, S. H. Jin, and A. Saxena, “Robobarista: Object part-based
transfer of manipulation trajectories from crowd-sourcing in 3d pointclouds,” in ISRR, 2015.
[7] D. Silver, et al., “Mastering the game of go with deep neural networks
and tree search,” Nature, vol. 529, no. 7587, pp. 484–489, 2016.
[8] M. C. Gemici and A. Saxena, “Learning haptic representation for
manipulating deformable food objects,” in IROS. IEEE, 2014.
[9] V. Chu, et al., “Using robotic exploratory procedures to learn the
meaning of haptic adjectives,” in ICRA, 2013.
[10] S. Bennett, “A brief history of automatic control,” IEEE Control
Systems Magazine, vol. 16, no. 3, pp. 17–25, 1996.
[11] J. K. Salisbury, “Active stiffness control of a manipulator in cartesian
coordinates,” in Decision and Control. IEEE, 1980.
[12] J. Barry, K. Hsiao, L. Kaelbling, and T. Lozano-Pérez, “Manipulation
with multiple action types,” in ISER, no. 1122374, 2012.
[13] S. Trimpe, A. Millane, S. Doessegger, and R. DAndrea, “A self-tuning
lqr approach demonstrated on an inverted pendulum,” in IFAC World
Congress, 2014, p. 11.
[14] I. Lenz, R. Knepper, and A. Saxena, “Deepmpc: Learning deep latent
features for model predictive control,” in RSS, 2015.
[15] F. Chaumette and S. Hutchinson, “Visual servo control. i. basic
approaches,” Robotics & Automation Magazine, IEEE, vol. 13, no. 4,
pp. 82–90, 2006.
[16] J. Park and O. Khatib, “A haptic teleoperation approach based on
contact force control,” IJRR, vol. 25, no. 5-6, pp. 575–591, 2006.
[17] Q. Li, C. Schürmann, R. Haschke, and H. J. Ritter, “A control
framework for tactile servoing.” in R:SS, 2013.
[18] P. Pastor, M. Kalakrishnan, S. Chitta, et al., “Skill learning and task
outcome prediction for manipulation,” in ICRA, 2011.
[19] S. Thrun, W. Burgard, D. Fox, et al., Probabilistic robotics. MIT
press Cambridge, 2005.
[20] K. Hsiao, L. P. Kaelbling, and T. Lozano-Perez, “Grasping pomdps,”
in ICRA, 2007.
[21] H. Kurniawati et al., “Sarsop: Efficient point-based pomdp planning
by approximating optimally reachable belief spaces.” in RSS, 2008.
[22] N. A. Vien and M. Toussaint, “Touch based pomdp manipulation via
sequential submodular optimization,” in Humanoids, 2015.
[23] B. Sallans, “Learning factored representations for partially observable
markov decision processes.” in NIPS. Citeseer, 1999, pp. 1050–1056.
[24] G. Contardo, L. Denoyer, T. Artieres, and P. Gallinari, “Learning states
representations in pomdp,” in ICLR, 2014.
[25] V. Vapnik and A. Vashist, “A new learning paradigm: Learning using
privileged information,” Neural Networks, 2009.
[26] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in NIPS, 2012.
[27] A. Hannun, et al., “Deep speech: Scaling up end-to-end speech
recognition,” arXiv preprint arXiv:1412.5567, 2014.
[28] Y. Gao, L. A. Hendricks, et al., “Deep learning for tactile understanding from visual and haptic data,” arXiv preprint 1511.06065, 2015.
[29] R. Hadsell, et al., “Deep belief net learning in a long-range vision
system for autonomous off-road driving,” in IROS, 2008.
[30] A. Valada, L. Spinello, and W. Burgard, “Deep feature learning for
acoustics-based terrain classification,” in ISRR, 2015.
[31] N. Wahlström, T. B. Schön, and M. P. Deisenroth, “From pixels to
torques: Policy learning with deep dynamical models,” arXiv preprint
arXiv:1502.02251, 2015.
[32] C. Finn, et al., “Deep spatial autoencoders for visuomotor learning,”
in ICRA, 2016.
[33] D. P. Kingma and M. Welling, “Auto-encoding variational bayes,”
arXiv preprint arXiv:1312.6114, 2013.
[34] D. J. Rezende, S. Mohamed, and D. Wierstra, “Stochastic backpropagation and approximate inference in deep generative models,” arXiv
preprint arXiv:1401.4082, 2014.
[35] R. G. Krishnan, U. Shalit, and D. Sontag, “Deep kalman filters,” arXiv
preprint arXiv:1511.05121, 2015.
[36] M. Hausknecht and P. Stone, “Deep recurrent q-learning for partially
observable mdps,” arXiv preprint arXiv:1507.06527, 2015.
[37] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction.
MIT press, 1998.
[38] M. D. Zeiler, “Adadelta: An adaptive learning rate method,” arXiv
preprint arXiv:1212.5701, 2012.
[39] “Jt cartesian controller.” [Online]. Available: http://wiki.ros.org/
robot mechanism controllers/JTCartesian%20Controller
[40] F. Bastien, et al., “Theano: new features and speed improvements,”
NIPS DLUFL Workshop, 2012.
[41] L. Van der Maaten and G. Hinton, “Visualizing data using t-sne,”
JMLR, vol. 9, no. 2579-2605, p. 85, 2008.
| 2cs.AI
|
arXiv:1108.2850v2 [math.AC] 8 Oct 2012
TORIC IDEALS OF FINITE GRAPHS AND
ADJACENT 2-MINORS
HIDEFUMI OHSUGI AND TAKAYUKI HIBI
Abstract. We study the problem when an ideal generated by adjacent 2-minors
is the toric ideal of a finite graph.
Let X = (xij ) i=1,...,m be a matrix of mn indeterminates , and let A = K[{xij } i=1,...,m ]
j=1,...,n
j=1,...,n
be the polynomial ring in mn variables over a field K. Given 1 ≤ a1 < a2 ≤ m and
1 ≤ b1 < b2 ≤ n, the symbol [a1 , a2 |b1 , b2 ] denotes the 2-minor xa1 b1 xa2 b2 − xa1 b2 xa2 b1
of X. In particular [a1 , a2 |b1 , b2 ] is a binomial of A. A 2-minor [a1 , a2 |b1 , b2 ] of X is
adjacent ([4]) if a2 = a1 + 1 and b2 = b1 + 1. Following [2], we say that a set M of
adjacent 2-minors of X is of chessboard type if the following conditions are satisfied:
• if [a, a + 1|b, b + 1] and [a, a + 1|b′ , b′ + 1] with b < b′ belong to M, then
b + 1 < b′ ;
• if [a, a + 1|b, b + 1] and [a′ , a′ + 1|b, b + 1] with a < a′ belong to M, then
a + 1 < a′ .
Given a set M of adjacent 2-minors of X of chessboard type, we introduce the finite
graph ΓM on the vertex set M, whose edges are {[a, a+1|b, b+1], [a′ , a′ +1|b′ , b′ +1]}
such that
• [a, a + 1|b, b + 1] 6= [a′ , a′ + 1|b′ , b′ + 1],
• {a, a + 1} ∩ {a′ , a′ + 1} =
6 ∅,
′ ′
• {b, b + 1} ∩ {b , b + 1} =
6 ∅.
For example, if M = {[1, 2|2, 3], [2, 3|3, 4], [3, 4|2, 3], [2, 3|1, 2]}, then ΓM is a cycle of
length 4. The ideal IM is generated by x12 x23 −x13 x22 , x23 x34 −x24 x33 , x32 x43 −x33 x42
and x21 x32 − x22 x31 . The binomial x32 (x13 x21 x34 x42 − x12 x24 x31 x43 ) belongs to IM
but neither x32 nor x13 x21 x34 x42 −x12 x24 x31 x43 belongs to IM . Thus IM is not prime.
A fundamental fact regarding ideals generated by adjacent 2-minors is
Lemma 1 ([2]). Let M be a set of adjacent 2-minors of X, and let IM be the ideal
of A generated by all 2-minors belonging to M. Then, IM is a prime ideal if and
only if M is of chessboard type, and ΓM possesses no cycle of length 4.
A finite graph G is said to be simple if G has no loop and no multiple edge.
Let G be a finite simple graph on the vertex set [d] = {1, . . . , d}, and let E(G) =
2010 Mathematics Subject Classification: Primary 13F20.
Keywords: Toric ideals, Ideals of 2-adjacent minors, finite graphs.
This research was supported by JST CREST .
1
{e1 , . . . , en } be its set of edges. Let K[t] = K[t1 , . . . , td ] denote the polynomial ring
in d variables over K, and let K[G] denote the subring of K[t] generated by the
squarefree quadratic monomials te = ti tj with e = {i, j} ∈ E(G). The semigroup
ring K[G] is called the edge ring of G. Let K[y] = K[y1 , . . . , yn ] denote the polynomial ring in n variables over K. The kernel IG of the surjective homomorphism
π : K[y] → K[G] defined by setting π(yi ) = tei for i = 1, . . . , n is called the toric
ideal of G. Clearly, IG is a prime ideal. It is known that IG is generated by the
binomials corresponding to even closed walks of G. See [7] , [6, Chapter 9] and [5,
Lemma 1.1] for details.
Example 2. Let G be a complete bipartite graph with the edge set E(G) = {{i, p +
j} | 1 ≤ i ≤ p, 1 ≤ j ≤ q}. Let X = (xij ) i=1,...,p be a matrix of pq indeterminates
j=1,...,q
and K[x] = K[{xij } i=1,...,p ]. Then, IG is the kernel of the surjective homomorphism
j=1,...,q
π : K[x] → K[G] defined by setting π(xij ) = ti tp+j for 1 ≤ i ≤ p, 1 ≤ j ≤ q. It
is known [6, Proposition 5.4] that IG is generated by the set of all 2-minors of X.
Note that each 2-minor xij xi′ j ′ − xij ′ xi′ j corresponds to the cycle {{i, p + j}, {p +
j, i′ }, {i′ , p + j ′ }, {p + j ′ , i}} of G.
In general, a toric ideal is the defining ideal of a homogeneous semigroup ring. We
refer the reader to [6] for detailed information on toric ideals. It is known [1] that a
binomial ideal I, i.e., an ideal generated by binomials, is a prime ideal if and only
if I is a toric ideal. An interesting research problem on toric ideals is to determine
when a binomial ideal is the toric ideal of a finite graph.
Example 3. The ideal I = hx1 x2 − x3 x4 , x1 x2 − x5 x6 , x1 x2 − x7 x8 i is the toric ideal
of the semigroup ring K[t1 t5 , t2 t3 t4 t5 , t1 t2 t5 , t3 t4 t5 , t2 t3 t5 , t1 t4 t5 , t1 t3 t5 , t2 t4 t5 ]. If there
exists a graph G such that I = IG , then three quadratic binomials correspond to
cycles of length 4. However, this is impossible since these three cycles must have
common two edges e1 and e2 such that e1 ∩ e2 = ∅. Thus, I cannot be the toric ideal
of a finite graph. This observation implies that the toric ideal of a finite distributive
lattice L (see [3]) is the toric ideal of a finite graph if and only if L is planar. In
fact, if L is planar, then it is easy to see that the toric ideal of L is the toric ideal of
a bipartite graph. If L is not planar, then L contains a sublattice that is isomorphic
to the Boolean lattice B3 of rank 3. Since the toric ideal of B3 has three binomials
above, the toric ideal of L cannot be the toric ideal of a finite graph.
Let M be a set of adjacent 2-minors. Now, we determine when a binomial ideal
IM generated by M is the toric ideal IG of a finite graph G. Since IG is a prime
ideal, according to Lemma 1, if there exists a finite graph G with IM = IG , then M
must be of chessboard type and ΓM possesses no cycle of length 4.
Theorem 4. Let M be a set of adjacent 2-minors. Then, there exists a finite graph
G such that IM = IG if and only if M is of chessboard type, ΓM possesses no cycle
of length 4, and each connected component of ΓM possesses at most one cycle.
2
Proof. We may assume that M is of chessboard type and ΓM possesses no cycle of
length 4. Let M = M1 ∪ · · · ∪ Ms , where ΓM1 , . . . , ΓMs is the set of connected
components of ΓM . If i 6= j, then f ∈ Mi and g ∈ Mj have no common variable.
Hence, there exists a finite graph G such that IM = IG if and only if for each
1 ≤ i ≤ s, there exists a finite graph Gi such that IMi = IGi . Thus, we may assume
that ΓM is connected. Let p be the number of vertices of ΓM , and let q be the
number of edges of ΓM . Since ΓM is connected, we have p ≤ q + 1.
[Only if] Suppose that there exists a finite graph G with IM = IG . From [2,
Theorem 2.3], the codimension of IM is equal to p. Let d be the number of vertices
of G, and let n be the number of edges of G. Then, we have d ≤ 4p − 2q and
n = 4p − q. The height of IG is given in [7]. If G is bipartite, then the codimension
of IG satisfies p ≥ n − d + 1 ≥ (4p − q) − (4p − 2q) + 1 = q + 1. Hence, we have
p = q + 1 and ΓM is a tree. On the other hand, if G is not bipartite, then the
codimension of IG satisfies p ≥ n − d ≥ (4p − q) − (4p − 2q) = q. Hence, we have
p ∈ {q, q + 1} and ΓM has at most one cycle.
[If] Suppose that ΓM has at most one cycle. Then, we have p ∈ {q, q + 1}.
Case 1. p = q + 1, i.e., ΓM is a tree.
Through induction on p, we will show that there exists a connected bipartite
graph G such that IM = IG . If p = 1, then IM = IG where G is a cycle of length
4. Let k > 1, and suppose that the assertion holds for p = k − 1. Suppose that
ΓM has k vertices. Since ΓM is a tree, ΓM has a vertex v = [a, a + 1|b, b + 1]
of degree 1. Let M′ = M \ {v}. Since ΓM′ is a tree, there exists a connected
bipartite graph G′ such that IM′ = IG′ by the hypothesis of induction. From [5,
Theorem 1.2], since IG′ is generated by quadratic binomials, any cycle of G′ of
length ≥ 6 has a chord. Let v ′ = [a′ , a′ + 1|b′ , b′ + 1] denote the vertex of ΓM that
is incident with v. Let e = {i, j} be the edge of G′ corresponding to the common
variable of v and v ′ . Let {1, 2, . . . , d} be the vertex set of G′ . We now define the
connected bipartite graph G on the vertex set {1, 2, . . . , d, d + 1, d + 2} with the edge
set E(G′ ) ∪ {{i, d + 1}, {d + 1, d + 2}, {d + 2, j}}. Then, any cycle of G of length
≥ 6 has a chord, and hence, IG is generated by quadratic binomials. Thus, IG is
generated by the quadratic binomials of IG′ together with v corresponding to the
cycle {{i, d + 1}, {d + 1, d + 2}, {d + 2, j}, {j, i}}. Therefore, IM = IG .
Case 2. p = q, i.e., ΓM has exactly one cycle.
Then, we have p ≥ 8. Through induction on p, we will show that there exists
a graph G such that IM = IG . If p = 8, then ΓM is a cycle of length 8. Then,
IM = IG where G is the graph shown in Figure 1.
Let k > 8 and suppose that the assertion holds for p = k − 1. Suppose that ΓM
has k vertices. If ΓM has a vertex v = [a, a + 1|b, b + 1] of degree 1, then ΓM′ where
M′ = M \ {v} has exactly one cycle, and hence, there exists a graph G′ such that
IM′ = IG′ by the hypothesis of induction. Let v ′ = [a′ , a′ + 1|b′ , b′ + 1] denote the
3
Figure 1. Graph for M such that ΓM is a cycle of length 8.
1
1
4
4
e
e’
3
2
−→
3
2
Figure 2. New graph G arising from G′ .
vertex of ΓM that is incident with v. Let e = {i, j} be the edge of G′ corresponding
to the common variable of v and v ′ . Suppose that the vertex set of G′ is {1, 2, . . . , d}.
We now define the graph G on the vertex set {1, 2, . . . , d, d + 1, d + 2} with the edge
set E(G′ ) ∪ {{i, d + 1}, {d + 1, d + 2}, {d + 2, j}}. Since G′ satisfies the conditions in
[5, Theorem 1.2], it follows that G satisfies the conditions in [5, Theorem 1.2]. Thus,
IG is generated by the quadratic binomials of IG′ together with v corresponding to
the cycle {{i, d + 1}, {d + 1, d + 2}, {d + 2, j}, {j, i}}. Therefore, IM = IG .
Suppose that ΓM has no vertex of degree 1. Then, ΓM is a cycle of length k. A
2-minor ad − bc ∈ M is called free if one of the following holds:
• Neither a nor d appears in other 2-minors of M,
• Neither b nor c appears in other 2-minors of M.
From [2, Lemma 1.6], M has at least two free 2-minors. Let v = [a, a + 1|b, b + 1]
be a free 2-minor of M. We may assume that neither xa,b nor xa+1,b+1 appears in
other 2-minors of M. Since ΓM is a cycle, xa+1,b appears in exactly two 2-minors
of M and xa,b+1 appears in exactly two 2-minors of M. Let M′ = M \ {v}. Since
ΓM′ is a tree, there exists a connected bipartite graph G′ such that IM′ = IG′ by the
argument in Case 1. Suppose that the edge {1, 3} corresponds to the variable xa+1,b
and the edge {2, 4} corresponds to the variable xa,b+1 . We now define the graph G
as shown in Figure 2, where vertices 1 and 2 belong to the same part of the bipartite
graph G′ . Note that G is not bipartite. Let e = {1, 2} and e′ = {3, 4}. Since G′ is
a bipartite graph, it follows that
4
(a) If either e or e′ is an edge of an even cycle C of G, then {e, e′ } ⊂ E(C).
(b) If C ′ is an odd cycle of G, then {e, e′ } ∩ E(C ′ ) 6= ∅.
Let I denote the ideal generated by all quadratic binomials in IG . Since each
quadratic binomial in IG corresponds to a cycle of G of length 4, it follows that
IM = I. Thus, it is sufficient to show that IG = I, i.e., IG is generated by quadratic
binomials. From [5, Theorem 1.2], since G′ is bipartite and since IG′ is generated
by quadratic binomials, all cycles of G′ of length ≥ 6 have a chord.
Let C be an even cycle of G of length ≥ 6. If E(C) ∩ {e, e′ } = ∅, then C has
an even-chord since all cycles of the bipartite graph G′ of length ≥ 6 have a chord.
Suppose that {e, e′ } ⊂ E(C) holds. Then, either {1, 3} or {2, 4} is a chord of C.
Moreover, such a chord is an even-chord of C from (b) above.
Let C and C ′ be odd cycles of G having exactly one common vertex. From (b)
above, we may assume that e ∈ E(C) \ E(C ′ ) and e′ ∈ E(C ′ ) \ E(C). If {1, 3}
does not belong to E(C) ∪ E(C ′ ), then {1, 3} satisfies the condition in [5, Theorem
1.2 (ii)]. If {1, 3} belongs to E(C) ∪ E(C ′ ), then {2, 4} ∈
/ E(C) ∪ E(C ′ ) since C
and C ′ have exactly one common vertex. Hence, {2, 4} satisfies the condition in [5,
Theorem 1.2 (ii)].
Let C and C ′ be odd cycles of G having no common vertex. Then, neither {1, 3}
nor {2, 4} belong to E(C) ∪ E(C ′ ). Hence, {1, 3} and {2, 4} satisfy the condition in
[5, Theorem 1.2 (iii)].
Thus, from [5, Theorem 1.2], IG is generated by quadratic binomials. Therefore,
IG = IM as desired.
References
[1] D. Eisenbud and B. Sturmfels, Binomial ideals, Duke Math. J. 84 (1996), 1–45.
[2] J. Herzog and T. Hibi, Ideals generated by adjacent 2-minors, J. Comm. Algebra, to appear.
[3] T. Hibi, Distributive lattices, affine semigroup rings and algebras with straightening laws, in
“Commutative Algebra and Combinatorics” (M. Nagata and H. Matsumura, Eds.), Adv. Stud.
Pure Math., Volume 11, North–Holland, Amsterdam, 1987, pp. 93–109.
[4] S. Hoşten and S. Sullivant, Ideals of adjacent minors, J. Algebra 277 (2004), 615–642.
[5] H. Ohsugi and T. Hibi, Toric ideals generated by quadratic binomials, J. Algebra 218 (1999),
509–527.
[6] B. Sturmfels, “Gröbner bases and convex polytopes,” Amer. Math. Soc., Providence, RI, 1996.
[7] R. Villarreal, Rees algebras of edge ideals, Comm. Algebra 23 (1995), 3513–3524.
Hidefumi Ohsugi, Department of Mathematics, College of Science, Rikkyo University, Toshima-ku, Tokyo 171-8501, Japan
E-mail address: [email protected]
Takayuki Hibi, Department of Pure and Applied Mathematics, Graduate School
of Information Science and Technology, Osaka University, Toyonaka, Osaka 5600043, Japan
E-mail address: [email protected]
5
| 0math.AC
|
High-Dimensional Vector Semantics
M. Andrecut
arXiv:1802.09914v1 [cs.CL] 23 Feb 2018
January 5, 2018
Calgary, Alberta, T3G 5Y8, Canada
[email protected]
Abstract
In this paper we explore the "vector semantics" problem from the perspective of "almost orthogonal" property of high-dimensional random vectors. We show that this intriguing property
can be used to "memorize" random vectors by simply adding them, and we provide an efficient
probabilistic solution to the set membership problem. Also, we discuss several applications to word
context vector embeddings, document sentences similarity, and spam filtering.
1
Introduction
In many natural language processing tasks the words and the documents are represented using the
"bag of words" model. In such a model, a document is represented by a high-dimensional vector, with
the components corresponding to the frequency of a particular word in the document (for a detailed
discussion see [1–3] and the references within). For example, assuming an English vocabulary of 25, 000
words, each document will be represented by a 25, 000 dimensional vector, where the component i is
the frequency of the ith word in the document. The vector representation is particularly useful in
text classification tasks, where the similarity of two documents can be simply estimated using the dot
product between the vectors. If the vectors are normalized, then their dot product is equal to the cosine
of the angle between the vectors, and therefore the more parallel the vectors are, the more similar the
documents are.
Another frequently encountered problem is the word vector embedding. In such a problem, the
words are represented by high-dimensional vectors, and their "meaning" is computed from their context,
which is modeled using the distribution of words around them (for a detailed discussion see [1–3] and
the references within). Several computational methods based on pointwise mutual information, (deep)
neural networks, matrix factorizations or agglomerative clustering have been developed to compute
the "meaning" of words. These efforts have culminated in identifying the words that share semantic
(dog, cat, cow) or syntactic (emptied, carried, danced) properties, or in solving more complex problems
like estimating the similarity between pairs of words [3, 4]. For example from the pairs (king, queen)
and (man, woman), one can roughly recover queen ≈ king − man + woman, by simply using linear
vector algebra [4].
1
A different approach to these problems is based on the random indexing method [5, 6]. In this
approach a d-dimensional sparse random vector called a random label is assigned to each different word
in the text data. These labels have a small number of randomly distributed -1s and +1s, with the rest
set to 0. In the next step, for any given word the labels for the words in its context window are added
to its context vector. This approach is motivated by an earlier observation, that in high-dimensional
spaces there are many more "almost orthogonal" directions than the dimensionality of the space [7].
Inspired by these ideas, here we explore the "vector semantics" problem from the perspective of "almost orthogonal" property of high-dimensional random vectors. More exactly, we extend the theoretical
justification of this method by providing a probabilistic solution to the set membership problem (bag
of words), and we discuss several potential applications to word context vector embeddings, document
sentences similarity, and spam filtering. Contrary to the "expensive" machine learning methods, this
method is very simple and it does not even require a "learning" process, however it exhibits similar
properties.
2
Almost orthogonal random vectors
Let us consider the set B d of random d−dimensional unit vectors:
1
γ = √ [γ1 , ..., γd ]T ,
d
kγk = 1,
(1)
with the components corresponding to independent Bernoulli variables, γi ∈ {−1, 1}, i = 1, ..., d, with
the probability p = 1/2.
Let us assume that ξ, ζ ∈ B d are two random vectors from B d . These two vectors are orthogonal
ξ ⊥ γ if their dot product:
d
ζT ξ =
1X
ζi ξi .
d i=1
(2)
is equal to zero. For high-dimensional vectors extracted from B d , the expectation value of the dot
product is obviously:
d
X
1
E(ζ ξ) = E
d
T
!
ζi ξi
= 0,
(3)
i=1
and the variance is:
!2
d
d
d
d
X
1
1 XX
1X 2
1
σ2 = E
ζi ξi =
ζi ζj E[ξi ξj ] =
ζi = .
d
d i=1 j=1
d i=1
d
i=1
(4)
Also, using the Chernoff bound [8] we obtain:
" #
2
δ
Pr(|ζ ξ| > δ) < exp −
= exp −dδ 2 .
σ
T
2
(5)
Thus, the probability that the two random vectors ζ, ξ ∈ B d are "almost orthogonal" is given by:
Prδ (ζ ⊥ ξ) > 1 − exp −dδ 2 .
(6)
This means that for a relatively large dimensionality d, the probability that two random vectors from B d
are "almost orthogonal" is quite high. For example, if δ = 0.05 and d = 1, 200 we have: Pr0.05 (ζ ⊥ ξ) >
0.95. In general, one can show that in a high dimensional space there is an exponentially large number
of "almost orthogonal" randomly chosen vectors [7, 9]. Following the random indexing approach, in the
next section we show that this "intriguing" property can be used to "memorize" random vectors from
B d by simply adding them.
3
Set membership problem
Let us now consider the following set membership problem: given a set of k random vectors, Ξ =
{ξ1 , ξ2 , ..., ξk | ξi ∈ B d , i = 1, ..., k} and a new random vector γ ∈ B d , we want to check if γ ∈ Ξ.
This is a typical binary decision problem, with the answer TRUE or FALSE. Normally, the solution
requires the calculation of the dot product of γ with each vector ξi ∈ Ξ, i = 1, 2, ..., K. If ∃ξi ∈ Ξ such
that γ T ξi = 1 then the answer is TRUE, otherwise the answer is FALSE. Thus, the solution to the set
membership problem is practically a binary classifier, which also acts as a "set filter" in B d .
The above set membership problem can be also reformulated as a "query" problem, by asking to
return the vector ξi∗ ∈ Ξ, which is most similar to the "query" vector γ. In this case the answer is
obtained by taking the k dot products, and searching for the index i∗ of the product with the highest
value:
i∗ = arg max ξiT γ.
(7)
i=1,...,k
We can see that statistically the set membership problem requires an average of k/2 operations (dot
products) in order to provide a correct answer, while the "query" problem requires k operations (dot
products) and a sorting procedure. However, here we will show that probabilistically the set membership
problem can be solved using only one operation (dot product).
Let us consider the sum of all random vectors from the given set Ξ:
ζ=
k
X
ξi ,
(8)
i=1
and the dot product of ζ with the "query" vector γ:
η = γ T ζ.
(9)
Let us first assume that ∃i such that ξi = γ. Then the expectation of the dot product is:
E(η) = E γ T ξi +
X
γ T ξj = 1 + E
j6=i
X
j6=i
3
γ T ξj = 1 +
X
j6=i
E(γ T ξj ) = 1,
(10)
members
non-members
2.0
1.5
1.0
0.5
0.0
0.5
1.0
0
200
400
600
i
800
1000
Figure 1: Probabilistic solutions for the set membership problem.
since γ and ξj6=i are random vectors from B d , and E(γ T ξj6=i ) = 0 according to the previous result.
Obviously, if @i such that ξi = γ, then the expectation of the dot product is:
E(η) = E
k
X
γ T ξj =
j=1
k
X
E(γ T ξj ) = 0.
(11)
j=1
Also, one can easily see that the variance of the dot product is:
2
k
X T
γ ξj = k/d.
σ 2 = E
(12)
j=1
In order to illustrate numerically this result we consider d = 104 and k = 103 , and we plot the value
of the dot product η in 103 cases where the "query" vector γ is a member, and respectively a nonmember, of the set Ξ. The results are shown in Figure 1. Here we have also included the distributions
of η for the distinct member and non-member situations.
Since the components of the vectors are Bernoulli distributed, the distribution of the η values is
√
binomial (scaled by a multiplying constant 1/ d). One can also approximate the binomial distribution
with a normal distribution with the mean µ = 1 for members, and respectively µ = 0 for non-members,
p
p
both having a standard deviation: σ = k/d. Therefore, we have two normal distributions N (1, k/d)
p
and respectively N (0, k/d). One can estimate the classification precision and recall by using the
p
overlap of these two normal distributions as a function of the standard deviation σ = k/d (as shown
in Figure 2). Since the intersection point of these two distributions is at η ∗ = 1/2, the overlap will be:
s(σ) = P(η > 1/2) + P(η < 1/2) = 1 − Φ
4
1
2σ
1
+Φ −
2σ
,
(13)
1.4
non-members
1.2
members
1.0
f
0.8
0.6
0.4
0.2
0.0
2
1
TP
FP FN
0
1
TN
2
3
Figure 2: Distributions overlap for the probabilistic set membership problem (σ = 1/3).
where Φ(x) is the cumulative distribution function:
Φ(x) =
1
2π
ˆ
x
exp(−t2 /2)dt.
(14)
−∞
In fact, the overlap s(σ) is an estimation of the sum of false positive (F P ) and false negative (F N )
classification cases:
s(σ) = F P (σ) + F N (σ),
(15)
also because of the perfect symmetry of the intersecting distributions we have:
F P (σ) = F N (σ) =
1
s(σ),
2
(16)
and respectively:
T P (σ) = T N (σ) = 1 − s(σ),
(17)
where T P and T N are the true positive, and respectively true negative classification cases. Therefore,
the classification precision and recall are equal to the following quantity:
ρ(σ) =
T P (σ)
T P (σ)
s(σ)
=
=1−
.
T P (σ) + F P (σ)
T P (σ) + F N (σ)
2 − s(σ)
(18)
In order to illustrate numerically this result we consider d = 103 and we let k = 2, ..., d. For each
k we compute the distributions from T = 103 samples with 103 cases where the "query" vector γ is
a member, and respectively a non-member, of the set Ξ. The obtained results for ρ(σ) are shown in
Figure 3, and they are in perfect agreement with the analytical estimation (13)-(18). It is interesting
to see that for σ ∈ (0, 0.215] the precision (recall) is ρ(σ) ≥ 0.99, and for σ ∈ (0, 0.375] we have
5
()
1.00
0.95
0.90
0.85
0.80
0.75
0.70
0.65
0.60
0.55
0.50
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Figure 3: Precision and recall ρ(σ) for the probabilistic set membership problem.
ρ(σ) ≥ 0.90. This means that for a large dimension d one can solve the set membership problem with
high probability for relatively large sets Ξ. For example if d = 1000 one can solve with high probability
(p > 0.9) set membership problems with up to k = 375 members, which is quite impressive, considering
the simplicity of the method.
4
Application to natural language processing tasks
4.1
Word-context vectors
We consider a vocabulary V of n unique words. With each word w ∈ V we associate a randomly
drawn vector ξw from B d . Thus, we associate the vocabulary V = {wi | i = 1, ..., n} with a set
Ξ = {ξi | ξi ∈ B d , i = 1, ..., n} of "almost orthogonal" random vectors.
Let us now consider a document D, containing |D| words from the vocabulary V . For each word
w(`) in D with the index i in the vocabulary V , w(`) ≡ wi , we also consider the context window of
length 2L:
W ` = {w(` − L), ..., w(` − 1), w(` + 1), ..., w(` + L) | ` = 1, ..., |D|}.
(19)
We define the context γi of the word wi ∈ V as the sum of the word vectors from all the corresponding
context windows extracted from D:
γi =
X
δ(χ(w(`)), i)
X
ξχ(w(`+j)) ,
(20)
w(`+j)∈W `
w(`)∈D
where χ(w) is the function that returns the index of the word w in the vocabulary V , and δ is Kronecker
6
delta function:
1 i = j
δ(i, j) =
.
0 i 6= j
(21)
Thus, given a document D, for each word wi we calculate its context as a sum γi over all windows at
the positions ` where wi appears in the document D. Since, the vectors ξi ∈ Ξ are "almost orthogonal",
and according to the previously obtained result for the "set membership problem", the context sum
can accommodate quite a large number of vectors until its "set filtering" properties will significantly
deteriorate. For example, let’s assume that the context of the word wi is the set Γi = {wi1 , ..., wim }
where each word wij appears a number of θij times. Then, the context vector of the word wi is:
mi
X
γi =
θij ξij .
(22)
j=1
One can easily check if a word wik (associated with the vector ξik ) is a "member" of the context γi by
simply taking the dot product:
ηik = ξiTk γi
(23)
which has the expectation:
X
X
E(ηik ) = E
θij ξiTk ξij = θik + E
θij ξiTk ξij = θik ≥ 1.
j
(24)
j6=k
Consequently, the variance is:
σ2 =
1X
θi .
d j j
(25)
Therefore, if σ ∈ (0, 0.375] we have a good precision and recall.
Now let’s assume that we have two words wi and wj with the context vectors γi and respectively
γj :
γi =
mi
X
θij ξij ,
(26)
θjk ξjk .
(27)
j=1
mj
γj =
X
k=1
One can check their similarity by taking their dot product:
ηij =
γiT γj
∈ [−1, 1].
kγi kkγj k
(28)
The expectation of ηij is:
E(ηij ) =
XX
1
θij θjk E(ξiTj ξjk ),
kγi kkγj k j
k
7
(29)
number of words in context vectors
number of unique words in context vectors
number of words
5000
4000
3000
2000
1000
0
0
1000
2000
3000
word index
4000
5000
6000
Figure 4: Distribution of the number of words, and of the unique number of words, in the context of
each word from the vocabulary in The Adventures of Sherlock Holmes.
where
1 i = j
j
k
E(ξiTj ξjk ) =
.
0 i 6= j
j
k
(30)
Thus, the similarity between wi and wj is determined by the number of identical words present in their
context vectors γi and γj .
Let us consider an example by using the book The Adventures of Sherlock Holmes by Sir Arthur
Conan Doyle, which can be downloaded from the Gutenberg Project [10]. The document is processed
by removing all the "stop words", which do not bring meaningful information to contexts (words with
very high frequency like: "the"), we also eliminate all the non-alphanumeric words, and the remaining
words are lemmatized. The resulted corpus has 38,812 words, with a vocabulary of 5,829 unique words.
In Figure 4 we show the (ordered) distribution of the number of words, and of the unique number of
words, in the context of each word from the resulted vocabulary, for a window length 2L = 10. One
can see that very few words (176 out of 5,829, or 3%) have a context with a number of words larger
than 375. Thus, by using random vectors with dimensionality d = 1000, we can still have a quite
high precision (recall) in most cases. Here are some interesting similarities derived from the resulted
word-context vectors: (addicted : college, theological); (administration : affairs, secretary); (advise :
watson); (americans : finns, germans); (answer : say); (arizona : montana); (arm : hand); (arrange
: extract); (artery : roadway, traffic). Here are also some relevant examples of word-context vector
operations: accent - german ' proficient; acid - pungent ' hydrochloric; aged - grizzle ' middle.
4.2
Sentence similarity
Another possible application is to search a document in order to find similar sentences to a given "query"
sentence. Again, we associate the vocabulary V = {wi | i = 1, ..., n} with a set Ξ = {ξi | ξi ∈ B d , i =
8
1, ..., n} of "almost orthogonal" random vectors, and we assume that each sentence from a document
D is represented by the sum of its words. Thus, the ith sentence si = {wi1 , ..., wimi } of a document D,
will be represented by the vector:
γi =
mi
X
ξ ij ,
(31)
j=1
where ij ≡ χ(wij ).
Let us also assume that the vector corresponding to the sentence query q = {w1 , ..., wk } is:
ζ=
k
X
ξ` ,
(32)
`=1
where ` ≡ χ(w` ). Finding the most similar sentence to q is equivalent to solving:
i∗ = arg max
i
ζ T γi
.
kζkkγi k
(33)
The expectation of the dot product ηi = νζ T γi is E(ηi ) = νni , where ni is the number of words the
sentences q and si have in common, and ν = (kζkkγi k)−1 . Also, the variance of the dot product is:
σi2 = νni /d. Thus, with a reasonable high dimensionality d we can obtain a very good precision (recall).
For example, let’s consider again the book The Adventures of Sherlock Holmes, and we ask the
following naive question: "Who is the woman Irene in the photograph, and what is her special connection
to Sherlock?" The question is naive because after processing the "query" sentence, only the following
words actually are used in the search process: woman, irene, photograph, special, connection, sherlock,
and they are also independent of eachother, since only their "sum" is used.
The first three sentences returned using the method described above are: (1) "And when he speaks
of Irene Adler, or when he refers to her photograph, it is always under the honourable title of the
woman."; (2) "And yet there was but one woman to him, and that woman was the late Irene Adler, of
dubious and questionable memory."; (3) "The photograph was of Irene Adler herself in evening dress,
the letter was superscribed to Sherlock Holmes, Esq."
We also used the dot product without normalization ηi = ζ T γi , and the results for the first three
returned sentences are also interesting: (1) "To Sherlock Holmes she is always THE woman."; (2) "And
when he speaks of Irene Adler, or when he refers to her photograph, it is always under the honourable
title of the woman."; (3) "And yet there was but one woman to him, and that woman was the late Irene
Adler, of dubious and questionable memory."
Therefore, the normalization of the sentence vectors before taking the dot product may or may not
be necessary, and both cases may return relevant results.
4.3
Spam filtering
Spam filters are built in order to protect email users from spam and phishing messages. Most spam
filters are word-based filters, which simply block any email that contains certain words or phrases.
Another approach is based on machine learning techniques such as Bayesian classifiers, which must be
trained on large sets of already classified spam and non-spam messages. Here we discuss a different
9
approach, based on the "almost orthogonal" property of random vectors.
As in the previously described applications, we associate the vocabulary V = {wi | i = 1, ..., n}
with a set Ξ = {ξi | ξi ∈ B d , i = 1, ..., n} of "almost orthogonal" random vectors, and we assume that
each message is represented by the sum of its words, and equivalently the sum of "almost orthogonal"
vectors representing the words. Therefore, we assume that we have m messages gj , j = 1, ..., m, already
classified, such that the associated vectors are:
γj =
mj
X
ξji ,
(34)
i=1
where mj is the number of words in the message gj . Also, the class of each message gj , j = 1, ..., m, is
known:
1 if g spam
j
class(gj ) =
.
0 if g non-spam
j
(35)
Now, let us assume that h is a new message, with the associated vector:
ζ=
k
X
ξ` ,
(36)
`=1
where k is the number of words in h.
In order to classify h as spam or non-spam we simply compute:
ζ T γj
,
j=1,...,m kζkkγj k
j ∗ = arg max
(37)
and we assign to h the class of gj ∗ :
class(h) = class(gj ∗ )
(38)
Thus, the class attributed to h is the class of the most similar, and already classified message gj ∗ .
In order to evaluate this very simple method we use the Ling-Spam corpus [11], as described in the
paper Ref. [12]. The data set contains four subdirectories, corresponding to four versions of the corpus:
(1) bare: lemmatiser disabled, stop-list disabled; (2) lemm: lemmatiser enabled, stop-list disabled; (3)
lemm-stop: lemmatiser enabled, stop-list enabled; (4) stop: lemmatiser disabled, stop-list enabled.
In our experiment we used the files from the first subdirectory: "bare: lemmatiser disabled, stoplist disabled". This directory contains 10 subdirectories (part1,..., part10), corresponding to the 10
partitions of the corpus used in the 10-fold cross validation experiment. In each repetition, one part is
reserved for testing and the other 9 are used for training. Each one of the 10 subdirectories contains
both spam and legitimate messages. The total number of files is 2,893. Files whose names have the
form "spmsg*.txt" are spam messages. All other files are legitimate messages.
We preprocessed the messages using the spaCy Python library [13]. The messages were processed
by removing the "stop words" and the remaining words were lemmatized, resulting in a vocabulary of
54,442 unique words. The results for 10-fold cross validation are shown in Figure 5, for three different
vector dimensionality values d = 1000, 2000, 3000. One can see that in all three cases the average values
are quite close, indicating that decreasing the dimensionality from 3000 to 1000 has only a slight effect
10
d=1000
d=2000
d=3000
spam precision
1.00
0.98
0.96
mean values
0.94
0.92
0.90
0.88
0.93 0.94 0.95 0.96 0.97 0.98 0.99 1.00
spam recall
Figure 5: Spam precision and recall for d = 1000, 2000, 3000.
on the classification precision and recall. Also, one can see that the described method based on "almost
orthogonal" random vectors gives better results (recall ≈ 0.967, precision ≈ 0.946 for d = 3000) than
the Bayesian approach described in Ref. [12], even though in this case there is no learning involved.
5
Conclusion
In this paper we have explored a different approach to the "vector semantics" problem, which is based
on the "almost orthogonal" property of high-dimensional random vectors. We have shown that the
"almost orthogonal" property can be used to "memorize" random vectors by simply adding them, and
we have provided an efficient probabilistic solution to the set membership problem. Also, we have
discussed several applications to word and context vector embeddings, document sentences similarity,
and spam filtering. One can easily extend this approach to other problems, like for example sentiment
analysis. Contrary to the "expensive" machine learning methods, this method is very simple and it
does not even require a "learning" process, however it exhibits similar properties.
References
[1] Y. Goldberg, Neural Network Methods in Natural Language Processing (Synthesis Lectures on
Human Language Technologies), Morgan and Claypool Pub. (2017).
[2] Y. Goldberg, A Primer on Neural Network Models for Natural Language Processing, Journal of
Artificial Intelligence Research 57, 345 (2016).
[3] O. Levy, Y. Goldberg, Neural Word Embedding as Implicit Matrix Factorization, NIPS’14 Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2
Page 2177, Montreal, Canada (2014).
11
[4] T. Mikolov, K. Chen, G. Corrado, J. Dean. 2013a. Efficient estimation of word representations in
vector space. CoRR, https://arxiv.org/pdf/1301.3781.pdf (2013).
[5] P. Kanerva, J. Kristofersson, A. Holst, Random Indexing of text samples for Latent Semantic
Analysis, Proceedings of the 22nd annual conference of the cognitive science society. New Jersey:
Erlbaum, 2000.
[6] M. Sahlgren, An introduction to random indexing. In Proceedings of the Methods and Applications
of Semantic Indexing Workshop at the 7th International Conference on Terminology and Knowledge
Engineering (TKE), Copenhagen, Denmark (2005).
[7] R. Hecht-Nielsen, Context vectors; general purpose approximate meaning representations selforganized from raw data. In J. M. Zurada et al. Computational intelligence: imitating life. IEEE
Press (1994).
[8] H. Chernoff, A Measure of Asymptotic Efficiency for Tests of a Hypothesis Based on the sum of
Observations, Annals of Mathematical Statistics 23, 493 (1952).
[9] A.N. Gorban, I.Yu. Tyukin, D.V. Prokhorov, K.I. Sofeikov, Approximation with Random Bases:
Pro et Contra, Information Sciences, 364-365, 129 (2016).
[10] http://www.gutenberg.org/ebooks/1661.
[11] Ling-Spam corpus, http://www.aueb.gr/users/ion/.
[12] I. Androutsopoulos, J. Koutsias, K.V. Chandrinos, George Paliouras, and C.D. Spyropoulos,
An Evaluation of Naive Bayesian Anti-Spam Filtering, in Potamias, G., Moustakis, V. and van
Someren, M. (Eds.), Proceedings of the Workshop on Machine Learning in the New Information
Age, 11th European Conference on Machine Learning (ECML 2000), Barcelona, Spain, pp. 9-17,
2000.
[13] spaCy - Industrial-strength Natural Language Processing in Python, https://spacy.io.
12
| 2cs.AI
|
Collaborative Learning for Weakly Supervised Object Detection
arXiv:1802.03531v1 [cs.CV] 10 Feb 2018
Jiajie Wang, Jiangchao Yao, Ya Zhang, Rui Zhang
Cooperative Madianet Innovation Center
Shanghai Jiao Tong University
{ww1024,sunarker,ya zhang,zhang rui}@sjtu.edu.cn
Abstract
Weakly supervised object detection has recently received much attention, since it only requires imagelevel labels instead of the bounding-box labels consumed in strongly supervised learning. Nevertheless, the save in labeling expense is usually at the
cost of model accuracy. In this paper, we propose a simple but effective weakly supervised collaborative learning framework to resolve this problem, which trains a weakly supervised learner and
a strongly supervised learner jointly by enforcing
partial feature sharing and prediction consistency.
For object detection, taking WSDDN-like architecture as weakly supervised detector sub-network
and Faster-RCNN-like architecture as strongly supervised detector sub-network, we propose an endto-end Weakly Supervised Collaborative Detection
Network. As there is no strong supervision available to train the Faster-RCNN-like sub-network, a
new prediction consistency loss is defined to enforce consistency of predictions between the two
sub-networks as well as within the Faster-RCNNlike sub-networks. At the same time, the two detectors are designed to partially share features to further guarantee the model consistency at perceptual
level. Extensive experiments on PASCAL VOC
2007 and 2012 data sets have demonstrated the effectiveness of the proposed framework.
1
Introduction
Learning frameworks with Convolutional Neural Network
(CNN) [Girshick, 2015; Ren et al., 2015; Redmon and
Farhadi, 2016] have persistently improved the accuracy and
efficiency of object detection over the recent years. However, most existing learning-based object detection methods
require strong supervisions in the form of instance-level annotations (e.g. object bounding boxes) which are labor extensive to obtain. As an alternative, weakly supervised object
detection explores image-level annotations that are more accessible from rich media data [Thomee et al., 2015].
A common practice for weakly supervised object detection is to model it as a multiple instance learning (MIL) problem, treating each image as a bag and the target proposals
as instances. Therefore, the learning procedure is alternating
between training an object classifier and selecting most confident positive instances [Bilen et al., 2015; Cinbis et al., 2017;
Zhang et al., 2006]. Recently, CNNs are leveraged for
the feature extraction and classification [Wang et al., 2014].
Some methods further integrate the instance selection step in
deep architectures by aggregating proposal scores to imagelevel predictions [Wu et al., 2015; Bilen and Vedaldi, 2016;
Tang et al., 2017] and build an efficient end-to-end network.
While the above end-to-end weakly supervised networks
have shown great promise for weakly supervised object detection, there is still a large gap in accuracy compared to
their strongly supervised counterparts. Several studies have
attempted to combine weakly and strongly supervised detectors in a cascaded manner, aiming to further refine coarse detection results by leveraging powerful strongly supervised detectors. Generally, instance-level predictions from a trained
weakly supervised detector are used as pseudo labels to train
a strongly supervised detector [Tang et al., 2017]. However,
these methods only consider a one-off unidirectional connection between two detectors, making the prediction accuracy
of the strongly supervised detectors depend heavily on that of
the corresponding weakly supervised detectors.
In this paper, we propose a novel weakly supervised collaborative learning (WSCL) framework which bridges weakly
supervised and strongly supervised learners in a unified learning process. The consistency of two learners, for both shared
features and model predictions, is enforced under the WSCL
framework. Focusing on object detection, we further develop
an end-to-end weakly supervised collaborative detection network, as illustrated in Fig. 1. A WSDDN-like architecture
is chosen for weakly supervised detector sub-network and a
Faster-RCNN-like architecture is chosen for strongly supervised detector sub-network. During each learning iteration,
the entire detection network takes only image-level labels
as the weak supervision and the strongly supervised detector sub-network is optimized in parallel to the weakly supervised detector sub-network by a carefully designed prediction
consistency loss, which enforces the consistency of instancelevel predictions between and within the two detectors. At
the same time, the two detectors are designed to partially
share features to further guarantee the model consistency at
perceptual level. Experimental results on the PASCAL VOC
2007 and 2012 data sets have demonstrated that the two de-
Classification
Loss
…
Image-level
Label
Weakly Supervised
Detector
Weakly
supervised
X
Image level
Predictions
X X
Learner
Weakly
Weakly
supervised
supervised
Learner
Learner
X
Strongly
supervised
Strongly
Strongly
Learner
supervised
X X
supervised
Learner
Learner
X1
Yweak
Yweak
Yweak
XX1 1
^
^
Y
^
YY
XX2 2
(a) WSCL
Feature
Sharing
Bounding-box
Predictions
Prediction
Consistency
Loss
Training Images
Strongly Supervised
Detector
Figure 1: The proposed weakly supervised collaborative learning
framework. A weakly supervised detector and a strongly supervised
detector are integrated into a unified architecture and trained jointly.
tectors mutually enhance each other through the collaborative
learning process. The resulting strongly supervised detector
manages to outperform several state-of-the-art methods. The
main contributions of the paper are summarized as follows.
• We propose a new collaborative learning framework for
weakly supervised object detection, in which two types of
detectors are trained jointly and mutually enhanced.
• To optimize the strongly supervised detector sub-network
without strong supervisions, a prediction consistency loss
is defined between the two sub-networks as well as within
the strongly supervised detector sub-network.
• We experiment with the widely used PASCAL VOC 2007
and 2012 data sets and show that the proposed approach
outperforms several state-of-the-art methods.
2
Weakly Supervised Collaborative Learning
Framework
Given two related learners, one weakly supervised learner
DW and one strongly supervised learner DS , we propose a
weakly supervised collaborative learning (WSCL) framework
to jointly train the two learners, leveraging the task similarity
between the two learners. As shown in Fig. 2(a), DW learns
from weak supervisions and generates fine-grained predictions such as object locations. Due to lack of strong supervisions, DS cannot be directly trained. But it is expected that
DS and DW shall output similar predictions for the same image if trained properly. Hence, DS learns by keeping its predictions consistent with that of DW . Meanwhile, DS and
DW are also expected to partially share feature representations as their tasks are the same. The WSCL framework thus
enforces DS and DW to partially share network structures
and parameters. Intuitively, DS with reasonable amount of
strong supervisions is expected to learn better feature representation than DW . By bridging the two learners under this
collaborative learning framework, we enable them to mutual
reinforcement each other through the joint learning process.
WSCL is similar to several learning frameworks such as
co-training and the EM-style learning as shown in Fig. 2.
Strongly
supervise
Learner A
Strongly
Strongly
supervise
semi
supervise
YYsemi
LearnerAA
Learner
Strongly
supervise
X 2 Strongly
Strongly
Learner Y
B
supervise
supervise
Learner B
Learner B
Ysemi
X
Ysemi
semi
Ysemi
(b) Co-training
XX
Strongly
Strongly
supervised
supervised
Learner
Learner
Strongly
supervised
Learner
^^
^
Y
YY
(c) EM-style
Figure 2: Comparison of WSCL with co-training and EM-style
frameworks. See text for details.
Co-training framework [Blum and Mitchell, 1998] is designed for semi-supervised settings, where two parallel learners are optimized with distinct views of data. Whenever the
labels in either learner are unavailable, its partner’s prediction can be used for auxiliary training. Compared with the
homogeneous collaboration in co-training, the WSCL framework is heterogeneous, i.e. the two learners have different
types of supervisions. Moreover, two learners in WSCL are
trained jointly rather than iteratively. EM-style framework
for weakly supervised object detection task [Jie et al., 2017;
Yan et al., 2017] usually utilizes a strongly supervised learner
to iteratively select training samples according to its own predictions. However, the strongly supervised learner in this
framework may not get stable training samples since it is sensitive to the initialization. By contrast, WSCL trains a weakly
supervised and a strongly supervised learner jointly and enables them to mutually enhance each other.
3
Weakly Supervised Collaborative Detection
In this section, we focus on the object detection applications.
Given a training set {(xn , yn ), n = 1, · · · , N }, where N is
the size of training set, xn is an image, and the image’s label yn ∈ RC is a C-dimensional binary vector indicating the
presence or absence of each category. The task is to learn an
object detector which predicts the locations of objects in an
image as {(pi , ti ), i = 1, · · · , B}, where B is the number of
proposal regions. And for the i-th proposal region x(i) , pi
is a vector of category probability, and ti is a vector of four
parameterized bounding box coordinates. The image-level
annotation y is considered as a form of weak supervisions,
because the detector is also expected to predict object categories and locations in terms of bounding boxes.
Under the weakly supervised collaborative learning framework, we propose a Weakly Supervised Collaborative Detection Network (WSCDN). A two-stream CNN similar to WSDDN [Bilen and Vedaldi, 2016] is chosen as the weakly supervised learner DW and Faster-RCNN [Ren et al., 2015] is
chosen as the strongly supervised learner DS . The two learners are integrated into an end-to-end collaborative learning
network as two sub-networks. The overall architecture is illustrated in Fig. 3.
3.1
Base Detectors
As shown in the blue area of Fig. 3, the weakly supervised
detector DW is composed of three parts. The first part (up to
FC7) takes pre-generated proposal regions and extracts features for each proposal. The middle part consists of two par-
Strongly Supervised Detector
SPPrpn
FC6
FC8_cls
{( pi , ti )}
FC7
Prediction
Consistency
Loss
FC8_reg
{( p j , t j )}
RPN
x
Conv1-5
SSW
SPPss
FC6
FC7
FC8_cls
FC8_loc
𝜎𝑐𝑙𝑎𝑠𝑠
𝜎𝑙𝑜𝑐
s cls
s loc
Weakly Supervised Detector
p
𝐀𝐠𝐠
ŷ
Multi-label
Classification
Loss
y
Figure 3: The architecture of our WSCDN model built based on VGG16. Red and blue lines are the forward paths for the strongly and weakly
supervised detectors respectively, while black solid and dashed lines indicate the shared parts of two detectors.
allel streams, one to compute classification score scls
jc and the
other to compute location score sloc
of
each
proposal
region
jc
x(j) for category c. The last part computes product over the
two scores to get a proposal’s detection score pjc , and then
aggregates the detection scores over all proposals to generate the image-level prediction ŷc . Suppose the weakly supervised detector DW has BW proposal regions, the aggregation
of prediction scores from the instance-level to the image-level
can be represented as
ŷc =
BW
X
pjc ,
where
loc
pjc = scls
jc · sjc .
(1)
j=1
With the above aggregation layer, DW can be trained in an
end-to-end manner given the image-level annotations y and
is able to give coordinate predictions directly from x(j) and
category predictions from pjc .
The network architecture of the strongly supervised detector DS is shown in the red area of Fig. 3. Region proposal
network (RPN) is used to extract proposals online. Then
bounding-box predictions {(p, t)} are made through classifying the proposals and refining their coordinates.
3.2
Collaborative Learning Network
For collaborative learning, the two learners are integrated into
an end-to-end architecture as two sub-networks and trained
jointly in each forward-backward iteration. Because the training data only have weak supervision in forms of classification
labels, we design the following two sets of losses for model
training. The first one is similar to WSDDN and many other
weakly supervised detectors and the second one focuses on
checking the prediction consistency, both between the two
detectors and within the strongly supervised detector itself.
For the weakly supervised detector sub-network DW , it
outputs category predictions on the image level as well as
location predictions on the object level. Given weak supervision y at the image level, we define a classification loss in
the form of a multi-label binary cross-entropy loss between y
and the image-level prediction yˆc from DW :
L (DW ) = −
C
X
c=1
(yc log ŷc + (1 − yc ) log(1 − ŷc )).
(2)
L(DW ) itself can be used to train a weakly supervised detector, as has been demonstrated in WSDDN. Under the
proposed collaborative learning framework, L(DW ) is also
adopted to train the weakly supervised sub-network DW .
Training the strongly supervised detector sub-network DS
independently usually involves losses consisting of a category
classification term and a coordinate regression term, which
requires instance-level bounding box annotations. However,
the strong supervisions in terms of instance-level labels are
not available in the weak settings. The major challenge
for training the weakly supervised collaborative detector network is how to define loss to optimize DS without requiring
instance-level supervisions at all. Considering both DW and
DS are designed to predict object bounding-boxes, we propose to leverage the prediction consistency in order to train
the strongly supervised sub-network DS . The prediction consistency consists of two parts: between both DW and DS and
only within DS . The former one enforces that the two detectors give similar predictions both in object classification and
object locations when converged. The latter one is included
because the output of DW is expected to be quite noisy, especially at the initial rounds of the training. Combining these
above two kinds of prediction consistency, we define the loss
function for training DS as
L(DS ) = −
BW X
BS X
C
X
j=1 i=1 c=1
Iij (β pjc log pic +(1 − β) pic log pic + pjc R(tjc − tic ))
| {z } |
| {z }
{z
}
P
Cinter
P
Cinner
L
Cinter
(3)
P
Cinter
P
Cinner
where the first two cross-entropy terms
and
consider the consistency of category predictions both on the
inter and inner level; pjc and pic are the category predicL
tions from DW and DS respectively; the last one Cinner
is
a regression term promising the consistency of only internetworks’ localization predictions, which measures the coordinate difference between proposals from DS and DW . Here,
R (·) is a smooth L1 loss [Girshick, 2015] and is weighted by
pj ; BW and BS are the numbers of proposal regions for DW
and DS in a mini-batch respectively; Iij is a binary indicator
with the value of 1 if the two proposal regions x(i) and x(j)
are closet and have a overlap ratio (IoU) more than 0.5, and
0 otherwise; β ∈ (0, 1) is a hyper parameter which balances
two terms of consistency loss for category predictions. If β is
larger than 0.5, DS will trust predictions from DW more than
from itself.
Max-out Strategy
The predictions of DS and DW could be inaccurate, especially in the initial rounds of training. For measuring the
prediction consistency, it is important to select only the most
confident predictions. We thus apply a Max-out strategy to
filter out most predictions. For each positive category, only
the region with highest prediction score by DW is chosen.
That is, if yc = 1, we have:
X
p̂jc∗ c = 1, s.j.
p̂jc = 1, where jc∗ = arg max pjc . (4)
j
j
If yc = 0, we have p̂jc = 0, ∀j, c. The category prediction
p̂jc is then used to replace pjc when calculating the consistency loss in L (DS ). The Max-out strategy can also reduce
the region numbers of DW used to calculate the prediction
consistency loss and thus can save much training time.
Feature Sharing
As the two detectors in WSCDN are designed to learn under different forms of supervision but for the same prediction task, the feature representations learned through the collaboration process are expected to be similar to each other.
We thus enforce the partial feature sharing between two subnetworks so as to ensuring the perceptual consistency of the
two detectors. Specifically, the weights of convolutional
(conv) layers and part of bottom fully-connected (fc) layers
are shared between DW and DS .
Network Training
With the image-level classification loss L (DW ) and instancelevel prediction consistency loss L (DS ), the parameters of
two detectors can be updated jointly with only image-level labels by the stochastic gradient descent (SGD) algorithm. The
gradients for individual layers of DS and DW are computed
only respect to L (DS ) and L (DW ) respectively, while the
shared layers’ gradients are produced by both loss functions.
4
4.1
Experimental Results
Data Sets and metrics
We experiment with two widely used benchmark data sets:
PASCAL VOC 2007 and 2012 [Everingham et al., 2010],
both containing 20 common object categories with a total of
9,962 and 22,531 images respectively. We follow the standard splits of the data sets and use the trainval set with only
image-level labels for training and the test set with groundtruth bounding boxes for testing.
Two standard metrics, Mean average precision (mAP) and
Correct localization (CorLoc) are adopted to evaluate different weakly supervised object detection methods. The mAP
measures the quality of bounding box predictions in test set.
Following [Everingham et al., 2010], a prediction is considered as true positive if its IoU with the target ground-truth is
Table 1: Comparison of detectors built with the WSCL framework
to their baselines and counterparts in terms of mAP and CorLoc on
PASCAL VOC 2007 data set.
Methods
mAP(%)
CorLoc(%)
IW
28.5
45.6
CLW
40.0
58.4
CLS
48.3
64.7
CSS
39.4
59.3
larger than 0.5. CorLoc of one category is computed as the
ratio of images with at least one object being localized correctly. It is usually used to measure the localization ability in
localization tasks where image labels are given. Therefore,
it is a common practice to validate the model’s CorLoc on
training set [Deselaers et al., 2012].
4.2
Implementation details
Both the weakly and strongly supervised detectors in the
WSCDN model are built on VGG16 [Simonyan and Zisserman, 2014], which is pre-trained on a large scale image classification data set, ImageNet [Russakovsky et al., 2015]. We
replace Pool5 layer with SPP layer [He et al., 2014] to extract region features. Two detectors share weights for convolutional (conv) layers and two fully-connected (fc) layers,
i.e., fc6, fc7. For the weakly supervised detector, we use
SelectiveSearch [Uijlings et al., 2013] to generate proposals and build network similar with WSDDN: the last fc layer
in VGG16 is replaced with a two-stream structure in 3.1, as
each stream consists a fc layer followed by a softmax layer
focusing on classification and localization respectively. For
the strongly supervised detector Faster-RCNN, we follow the
model structure and setting of its original implementation.
At training time, we apply image multi-scaling and random
horizontal flipping for data augmentation, with the same parameters in [Ren et al., 2015]. We empirically set the hyper
parameter β to 0.8. RPN and the following region-based detectors in Fasrer-RCNN are trained simultaneously. We train
our networks for total 20 epochs, setting the learning rate of
the first 12 epochs to 1e-3, and the last 8 epochs to 1e-4. At
test time, we obtain two sets of predictions for each image
from the weakly and strongly supervised detectors, respectively. We apply non-maximum suppression to all predicted
bounding boxes, with the IoU threshold set to 0.6.
4.3
Influence of Collaborative Learning
To investigate the effectiveness of the collaborative learning
framework for weakly supervised object detection, we compare the following detectors: 1) the weakly and strongly supervised detectors built with the collaborative learning framework, denoted as CLW and CLS , respectively; 2) The initial weakly supervised detector built above before collaborative learning as the baseline, denoted as IW ; 3) The same
weakly supervised and strongly supervised detector networks
trained in cascaded manner similar to [Tang et al., 2017;
Yan et al., 2017]. The resulting strongly supervised detector is denoted as CSS .
The mAPs and CorLoc on PASCAL VOC 2007 data set are
presented in Table 1. Among the four detectors under comparison, CLS achieves the best performance in terms of mAP
and CorLoc and outperforms the baseline IW , its collaborator
IW
CLW
CS S
CLS
Figure 4: Visualization of the detection results of four detectors in Table 1. Images from the 1st to 4th row are results from the IW , CLW ,
CSS and CLS respectively.
50
48
46
44
mAP (%)
CLW , and its cascade counterpart, CSS . Compared to CSS ,
the mAP and CorLoc are improved from 39.4% to 48.3%
and from 59.3% to 64.7%, respectively, suggesting the effectiveness of the proposed collaborative learning framework.
Furthermore, CLW outperforms IW in terms of mAP and
CorLoc by a large margin of 11.5% and 12.8%, respectively,
showing that the parameters sharing between the two detectors enables a better feature representation and thus leading
to significantly improve of the weakly supervised detector.
We also qualitatively compare the detection results of IW ,
CLW , CSS and CLS . As can be seen in Fig. 4, the strongly
supervised detector CLS clearly outperforms the other three
detectors, with more objects correctly detected. For example, in the first column and the fifth column where there are
multiple objects in one images, only CLS is able to correctly detect all of them, while the other three detectors all
missed one or more objects. Moreover, CLS generates more
accurate bounding boxes. Weakly supervised detectors are
known for often generating bounding boxes that only cover
the most discriminate part of an object (e.g. face of a person
or wings/head of a bird). CLS can generate more bounding
boxes that tightly cover the entire objects as shown in the third
and fifth column of Fig. 4, indicating the collaborative learning framework is able to learn a better feature representation
for objects. Compared to IW , CLW generates tighter object bounding box in the second and fourth columns, i.e. the
performance of the weakly supervised detector is improved
after collaborative learning, suggesting that feature sharing
between the two detectors helps optimizing the weakly supervised detector.
To show how CLW and CLS improve during the collaborative learning, we plot mAP of the two detectors for different training iterations. As shown in Fig. 5, both detectors
get improved with increasing training iterations. Initially, the
strongly supervised detector CLS has a smaller mAP than
42
40
38
36
34
CL-W
CLW
32
CLS
CL-S
30
1
2
4
5
7
8
9
Number of iterations (10k)
Figure 5: The changes of mAP for CLS and CLW on PASCAL
VOC 2007 data set during the process of collaborative learning.
the weakly supervised detector CLW . But in a dozen thousands iterations, CLS surpasses CLW and further outperforms CLW by a large margin in the end, suggesting the effectiveness of the prediction consistency loss we proposed.
4.4
Comparison with the state-of-the-arts
In general, two types of weakly supervised object detection methods are compared. The first includes the MIL
methods [Cinbis et al., 2017; Wang et al., 2014] and various end-to-end MIL-CNN models [Bilen and Vedaldi, 2016;
Kantorov et al., 2016; Tang et al., 2017] following the twostream structure of WSDDN [Bilen and Vedaldi, 2016]. The
second type of methods build a curriculum pipeline to find
confident regions online, and train an instance-level modern
detector in a strongly supervised manner [Li et al., 2016;
Jie et al., 2017]. So the detectors they used to report the
results share a similar structure and characteristics with our
strongly supervised detector.
Table 2: Comparison of WSCDN to the state-of-the-art on PASCAL VOC 2007 test set in terms of average precision (AP) (%).
Methods
[Cinbis et al., 2017]
[Wang et al., 2014]
[Bilen and Vedaldi, 2016]
[Kantorov et al., 2016]
[Tang et al., 2017]
[Li et al., 2016]
[Jie et al., 2017]
CLW
CLS
aer
38.1
48.9
39.4
57.1
58.0
54.5
52.2
59.7
61.2
bik
47.6
42.3
50.1
52.0
62.4
47.4
47.1
54.7
66.6
brd
28.2
26.1
31.5
31.5
31.1
41.3
35.0
31.6
48.3
boa
13.9
11.3
16.3
7.6
19.4
20.8
26.7
24.1
26.0
btl
13.2
11.9
12.6
11.5
13.0
17.7
15.4
13.2
15.8
bus
45.2
41.3
64.5
55.0
65.1
51.9
61.3
59.6
66.5
car
48.0
40.9
42.8
53.1
62.2
63.5
66.0
53.2
65.4
cat
19.3
34.7
42.6
34.1
28.4
46.1
54.3
39.0
53.9
cha
17.1
10.8
10.1
1.7
24.8
21.8
3.0
19.3
24.7
cow
27.7
34.7
35.7
33.1
44.7
57.1
53.6
49.9
61.2
tbl
17.3
18.8
24.9
49.2
30.6
22.1
24.7
35.8
46.2
dog
19.0
34.4
38.2
42.0
25.3
34.4
43.6
45.0
53.5
hrs
30.1
35.4
34.4
47.3
37.8
50.5
48.4
38.2
48.5
mbk
45.4
52.7
55.6
56.6
65.5
61.8
65.8
63.6
66.1
prs
13.5
19.1
9.4
15.3
15.7
16.2
6.6
7.1
12.1
plt
17.0
17.4
14.7
12.8
24.1
29.9
18.8
16.9
22.0
shp
28.8
35.9
30.2
24.8
41.7
40.7
51.9
36.6
49.2
sfa
24.8
33.3
40.7
48.9
46.9
15.9
43.6
47.9
53.2
trn
38.2
34.8
54.7
44.4
64.3
55.3
53.6
54.9
66.2
tv
15.0
46.5
46.9
47.8
62.6
40.2
62.4
50.0
59.4
Avg.
27.4
31.6
34.8
36.3
41.2
39.5
41.7
40.0
48.3
Table 3: Comparison of WSCDN to the state-of-the-art on PASCAL VOC 2007 trainval set in terms of Correct Localization (CorLoc) (%).
Methods
[Cinbis et al., 2017]
[Wang et al., 2014]
[Bilen and Vedaldi, 2016]
[Kantorov et al., 2016]
[Tang et al., 2017]
[Li et al., 2016]
[Jie et al., 2017]
CLW
CLS
aer
57.2
80.1
65.1
83.3
81.7
78.2
72.7
82.5
85.8
bik
62.2
63.9
58.8
68.6
80.4
67.1
55.3
75.7
80.4
brd
50.9
51.5
58.5
54.7
48.7
61.8
53.0
63.1
73.0
boa
37.9
14.9
33.1
23.4
49.5
38.1
27.8
44.1
42.6
btl
23.9
21.0
39.8
18.3
32.8
36.1
35.2
32.4
36.6
bus
64.8
55.7
68.3
73.6
81.7
61.8
68.6
72.1
79.7
car
74.4
74.2
60.2
74.1
85.4
78.8
81.9
76.7
82.8
cat
24.8
43.5
59.6
54.1
40.1
55.2
60.7
50.3
66.0
cha
29.7
26.2
34.8
8.6
40.6
28.5
11.6
35.0
34.1
cow
64.1
53.4
64.5
65.1
79.5
68.8
71.6
74.0
78.1
tbl
40.8
16.3
30.5
47.1
35.7
18.5
29.7
30.8
36.9
dog
37.3
56.7
43.0
59.5
33.7
49.2
54.3
57.9
68.6
hrs
55.6
58.3
56.8
67.0
60.5
64.1
64.3
57.5
72.4
mbk
68.1
69.5
82.4
83.5
88.8
73.5
88.2
82.3
91.6
prs
25.5
14.1
25.5
35.3
21.8
21.4
22.2
19.1
22.2
plt
38.5
38.3
41.6
39.9
57.9
47.4
53.7
47.6
51.3
shp
65.2
58.8
61.5
67.0
76.3
64.6
72.2
76.3
79.4
sfa
35.8
47.2
55.9
49.7
59.9
22.3
52.6
50.0
63.7
trn
56.6
49.1
65.9
63.5
75.3
60.9
68.9
71.1
74.5
tv
33.5
60.9
63.7
65.2
81.4
52.3
75.5
69.5
74.6
Avg.
47.3
48.5
53.5
55.1
60.6
52.4
56.1
58.4
64.7
Table 4: Comparison of WSCDN to the state-of-the-art on PASCAL VOC 2012 test set in terms of average precision (AP) (%).
Methods
[Kantorov et al., 2016]
[Tang et al., 2017]
[Jie et al., 2017]
CLW
CLS
aer
64.0
67.7
60.8
64.0
70.5
bik
54.9
61.2
54.2
60.3
67.8
brd
36.4
41.5
34.1
40.1
49.6
boa
8.1
25.6
14.9
18.5
20.8
btl
12.6
22.2
13.1
15.0
22.1
bus
53.1
54.6
54.3
57.4
61.4
car
40.5
49.7
53.4
38.3
51.7
cat
28.4
25.4
58.6
25.3
34.7
cha
6.6
19.9
3.7
17.3
20.3
cow
35.3
47.0
53.1
32.4
50.3
tbl
34.4
18.1
8.3
16.5
19.0
dog
49.1
26.0
43.4
33.1
43.5
hrs
42.6
38.9
49.8
28.6
49.3
mbk
62.4
67.7
69.2
64.8
70.8
prs
19.8
2.0
4.1
6.9
10.2
plt
15.2
22.6
17.5
16.6
20.8
shp
27.0
41.1
43.8
34.3
48.1
sfa
33.1
34.3
25.6
41.4
41.0
trn
33.0
37.9
55.0
52.4
56.5
tv
50.0
55.3
50.1
51.2
56.7
Avg.
35.3
37.9
38.3
35.7
43.3
Table 5: Comparison of WSCDN to the state-of-the-art on PASCAL VOC 2012 trainval set in terms of Correct Localization (CorLoc) (%).
Methods
[Kantorov et al., 2016]
[Tang et al., 2017]
[Jie et al., 2017]
CLW
CLS
aer
78.3
86.2
82.4
88.0
89.2
bik
70.8
84.2
68.1
79.7
86.0
brd
52.5
68.7
54.5
66.4
72.8
boa
34.7
55.4
38.9
51.0
50.4
btl
36.6
46.5
35.9
40.9
40.1
bus
80.0
82.8
84.7
84.0
87.7
car
58.7
74.9
73.1
65.4
72.6
cat
38.6
32.2
64.8
35.6
37.0
For the PASCAL VOC 2007 dataset, the mAP and CorLoc
results are shown in Table 2 and Table 3, respectively. The
propose model gets 39.4% and 49.4% in terms of map for
the weakly supervised detector and the strongly supervised
detector respectively. On CorLoc, our two detectors also perform well, get 61.1% and 67.5%. In particular, the strongly
supervised detector CLS in our model receives best results
among those methods by both mAP and CorLoc.
Compared to the first type of methods, CLS improves detection performance by more than 7.1% on mAP and 4.1% on
CorLoc. Our CLW that has a similar but the simplest structure, also gets comparable results with regard to other models,
revealing the effectiveness of collaborative learning. With respect to the second set of methods under comparison, we use
a weakly supervised detector to achieve confident region selection in a collaboration learning process, compared to those
complicated schemes. The collaborative learning framework
enables the strongly supervised detector CLS to outperform
[Jie et al., 2017] by 6.6% and 8.6% on mAP and CorLoc respectively.
Similar results are obtained on PASCAL VOC 2012 dataset
as shown in Table 4 and Table 5. CLS achieved 43.3% on
mAP and 65.2% on CorLoc, both of which outperform the
other state-of-the-art methods, indicating the effectiveness of
cha
27.7
46.7
17.1
46.5
48.2
cow
71.2
82.8
78.3
69.9
80.3
tbl
32.3
42.9
22.5
46.6
49.3
dog
48.7
41.0
57.0
49.7
54.4
hrs
76.2
68.1
70.8
52.4
72.7
mbk
77.4
89.6
86.6
89.2
88.8
prs
16.0
9.2
18.7
21.2
21.6
plt
48.4
53.9
49.7
47.2
48.9
shp
69.9
81.0
80.7
73.3
85.6
sfa
47.5
52.9
45.3
54.8
61.0
trn
66.9
59.5
70.1
70.5
74.5
tv
62.9
83.2
77.3
75.5
82.2
Avg.
54.8
62.1
58.8
60.4
65.2
the collaborative learning framework.
5
Conclusion
In this paper, we propose a simple but efficient framework
WSCL for weakly supervised object detection, in which two
detectors with different mechanics and characteristics are integrated in a unified architecture. With WSCL, a weakly supervised detector and a strongly supervised detector can benefit each other in the collaborative learning process. In particular, we propose an end-to-end Weakly Supervised Collaborative Detection Network for object detection. The two subnetworks are required to partially share parameters to achieve
feature sharing. The WSDDN-like sub-network is trained
with the classification loss. As there is no strong supervision
available to train the Faster-RCNN-like sub-network, a new
prediction consistency loss is defined to enforce consistency
between the fine-grained predictions of the two sub-networks
as well as the predictions within the Faster-RCNN-like subnetworks. Extensive experiments on benchmark dataset have
shown that under the collaborative learning framework, two
detectors achieve mutual enhancement. As a result, the detection performance exceeds state-of-the-arts methods.
References
[Bilen and Vedaldi, 2016] Hakan Bilen and Andrea Vedaldi.
Weakly supervised deep detection networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2846–2854, 2016.
[Bilen et al., 2015] Hakan Bilen, Marco Pedersoli, and
Tinne Tuytelaars. Weakly supervised object detection with
convex clustering. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pages 1081–
1089, 2015.
[Blum and Mitchell, 1998] Avrim Blum and Tom Mitchell.
Combining labeled and unlabeled data with co-training. In
Proceedings of the eleventh annual conference on Computational learning theory, pages 92–100. ACM, 1998.
[Cinbis et al., 2017] Ramazan Gokberk Cinbis, Jakob Verbeek, and Cordelia Schmid. Weakly supervised object localization with multi-fold multiple instance learning. IEEE
transactions on pattern analysis and machine intelligence,
39(1):189–203, 2017.
[Deselaers et al., 2012] Thomas Deselaers, Bogdan Alexe,
and Vittorio Ferrari. Weakly supervised localization and
learning with generic knowledge. International journal of
computer vision, 100(3):275–293, 2012.
[Everingham et al., 2010] Mark Everingham, Luc Van Gool,
Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge.
International journal of computer vision, 88(2):303–338,
2010.
[Girshick, 2015] Ross Girshick. Fast r-cnn. In Proceedings
of the IEEE international conference on computer vision,
pages 1440–1448, 2015.
[He et al., 2014] Kaiming He, Xiangyu Zhang, Shaoqing
Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. In European
Conference on Computer Vision, pages 346–361. Springer,
2014.
[Jie et al., 2017] Zequn Jie, Yunchao Wei, Xiaojie Jin, Jiashi Feng, and Wei Liu. Deep self-taught learning for
weakly supervised object localization. arXiv preprint
arXiv:1704.05188, 2017.
[Kantorov et al., 2016] Vadim Kantorov, Maxime Oquab,
Minsu Cho, and Ivan Laptev. Contextlocnet: Contextaware deep network models for weakly supervised localization. In European Conference on Computer Vision,
pages 350–365. Springer, 2016.
[Li et al., 2016] Dong Li, Jia-Bin Huang, Yali Li, Shengjin
Wang, and Ming-Hsuan Yang. Weakly supervised object
localization with progressive domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pages 3512–3520, 2016.
[Redmon and Farhadi, 2016] Joseph Redmon and Ali
Farhadi. Yolo9000: better, faster, stronger. arXiv preprint
arXiv:1612.08242, 2016.
[Ren et al., 2015] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances
in neural information processing systems, pages 91–99,
2015.
[Russakovsky et al., 2015] Olga Russakovsky, Jia Deng,
Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma,
Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael
Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision,
115(3):211–252, 2015.
[Simonyan and Zisserman, 2014] Karen Simonyan and Andrew Zisserman.
Very deep convolutional networks
for large-scale image recognition.
arXiv preprint
arXiv:1409.1556, 2014.
[Tang et al., 2017] Peng Tang, Xinggang Wang, Xiang Bai,
and Wenyu Liu. Multiple instance detection network
with online instance classifier refinement. arXiv preprint
arXiv:1704.00138, 2017.
[Thomee et al., 2015] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas
Poland, Damian Borth, and Li Jia Li. The new data and
new challenges in multimedia research. Communications
of the Acm, 59(2):64–73, 2015.
[Uijlings et al., 2013] Jasper RR Uijlings, Koen EA Van
De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal
of computer vision, 104(2):154–171, 2013.
[Wang et al., 2014] Chong Wang, Weiqiang Ren, Kaiqi
Huang, and Tieniu Tan. Weakly supervised object localization with latent category learning. In European Conference on Computer Vision, pages 431–445. Springer, 2014.
[Wu et al., 2015] Jiajun Wu, Yinan Yu, Chang Huang, and
Kai Yu. Deep multiple instance learning for image classification and auto-annotation. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pages 3460–3469, 2015.
[Yan et al., 2017] Ziang Yan, Jian Liang, Weishen Pan, Jin
Li, and Changshui Zhang. Weakly-and semi-supervised
object detection with expectation-maximization algorithm.
arXiv preprint arXiv:1702.08740, 2017.
[Zhang et al., 2006] Cha Zhang, John C Platt, and Paul A Viola. Multiple instance boosting for object detection. In
Advances in neural information processing systems, pages
1417–1424, 2006.
| 1cs.CV
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.